Scaling and uncertainty in landsat remote sensing of biophysical attributes

Abstract

Monitoring environmental change is of high importance in our time of global change. Remote sensing technology provides the tools to view the ecological dynamics at a landscape scale and review the change through time with time series data availability. Creating congruence between data scales and functional scales is a long standing challenge for Earth system scientists. In this research we evaluate methods for change detection and scaling data in a discontinuous permafrost zone of central Alaska and is characterized by pronounced permafrost thaw and methane release over decadal to century timescales. The primary goal is to evaluate the applicability of Landsat satellite remote sensing for detecting bog thermal expansion over time. We implement a Random Forests classification scheme in order to separate the landscape into its various land features and bog types, many features in this landscape are developed through an aged-stage transition of thermal expansion. The results of this classification were dominated by hydrologic features, with a 0.05 increase in mean albedo, providing essentially no change in both mean Normalized Difference Vegetation Index (NDVI) and mean Brightness Temperature (BT). In addition, we attempt to capture the scales of variation within the landscape using multi-resolution methods. The scale of variance as illustrated by a wavelet analysis for NDVI show the greatest amount of variance around 4.5 km to 5 km. Brightness Temperature had three peaks of high variance between 0.06 km - 1 km including maximum variance at about 0.5 km and a pair of peaks between 3 km and 4 km. An important component of any data analysis is quantifying the uncertainty. Uncertainty quantification in remote sensing data analysis is often over looked. In a second analysis we attempt to quantify the primary sources of uncertainty in Landsat remote sensing data via simulation methods. Specifically, we evaluate the level of uncertainty contributed to the data by applying a typical atmospheric correction through Monte Carlo simulation approach to estimate the total variance within several Landsat scenes. We find the contribution of uncertainty due to the MODTRAN conversion to be between 7-27% differing by total scene variance per image.