Winter tree and terrain shadow correction reveals land cover mapping errors in Sentinel-2 satellite imagery

Abstract

This paper presents an approach for detecting and correcting winter shadows cast by trees and terrain. The approach is geographically adaptable and was evaluated across 7 international Sentinel-2 10 m images. The detection method employs a physical-based approach using a Digital Elevation Model (DEM), solar angles, tree height layer, and adaptive thresholds to reduce errors. We evaluated two tree canopy height models (CHMs), the Meta CHM and UMD CHM, and our approach using Meta was most accurate, yielding an overall accuracy of 83.6% (±1.4%) over 5 non-snow images, and 82.8% (±1.6%) over 2 snow images. We found shadows occupy a substantial land surface area ranging between 8% and 30% within each image, indicating detection and correction importance for environmental monitoring studies. Subsequently, we tested several deep learning and ensemble regression models for correcting shadow-afflicted pixels via a gap-filling approach, including the use of multitemporal SAR, texture, and a novel shadow-independent focal layer, all combined with a hyperlocal training approach enabling locally-relevant predictions. The LightGBM and XGBoost tuned models offered the best combination of accuracy and prediction speed making them scalable for broader-scale analysis. The Mean Absolute Error (MAE) for LightGBM model corrections were: MAE = 440 (Near-IR), MAE = 302 Red, MAE = 259 green, and MAE = 0.037 NDVI. Evaluation of the Dynamic World land cover model indicated approximately 10–18% of water pixels in two selected images were erroneous and caused by tree or terrain shadows. This manuscript highlights the importance of accounting for winter terrain and tree shadows when using Sentinel-2 for environmental monitoring in moderate-to-high latitudes. • Scalable approach to detect shadows cast by trees or terrain with 83% accuracy. • Shadows covered 8%–30% of each image, indicating importance for monitoring. • Evaluation of machine and deep learning models for correcting shadow pixels. • Creation of novel shadow-independent focal layer to improve shadow correction. • Tree and terrain shadows linked to errors in Dynamic World land cover product.