SNOW AVALANCHE IDENTIFICATION USING SENTINEL-1: DETECTION RATES AND CONTROLLING FACTORS by Zachary Marshall Keskinen A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Earth Sciences MONTANA STATE UNIVERSITY Bozeman, Montana May 2021 ©COPYRIGHT by Zachary Marshall Keskinen 2021 All Rights Reserved ii ACKNOWLEDGEMENTS This thesis would not have been possible without support from many friends. I am extremely grateful to Jordy Hendrikx for his guidance on so many subjects. I am also indebted to Karl Birkeland and Markus Eckerstorfer. Their experience, knowledge and thoughtful advice was critical to this project’s success. I want to thank the entire Gallatin National Forest Avalanche Center for their mentorship and insights. I also want to thank Zachary Miller, Ross Palomaki, and Gabrielle Antonoli for excellent conversations and field days. Finally, I want to thank my parents and my partner, Betsie, for their support and love. iii TABLE OF CONTENTS 1. INTRODUCTION ....................................................................................................................... 1 Significance ................................................................................................................................. 1 Literature Review ........................................................................................................................ 5 Proof of Concepts Studies ................................................................................................... 5 Automated Detection Algorithms ........................................................................................ 9 Multiorbital Sentinel-1 Compositing ................................................................................. 12 The Missing Piece ............................................................................................................. 15 Reference Cited ................................................................................................................. 16 2. SNOW AVALANCHE IDENTIFICATION USING SENTINEL-1 BACKSCATTER IMAGERY: DETECTION RATES AND CONTROLLING FACTORS .................................... 19 Contribution of Authors and Co-Authors .................................................................................. 19 Manuscript Information ............................................................................................................. 20 Introduction ............................................................................................................................... 21 Background ................................................................................................................................ 23 SAR Background ........................................................................................................... 23 Microwave’s Interaction with Snow Cover ........................................................... 23 Temporal Change Analysis ................................................................................... 24 Image Artifacts ...................................................................................................... 27 Literature Review ...................................................................................................................... 28 Research Questions ................................................................................................................... 31 Methods ..................................................................................................................................... 32 Datasets Used .................................................................................................................... 32 Field Observed Avalanches ................................................................................... 32 Sentinel-1 Imagery ................................................................................................ 35 Processing and Analysis .................................................................................................... 38 Avalanche Path Identification ............................................................................... 39 Manual Avalanche Detection ................................................................................ 40 Supplementation of Avalanche Records ............................................................... 42 Statistical Analysis ................................................................................................ 44 Results ....................................................................................................................................... 46 Image Pairs and Field Location Detection Rates .............................................................. 46 Destructive Size ................................................................................................................. 47 Covariation Analysis ......................................................................................................... 48 Kruskal-Wallis (KW) ........................................................................................................ 49 Chi-Squared Analysis ........................................................................................................ 50 Random Forest ................................................................................................................... 51 Discussion .................................................................................................................................. 52 How many avalanches was Sentinel-1 backscatter temporal iv TABLE OF CONTENTS CONTINUED change analysis able to detect? .......................................................................................... 52 How did detection rates vary between the two regions? ....................................... 52 How did detection rates vary between the image pairs? ....................................... 52 What factors correlated with successful detections? ......................................................... 53 Destructive Size ..................................................................................................... 55 Avalanche Type ..................................................................................................... 55 Path Length ............................................................................................................ 56 Local Incidence Angle ........................................................................................... 57 Layover Percentage ............................................................................................... 59 Slope Angle ........................................................................................................... 60 Curvature ............................................................................................................... 61 Tree Percentage ..................................................................................................... 61 Lag Days ................................................................................................................ 58 Limitations and Future Work ............................................................................................ 61 Conclusion ................................................................................................................................. 64 Reference Cited ......................................................................................................................... 67 3. CONCLUSIONS ....................................................................................................................... 70 Summary .................................................................................................................................... 70 Future Work ............................................................................................................................... 71 REFERENCES CITED ................................................................................................................. 73 APPENDICES ............................................................................................................................... 77 APPENDIX A: COVARIANCE FIGURE ........................................................................ 78 APPENDIX B: AVALANCHE DATABASES ................................................................ 81 APPENDIX C: PYTHON SCRIPTS USED IN ANALYSIS ......................................... 109 v LIST OF TABLES Table Page 1. Avalanche Cycles and Imagery Sets Used .................................................................... 35 2. Explanatory Factors with data sources, spatial resolutions, and shortened names. ........................................................................................................... 42 3. Image Pairs with Detection Rates. ................................................................................. 45 4. Detection Rates by destructive size.. ............................................................................. 46 5. KW Results for Full Dataset, UDOT, BTAC. ............................................................... 47 6. Feature individual and cumulative importance in the random forest model. ................ 50 7. Summary of factors related to increased detection rates ............................................... 53 vi LIST OF FIGURES Figure Page 1. Avalanche deaths in the United States ............................................................................ 2 2. RGB Composite of avalanche debris from ERS 1/2 and aerial photo of the highlighted avalanche. ............................................................................................... 6 3. Conceptual model of backscatter contributions for undisturbed snow vs avalanche debris. ............................................................................................... 8 4. Destructive size, avalanche type, snow moisture relative to detection rates. ................ 11 5. TerraSAR-X RGB composite image, Sentinel-1 RGB composite image, multi-orbit composite of Sentinel-1 images, and a multi-orbit composite of Sentinel-1 images with nonlocal mean filter applied to reduce speckle effects. ........... 14 6. Conceptual model for the interaction between dry avalanche debris and SAR microwaves ......................................................................................... 23 7. Reference, activity, and RGB composite images for two field reported avalanches .... 25 8. SAR image showing radar shadow and layover in mountainous terrain and conceptual diagram showing the areas affected by radar shadow and layover. ..... 27 9. Avalanche locations for the BTAC and UDOT avalanche databases. .......................... 31 10. Avalanche activity index for the two databases from 2016-2020. .............................. 32 11. Avalanche destructive size distributions for each database. ........................................ 33 12. Flowchart of data collection and processing steps. ..................................................... 37 vii LIST OF FIGURES CONTINUED Figure Page 13. An annotated example of the avalanche path identification process for the BTAC avalanche paths. ........................................................................ 39 14. RGB Composite image of a detectable avalanche. ...................................................... 41 15. Boxplots showing the means, interquartile ranges, whiskers, and outliers for detected and undetected avalanches in the full dataset. .............................. 49 16. Proposed conceptual model explaining the increased detection rates of avalanche debris at higher incidence angles ..................................................... 57 17. The same BTAC reported avalanche in an image pair unaffected by layover and in an image pair affected by layover. ......................................................... 59 viii ABSTRACT Snow avalanches present a significant hazard that endangers lives and infrastructure. Consistent and accurate datasets of avalanche events is valuable for improving forecasting ability and furthering knowledge of avalanches' spatial and temporal patterns. Remote sensing-based techniques of identifying avalanche debris allow for continuous and spatially consistent datasets of avalanches to be acquired. This study utilizes expert manual interpretations of Sentinel-1 synthetic aperture radar (SAR) satellite backscatter images to identify avalanche debris and compares those detections against historical field records of avalanches in the transitional snow climates of Wyoming and Utah. This study explores the utility of Sentinel-1 (a SAR satellite) images to detect avalanche debris on primarily dry slab avalanches. The overall probability of detection (POD) rate for avalanches large enough to destroy trees or bury a car (i.e., D3 on the Destructive Size Scale) was 64.6%. There was a significant variance in the POD among the 13 individual SAR image pairs (15.4 – 87.0%). Additionally, this study investigated the connection between successful avalanche detections and SAR-specific, topographic, and avalanche type variables. The most correlated variables with higher detection rates were avalanche path lengths, destructive size of the avalanche, incidence angles for the incoming microwaves, slope angle, and elapsed time between the avalanche and a Sentinel-1 satellite passing over. This study provides an initial exploration of the controlling variables in the likelihood of detecting avalanches using Sentinel-1 backscatter change detection techniques. This study also supports the generalizability of SAR backscatter difference analysis by applying the methodology in different regions with distinct snow climates from previous studies. 1 CHAPTER ONE INTRODUCTION Significance Snow avalanches are a complex and dangerous natural hazard that kills hundreds of people every year and damages infrastructure across the world. Schweizer (2008) estimated the number of worldwide deaths to be around 250. In the United States there have been 1081 people killed in avalanches from 1950-2019 (Figure 1), while Europe saw 4750 fatalities between 1970-2015 (Techel et al. 2016, Peitzsch et al. 2019). These statistics are certainty an underestimation of the total deaths from avalanches as there are many tragic avalanches reported across the world resulting in major losses of life (Chabot, 2017). Figure 1: Avalanche deaths in the United States 1950-2019. The red line represents a 5-year moving average. Figure from (Colorado Avalanche Information Center, 2021). 2 In addition to the tragic loss of lives from avalanches, there are also significant infrastructure impacts worldwide: direct damage, necessary defensive measures, and closed or restricted transportation corridors. A comprehensive assessment of avalanche damage to infrastructure in the United States has not been undertaken. However, a 1990 commission on Avalanche Hazard in the United States cited 11.4 million dollars, 23.5 million dollars in 2021 dollars, in direct infrastructure damages in Alaska in the period from 1977-1986 (National Research Council, 1990). An assessment in Iceland found that direct infrastructure damage across the country from 1970-2000 amounted to 41 million dollars, 63.9 million dollars in 2021 (Jóhannesson and Arnalds, 2000). A further study in Grisons, Switzerland found that direct damage in the valley from 573 avalanches over 50 years amounted to 63.3 million euros, or approximately 117.8 million dollars in 2021 (Fuchs and Bründl, 2005). Clearly, infrastructure damage and associated defensive measures against avalanches are major economic costs to all countries with infrastructure crossing through mountainous terrain. The loss of life and infrastructure damage caused by avalanches provides a compelling argument for continued research and improved forecasting techniques for this natural hazard. However, current research and forecasting of avalanches is limited to by datasets that are opportunistically collected and thus spatial and temporal incomplete. Current avalanche datasets come primarily from field observations that are either opportunistically collected or limited in their spatial and temporal scopes. This limits their utility for improving forecasting of avalanche hazard relative to remotely sensed datasets. Opportunistic field observations of avalanches are incomplete, tending to be spatially biased toward accessible locations (lower elevations, near transportation corridors, near population centers), and 3 temporally biased to convenient times (clear weather, low avalanche danger) (Eckerstorfer et al., 2016). Other datasets from operational road or ski forecasting are relatively complete but are limited in their spatial scale to single mountains or transportation corridors. These small spatial extents relative to backcountry regional forecasting limits the use of operational avalanche datasets as validation for backcountry forecasts or for trend analysis. Additionally, these datasets tend to be dominated by artificially triggered avalanches from mitigation efforts. This limits their use for data in developing machine learning techniques that are focused on naturally releasing avalanches. Finally, for regions with less avalanche awareness there is often no established record of field observed datasets. These limitations of the field observed datasets mean that comparisons between mountain ranges, over long time periods, and using datasets of avalanche occurrences as spatially complete is challenging. Developing records of avalanche occurrences from remotely sensed imagery would greatly improve our research into avalanche hazard and consequently our forecasting ability of avalanches, especially in developing parts of the world. Researchers are currently exploring remotely sensing avalanche debris as a more spatially complete, consistent method of inventorying and identifying avalanches (Eckerstorfer et al. 2016, Bühler et al. 2019). Remotely sensed datasets of avalanche occurrences would allow for improved research and forecasting of avalanche hazard. Current field observed datasets, with their spatial and temporal gaps and biases, are limiting meaningful statistical analyses of avalanche occurrences. On the research side a more complete dataset of avalanche occurrences would allow for machine learning based techniques to identify explanatory variables and model- based methods of assessing avalanche occurrence probability. Previous research has been promising, but has been limited by spatially limited and incomplete datasets of avalanche 4 occurrences (Hendrikx et al. 2014, Pozdnoukhov et al. 2011). Additionally, operational avalanche forecasters would benefit from real-time information on avalanches’ locations and timings (Eckerstorfer et al., 2019). These benefits, both to scientific research and operational hazard analysis, demonstrate the need for continued research into methods of remotely sensing avalanches. A remote sensing sensor that has shown considerable promise in the detection of avalanche debris is synthetic aperture radar (SAR) (Eckerstorfer et al., 2016). SAR sensors actively emit microwaves towards the surface and measure the returning backscattered energy from the surface. Since these sensors utilize microwaves, they are minimally affected by precipitation and cloud cover. This means that, unlike sensors operating in the visual spectrum, these sensors can provide information during periods of cloud cover. This is invaluable for avalanche detection, as avalanches often occur during periods of snow fall and cloud cover. SAR sensors also utilized the movement of the sensor, either along a fixed track or an orbital/flight path, to generate higher spatial resolutions. This higher resolution is important for avalanches, which can be quite small. These SAR sensors have been studied over the previous twenty years and numerous studies have confirmed their utility for avalanche detection (Wiesmann et al., (2001), Martinez-Vazquez and Fortuny-Guasch, (2008), Malnes et al., (2013), Eckerstorfer and Malnes, (2015), Malnes et al., (2015)). 5 Literature Review SAR based monitoring for avalanche has been explored since the early 2000s as a method of detection avalanche debris (Wiesmann et al., 2001). Since our study focused on satellite-based SAR avalanche detections this literature review will focus on the development of satellite-based detections, but both ground and airborne based methods for avalanche detection also show promise (Martinez-Vazquez and Fortuny-Guasch 2008, Bühler et al. 2009). Proof of Concept Studies Wiesmann et al. (2001) illustrated the first space borne detection of avalanches from a large cycle of natural avalanches in Switzerland. They used imagery from the SAR sensors aboard ERS 1/2 satellites to create RGB composite images. In these composite images they identified large increases in backscatter from the debris of an avalanche. The composite images were created with the reference image, prior to the avalanche, in the red and blue channels and the green channel contained an activity image, from after the avalanche. Areas in green hence represent increases in backscatter and purple represents decreases in backscattered energy (Figure 2). Wiesmann et al. (2001) proposed that the increased backscatter was the result of the “compacted rough snow of an avalanche cone”. This study provided our first analysis of the 6 impact of avalanche debris in SAR imagery and the possibilities of detecting avalanche debris with SAR imagery. Figure 2: RGB Composite of avalanche debris from ERS 1/2 (left) and aerial photo of the highlighted avalanche (right). Green regions in the SAR composite represent increased backscatter and correspond to the field observed avalanche debris. SAR image from Wiesmann et al. (2001) and the aerial photo is property of the WSL0SLF Swiss Institute for Snow and Avalanche Research. Malnes et al. (2013) used Radarsat-2 images to successfully identify two of three avalanches and compare their extents from ortho-photos to the delineation from the SAR RGB composites. This use of RadarSat-2 was continued by Eckerstorfer and Malnes (2015) when they used a set of 12 SAR images to identify 467 features that were likely avalanche debris and could successfully verify 173 (37%) of those features through optical satellite imagery (Landsat 8) and field campaigns. This larger scale study also discussed sources of possible false alarms: wind drifted snow and areas of wet or refreezing snow. Liquid water in the snowpack absorbs microwaves and results in significantly reduced backscatter. These studies confirmed the ability 7 of SAR satellites to detect avalanche debris, but the high acquisition costs of RadarSat-2 prevented this from being a viable continual method of monitoring for avalanches Eckerstorfer and Malnes (2015) provided the first conceptual model of the differences between undisturbed snow and avalanche debris for backscattered SAR energy (Figure 3). Since avalanche debris was increasing the backscattered energy relative to undisturbed snow, they focused on mechanisms that would increase backscattering in avalanche debris. First, they identified that the rougher surface of avalanche debris would increase the amount of energy backscattered from the air-snow interface. The rougher surface would increase diffuse omnidirectional over specular forward scattering of the energy. This would increase the energy backscattered towards the SAR sensor instead of forwards and away from the sensor. Secondly the increased density and depth of avalanche debris would also increase diffuse volumetric scattering. This increased volumetric scattering would increase the omnidirectional scattering and minimize the amount of specular scattering by the ground surface. The exact contributions between the two mechanism are still unclear. In the case of wet avalanches, they proposed that since SAR energy is rapidly absorbed in wet snow, that the surface scattering mechanism may be the dominant control on increased backscatter. 8 Figure 3: Conceptual model of backscatter contributions for undisturbed snow vs avalanche debris. The increasing surface roughness and (for dry avalanches) volumetric backscatter is proposed as the primary controls on backscatter differences. Figure from Eckerstorfer and Malnes (2015). The beginning of operational use of Sentinel-1A in October 2014, equipped with a SAR sensor with reasonably high spatial resolution (20 meters), reasonable return periods (6 to 12 days) and free data availability, made the continual monitoring of avalanches feasible. Malnes et al. (2015) detected the first avalanches with Sentinel-1 detections. They used temporal change analysis to identify increased backscatter in Sentinel-1 imagery. Using this method, they detected 9 and verified (using field observations and Radarsat-2 imagery) 489 avalanche deposits from a cycle in Norway in January 2015. This demonstrated the ability of the slightly lower resolution (20 meter) Sentinel-1 SAR sensor to still detect avalanche debris using temporal change analysis. Automated Detection Algorithms With the feasibility of manual Sentinel-1 detections confirmed some research efforts switched to attempting to automate detections of avalanche debris in Sentinel-1 imagery. This automation would allow for continual monitoring for avalanche activity. The first attempt at creating an automated detection system for Sentinel-1 imagery was by Vickers et al. (2016). Working with Sentinel-1 imagery captured over Norway they used k-means classification, an unsupervised classification method, to segment pixels into avalanche and non-avalanche classes. Various masks were also used to increase the accuracy – radar layover, radar shadow, water bodies, and slope angle that would not collect avalanche debris (masking pixels over 35 degrees). This method used both an initial thresholding step (1% of pixels in a scene increasing by more than 6 decibels) to eliminate scenes that had insufficient backscatter increases and then k-means to differentiate the pixels with the largest increase. The probability of correctly detect avalanche debris was over 60%, but there were issues with both missed classifications and falsely detected avalanches. This study showed that automated detections were possible but also the significant issues that arose from the variable and dynamic nature of the winter snowpack. This automated detection method (with some adjustments) was applied to multiple winters of Sentinel-1 imagery and the output compared to manually identified avalanches from Sentinel-1 imagery by Eckerstorfer et al. (2019). Following parameter optimization, a probability of detection (POD) of 64.7% and a False Alarm Rate (FAR) of 45.9% were achieved. These 10 metrics were quite variable with ranges for POD from 36.4% to 89.5% and FAR from 25.9% to 81.7%. This study proposed that the high variability in detection rates was a function of snow conditions during imaging, avalanche types, and destructive sizes. Image pairs that had decreased backscattered between the images had higher detection rates and fewer false alarms. This overall decrease in backscatter led to clear contrast with the areas of increased backscatter from the avalanche debris. This situation was most common when “the reference image is primarily dry and turned wet in the activity image” (Eckerstorfer et al., 2019). High false alarm rates and lower detection rates occurred when there was a wet to dry snow transition between the reference to activity image. This study also examined the manual and automated detection rates against destructive size, avalanche types and snow type. About 90% of the avalanches were slabs avalanches limiting an analysis of the effect of avalanche type on detection rates. Both the manual and automated detection had lower detection rate for loose avalanches than slab avalanches, but this may have been complicated by the fact that loose avalanches were on average smaller than slab avalanches. Looking at wetness 90% were predominately wet debris again limiting the analysis of avalanche type’s effect on detection rates. They found that wet avalanches were easier to 11 detect than dry avalanches regardless of avalanche size, most likely due to the higher surface roughness exhibited by wet avalanche relative to dry avalanche debris (Figure 4). Figure 4: Destructive size, avalanche type, snow moisture relative to detection rates. Figure from Eckerstorfer et al. (2019). The next step in automated detection algorithms for avalanche debris focused on more complex machine learning techniques, specifically neural networks. Kummervold et al. (2018) trained a neural network using eight Sentinel-1 images with manually detected avalanche labels. This neural network was trained to classify 60x60 pixel images as containing avalanche debris or not. There were equal number of avalanche (n = 817) and non-avalanche slices and used an 80%/20% training/test split. The validation results were very promising with accuracies around 12 90%. However, the neural network relied on a dataset which did not represent the skewed class reality, where avalanche debris are a very small percentage of the overall scenes. This neural network approach was continued in Bianchi et al. (2019) when the training dataset was expanded to include 118 Sentinel-1 scenes with manually labeled avalanche debris. This method was a pixel-based analysis instead of a scene classification and it also more accurately represented the proportion of avalanche pixels, with most of the dataset comprising non-avalanche pixels. The neural network improved on Eckerstorfer et al.'s (2019) baseline segmentation f1 score, the weighted average of precision and recall, of 38.1% to an f1 of 66.6%. These neural network-based techniques show great promise but their computational expense and complication has prevented their wide-scale into automated avalanche detections. Multiorbital Sentinel-1 Compositing Following a major avalanche cycle in central Switzerland, Leinss et al. (2020) mapped avalanche debris using SPOT-6 (1.5- 6m resolution optical sensor), TerraSAR-X (1m resolution SAR satellite), and Sentinel-1. This study did not use field observed data to compare to the manually and automatic detections and only compared between sensors. This limited the ability of this study to analyze type and size controls on detection. However, they qualitatively noted that “avalanches reaching below the wet-snow line were much more visible than avalanche from the dry snow zone”. When comparing the three satellite sensors Leinss et al. (2020) found the optical SPOT-6 missed avalanches in areas of shadow from tall peaks and ridges but had higher overall resolution than the SAR images. The optical sensor detected 104 avalanches that could not be seen in either SAR image set. They also point out the additional power of being able to identify 13 start zones and tracks in optical imagery while only the debris is identifiable in most SAR images. The higher resolution TerraSAR-X was also able detect around a third more avalanches than Sentinel-1. An interesting technique developed by Leinss et al. (2020) was the compositing of multiple Sentinel-1 images from multiple orbits into a single composite image. This technique reduced the effects from layover and radar shadow. They found that about twice as many avalanches could be detected by this method relative to a single Sentinel-1 image. However some avalanches were “averaged” out. They were visible in a few individual images but disappeared (most likely due to weathering of the surface roughness) in later images and did not appear in the composite. Figure 5 shows a TerraSAR-X RGB image, a Sentinel-1 RGB image. Overall, this compositing seemed to increase the resolution and is a potentially valuable method of avalanche identification. 14 Figure 5: (a) TerraSAR-X RGB composite image (b) Sentinel-1 RGB composite image, (c) multi-orbit composite of Sentinel-1 images (d) a multi-orbit composite of Sentinel-1 images with nonlocal mean filter applied to reduce speckle effects. This figure places the activity image in the red channel with the reference in the blue and green channels meaning that increased backscattered is represented by red regions. Figure from Leinss et al. (2020) Leinss et al. (2020) composited by simply averaging the change detection images from each image pair. However, they proposed that using a local-resolution weighted averaging method that would favor slopes with higher incidence angles would be a potentially useful. This local resolution weighted averaging would prioritize the values from pixels with higher incidence 15 angles, which due to SAR geometries tend to have higher resolution and (they theorized based on the current conceptual model of avalanche debris) a greater increase in backscatter from avalanche debris. The missing piece While the body of research on SAR-based avalanche detections has expanded immensely over the last decade there remains several unanswered questions. Are the detection rates found in papers such as Eckerstorfer et al., (2019) replicated in other snow climates and mountain ranges? Do higher incidence angles actually lead to higher detection rates for avalanche debris as suggested by Leinss et al., (2020)? What are the primary drivers (in addition to the type and size of the avalanche) that control when we can detect avalanche debris? Without a clear answer to when we can and cannot expect to detect avalanches with SAR sensors, we are unable to assess how useful SAR avalanche detections will be to researchers and practitioners worldwide. Answering these questions will provide important context for utilizing this technique to answer other avalanche and snow related questions. The manuscript presented in Chapter 2 of this thesis provides an in-depth analysis of detection rates for avalanche debris in Sentinel-1 imagery for the intercontinental snow climates of Utah and Wyoming. Utilizing records of field observed avalanches, we explore not only the overall detection rates in Sentinel-1 imagery, but also an analysis of factors that are correlated with missed detections. Understanding when and why avalanche debris is likely to be visible or not in SAR imagery will increase the power of this technique in analyzing snow avalanche hazards and improve our understanding of its utility globally. 16 References Cited Abermann, J., Eckerstorfer, M., Malnes, E., Hansen, B.U., 2019. A large wet snow avalanche cycle in West Greenland quantified using remote sensing and in situ observations. Nat Hazards 97, 517–534. https://doi.org/10.1007/s11069-019-03655-8 Bianchi, F.M., Grahn, J., Eckerstorfer, M., Malnes, E., Vickers, H., 2021. Snow avalanche segmentation in SAR images with Fully Convolutional Neural Networks. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 75–82. https://doi.org/10.1109/JSTARS.2020.3036914 Birkeland, K., 2017. In Response to Avalanche Fatalities in the United States by Jekich et al. Wilderness and Environmental Medicine 28. https://doi.org/10.1016/j.wem.2017.06.009 Breiman, L., 2001. Random Forests. Machine Learning 45, 5–32. https://doi.org/10.1023/A:1010933404324 Bühler, Y., Hafner, E., Zweifel, B., Zesiger, M., Heisig, H., 2019. Where are the avalanches? Rapid mapping of a large snow avalanche period with optical satellites. The Cryosphere Discussions 1–21. https://doi.org/10.5194/tc-2019-119 Bühler, Y., Hüni, A., Christen, M., Meister, R., Kellenberger, T., 2009. Automated detection and mapping of avalanche deposits using airborne optical remote sensing data. Cold Regions Science and Technology 57, 99–106. https://doi.org/10.1016/j.coldregions.2009.02.007 Chabot, D., 2017. Avalanche Forecasting in Afghanistan, Pakistan and Tajikistan. The Avalanche Review 35, 30–31. Colorado Avalanche Information Center, 2021. US Avalanche Accident Reports. Eckerstorfer, M., Bühler, Y., Frauenfelder, R., Malnes, E., 2016. Remote sensing of snow avalanches: Recent advances, potential, and limitations. Cold Regions Science and Technology 121, 126–140. https://doi.org/10.1016/j.coldregions.2015.11.001 Eckerstorfer, M., Malnes, E., 2015. Manual detection of snow avalanche debris using high- resolution Radarsat-2 SAR images. Cold Regions Science and Technology 120. https://doi.org/10.1016/j.coldregions.2015.08.016 Eckerstorfer, M., Vickers, H., Malnes, E., Grahn, J., 2019. Near-Real Time Automatic Snow Avalanche Activity Monitoring System Using Sentinel-1 SAR Data in Norway. Remote Sensing 11, 2863. https://doi.org/10.3390/rs11232863 Fischer, J.-T., Kaitna, R., Heil, K., Reiweger, I., 2018. The Heat of the Flow: Thermal Equilibrium in Gravitational Mass Flows. Geophysical Research Letters 45, 11,219- 11,226. https://doi.org/10.1029/2018GL079585 Fuchs, S., Bründl, M., 2005. Damage Potential and Losses Resulting from Snow Avalanches in Settlements of the Canton of Grisons, Switzerland. Natural Hazards 34, 53–69. https://doi.org/10.1007/s11069-004-0784-y Gesch, D., Oimoen, M., Greenlee, S., Nelson, C., Steuck, M., Tyler, D., 2002. The National Elevation Dataset. Photogrammetric Engineering and Remote Sensing 68, 5–11. Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, Big Remotely Sensed Data: tools, applications and experiences 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031 17 Greene, E., Birkeland, K., Elder, K., Mccammon, I., Staples, M., Sharaf, D., 2016. Snow, Weather, and Avalanches: Observational Guidelines for Avalanche Programs in the United States, 3rd ed. American Avalanche Association, Victor, ID. Hendrikx, J., Murphy, M., Onslow, T., 2014. Classification trees as a tool for operational avalanche forecasting on the Seward Highway, Alaska. Cold Regions Science and Technology 97, 113–120. https://doi.org/10.1016/j.coldregions.2013.08.009 Jóhannesson, T., Arnalds, Þ., 2000. Accidents and economic damage due to snow avalanches and landslides in Iceland. Jökull 50. Kummervold, P., Malnes, E., Eckerstorfer, M., Arntzen, I., Bianchi, F.M., 2018. Avalanche detection in Sentinel-1 radar images using convolutional neural networks. Presented at the International Snow Science Workshop, Innsbruck, Austria. Larsen, H., Hendrikx, J., Slåtten, M.S., Engeset, R.V., 2020a. Developing nationwide avalanche terrain maps for Norway. Nat Hazards 103, 2829–2847. https://doi.org/10.1007/s11069- 020-04104-7 Larsen, H., Sykes, J., Hendrikx, J., Schauer, A., Langford, R., Statham, G., Campbell, C., Neuhauser, M., Fischer, J.-T., 2020b. Development of Automated Avalanche Terrain Exposure Scale Maps: Current and Future. Leinss, S., Wicki, R., Holenstein, S., Baffelli, S., Bühler, Y., 2020. Snow avalanche detection and mapping in multitemporal and multiorbital radar images from TerraSAR-X and Sentinel-1. Natural Hazards and Earth System Sciences 20, 1783–1803. https://doi.org/10.5194/nhess-20-1783-2020 Liu, Y., Zhou, Y., Wen, S., Tang, C., 2014. A Strategy on Selecting Performance Metrics for Classifier Evaluation. International Journal of Mobile Computing and Multimedia Communications 6, 20–35. https://doi.org/10.4018/IJMCMC.2014100102 Malnes, E., Eckerstorfer, M., Larsen, Y., Frauenfelder, R., Jonsson, A., Jaedicke, C., Solbø, S., 2013. Remote sensing of avalanches in northern Norway using Synthetic Aperture Radar. Presented at the International Snow Science Workshop, Grenoble, France. https://doi.org/10.13140/2.1.2462.0481 Malnes, E., Eckerstorfer, M., Vickers, H., 2015. First Sentinel-1 detections of avalanche debris. The Cryosphere Discussions 9, 1943–1963. https://doi.org/10.5194/tcd-9-1943-2015 Martinez-Vazquez, A., Fortuny-Guasch, J., 2008. A GB-SAR Processor for Snow Avalanche Identification. IEEE Transactions on Geoscience and Remote Sensing 46, 3948–3956. https://doi.org/10.1109/TGRS.2008.2001387 McClung, D.M., 2002a. The Elements of Applied Avalanche Forecasting, Part I: The Human Issues. Natural Hazards 26, 111–129. https://doi.org/10.1023/A:1015665432221 McClung, D.M., 2002b. The Elements of Applied Avalanche Forecasting, Part II: The Physical Issues and the Rules of Applied Avalanche Forecasting. Natural Hazards 26, 131–146. https://doi.org/10.1023/A:1015604600361 McClung, D.M., Schaerer, P., 2006. The Avalanche Handbook. The Mountaineers. Mock, C.J., Birkeland, K.W., 2000. Snow Avalanche Climatology of the Western United States Mountain Ranges. Bull. Amer. Meteor. Soc. 81, 2367–2392. https://doi.org/10.1175/1520-0477(2000)081<2367:SACOTW>2.3.CO;2 National Research Council, 1990. Snow Avalanche Hazards and Mitigation in the United States. The National Academies Press, Washington, DC. 18 Peitzsch, E., Boilen, S., Logan, S., Birkeland, K., Greene, E., 2020. Research note: How old are the people who die in avalanches? A look into the ages of avalanche victims in the United States (1950–2018). Journal of Outdoor Recreation and Tourism 29. https://doi.org/10.1016/j.jort.2019.100255 Pozdnoukhov, A., Matasci, G., Kanevski, M., Purves, R., 2011. Spatio-temporal avalanche forecasting with Support Vector Machines. Natural Hazards and Earth System Sciences 11. https://doi.org/10.5194/nhess-11-367-2011 Schweizer, J., 2008. Snow avalanche formation and dynamics. Cold Regions Science and Technology 54, 153–154. https://doi.org/10.1016/j.coldregions.2008.08.005 Schweizer, J., Mitterer, C., Techel, F., Stoffel, A., Reuter, B., 2018. Quantifying the obvious: the avalanche danger level. International Snow Science Workshop Proceedings 2018, Innsbruck, Austria 1052–1058. Sexton, J., Song, X.-P., Feng, M., Noojipady, P., Anand, A., Huang, C., Kim, D.-H., Collins, K., Channan, S., Dimiceli, C., Townshend, J., 2013. Global, 30-m resolution continuous fields of tree cover: Landsat-based rescaling of MODIS vegetation continuous fields with lidar-based estimates of error. International Journal of Digital Earth 6, 1–22. https://doi.org/10.1080/17538947.2013.786146 Techel, F., Jarry, F., Kronthaler, G., Mitterer, S., Nairz, P., Pavšek, M., Valt, M., Darms, G., 2016. Avalanche fatalities in the European Alps: Long-term trends and statistics. Geographica Helvetica 71, 147–159. https://doi.org/10.5194/gh-71-147-2016 Vickers, H., Eckerstorfer, M., Malnes, E., Larsen, Y., Hindberg, H., 2016. A method for automated snow avalanche debris detection through use of Synthetic Aperture Radar (SAR) imaging: Automated avalanche detection. Earth and Space Science 3. https://doi.org/10.1002/2016EA000168 Vollrath, A., Mullissa, A., Reiche, J., 2020. Angular-Based Radiometric Slope Correction for Sentinel-1 on Google Earth Engine. Remote Sens 12, 1867. Wiesmann, A., Wegmuller, U., Honikel, M., Strozzi, T., Werner, C.L., 2001. Potential and methodology of satellite based SAR for hazard mapping, in: IGARSS 2001. Scanning the Present and Resolving the Future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Symposium (Cat. No.01CH37217). Presented at the IGARSS 2001. Scanning the Present and Resolving the Future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Symposium (Cat. No.01CH37217), pp. 3262–3264 vol.7. https://doi.org/10.1109/IGARSS.2001.978322 Wilks, D., 2019. Statistical Methods in the Atmospheric Sciences. Elsevier. https://doi.org/10.1016/C2017-0-03921-6 Wynne, R., Campbell, J., 2011. Introduction to Remote Sensing, 5th ed. The Guilford Press. Yang, J., Li, C., Li, L., Ding, J., Zhang, R., Han, T., Liu, Y., 2020. Automatic Detection of Regional Snow Avalanches with Scattering and Interference of C-band SAR Data. Remote Sensing 12, 2781. https://doi.org/10.3390/rs12172781 19 CHAPTER TWO SNOW AVALANCHE IDENTIFICATION USING SENTINEL-1 BACKSCATTER IMAGERY: DETECTION RATES AND CONTROLLING FACTORS Contribution of Authors and Co-Authors Manuscript in Chapter 2 Author: Zachary Keskinen Contributions: ZK conceived of the presented idea, downloaded and processed imagery, detected avalanches, performed all analysis, wrote the paper. Co-Author: Dr. Jordy Hendrikx Contributions: JH provided guidance and advice in conceiving the original idea. JH facilitated database and atlas acquisition from the Utah Department of Transportation. JH suggested using the auto-ATES map for avalanche path identification and helped with figure ideas. Finally, JH provided valuable editing of all drafts. Co-Author: Dr. Karl Birkeland Contributions: KB assisted in acquiring the Bridger Teton Avalanche Center dataset. KB also provided valuable editing on the manuscript. Co-Author: Dr. Markus Eckerstorfer Contributions: ME provided technical guidance with SAR imagery and detections. ME provided sample detections. Finally, ME provided valuable editing on the manuscript. 20 Manuscript Information Zachary Keskinen, Jordy Hendrikx, Karl Birkeland, Markus Eckerstorfer Natural Hazards and Earth System Sciences Status of Manuscript: __x__ Prepared for submission to a peer-reviewed journal ____ Officially submitted to a peer-reviewed journal ____ Accepted by a peer-reviewed journal ____ Published in a peer-reviewed journal Published by the European Geosciences Union 21 SNOW AVALANCHE IDENTIFICATION USING SENTINEL-1 BACKSCATTER IMAGERY: DETECTION RATES AND CONTROLLING FACTORS Snow avalanches present a significant hazard that endangers lives and infrastructure. Consistent and accurate datasets of avalanche events is valuable for improving forecasting ability and furthering knowledge of avalanches' spatial and temporal patterns. Remote sensing-based techniques of identifying avalanche debris allow for continuous and spatially consistent datasets of avalanches to be acquired. This study utilizes expert manual interpretations of Sentinel-1 synthetic aperture radar (SAR) satellite backscatter images to identify avalanche debris and compares those detections against historical field records of avalanches in the transitional snow climates of Wyoming and Utah. This study explores the utility of Sentinel-1 (a SAR satellite) images to detect avalanche debris on primarily dry slab avalanches. The overall probability of detection (POD) rate for avalanches large enough to destroy trees or bury a car (i.e., D3 on the Destructive Size Scale) was 64.6%. There was a significant variance in the POD among the 13 individual SAR image pairs (15.4 – 87.0%). Additionally, this study investigated the connection between successful avalanche detections and SAR-specific, topographic, and avalanche type variables. The most correlated variables with higher detection rates were avalanche path lengths, destructive size of the avalanche, incidence angles for the incoming microwaves, slope angle, and elapsed time between the avalanche and a Sentinel-1 satellite passing over. This study provides an initial exploration of the controlling variables in the likelihood of detecting avalanches using Sentinel-1 backscatter change detection techniques. This study also supports the generalizability of SAR backscatter difference analysis by applying the methodology in different regions with distinct snow climates from previous studies. 1 Introduction Snow avalanches are a complex mountain hazard that endangers lives, threatens infrastructure, and closes transportation corridors (McClung and Schaerer 2006, Birkeland 2017). Timely knowledge about when and where avalanches occur is vital to avalanche forecasters and researchers (McClung, 2002a, 2002b). Currently, gathering data about the spatial distribution and magnitude of avalanche events relies on human field observers and is consequently spatially and temporally limited. Furthermore, human observations are opportunistic, rarely cover an entire mountain range, cannot be collected during low visibility or high danger conditions, and are biased towards accessible locations and convenient collection 22 periods (Eckerstorfer and Malnes, 2015). Recent research has shown promising results with satellite-based avalanche detections as an alternative method to provide near real-time, consistent data about avalanche occurrences (Bianchi et al. 2019; Leinss et al. 2020). Satellite based detections provide a valuable supplement to field-based observations due to their large spatial coverage and the consistency with which all avalanches are observed, allowing for more reliable comparisons in space and time. A form of satellite sensor that has demonstrated specific utility in detecting avalanche debris is Synthetic Aperture Radar (SAR) (Eckerstorfer et al., 2016). SAR sensors utilize the microwave spectrum (wavelengths: 1- 300mm) and provide higher resolution imagery than passive microwave sensors. Since SAR sensors utilize the microwave spectrum they are not limited by cloud coverage or polar darkness, unlike visible- based sensors (Eckerstorfer et al., 2016). SAR based detections of avalanche debris are a powerful tool in hazard analysis. However, without a clear understanding of the main concepts behind this method the limitations and potential of the technique are hard to understand. Consequently, a brief discussion of these key concepts is presented below. 23 2 Background 2.1 SAR Background 2.1.1 Microwave's Interaction with Snow Cover SAR sensors operate by emitting microwaves and measuring the phase and intensity of energy returning (backscattered) to the SAR sensor (Wynne and Campbell, 2011). Eckerstorfer and Malnes (2015a) propose a conceptual model for how SAR microwaves interact with dry avalanche debris (Figure 6). Snow- free ground (Figure 6a), for incidence angles further from zero, tends to specularly reflect most incident energy away from the sensor. Once this ground has been covered by snow (Figure 6b) the snowpack increases the diffuse scattering at both the snow-surface and volumetrically within the snowpack. This increases the backscatter as less energy is reflected away from the sensor by the ground. Finally, if avalanche debris covers the pre-existing snowpack (Figure 6c), this leads to a dramatic increase in backscatter due to the increased surface roughness of avalanche debris relative to undisturbed snow and due to the increased volume of snow. This relationship between avalanche debris and increases in backscattered energy for SAR imagery has been established in multiple studies (Wiesmann et al. (2001), Malnes et al. (2013), Eckerstorfer and Malnes (2015), Leinss et al. (2020)). 24 Figure 6: Conceptual model for the interaction between dry avalanche debris and SAR microwaves. Adapted from Eckerstorfer and Malnes (2015a). 2.1.2 Temporal Change Analysis SAR temporal change analysis compares SAR images captured from the same orbital geometries at two different times to identify regions of increased or decreased intensity of backscattered energy (Eckerstorfer and Malnes, 2015). By examining the lower portions of avalanche paths for increases in backscatter, we identify avalanche debris. 25 SAR temporal change analysis is the fundamental method utilized in this study to identify avalanche debris. We start by examining the more recent (activity) image for regions with increased backscattered energy relative to the previous (reference) image. Researchers often use RGB composite images to identify these changes in backscatter. We place the reference in the red and blue channels and the activity image in the green channel. Increased backscatter areas then appear as green pixels and zones of decreased backscatter in purple (Figure 7). Analyzed SAR images must be from the same orbital geometries. Matching orbital geometries of the SAR images assures that changes in backscatter result from changing scattering properties across the landscape instead of different incidence angles (the angle between the incoming microwave energy and normal to the ground surface) (Wynne and Campbell, 2011). Incidence angles nearer zero degrees have high backscatter because the ground surface reflects energy directly back to the sensor. In contrast, slopes with high incidence angles have lower backscatter because energy reflects off the ground surface and away from the sensor. Therefore, we only compare SAR images from the same orbital geometry, so any backscatter increase is due to changing snowpack scattering properties rather than reduced incidence angle. 26 Figure 7: Reference, activity, and RGB composite images for two field reported avalanches in the BTAC region. Red dots show the field reported crown locations. Temporal change analysis compares the reference image to the activity image to identify increased backscatter areas in the avalanche paths (marked in orange). Only the specific avalanches reported were highlighted, but more avalanche debris is present in the images. 27 2.1.3 Image Artifacts SAR sensors operate in a side-looking geometry. These sensors emit energy at an angle across the terrain instead of at nadir (i.e., vertical). This side-looking geometry results in image artifacts. Two common artifacts are radar shadow and radar layover. Radar shadow is a terrain-induced phenomenon where steep topography blocks the incoming microwaves. Since there is no energy illuminating the region behind the steep topography, no backscatter returns to the sensor. This no-data region means that avalanche debris cannot be identified (Wynne and Campbell, 2011) (Figure 8). Radar layover occurs on slopes with incidence angles close to zero and causes slopes facing the SAR sensor to foreshorten towards the sensor. For steep slopes, the SAR sensor's emitted energy returns from the top of the hillslope before energy from the bottom. Since the energy returns from the top of the hill first, the sensor incorrectly identifies the slope's top as closer to the sensor than an accurate representation would show (Wynne and Campbell, 2011). While terrain correction can minimize this effect, layover still results in overexposed slopes that prevent avalanche debris detections (Figure 8). Fortunately, SAR sensors capture the same area from different orbital geometries and flight directions, which means that images from other geometries capture the region without layover. 28 Figure 8: a) SAR image showing radar shadow and layover in mountainous terrain. b) conceptual diagram showing the areas affected by radar shadow and layover. 2.2 Previous Research SAR-equipped satellites have been used to detect avalanche debris since the early 2000s (Wiesmann et al., 2001). However, the recent introduction of freely available imagery from the Sentinel-1 constellation has vastly expanded opportunities for SAR avalanche detection. The 29 Sentinel-1 constellation is a pair of polar-orbiting, sun-synchronous satellites with a maximum 6- day revisit interval (down to 1 day at higher latitudes). Both satellites are equipped with C-band SAR sensors (wavelength ~5.54 cm) that image at 10-meter pixel spacing. The sensor can operate in multiple polarization, VV (vertical transmit, vertical receive), VH (vertical transmit, horizontal receive), HV (horizontal transmit, vertical receive), and HH (horizontal transmit, horizontal receive) modes. Multiple studies exploit Sentinel-1 imagery to map avalanche cycles and create automatic detection systems (Malnes et al. 2015, Abermann et al. 2019 , Eckerstorfer et al. 2019; Leinss et al. 2020). These studies identify avalanche debris using the temporal backscatter change analysis discussed earlier. While backscatter change detection is valuable for detecting avalanche debris, important questions remain about the technique. How many avalanches are manually (i.e., using SAR avalanche expert interpretation) detectable with Sentinel-1 backscatter change analysis? Without a thorough understanding of the benchmark avalanche detection rate and how it varies as a function of topographic and meteorologic variables, we cannot accurately communicate the global utility of this technique or accurately assess automated detection systems. Eckerstorfer et al. (2019) provide one of the only studies comparing manual detections using SAR backscatter change detection against avalanches field observations. Their study compared field datasets of avalanches against manual and automated detections. They reported an overall probability of detection (POD) of 77.3%. The POD is the ratio of correct detections to total number of avalanches. They reported multiple factors correlated with detection rates including destructive size, avalanche release type, and avalanche wetness. Smaller avalanches, destructive size (D) of 1.5 30 and smaller, were detected at a rate of 64.9%, and larger avalanches (>D3) 100% of the time. The destructive size measures the destructive potential of an avalanche ranging from D1 (harmless to humans) up to D5 (capable of destroying an entire village) (Greene et al., 2016). This difference is due to large avalanches producing more extensive and deeper debris piles, thereby increasing backscatter over a larger area. Dry avalanches are also harder to detect. A possible explanation is that higher surface roughness of wet avalanche debris causes increased backscatter. Detection rates in Eckerstorfer et al. (2019) are from a field area with a maritime snow climate where 90% of the reported avalanches were wet slab avalanches. Thus, it is unclear how well this technique works in continental or transitional snow climates (Mock and Birkeland, 2000) dominated by dry snow avalanches. By exploring avalanche data from the western United States, this study provides essential information on the utility of SAR change detection analysis for mountain ranges in other snow climates. Leinss et al. (2020) provides an important first exploration of the utility of Sentinel-1 in transitional snow climates but was limited by a lack of field observations to compare to detections. This study compared avalanche detections from Sentinel-1 against costly 3m resolution TerraSAR-X imagery. They looked at a major avalanche cycle from January 2018 in the swiss alps and were able to manually delineate 7361 possible avalanche features in Sentinel- 1. Compared to the limited subset of 164 detections from TerraSAR-X detections Sentinel-1 missed around 1/3 of the avalanches, mostly due to their smaller size. They proposed that 2000m2 was the smallest detectable avalanches with Sentinel-1. For avalanche type, they could not know the exact types of avalanches due to the lack of field data. However, they did note that 31 avalanches that descended “below the wet-snow line were much more visible than avalanches from the dry snow-zone”. Finally, they suggested that higher incidence angles should provide higher detection rates but could not confirm without field observations. The important shortcoming of this powerful study was the lack of a field observed dataset to validate detections. This lack of ground data meant that calculations of the probability of detections was “difficult or even impossible”. The understanding of how many avalanches are even potentially detectable with this technique is valuable context in transitional snow climates. Our study hopes to provide this important knowledge about the utility of this technique in transitional snow climates in general and in the western United State specifically. 2.4 Research Questions The purpose of this study therefore is to compare Sentinel-1 SAR manual avalanche detections to field-observations of primarily dry avalanches in two different regions of the western United States. We address the following questions: 1. What are the manual detection rates of avalanches using change detection techniques on Sentinel-1 SAR backscatter images in the transitional snow climates of Utah and Wyoming, U.S.A.? 2. How do topographic and avalanche characteristics affect SAR avalanche detection rates? 32 3 Methods 3.1 Datasets Used 3.1.1 Field Observed Avalanches We used records of field-observed avalanches from the Utah Department of Transportation (UDOT) and the Bridger Teton Avalanche Center (BTAC). The BTAC records covered their avalanche forecasting region near Jackson, Wyoming. UDOT avalanches were from Little Cottonwood Canyon outside of Salt Lake City, Utah (Figure 9). These records included location, date, avalanche type, and destructive size. We filtered avalanche type to identify avalanches with wet or dry debris. We also divided avalanche type into loose (avalanches resulting from cohesionless surface snow) or slab (resulting from failure of an underlying weak layer) avalanches (McClung and Schaerer, 2006). We examined only avalanche records between 2016-2020, covering the operational period for both satellites of the Sentinel-1 constellation. 33 Figure 9: Avalanche locations for the a) BTAC (n = 136) and b) UDOT (n =143) avalanche databases. We calculated avalanche activity indexes (AAIs) for each of the databases to identify significant avalanche cycles. The AAI is the sum of observed avalanches weighted by their destructive size (D1: 0.01, D2: 0.1, D3:1, D4:10) for each day (Figure 10) (Schweizer et al. 2018). Our study investigated the two most significant cycles from the BTAC (a’s AAI: 120, b: 41) and UDOT (c: 80, d: 111). We downloaded Sentinel-1 images that spatially overlapped all four avalanche cycles. Due to the 6-day revisit interval of the Sentinel-1 constellation, the image pairs with matching orbital geometries (same path #) were all 6, 12, or 24 days apart. 34 Figure 10: Avalanche activity index for the two databases from 2016-2020 with the two largest cycles for each region (four total) noted. The 99th percentile of the avalanche activity index for UDOT (73) and BTAC (35) shows these events’ extreme nature. Insets show each avalanche cycle with capped lines showing the imagery dates of all utilized Sentinel-1 image pairs. Note that insets a-d have different y-axes. We analyzed all the avalanches that released between a pair of Sentinel-1 images. Consequently, we analyzed 279 avalanches between the two regions. There were 143 UDOT avalanches. Of those avalanches, 126 (88.1%) were dry, and 17 (11.9%) were wet. One hundred and twenty-four (86.7%) were slab avalanches, and 19 (13.3%) were loose avalanches. Forty-two (29.4%) were natural avalanches, and 101 (70.6%) were artificially triggered. For the BTAC dataset, there were 136 avalanches. Of those, 135 (99.3%) were dry, and one was wet (0.7%). One hundred and thirty-five (99.3%) were classified as slabs and one as loose (0.7%). Ninety-four were naturally triggered (69.1%), and 40 (30.9%) were artificially 35 triggered. The destructive size distribution of the UDOT and BTAC was primarily centered on D2 through D3 (Figure 11). Figure 11: Avalanche destructive size distributions for each database. Cross-hatching indicates loose avalanches. 3.1.2 Sentinel-1 Imagery We used imagery from the Sentinel-1 constellation, courtesy of the European Space Agency and the European Commission. This study utilized Sentinel-1 images that were high resolution (~20 x 22.5 meter) with VV and VH bands. We downloaded Sentinel 1 backscatter intensity images from Google Earth Engine (GEE; Gorelick et al. 2017) and processed them as follows: 1. GEE Sentinel-1 imagery is preprocessed via the application of a precise orbit file, removal of border and thermal noise, application of radiometric calibration, and terrain correction. Specific information on this preprocessing is found at https://developers.google.com/earth- engine/guides/sentinel1. 36 2. For each Sentinel-1 image, we created and applied layover and shadow masks using the United State Geophysical Survey's (USGS) National Elevation Dataset at 1/3 arc-second resolution (Gesch et al., 2002). During this step, we also extracted local incidence angles for each pixel. This processing workflow used a python script adapted from Vollrath et al. (2020) (Appendix A). We began by selecting pairs of Sentinel-1 images with matching orbital geometries that spatially overlapped the reported avalanche datasets. This pair of images included: a reference image, that preceded the avalanche cycle, and an activity image, from after, with the same orbital geometry. Matching orbital geometry implies images with the same track # and orbit direction (ascending or descending). The orbit direction expresses whether the satellite traveled from the south pole to the north – ascending or north to south – descending. We used 14 Sentinel-1 image pairs for this study (Table 1). 37 Table 1: Avalanche Cycles and Imagery Sets Used Activity Image Date Reference Image Date Orbit Direction Path # BTAC February 5th, 2018 Cycle February 8th, 2018 January 27th, 2018 Ascending 49 February 12th, 2018 January 31st, 2018 Descending 27 February 17th, 2018 1 January 24th, 2018 1 Descending 100 January 14th, 2020 Cycle January 14th, 2020 January 2nd, 2020 Descending 100 January 17th, 2020 January 5th, 2020 Ascending 49 January 21st, 2020 January 9th, 2020 Descending 27 January 22nd, 2020 January 10th, 2020 Ascending 122 UDOT January 2nd, 2020 Cycle January 3rd, 2020 December 22nd, 2019 Ascending 20 January 9th, 2020 December 28th, 2019 Descending 27 January 10th, 2020 December 29th, 2019 Ascending 122 February 7th, 2020 Cycle February 7th, 2020 January 26th, 2020 Descending 27 February 8th, 2020 January 27th, 2020 Ascending 20 February 14th, 2020 February 2nd, 2020 Descending 27 February 15th, 2020 February 3rd, 2020 Ascending 122 1 Only pair with 24 days instead of 12 days between imaging dates 38 3.2 Processing and Analysis This study utilized two primary input sources (historical records of field observed avalanches and Sentinel-1 imagery) along with three ancillary sources (Global Forest Cover Change Tree Cover, USGS National Elevation Dataset, and auto-Avalanche Terrain Exposure) to identify avalanche paths, generate manual detection of field observed avalanches, and extract ancillary data about the avalanches (Figure 12) Figure 12: Flowchart of data collection and processing steps. 39 3.2.1 Avalanche Path Identification We identified the outline of the avalanche paths for every field observed avalanche (whether detected or not). We used these avalanche paths, defined as the maximum possible affected area from a specific starting zone, to find path lengths and widths and determine the zones to extract other explanatory variables. We used a pre- existing avalanche atlas for the UDOT avalanches. This atlas defined the area affected by a 100- year avalanche (McClung and Schaerer, 2006). An avalanche atlas was not available for the BTAC center, so we applied expert interpretation of likely avalanche paths based on locations, and descriptions of the avalanches supplemented with automatic Avalanche Terrain Exposure maps of the regions (Larsen et al., 2020). We defined the starting zones and tracks by followed topographic constraining features (ridgelines). We drew the runout extent using auto-ATES maps and ending at the boundary of "challenging" and "simple" terrain, defined by an 23-degree alpha angle from the nearest starting zone (Figure 13) (Larsen et al. 2020b, McClung and Schaerer 2006). The auto-ATES maps were generated using alpha angles from the nearest starting zones but did not incorporate vegetation or use physically based modeling of likely avalanches. We drew these paths before SAR debris detections to avoid any biasing of the path characteristics. 40 Figure 13: An annotated example of the avalanche path identification process for the BTAC avalanche paths. 3.2.2 Manual Avalanche Detection For every field observed avalanche, we inspected the avalanche path for increased backscatter areas in either the VV or VH polarization. These increases had to morphologically match the description and photos of the field observed 41 avalanche. If we detected backscatter increases of at least 2-3 decibels (dB) based on Eckerstorfer and Malnes (2015), we marked the avalanche as "Detected." We used a checklist of factors to identify avalanche debris, which included: increased backscatter relative to the previous image in the avalanche path, similar shape of the increased backscatter to the observed avalanches, and higher backscatter values in the debris zone of the path relative to surrounding snow (Figure 14). The UDOT dataset included multiple reported avalanches per avalanche path, while the BTAC dataset did not. These avalanches were either all marked "Detected" or "Not detected." The average avalanches per path in the UDOT dataset was 2.3. Figure 14: RGB Composite image of a detectable avalanche (reported on February 6th, 2018) from the BTAC Dataset. The red and blue channels are a reference image (January 31st, 2018), and the green channel is an activity image (February 12th, 2018). Areas of increased backscatter appear green. 42 3.2.3 Supplementation of Avalanche Records We extracted possible explanatory variables for every avalanche event. The avalanche databases from the BTAC and UDOT contained standard SWAG avalanche descriptive data, including dates, D-size, wet vs. dry debris, slab vs. loose type, and natural vs. artificial release for all included avalanche observations (Greene et al., 2016). To explore topographic effects on detection, we extracted: average elevation of the avalanche path, the average slope, average profile curvature, path length, and path width (using a minimum bounding geometry on the paths from step 3.2.2) (Gesch et al., 2002). The Global Forest Cover Change Tree Cover (GFCC) Multi-Year Global Dataset provided the average tree coverage within each avalanche path (Sexton et al., 2013). C- band SAR is heavily affected by tree cover, so we chose this variable to explore if tree cover in the debris zone negatively affected the detection rates (Eckerstorfer et al., 2016). Finally, we extracted the percentage of the path affected by radar layover, percentage affected by radar shadow, lag time (the number of days between the avalanche observation and the acquisition of the activity image), and the average local incidence angle within each avalanche path from the Sentinel-1 image pair (Table 2). 43 Table 2: Explanatory Factors with data sources, spatial resolutions, and shortened names. Explanatory Factor Source Resolution Alias Parameter Range Destructive size of avalanche BTAC, UDOT D-size 1-4.5 Avalanche release type BTAC, UDOT Type Slab, Loose Avalanche wetness BTAC, UDOT Type Dry, Wet Mean tree coverage percentage in GFCC 30m Tree Perc 0 - 47% the path Average elevation of avalanche USGS NED 10m Elevation 1857- 3823 m path Average slope in the path USGS NED 10m Slope Angle 15 - 55° Avalanche path length USGS NED 10m Path Length 0.19 - 3.2 km Avalanche path width USGS NED 10m Path Width .13 - 1.8 km Average profile curvature in the USGS NED 10m Curvature 0 – 0.175 path Average local incidence angle USGS NED, Incidence 20m 11 - 83° within the path Sentinel-1 Angle Percentage of the path affected by Sentinel-1 20m Layover Perc 0 - 0.77 radar layover Percentage of the path affected by Sentinel-1 20m Shadow Perc 0 - 0.54 radar shadow Days between avalanche and the Sentinel-1 1 day Lag Days 0-11 activity image acquisition 44 3.2.4 Statistical Analysis For our statistical analysis, we used the entire dataset and the two regional subsets. We calculated the covariance of our quantitative variables using Spearman's rank correlation test and our categorical variables' covariances using chi-squared contingency tables (Wilks, 2019). If two variables were strongly correlated (p <0.05 or |rs| > 0.5), we retained the variable that appeared most directly in the conceptual model of avalanche debris backscatter (Figure 6) (Eckerstorfer and Malnes, 2015). We then identified explanatory variables with significant variance between the detectable and undetected avalanches. We used thresholds of p <0.05 as significant and p<0.1 for marginally significant. We analyzed categorical values using chi-squared values. We performed Shapiro Wilk’s test for normality on all quantitative variables. Since all variables failed the test for normality (results in appendix A) we performed the non-parametric Kruskal Wallis (KW) tests to assess for significant variation in means between detected and non-detected avalanches (Wilks, 2019). Finally, we wanted to investigate the dataset for non-linear variable importance. Consequently, we trained a random forest model to identify detectable vs. non-detectable avalanches (Breiman, 2001). The random forest used the entire dataset with 10-fold cross- validation (n = 545). Additionally, we inserted a random variable to assess which variables were more important than a random variable. We tuned the hyperparameter based on maximizing validation accuracy to optimize the number of decision trees (250) and the network’s max depth (7). The random forest's performance was assessed with an accuracy score and F1 score, a 45 weighted average of precision and recall (Liu et al., 2014). We finished by extracting feature importance from the trained random forest. 46 4 Results There were 274 detected avalanches and 271 undetected avalanches in the overall dataset for an overall detection rate of 50%. However, this detection rate varied significantly between image pairs and between the two field locations. 4.1 Image Pairs and Field Location Detection Rates The BTAC had a total detection rate of 40% and the UDOT was 58%. As mentioned earlier, the UDOT avalanches had multiple avalanches reported per path. This meant we could not know which avalanche’s debris we were noting in the Sentinel-1 images, so we marked all avalanches in the path as “Detected” or “Not Detected”. To assess the impact of these multiple reported avalanches in a path on detection rates, we extracted UDOT avalanches events that had only one avalanche in the path (n = 47). The detection rate of this subset of single UDOT avalanches was 43%. Detection rates for individual image pairs for both regions ranged from 15% up to 87%. Ascending image pairs averaged 53%, and while descending image pairs averaged 38%. We extracted the following explanatory variables for image pair variability: the number of avalanches, average destructive size of avalanches, and the average local incidence angle (Table 3). 47 Table 3: Image Pairs with Detection Rates. Darker colors show larger incidence angles or detection rates. State Ref Date Act Date Direction Orbit # Avalanches Mean D-Size Incidence Angle Detection % Bridger Teton 1/24/18 2/5/18 Descending 100 19 1.74 37 20 Bridger Teton 1/27/18 2/8/18 Ascending 49 32 2.52 49 48 Bridger Teton 1/31/18 2/12/18 Descending 27 45 2.63 39 43 Bridger Teton 1/2/20 1/14/20 Descending 100 41 2.11 36 19 Bridger Teton 1/5/20 1/17/20 Ascending 49 28 2.16 49 39 Bridger Teton 1/9/20 1/21/20 Descending 27 41 2.45 44 54 Bridger Teton 1/10/20 1/22/20 Ascending 122 40 2.46 53 48 UDOT 12/22/19 1/3/20 Ascending 20 46 2.46 58 87 UDOT 12/28/19 1/9/20 Descending 27 18 2.44 47 50 UDOT 12/29/19 1/10/20 Ascending 122 42 2.52 50 52 UDOT 1/26/20 2/7/20 Descending 100 58 2.03 45 66 UDOT 1/27/20 2/8/20 Ascending 20 62 2.42 58 82 UDOT 2/2/20 2/14/20 Descending 27 26 2.73 50 15 UDOT 2/3/20 2/15/20 Ascending 122 47 2.66 50 19 Multiple pairs of Sentinel-1 images captured individual field observed avalanches. Of the 279 field reported avalanches, 89 (32%) fell between only 1 pair of Sentinel-1 images, 118 (42%) between 2, 65 (23%) between 3, and 7 (3%) between 4. Of the 190 avalanches captured between more than 1 image pair, the average percentage of images with a successful detection was 64%. 4.2 Destructive Size Destructive size has previously been established as a controlling variable on detection rate (Eckerstorfer et al., 2019). We also found a strong relationship between increased detection rates and larger avalanches (Table 4). 48 Table 4: Detection Rates by destructive size. Darker green represents higher detection rates. Destructive Size # Detected # Not Detection Rate 1 4 23 15% 1.5 26 43 38% 2 78 100 44% 2.5 60 47 56% 3 56 41 58% 3.5 17 2 89% 4 31 15 67% 4.5 2 0 100% 4.3 Covariation Analysis We present graphs of variable correlations in appendix A. We evaluated quantitative variable covariance with a Spearman's rank correlation test. Three pairs of variables showed correlations over our threshold (p <0.05 or |rs| > 0.5): 1) the percentage of the path in shadow with local incidence angle (r = 0.65), 2) average elevation with tree percentage (r = 0.50), and 3) path width with path length (r = 0.56). We retained path length due to its higher correlation with detection rates. We also kept tree percentage due to its higher correlation to detection rates relative to average elevation. For the percentage of the path in shadow vs. local incidence angle – we kept local incidence angle because the percentage of path in shadow was less correlated to detection rates and had a nonintuitive sign of correlation. As the percentage of the path in shadow increased, the detection rates also increased, which is not a logical trend for these two variables. This nonintuitive sign suggested the higher detection rates with the increasing percentage of path in radar shadow resulted from some correlated variable with shadow rather than a causal relationship between percentage of shadow and probability of detection. 49 4.4 Kruskal-Wallis (KW) Since our variables were not normally distributed (appendix A) we performed the non- parametric KWs on the full dataset and separate subsets for BTAC and UDOT regions (Table 5). Table 5: KW Results for Full Dataset, UDOT, BTAC. For all marginally significant or significant results, the detected and undetected mean relative values are noted below the p-value. Darker green shows higher significance. variable D-Size Tree Perc Incidence Angle Slope Curvature Path Length Layover Perc Lag Days Full Dataset p-value 3.10E-09 0.69 6.60E-07 0.0013 0.62 6.70E-19 0.86 1.90E-05 det > undet det > undet det > undet det > undet det < undet Bridger Teton Avalanche Center p-value 1.20E-12 0.77 0.0056 0.51 0.2 3.50E-08 0.25 0.43 det > undet det > undet det > undet Utah Department of Tranportation p-value 0.22 0.09 0.0011 0.0013 0.062 5.60E-11 0.31 3.40E-05 det > undet det > undet det > undet det < undet The variables with significant variance in the complete dataset were destructive size, incidence angle, path length, slope, and lag days (Figure 15). Our regional datasets showed similar patters to the overall dataset. 50 Figure 15: Boxplots showing the means, interquartile ranges, whiskers, and outliers for detected and undetected avalanches in the full dataset. Whiskers are 1.5 times the interquartile range above and below the quartiles, and outliers are data points beyond the whiskers. 4.5 Chi-Squared Analysis Dry (n = 509) vs wet (n = 36) avalanches did not have statistically significant differences (p = 0.31) between detected and undetected avalanches. The detection rate difference between slab 51 (n = 508) vs loose (n = 37) avalanches were marginally statistically significant (p = 0.066). Loose avalanches were easier to detect despite being smaller on average (1.77 vs. 2.43). Of note is that of the loose avalanches recorded, 34 were wet, and 3 were dry avalanches. 4.6 Random Forest Following hyperparameter tuning, the random forest attained a validation accuracy of 67% and an F1 score of 62% with 10-fold cross-validation. Random forest results aligned with the findings from the KW analysis with path-length, incidence angle, lag days, and D-size as the most significant variables (Table 6). Table 6: Feature individual and cumulative importance in the random forest model. Green indicates significant variables (>8%), yellow is marginally important, and red is less effective than a random variable. Feature Individual Importance Cumulative Path Length 27.47% 27.47% Incidence Angle 16.49% 43.96% Lag Days 12.82% 56.78% D- Size 8.92% 65.70% Curvature 7.60% 73.30% Slope 7.60% 80.90% Tree Precentage 6.78% 87.68% Layover Percentage 5.90% 93.58% Random Variable 4.63% 98.21% Natural/Artifical 0.07% 98.28% Slab/Loose 0.07% 98.35% Dry/Wet 0.02% 98.37% 52 5 Discussion 5.1 How many avalanches was Sentinel-1 backscatter temporal change analysis able to detect? Our techniques detected 50% of the field observed avalanches (n = 545). Thus, about half of all avalanches will be missed when detecting primarily dry snow avalanches. However, when considering avalanches equal to or larger than D3 the manual detection rate rises to 65%. This technique is effective for most avalanches large enough to do substantial damage to infrastructure or vegetation. However, we still miss around 35% of even these large avalanches. 5.1.1 How did detection rates vary between the two regions? The detection rate for the UDOT avalanches (POD = 53%) is substantially higher than for the BTAC avalanches (POD = 39%). This could be due to two factors. A possible explanation may be that the UDOT reported multiple avalanches within single paths (average 2.3 avalanches per path), which the BTAC dataset did not do. When we extracted only UDOT avalanches with a single avalanche in their path (n = 47) the detection rate (43%) dropped closer to the BTAC rate. This suggests the actual detection rates is similar for these two regions at around 40%.A second explanation is avalanche type, since 12% of the UDOT avalanches reported as wet (which previous studies have shown to have improved detection rates) compared with only 0.7% of the BTAC avalanches. 5.1.2 How did detection rates vary between image pairs? There were significant differences in the detection rates between Sentinel-1 image pairs, with the lowest detection rate being 15% and the highest being 87% (Table 3). This massive difference means that certain pairs of images are significantly better for detecting avalanche debris than others. Since there is no clear relationship between average destructive size for the image pairs and detection rates image 53 pairs are not skewed with anomalously small or large slides. However, the clear relationship between incidence angle and detection rates suggests these varying detection rates might be due to trends in the aspect of the active avalanche paths relative to the SAR sensor. Ascending image pairs had higher detection rates (53%) compared to descending image pairs (38%), and ascending image pairs also had higher average incidence angles (52.4°) relative to descending image pairs (42.5°). In ascending images, more paths that avalanched in this study faced away from the sensor (east facing) leading to higher incidence angles and potentially detection rates. As explored in section 4.2.3, there is a strong relationship between high incidence angles and higher detection rates in this study. This aspect dependence suggests that depending on the predominant aspects of avalanches in a region, either ascending or descending, will yield higher detection rates. Specifically, to Wyoming and Utah where, due to prevailing wind direction and less solar warming, more avalanches are noted in field observations on east facing aspects. This suggests that ascending image pairs may have generally higher detections rates for this region. 5.2 What factors correlated with successful detections? This study demonstrates several factors that are important for detecting avalanches. The variables are highly significant in our KWs (Table 5) for one or both datasets are destructive size, tree percentage, lag days, average local incidence angle, average slope, and path length. The chi-squared analysis shows only a marginal significance for loose vs. slab avalanches, with loose being more detectable (Table 7), though this relationship is complicated by the increased detection of wet loose avalanches in comparison with dry loose avalanches. 54 Table 7: Summary of factors related to increased detection rates sorted by KW p-values. Factor Increased Detection Rates Section KW P-Value RF Importance Path length Longer path length 5.2.3 <0.001 1st Local incidence Increased local incidence angle angles 5.2.4 <0.001 2nd Lag Days Fewer days elapsed 5.2.9 <0.001 3rd Destructive size Larger destructive size 5.2.1 <0.001 4th Slope angle Steeper average slope angle 5.2.6 0.0013 6th Curvature More convex curvature 5.2.7 0.062 (UDOT only) 5th Avalanche Less than random Type Loose avalanches 5.2.2 0.066 variable Wet avalanche more Avalanche detectable relative to Less than random Wetness previous research 5.2.2 >0.1 variable Layover Qualitative examples of Percentage missed detections 5.2.5 >0.1 8th Tree Coverage 5.2.8 >0.1 7th 55 5.2.1 Destructive Size Our results show a strong and positive correlation between the field reported destructive size and Sentinel-1 based probability of detection (Table 4), matching previous results reported by Eckerstorfer et al. (2019). Their study found PODs around 60% for size D1.5, 80%+ for avalanches of D2 to 2.5 and 100% detection rates for all avalanches over D3. Our results are significantly lower, with a 64% detection rates for avalanches D3 and larger. A probable explanation is our dataset was almost entirely dry avalanches, unlike the dataset in Eckerstorfer et al. (2019) that was over 90% wet avalanches. The correlation between destructive size and probability of detection is likely due to the spatial resolution of the Sentinel-1 sensor. Larger avalanches will typically result in wider and deeper debris piles. These expanded debris piles are easier to detect with the relatively coarse resolution of the Sentinel-1 sensor. Leinss et al. (2020) suggested a smallest detectable size of 1000m2 for avalanches in their Sentinel-1 automated avalanche detection system. A key takeaway is that avalanche inventories using Sentinel-1 are biased towards larger avalanches. Any future avalanche inventories or automated avalanche detection system should take this into account. Higher resolution sensors may improve the ability to detect smaller avalanches, but this technique likely continue to overrepresent larger avalanches. Fortunately for many applications, especially if infrastructure damage is the primary concern, accurate mapping of more significant slides will be sufficient. 5.2.2 Avalanche Type Our field observed avalanches dataset contained very few loose avalanches (n = 37) and very few wet avalanches (n=20). This made an analysis of the connection between avalanche type and detection rates problematic. Our random forest found avalanche type to be less significant than a random variable and the only avalanche type factor 56 with a significant chi-squared value is slab versus loose avalanches. This relationship for loose and slab avalanches is probably because 34 of the 37 loose avalanches in our dataset were wet and previous research has shown that wet avalanches tend to be easier to detect (Eckerstorfer et al., 2019). Further exploration with a more diverse avalanche dataset will be necessary to constrain avalanche type importance on detectability. When compared to previous studies, our results provide strong evidence that dry avalanches are more difficult to detect than wet avalanches. Our dataset was over 90% dry avalanches and consistently found lower POD rates than those in previous research on the detection rates for wet slabs (Eckerstorfer et al. 2019). Consequently, this technique may be best suited to regions dominated by wet avalanches and should include these lower detection rates as context when applied to continental and transitional snow climates. An explanation for these lower detection rates may be that wet avalanches debris is typically rougher than dry avalanche debris. This more irregular surface will backscatter more energy towards the Sentinel-1 sensor leading to increased backscatter changes. Additionally, the higher density of wet snow debris may also increase volumetric backscatter relative to lower density dry debris. Finally, the softer debris of dry avalanches and the availability of loose snow for wind redistribution might lead to surface smoothing and possibly missed detections. Regardless of the reasons, this study provides evidence for lower detection rates in regions dominated by dry slabs. 5.2.3 Path Length While previous research established that destructive size affected detection rates, this study found a stronger relationship between path length and detection rates (KW p-value <0.001 and first in the random forest feature importance), with longer paths having higher detection rates. The first possible explanation is that path boundaries for the BTAC were 57 drawn in by expert interpretation, which may have introduced error and bias. However, this strong relationship persists in the UDOT avalanches, which utilized an independent avalanche atlas for path lengths, suggesting this is not the whole story. Secondly, longer paths may produce larger avalanches. However, we did not find path length to be significantly correlated with destructive size. This lack of correlation again suggests that path length is an independent factor in detection rates. Finally, longer paths have a longer distance for debris to run before starting to slow, which may have increased the surface roughness of the debris by causing wetter debris through particle collisions. A recent study that mechanically synthesized snow avalanche flows found significant warming of avalanche debris with longer flow times (Fischer et al., 2018). Regardless, if this relationship persists in future research, our work suggests that certain regions or topographies with exceptionally long path lengths might have higher detection rates. 5.2.4 Local Incidence Angle The average local incidence angle was significant in the KW comparisons for every dataset and was the second most important feature in our random forest model. This relationship between high local incidence angle and high detection rates could be based on a few factors. First, due to the SAR images' geometry, there is higher resolution in areas where the incidence angle is farther from zero. Also, this limits the effects of layover, which can obscure avalanche debris piles. Finally, Leinss et al. (2020) proposed that slopes facing away from the radar should exhibit increased scattering from the rough surface of the avalanche debris and less specular reflection from the ground surface – both of which increase the backscatter contrast of avalanche debris (Figure 16). 58 Figure 16: Proposed conceptual model explaining the increased detection rates of avalanche debris at higher incidence angles While incidence angle effects have been theorized, they have not been previously demonstrated. Our results show a clear difference in local incidence angle between avalanches that are detected and not detected. The significantly higher local incidence angles (Figure 15), KW results (Table 5), and random forest feature importance (Table 6) suggest that this effect is significant. This improved understanding of the effect of incidence angles means future research should utilize spatial averaging methods that prioritize areas of high local incidence angles. Future applications could also use this knowledge to differentiate paths with a high probability of detection from relatively low detectability probability. Identifying paths with a high probability of detection is valuable for suggesting when avalanches are not occurring instead of just being missed by this technique. 5.2.5 Lag days The final variable explored was the elapsed days between the avalanche field observation and the acquisition of the Sentinel-1 activity image. This variable was highly significant (p <0.0001) for our complete and UDOT datasets, with fewer lag days correlated with 59 higher detections. Additionally, it was the 3rd more important variable in our random forest. This connection between fewer days elapsed and higher detection rates makes intuitive sense. The increased number of days leads to greater exposure to ablative meteorological factors such as: 1) increased temperatures that might reduce the chances of detection due to melting and smoothing of the rough avalanche surface, 2) increased cumulative precipitation, which might cover the rough surface of the debris and lead to a less dramatic contrast between the increased volumetric backscatter from the deeper avalanche debris and the surrounding snowpack, and 3) more exposure to wind redistribution, which might smooth the avalanche debris surface. These factors and other weather factors might explain the strong correlation between increased lag days and the detectability of avalanche debris. In regions with less frequent Sentinel-1 imaging (such as lower latitudes and outside of the European Union), atmospheric weathering will likely cause lower detection rates. 5.2.6 Layover Percentage The percentage of the path affected by layover did not significant KW results and was only slightly more important than a random variable in our random forest. Thus, our results do not provide strong evidence for a connection between layover effect and detection rates. However, layover can effectively obscure large portions of an avalanche path, leading to missed detections. Therefore, the conclusion that paths highly affected by layover would have lower detection rates seems reasonable. While this study did not find evidence for layover's impacts on detection, we noted a few examples (Figure 17). 60 Figure 17: A BTAC reported avalanche from February 5th (red dot) in an image pair unaffected by layover (a) and in an image pair affected by layover (b). 5.2.7 Slope Angle The average slope (across the entire path) was significant in the overall and UDOT dataset (p = 0.0013) and the 6th most important feature in our random forest. The average slope was 31.5 degrees for detected avalanches and 30.5 for the undetected avalanches. These results suggest a slight relationship between steeper slopes and higher detection rates but may also be due to steeper slopes generally having higher incidence angles in this study. 61 5.2.8 Curvature The curvature values displayed marginal correlation (p = 0.062) with detectability in our KW for the BTAC dataset and was the 5th most importance variable in our random forest. The mean curvature for non-detections was concave (mean = 0.0012) and the mean curvature for detected avalanches was convex (mean = -0.0040). This weak evidence suggests that convex slopes may have a higher probability of detections relative to concave. The curvature effect could be due to fewer terrain distortions (layover or shadow) with concave paths. Future analysis with a higher resolution digital elevation model may provide more conclusive evidence for this relationship. 5.2.9 Tree coverage The tree percentage was only marginally significance (p = 0.09) for the UDOT dataset and was only the 7th in our feature importance. This result is surprising given that trees have a considerable impact on C-band SAR backscatter and, if thick enough, can completely prevent scattering from beneath them. However, since avalanche paths are generally treeless, there may not be enough debris ending up beneath trees for this to be a significant factor. This study had a mean tree percentage of 12%, with a standard deviation of 10% and a maximum of 47%. Potentially a future investigation in a region with higher tree percentage coverage in the paths might show a more significant control on detection rates. 5.3 Limitations and future work This study assessed the detectability of avalanches but had several limitations to note that may help direct future research. First, our avalanches datasets covered only two regions that are both transitional snow climates. Future research focused on confirming the general detection 62 rates in these regions and snow climates will be necessary. Also, an analysis in continental and maritime snow climates is necessary to determine similarities or differences with this study. This research utilizes only four avalanche cycles and 545 avalanches. These four avalanche cycles were selected to maximize the number of avalanches in each pair of Sentinel-1 images but choosing these extreme events may have introduced other factors that skewed our results. Studies in the future with additional avalanche cycles may help to clarify if our findings hold for less extreme avalanche cycles. Additionally, choosing spring avalanche cycles will be valuable. Since our study only utilizes avalanches cycles from January and February, our results are biased towards dry avalanches (11.9% of UDOT and 0.7% of the BTAC avalanches were classified as wet). Confirming if these detection rates and controlling factors continue into the spring with different meteorological conditions and avalanche types would be valuable. Next, this study utilized avalanche records that were opportunistically collected. The BTAC records may have been incorrect due to non-expert public sources reporting. These records were potentially also skewed towards accessible locations, periods of high visibility, and larger events. Future studies could limit these biasing effects by performing randomized observations or confirming complete coverage of a region. This study also focused exclusively on backscatter analysis of a single sensor. Future studies should explore other SAR sensors for avalanche detection, operating at different bandwidths and spatial resolutions. An especially interesting SAR satellite is the NISAR satellite, launching in 2022, that will provide increased numbers of SAR images over the western US. These could then be combined with Sentinel-1 to improve detection rates. Additionally, other parameters from SAR sensors, not covered here, such as those produced by considering the 63 differences in backscattered intensity between polarization and phase changes of the returning energy, will be valuable additions. The concepts behind polarimetric decomposition and interferometric SAR analysis are too complex for this discussion, but Yang et al. (2020) provides an exciting study of polarization and interferometric analysis of avalanche debris. Finally, this study utilized manual detections, which is currently considered the gold standard of avalanche detections using SAR imagery (Bianchi et al., 2021). However, since only one person (the first author) performed the manual detections, there may have been biases that systematically or randomly affected the results. Future studies to confirm or refute these manual detection results will be useful. In addition, further advancements in automated detections will help move this technology in the direction of being a useful operation tool for avalanche forecasters. Currently, machine learning techniques show promise in automating detections, but continued work matching manual detection rates with complex image classifiers will be critically important. Reliable automated avalanche detections will allow for near-real-time detection systems and broader-scale monitoring of avalanche activity in the coming years. 64 6 Conclusion This study examined the manual detection rates relative to a field observed avalanche dataset from two locations in the western United States of primarily dry snow avalanches using change detection analysis in Sentinel-1 backscatter images. For the 545 avalanches examined in this study, 274 (50%) were manually detectable in the Sentinel-1 images. This detection rate increased to 65% when considering only avalanches large enough to bury cars or break trees (D3). Thus, the majority dry avalanches larger than >D3 can be detected by SAR-based change detection analysis, though this technique still missed about 35% of these large avalanches in our combined database. The primary variables showing a strong correlation with detection rates (both in our KW and random forest analysis) are local incidence angle, path length, destructive size, slope angle, and elapsed days between the avalanche and the activity image. We also established a weaker relationship for the percentage of path affected by layover and avalanche type (slab vs. loose). These variables were either only significant for a regional subset of the avalanches or only reached a marginally significant p-value. Avalanche wetness, tree coverage, average slope steepness, and the slope's curvature failed to reach a marginal significance. Further investigation with a more diverse avalanche dataset, higher resolution DEM and tree cover maps might reveal a connection between these factors and detection rates. Future applications of this technique could use these factors to identify specific storms or avalanche paths with a higher probability of detection. That understanding would give practitioners a better understanding of when SAR change detection analysis is likely to produce reliable avalanche debris detections. 65 Previous research on wet slab avalanches detected 100% of avalanches larger than D3 (Eckerstorfer et al., 2019). In comparison, this study of over 90% dry slabs found a much lower detection rate (64%) for comparably sized avalanches. While there are topographic and climatic factors that may be influencing these results, our results suggest that regions dominated by wet slabs may be better suited to Sentinel-1 temporal change analysis techniques. These findings provide further evidence about this method's utility for identifying avalanche debris, suggest controlling factors on the detectability of avalanche debris, and provide a baseline for comparisons between automated detection results and field observed avalanches. Our study is the first to provide a large-scale analysis of detection rates for Sentinel-1 images relative to field observed datasets in the continental United States, which has a markedly different topography and snow climate from previous avalanche detection studies. Future work should utilize larger datasets across more climatic and topographic regions that incorporate more sensors and techniques into the detection process. The SAR avalanche detections in the United States has the potential to provide avalanche practitioners with much-needed feedback and researchers with spatially extensive and relatively complete datasets of avalanche occurrences for machine learning techniques and trend analysis. While future research is necessary, there are regions that may be most suited to immediate applications of this technology including manual mapping of major cycles and automated near- real time detection techniques. This study suggests that regions dominated by wet avalanches (such as the maritime snow climates of California, Washington, and Alaska) may have quite high detection rates for avalanche D3 and larger. Higher latitudes (Alaska) with fewer lag days that could reduce surface weathering, especially for dry snow avalanches that are prone to wind and 66 sublimation, may also improve detection rates. These combined with the benefits of longer paths suggest that coastal Alaska may be a region well suited to applying automated detection methods and manual avalanche inventories over the coming years. 67 7 References Cited Abermann, J., Eckerstorfer, M., Malnes, E., Hansen, B.U., 2019. A large wet snow avalanche cycle in West Greenland quantified using remote sensing and in situ observations. Nat Hazards 97, 517–534. https://doi.org/10.1007/s11069-019-03655-8 Bianchi, F.M., Grahn, J., Eckerstorfer, M., Malnes, E., Vickers, H., 2021. Snow avalanche segmentation in SAR images with Fully Convolutional Neural Networks. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 75–82. https://doi.org/10.1109/JSTARS.2020.3036914 Birkeland, K., 2017. In Response to Avalanche Fatalities in the United States by Jekich et al. Wilderness and Environmental Medicine 28. https://doi.org/10.1016/j.wem.2017.06.009 Breiman, L., 2001. Random Forests. Machine Learning 45, 5–32. https://doi.org/10.1023/A:1010933404324 Bühler, Y., Hafner, E., Zweifel, B., Zesiger, M., Heisig, H., 2019. Where are the avalanches? Rapid mapping of a large snow avalanche period with optical satellites. The Cryosphere Discussions 1–21. https://doi.org/10.5194/tc-2019-119 Bühler, Y., Hüni, A., Christen, M., Meister, R., Kellenberger, T., 2009. Automated detection and mapping of avalanche deposits using airborne optical remote sensing data. Cold Regions Science and Technology 57, 99–106. https://doi.org/10.1016/j.coldregions.2009.02.007 Chabot, D., 2017. Avalanche Forecasting in Afghanistan, Pakistan and Tajikistan. The Avalanche Review 35, 30–31. Colorado Avalanche Information Center, 2021. US Avalanche Accident Reports. Eckerstorfer, M., Bühler, Y., Frauenfelder, R., Malnes, E., 2016. Remote sensing of snow avalanches: Recent advances, potential, and limitations. Cold Regions Science and Technology 121, 126–140. https://doi.org/10.1016/j.coldregions.2015.11.001 Eckerstorfer, M., Malnes, E., 2015. Manual detection of snow avalanche debris using high- resolution Radarsat-2 SAR images. Cold Regions Science and Technology 120. https://doi.org/10.1016/j.coldregions.2015.08.016 Eckerstorfer, M., Vickers, H., Malnes, E., Grahn, J., 2019. Near-Real Time Automatic Snow Avalanche Activity Monitoring System Using Sentinel-1 SAR Data in Norway. Remote Sensing 11, 2863. https://doi.org/10.3390/rs11232863 Fischer, J.-T., Kaitna, R., Heil, K., Reiweger, I., 2018. The Heat of the Flow: Thermal Equilibrium in Gravitational Mass Flows. Geophysical Research Letters 45, 11,219- 11,226. https://doi.org/10.1029/2018GL079585 Fuchs, S., Bründl, M., 2005. Damage Potential and Losses Resulting from Snow Avalanches in Settlements of the Canton of Grisons, Switzerland. Natural Hazards 34, 53–69. https://doi.org/10.1007/s11069-004-0784-y Gesch, D., Oimoen, M., Greenlee, S., Nelson, C., Steuck, M., Tyler, D., 2002. The National Elevation Dataset. Photogrammetric Engineering and Remote Sensing 68, 5–11. Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, Big Remotely Sensed Data: tools, applications and experiences 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031 68 Greene, E., Birkeland, K., Elder, K., Mccammon, I., Staples, M., Sharaf, D., 2016. Snow, Weather, and Avalanches: Observational Guidelines for Avalanche Programs in the United States, 3rd ed. American Avalanche Association, Victor, ID. Hendrikx, J., Murphy, M., Onslow, T., 2014. Classification trees as a tool for operational avalanche forecasting on the Seward Highway, Alaska. Cold Regions Science and Technology 97, 113–120. https://doi.org/10.1016/j.coldregions.2013.08.009 Jóhannesson, T., Arnalds, Þ., 2000. Accidents and economic damage due to snow avalanches and landslides in Iceland. Jökull 50. Kummervold, P., Malnes, E., Eckerstorfer, M., Arntzen, I., Bianchi, F.M., 2018. Avalanche detection in Sentinel-1 radar images using convolutional neural networks. Presented at the International Snow Science Workshop, Innsbruck, Austria. Larsen, H., Hendrikx, J., Slåtten, M.S., Engeset, R.V., 2020a. Developing nationwide avalanche terrain maps for Norway. Nat Hazards 103, 2829–2847. https://doi.org/10.1007/s11069- 020-04104-7 Larsen, H., Sykes, J., Hendrikx, J., Schauer, A., Langford, R., Statham, G., Campbell, C., Neuhauser, M., Fischer, J.-T., 2020b. Development of Automated Avalanche Terrain Exposure Scale Maps: Current and Future. Leinss, S., Wicki, R., Holenstein, S., Baffelli, S., Bühler, Y., 2020. Snow avalanche detection and mapping in multitemporal and multiorbital radar images from TerraSAR-X and Sentinel-1. Natural Hazards and Earth System Sciences 20, 1783–1803. https://doi.org/10.5194/nhess-20-1783-2020 Liu, Y., Zhou, Y., Wen, S., Tang, C., 2014. A Strategy on Selecting Performance Metrics for Classifier Evaluation. International Journal of Mobile Computing and Multimedia Communications 6, 20–35. https://doi.org/10.4018/IJMCMC.2014100102 Malnes, E., Eckerstorfer, M., Larsen, Y., Frauenfelder, R., Jonsson, A., Jaedicke, C., Solbø, S., 2013. Remote sensing of avalanches in northern Norway using Synthetic Aperture Radar. Presented at the International Snow Science Workshop, Grenoble, France. https://doi.org/10.13140/2.1.2462.0481 Malnes, E., Eckerstorfer, M., Vickers, H., 2015. First Sentinel-1 detections of avalanche debris. The Cryosphere Discussions 9, 1943–1963. https://doi.org/10.5194/tcd-9-1943-2015 Martinez-Vazquez, A., Fortuny-Guasch, J., 2008. A GB-SAR Processor for Snow Avalanche Identification. IEEE Transactions on Geoscience and Remote Sensing 46, 3948–3956. https://doi.org/10.1109/TGRS.2008.2001387 McClung, D.M., 2002a. The Elements of Applied Avalanche Forecasting, Part I: The Human Issues. Natural Hazards 26, 111–129. https://doi.org/10.1023/A:1015665432221 McClung, D.M., 2002b. The Elements of Applied Avalanche Forecasting, Part II: The Physical Issues and the Rules of Applied Avalanche Forecasting. Natural Hazards 26, 131–146. https://doi.org/10.1023/A:1015604600361 McClung, D.M., Schaerer, P., 2006. The Avalanche Handbook. The Mountaineers. Mock, C.J., Birkeland, K.W., 2000. Snow Avalanche Climatology of the Western United States Mountain Ranges. Bull. Amer. Meteor. Soc. 81, 2367–2392. https://doi.org/10.1175/1520-0477(2000)081<2367:SACOTW>2.3.CO;2 National Research Council, 1990. Snow Avalanche Hazards and Mitigation in the United States. The National Academies Press, Washington, DC. 69 Peitzsch, E., Boilen, S., Logan, S., Birkeland, K., Greene, E., 2020. Research note: How old are the people who die in avalanches? A look into the ages of avalanche victims in the United States (1950–2018). Journal of Outdoor Recreation and Tourism 29. https://doi.org/10.1016/j.jort.2019.100255 Pozdnoukhov, A., Matasci, G., Kanevski, M., Purves, R., 2011. Spatio-temporal avalanche forecasting with Support Vector Machines. Natural Hazards and Earth System Sciences 11. https://doi.org/10.5194/nhess-11-367-2011 Schweizer, J., 2008. Snow avalanche formation and dynamics. Cold Regions Science and Technology 54, 153–154. https://doi.org/10.1016/j.coldregions.2008.08.005 Schweizer, J., Mitterer, C., Techel, F., Stoffel, A., Reuter, B., 2018. Quantifying the obvious: the avalanche danger level. International Snow Science Workshop Proceedings 2018, Innsbruck, Austria 1052–1058. Sexton, J., Song, X.-P., Feng, M., Noojipady, P., Anand, A., Huang, C., Kim, D.-H., Collins, K., Channan, S., Dimiceli, C., Townshend, J., 2013. Global, 30-m resolution continuous fields of tree cover: Landsat-based rescaling of MODIS vegetation continuous fields with lidar-based estimates of error. International Journal of Digital Earth 6, 1–22. https://doi.org/10.1080/17538947.2013.786146 Techel, F., Jarry, F., Kronthaler, G., Mitterer, S., Nairz, P., Pavšek, M., Valt, M., Darms, G., 2016. Avalanche fatalities in the European Alps: Long-term trends and statistics. Geographica Helvetica 71, 147–159. https://doi.org/10.5194/gh-71-147-2016 Vickers, H., Eckerstorfer, M., Malnes, E., Larsen, Y., Hindberg, H., 2016. A method for automated snow avalanche debris detection through use of Synthetic Aperture Radar (SAR) imaging: Automated avalanche detection. Earth and Space Science 3. https://doi.org/10.1002/2016EA000168 Vollrath, A., Mullissa, A., Reiche, J., 2020. Angular-Based Radiometric Slope Correction for Sentinel-1 on Google Earth Engine. Remote Sens 12, 1867. Wiesmann, A., Wegmuller, U., Honikel, M., Strozzi, T., Werner, C.L., 2001. Potential and methodology of satellite based SAR for hazard mapping, in: IGARSS 2001. Scanning the Present and Resolving the Future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Symposium (Cat. No.01CH37217). Presented at the IGARSS 2001. Scanning the Present and Resolving the Future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Symposium (Cat. No.01CH37217), pp. 3262–3264 vol.7. https://doi.org/10.1109/IGARSS.2001.978322 Wilks, D., 2019. Statistical Methods in the Atmospheric Sciences. Elsevier. https://doi.org/10.1016/C2017-0-03921-6 Wynne, R., Campbell, J., 2011. Introduction to Remote Sensing, 5th ed. The Guilford Press. Yang, J., Li, C., Li, L., Ding, J., Zhang, R., Han, T., Liu, Y., 2020. Automatic Detection of Regional Snow Avalanches with Scattering and Interference of C-band SAR Data. Remote Sensing 12, 2781. https://doi.org/10.3390/rs12172781 70 CHAPTER THREE CONCLUSION Summary This study examined the detectability of snow avalanches using change detection analysis on Sentinel-1 Imagery. It attempted to answer the following questions: 1. What are the manual detection rates of avalanches using change detection techniques on Sentinel-1 backscatter images in the transitional snow climate of the western United States? 2. What are the dominant SAR-specific, topographic, and avalanche characteristic controls on those detection rates for avalanches in Sentinel-1 images? Utilizing historical avalanche records, we examined four avalanche cycles. (two in Little Cottonwood Canyon, Utah, and two in the Jackson Hole, Wyoming region). We were able to detect 274 of the 545 field-reported avalanches. Destructive size, avalanche path length, slope angle, local incidence angle, and lag days are correlated with detection probability. Our results indicated that SAR temporal change analysis is a viable technique for avalanche detections in the transitional snow climates of the western United States but may be less effective than regions dominated by wet slabs. 71 Future Work Future research should address 1) how detection rates of avalanches vary across other mountain ranges and snow climates, 2) confirming controlling factors on detection rates, and 3) exploring the utility of other SAR sensors and parameters for avalanche detection. First, exploring overall detection rates for avalanches would be valuable. Future studies on detection rates relative to field observed datasets would help us understand how many avalanches will be missed by this technique and trends in missed detections. A large study utilizing avalanche records and SAR imagery that spanned multiple mountain ranges and snow climates would establish if varying snow climates and slope-specific topographic factors significantly detection rates. Additionally, these future studies should utilize a spatially and temporally structured field campaign rather than opportunistically collecting avalanche observations to allow for more rigorous statistical analysis. Another question for future studies is the factors that impact detection rates. This study found that avalanche type, its destructive size, lag days, avalanche path length and slope, and local incidence angle were all linked to detection rates. Future studies should confirm these results in other snow climates and with larger datasets. Additionally, this study found lag days to be a correlated variable with detection rates. Future studies could constrain specific meteorological factors, such as debris being obscured by significant amounts of precipitation, ablation from above freezing temperatures, or high wind speeds as the cause of decreasing detection rates. Working with a higher resolution digital elevation model would also allow a future analysis to study the effect of topography. We found a marginally significant relationship between the avalanche path’s curvature and detection rates. Future studies could confirm this 72 relationship. Knowing what factors control detection rates and if different storm cycles or topographies are better suited to detecting avalanche debris will be invaluable. This study focused on analyzing the changes in the amount of backscattered energy. Other research utilizing other SAR-specific parameters such as the ratio between polarizations, phase, and coherence (how much the phase shifts between images) could provide valuable data to detect avalanches. Finally, combining the Sentinel-1 satellite imagery used in this study with other sensors, especially in other microwave spectrum bands, could improve the detection rates. Combining Sentinel-1 imagery with overlapping imagery from other satellites could also improve detections. Additional sensors may be operating with different imaging geometries, minimizing the effects of radar layover and shadow, or in different ranges of the microwave spectrum, possibly with greater sensitivity to varying aspects of avalanche debris. These future studies will continue to improve our methods and understanding of how to remotely detect avalanches. By providing researchers and practitioners with improved knowledge of when and where avalanches are occurring, we can hopefully increase our awareness of avalanche hazards. Additionally, the spatially complete nature of many satellite- based methods may allow for improved forecasting in remote and less developed regions. These remote sensing techniques will continue to improve and reduce the number of deaths from avalanches worldwide. 73 REFERENCES CITED 74 Abermann, J., Eckerstorfer, M., Malnes, E., Hansen, B.U., 2019. A large wet snow avalanche cycle in West Greenland quantified using remote sensing and in situ observations. Nat Hazards 97, 517–534. https://doi.org/10.1007/s11069-019-03655-8 Bianchi, F.M., Grahn, J., Eckerstorfer, M., Malnes, E., Vickers, H., 2021. Snow avalanche segmentation in SAR images with Fully Convolutional Neural Networks. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 75–82. https://doi.org/10.1109/JSTARS.2020.3036914 Birkeland, K., 2017. In Response to Avalanche Fatalities in the United States by Jekich et al. Wilderness and Environmental Medicine 28. https://doi.org/10.1016/j.wem.2017.06.009 Breiman, L., 2001. Random Forests. Machine Learning 45, 5–32. https://doi.org/10.1023/A:1010933404324 Bühler, Y., Hafner, E., Zweifel, B., Zesiger, M., Heisig, H., 2019. Where are the avalanches? Rapid mapping of a large snow avalanche period with optical satellites. The Cryosphere Discussions 1–21. https://doi.org/10.5194/tc-2019-119 Bühler, Y., Hüni, A., Christen, M., Meister, R., Kellenberger, T., 2009. Automated detection and mapping of avalanche deposits using airborne optical remote sensing data. Cold Regions Science and Technology 57, 99–106. https://doi.org/10.1016/j.coldregions.2009.02.007 Chabot, D., 2017. Avalanche Forecasting in Afghanistan, Pakistan and Tajikistan. The Avalanche Review 35, 30–31. Colorado Avalanche Information Center, 2021. US Avalanche Accident Reports. Eckerstorfer, M., Bühler, Y., Frauenfelder, R., Malnes, E., 2016. Remote sensing of snow avalanches: Recent advances, potential, and limitations. Cold Regions Science and Technology 121, 126–140. https://doi.org/10.1016/j.coldregions.2015.11.001 Eckerstorfer, M., Malnes, E., 2015. Manual detection of snow avalanche debris using high- resolution Radarsat-2 SAR images. Cold Regions Science and Technology 120. https://doi.org/10.1016/j.coldregions.2015.08.016 Eckerstorfer, M., Vickers, H., Malnes, E., Grahn, J., 2019. Near-Real Time Automatic Snow Avalanche Activity Monitoring System Using Sentinel-1 SAR Data in Norway. Remote Sensing 11, 2863. https://doi.org/10.3390/rs11232863 Fischer, J.-T., Kaitna, R., Heil, K., Reiweger, I., 2018. The Heat of the Flow: Thermal Equilibrium in Gravitational Mass Flows. Geophysical Research Letters 45, 11,219- 11,226. https://doi.org/10.1029/2018GL079585 Fuchs, S., Bründl, M., 2005. Damage Potential and Losses Resulting from Snow Avalanches in Settlements of the Canton of Grisons, Switzerland. Natural Hazards 34, 53–69. https://doi.org/10.1007/s11069-004-0784-y Gesch, D., Oimoen, M., Greenlee, S., Nelson, C., Steuck, M., Tyler, D., 2002. The National Elevation Dataset. Photogrammetric Engineering and Remote Sensing 68, 5–11. Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, Big Remotely Sensed Data: tools, applications and experiences 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031 75 Greene, E., Birkeland, K., Elder, K., Mccammon, I., Staples, M., Sharaf, D., 2016. Snow, Weather, and Avalanches: Observational Guidelines for Avalanche Programs in the United States, 3rd ed. American Avalanche Association, Victor, ID. Hendrikx, J., Murphy, M., Onslow, T., 2014. Classification trees as a tool for operational avalanche forecasting on the Seward Highway, Alaska. Cold Regions Science and Technology 97, 113–120. https://doi.org/10.1016/j.coldregions.2013.08.009 Jóhannesson, T., Arnalds, Þ., 2000. Accidents and economic damage due to snow avalanches and landslides in Iceland. Jökull 50. Kummervold, P., Malnes, E., Eckerstorfer, M., Arntzen, I., Bianchi, F.M., 2018. Avalanche detection in Sentinel-1 radar images using convolutional neural networks. Presented at the International Snow Science Workshop, Innsbruck, Austria. Larsen, H., Hendrikx, J., Slåtten, M.S., Engeset, R.V., 2020a. Developing nationwide avalanche terrain maps for Norway. Nat Hazards 103, 2829–2847. https://doi.org/10.1007/s11069- 020-04104-7 Larsen, H., Sykes, J., Hendrikx, J., Schauer, A., Langford, R., Statham, G., Campbell, C., Neuhauser, M., Fischer, J.-T., 2020b. Development of Automated Avalanche Terrain Exposure Scale Maps: Current and Future. Leinss, S., Wicki, R., Holenstein, S., Baffelli, S., Bühler, Y., 2020. Snow avalanche detection and mapping in multitemporal and multiorbital radar images from TerraSAR-X and Sentinel-1. Natural Hazards and Earth System Sciences 20, 1783–1803. https://doi.org/10.5194/nhess-20-1783-2020 Liu, Y., Zhou, Y., Wen, S., Tang, C., 2014. A Strategy on Selecting Performance Metrics for Classifier Evaluation. International Journal of Mobile Computing and Multimedia Communications 6, 20–35. https://doi.org/10.4018/IJMCMC.2014100102 Malnes, E., Eckerstorfer, M., Larsen, Y., Frauenfelder, R., Jonsson, A., Jaedicke, C., Solbø, S., 2013. Remote sensing of avalanches in northern Norway using Synthetic Aperture Radar. Presented at the International Snow Science Workshop, Grenoble, France. https://doi.org/10.13140/2.1.2462.0481 Malnes, E., Eckerstorfer, M., Vickers, H., 2015. First Sentinel-1 detections of avalanche debris. The Cryosphere Discussions 9, 1943–1963. https://doi.org/10.5194/tcd-9-1943-2015 Martinez-Vazquez, A., Fortuny-Guasch, J., 2008. A GB-SAR Processor for Snow Avalanche Identification. IEEE Transactions on Geoscience and Remote Sensing 46, 3948–3956. https://doi.org/10.1109/TGRS.2008.2001387 McClung, D.M., 2002a. The Elements of Applied Avalanche Forecasting, Part I: The Human Issues. Natural Hazards 26, 111–129. https://doi.org/10.1023/A:1015665432221 McClung, D.M., 2002b. The Elements of Applied Avalanche Forecasting, Part II: The Physical Issues and the Rules of Applied Avalanche Forecasting. Natural Hazards 26, 131–146. https://doi.org/10.1023/A:1015604600361 McClung, D.M., Schaerer, P., 2006. The Avalanche Handbook. The Mountaineers. Mock, C.J., Birkeland, K.W., 2000. Snow Avalanche Climatology of the Western United States Mountain Ranges. Bull. Amer. Meteor. Soc. 81, 2367–2392. https://doi.org/10.1175/1520-0477(2000)081<2367:SACOTW>2.3.CO;2 National Research Council, 1990. Snow Avalanche Hazards and Mitigation in the United States. The National Academies Press, Washington, DC. 76 Peitzsch, E., Boilen, S., Logan, S., Birkeland, K., Greene, E., 2020. Research note: How old are the people who die in avalanches? A look into the ages of avalanche victims in the United States (1950–2018). Journal of Outdoor Recreation and Tourism 29. https://doi.org/10.1016/j.jort.2019.100255 Pozdnoukhov, A., Matasci, G., Kanevski, M., Purves, R., 2011. Spatio-temporal avalanche forecasting with Support Vector Machines. Natural Hazards and Earth System Sciences 11. https://doi.org/10.5194/nhess-11-367-2011 Schweizer, J., 2008. Snow avalanche formation and dynamics. Cold Regions Science and Technology 54, 153–154. https://doi.org/10.1016/j.coldregions.2008.08.005 Schweizer, J., Mitterer, C., Techel, F., Stoffel, A., Reuter, B., 2018. Quantifying the obvious: the avalanche danger level. International Snow Science Workshop Proceedings 2018, Innsbruck, Austria 1052–1058. Sexton, J., Song, X.-P., Feng, M., Noojipady, P., Anand, A., Huang, C., Kim, D.-H., Collins, K., Channan, S., Dimiceli, C., Townshend, J., 2013. Global, 30-m resolution continuous fields of tree cover: Landsat-based rescaling of MODIS vegetation continuous fields with lidar-based estimates of error. International Journal of Digital Earth 6, 1–22. https://doi.org/10.1080/17538947.2013.786146 Techel, F., Jarry, F., Kronthaler, G., Mitterer, S., Nairz, P., Pavšek, M., Valt, M., Darms, G., 2016. Avalanche fatalities in the European Alps: Long-term trends and statistics. Geographica Helvetica 71, 147–159. https://doi.org/10.5194/gh-71-147-2016 Vickers, H., Eckerstorfer, M., Malnes, E., Larsen, Y., Hindberg, H., 2016. A method for automated snow avalanche debris detection through use of Synthetic Aperture Radar (SAR) imaging: Automated avalanche detection. Earth and Space Science 3. https://doi.org/10.1002/2016EA000168 Vollrath, A., Mullissa, A., Reiche, J., 2020. Angular-Based Radiometric Slope Correction for Sentinel-1 on Google Earth Engine. Remote Sens 12, 1867. Wiesmann, A., Wegmuller, U., Honikel, M., Strozzi, T., Werner, C.L., 2001. Potential and methodology of satellite based SAR for hazard mapping, in: IGARSS 2001. Scanning the Present and Resolving the Future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Symposium (Cat. No.01CH37217). Presented at the IGARSS 2001. Scanning the Present and Resolving the Future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Symposium (Cat. No.01CH37217), pp. 3262–3264 vol.7. https://doi.org/10.1109/IGARSS.2001.978322 Wilks, D., 2019. Statistical Methods in the Atmospheric Sciences. Elsevier. https://doi.org/10.1016/C2017-0-03921-6 Wynne, R., Campbell, J., 2011. Introduction to Remote Sensing, 5th ed. The Guilford Press. Yang, J., Li, C., Li, L., Ding, J., Zhang, R., Han, T., Liu, Y., 2020. Automatic Detection of Regional Snow Avalanches with Scattering and Interference of C-band SAR Data. Remote Sensing 12, 2781. https://doi.org/10.3390/rs12172781 77 APPENDICES 78 APPENDIX A COVARIANCE AND NORMALITY 79 The correlation of all explanatory and target variables was extracted and graphed to look for strongly correlated variables. 80 Additionally, the normality of the quantitative variables was tested using a Shapiro-Wilks test. The results are presented below, with p-values representing the probability the variable in not normally distributed: Variable Statistic P-Value Lag Days 0.88 4.35E-20 Layover 0.53 1.26E-35 D-Size 0.93 1.48E-15 Tree % 0.8 3.68E-25 Incidence Angle 0.966 6.80E-10 Slope 0.952 2.25E-12 Curvature 0.919 1.66E-16 Path Length 0.945 4.18E-13 81 APPENDIX B AVALANCHE DATABASES 82 This appendix contains the utilized records for the UDOT and BTAC. The column names are explained below: Detected Avalanche detection: 1 – detected, 0 – not detected, -1 – did not overlap with image Date Estimated avalanche release date. State UDOT (‘udot’) or BTAC (‘jh’) Orbit Sentinel-1 path number for images Ref_date_img Date of reference image acquisition Act_date_img Date of activity image acquisition Latitude Field approximated latitude Longitude Field approximated longitude Detected Avalanche detection: 1 – detected, 0 – not detected, -1 – did not overlap with image Tree_perc The tree percentage in the avalanche path description Field observers notes on avalanche Shadow_perc The percentage of path affected by radar shadow Layover_perc The percentage of path affected by radar layover Ave_incidence_angle Average local incidence angle (degrees) within the avalanche path 83 Ave_slope Average slope (degrees) within the avalanche path Ave_elev Average elevation (m) of the avalanche path Ave_curvature Average curvature of the avalanche path. Negative indicates concave and positive convex Mbg_length Length of the minimum bounding geometry (km) around avalanche path Mbg_width Width of the minimum bounding geometry (km) around avalanche path Path_id BTAC specific field showing the avalanche path associated with that observation Region UDOT specific field showing the avalanche path associated with that observation Ref_vv_ave_value The average backscatter value (dB) within the avalanche path for the reference image in VV Ref_vh_ave_value The average backscatter value (dB) within the avalanche path for the reference image in VH Act_vv_ave_value The average backscatter value (dB) within the avalanche path for the activity image in VV Act_vh_ave_value The average backscatter value (dB) within the avalanche path for the activity image in VH 84 Lag Days The number of days between the estimated avalanche release date and the acquisition of the activity image. 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 APPENDIX C PYTHON SCRIPTS USED IN ANALYSIS 110 This appendix contains python scripts used to filter the avalanche databases, calculate AAIs, process, and download Sentinel-1 imagery, and perform analysis on the avalanche records. Python version 3.7.10 was used and other environment data is presented below: !pip list Package Version ----------------------------- -------------- absl-py 0.10.0 alabaster 0.7.12 albumentations 0.1.12 altair 4.1.0 appdirs 1.4.4 argon2-cffi 20.1.0 asgiref 3.3.1 astor 0.8.1 astropy 4.2 astunparse 1.6.3 async-generator 1.10 atari-py 0.2.6 atomicwrites 1.4.0 attrs 20.3.0 audioread 2.1.9 autograd 1.3 Babel 2.9.0 backcall 0.2.0 beautifulsoup4 4.6.3 bleach 3.3.0 blis 0.4.1 bokeh 2.1.1 Bottleneck 1.3.2 branca 0.4.2 bs4 0.0.1 CacheControl 0.12.6 cachetools 4.2.1 catalogue 1.0.0 certifi 2020.12.5 cffi 1.14.5 chainer 7.4.0 chardet 3.0.4 click 7.1.2 cloudpickle 1.3.0 cmake 3.12.0 cmdstanpy 0.9.5 colorlover 0.3.0 community 1.0.0b1 contextlib2 0.5.5 convertdate 2.3.1 coverage 3.7.1 coveralls 0.5 111 crcmod 1.7 cufflinks 0.17.3 cvxopt 1.2.6 cvxpy 1.0.31 cycler 0.10.0 cymem 2.0.5 Cython 0.29.22 daft 0.0.4 dask 2.12.0 datascience 0.10.6 debugpy 1.0.0 decorator 4.4.2 defusedxml 0.7.1 descartes 1.1.0 dill 0.3.3 distributed 1.25.3 Django 3.1.7 dlib 19.18.0 dm-tree 0.1.5 docopt 0.6.2 docutils 0.16 dopamine-rl 1.0.5 earthengine-api 0.1.255 easydict 1.9 ecos 2.0.7.post1 editdistance 0.5.3 en-core-web-sm 2.2.5 entrypoints 0.3 ephem 3.7.7.1 et-xmlfile 1.0.1 fa2 0.3.5 fancyimpute 0.4.3 fastai 1.0.61 fastdtw 0.3.4 fastprogress 1.0.0 fastrlock 0.5 fbprophet 0.7.1 feather-format 0.4.1 filelock 3.0.12 firebase-admin 4.4.0 fix-yahoo-finance 0.0.22 Flask 1.1.2 flatbuffers 1.12 folium 0.8.3 future 0.16.0 gast 0.3.3 GDAL 2.2.2 gdown 3.6.4 gensim 3.6.0 geographiclib 1.50 geopy 1.17.0 gin-config 0.4.0 glob2 0.7 google 2.0.3 112 google-api-core 1.26.1 google-api-python-client 1.12.8 google-auth 1.27.1 google-auth-httplib2 0.0.4 google-auth-oauthlib 0.4.3 google-cloud-bigquery 1.21.0 google-cloud-bigquery-storage 1.1.0 google-cloud-core 1.0.3 google-cloud-datastore 1.8.0 google-cloud-firestore 1.7.0 google-cloud-language 1.2.0 google-cloud-storage 1.18.1 google-cloud-translate 1.5.0 google-colab 1.0.0 google-pasta 0.2.0 google-resumable-media 0.4.1 googleapis-common-protos 1.53.0 googledrivedownloader 0.4 graphviz 0.10.1 grpcio 1.32.0 gspread 3.0.1 gspread-dataframe 3.0.8 gym 0.17.3 h5py 2.10.0 HeapDict 1.0.1 heatmapz 0.0.4 hijri-converter 2.1.1 holidays 0.10.5.2 holoviews 1.13.5 html5lib 1.0.1 httpimport 0.5.18 httplib2 0.17.4 httplib2shim 0.0.3 humanize 0.5.1 hyperopt 0.1.2 ideep4py 2.0.0.post3 idna 2.10 image 1.5.33 imageio 2.4.1 imagesize 1.2.0 imbalanced-learn 0.4.3 imblearn 0.0 imgaug 0.2.9 importlib-metadata 3.7.2 importlib-resources 5.1.2 imutils 0.5.4 inflect 2.1.0 iniconfig 1.1.1 intel-openmp 2021.1.2 intervaltree 2.1.0 ipykernel 4.10.1 ipython 5.5.0 ipython-genutils 0.2.0 ipython-sql 0.3.9 113 ipywidgets 7.6.3 itsdangerous 1.1.0 jax 0.2.10 jaxlib 0.1.62+cuda110 jdcal 1.4.1 jedi 0.18.0 jieba 0.42.1 Jinja2 2.11.3 joblib 1.0.1 jpeg4py 0.1.4 jsonschema 2.6.0 jupyter 1.0.0 jupyter-client 5.3.5 jupyter-console 5.2.0 jupyter-core 4.7.1 jupyterlab-pygments 0.1.2 jupyterlab-widgets 1.0.0 kaggle 1.5.10 kapre 0.1.3.1 Keras 2.4.3 Keras-Preprocessing 1.1.2 keras-vis 0.4.1 kiwisolver 1.3.1 knnimpute 0.1.0 korean-lunar-calendar 0.2.1 librosa 0.8.0 lightgbm 2.2.3 llvmlite 0.34.0 lmdb 0.99 lucid 0.3.8 LunarCalendar 0.0.9 lxml 4.2.6 Markdown 3.3.4 MarkupSafe 1.1.1 matplotlib 3.2.2 matplotlib-venn 0.11.6 missingno 0.4.2 mistune 0.8.4 mizani 0.6.0 mkl 2019.0 mlxtend 0.14.0 more-itertools 8.7.0 moviepy 0.2.3.5 mpmath 1.2.1 msgpack 1.0.2 multiprocess 0.70.11.1 multitasking 0.0.9 murmurhash 1.0.5 music21 5.5.0 natsort 5.5.0 nbclient 0.5.3 nbconvert 5.6.1 nbformat 5.1.2 nest-asyncio 1.5.1 114 networkx 2.5 nibabel 3.0.2 nltk 3.2.5 notebook 5.3.1 np-utils 0.5.12.1 numba 0.51.2 numexpr 2.7.3 numpy 1.19.5 nvidia-ml-py3 7.352.0 oauth2client 4.1.3 oauthlib 3.1.0 okgrade 0.4.3 opencv-contrib-python 4.1.2.30 opencv-python 4.1.2.30 openpyxl 2.5.9 opt-einsum 3.3.0 osqp 0.6.2.post0 packaging 20.9 palettable 3.3.0 pandas 1.1.5 pandas-datareader 0.9.0 pandas-gbq 0.13.3 pandas-profiling 1.4.1 pandocfilters 1.4.3 panel 0.9.7 param 1.10.1 parso 0.8.1 pathlib 1.0.1 patsy 0.5.1 pexpect 4.8.0 pickleshare 0.7.5 Pillow 7.0.0 pip 19.3.1 pip-tools 4.5.1 plac 1.1.3 plotly 4.4.1 plotnine 0.6.0 pluggy 0.7.1 pooch 1.3.0 portpicker 1.3.1 prefetch-generator 1.0.1 preshed 3.0.5 prettytable 2.1.0 progressbar2 3.38.0 prometheus-client 0.9.0 promise 2.3 prompt-toolkit 1.0.18 protobuf 3.12.4 psutil 5.4.8 psycopg2 2.7.6.1 ptyprocess 0.7.0 py 1.10.0 pyarrow 3.0.0 pyasn1 0.4.8 115 pyasn1-modules 0.2.8 pycocotools 2.0.2 pycparser 2.20 pyct 0.4.8 pydata-google-auth 1.1.0 pydot 1.3.0 pydot-ng 2.0.0 pydotplus 2.0.2 PyDrive 1.3.1 pyemd 0.5.1 pyerfa 1.7.2 pyglet 1.5.0 Pygments 2.6.1 pygobject 3.26.1 pymc3 3.7 PyMeeus 0.5.9 pymongo 3.11.3 pymystem3 0.2.0 pynndescent 0.5.2 PyOpenGL 3.1.5 pyparsing 2.4.7 pyrsistent 0.17.3 pysndfile 1.3.8 PySocks 1.7.1 pystan 2.19.1.1 pytest 3.6.4 python-apt 0.0.0 python-chess 0.23.11 python-dateutil 2.8.1 python-louvain 0.15 python-slugify 4.0.1 python-utils 2.5.6 pytz 2018.9 pyviz-comms 2.0.1 PyWavelets 1.1.1 PyYAML 3.13 pyzmq 22.0.3 qdldl 0.1.5.post0 qtconsole 5.0.3 QtPy 1.9.0 regex 2019.12.20 requests 2.23.0 requests-oauthlib 1.3.0 resampy 0.2.2 retrying 1.3.3 rpy2 3.4.2 rsa 4.7.2 scikit-image 0.16.2 scikit-learn 0.22.2.post1 scipy 1.4.1 screen-resolution-extra 0.0.0 scs 2.1.2 seaborn 0.11.1 Send2Trash 1.5.0 116 setuptools 54.1.2 setuptools-git 1.2 Shapely 1.7.1 simplegeneric 0.8.1 six 1.15.0 sklearn 0.0 sklearn-pandas 1.8.0 smart-open 4.2.0 snowballstemmer 2.1.0 sortedcontainers 2.3.0 SoundFile 0.10.3.post1 spacy 2.2.4 Sphinx 1.8.5 sphinxcontrib-serializinghtml 1.1.4 sphinxcontrib-websupport 1.2.4 SQLAlchemy 1.3.23 sqlparse 0.4.1 srsly 1.0.5 statsmodels 0.10.2 sympy 1.7.1 tables 3.4.4 tabulate 0.8.9 tblib 1.7.0 tensorboard 2.4.1 tensorboard-plugin-wit 1.8.0 tensorflow 2.4.1 tensorflow-datasets 4.0.1 tensorflow-estimator 2.4.0 tensorflow-gcs-config 2.4.0 tensorflow-hub 0.11.0 tensorflow-metadata 0.28.0 tensorflow-probability 0.12.1 termcolor 1.1.0 terminado 0.9.2 testpath 0.4.4 text-unidecode 1.3 textblob 0.15.3 textgenrnn 1.4.1 Theano 1.0.5 thinc 7.4.0 tifffile 2021.3.17 toml 0.10.2 toolz 0.11.1 torch 1.8.0+cu101 torchsummary 1.5.1 torchtext 0.9.0 torchvision 0.9.0+cu101 tornado 5.1.1 tqdm 4.41.1 traitlets 5.0.5 tweepy 3.10.0 typeguard 2.7.1 typing-extensions 3.7.4.3 tzlocal 1.5.1 117 umap-learn 0.5.1 uritemplate 3.0.1 urllib3 1.24.3 vega-datasets 0.9.0 wasabi 0.8.2 wcwidth 0.2.5 webencodings 0.5.1 Werkzeug 1.0.1 wheel 0.36.2 widgetsnbextension 3.5.1 wordcloud 1.5.0 wrapt 1.12.1 xarray 0.15.1 xgboost 0.90 xkit 0.0.0 xlrd 1.1.0 xlwt 1.3.0 yellowbrick 0.9.1 zict 2.0.0 zipp 3.4.1 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206