Comparison of Speckle Noise Filters on Crop Classification Based on Sentinel-1 Sar Time-Series

Knowing the spatial distribution of crops is key for the estimation of water consumption, crop yield, food policy, among others. In this study, dual-polarization (VV/VH) intensity time-series C-band Sentinel-1 sensor are used to perform crop classification with two models: Spectral Similarity Value SSV and Random Forest RF. Their performance was evaluated under three SAR speckle filter scenarios, for each polarization. Scenario 1 Sigma Lee, Scenario 2 IDAN, and Scenario 3 Non-Local Means. Crop classification was performed with 24 images of 2018. Ground-truth data for training/validation was obtained from the United States Department of Agriculture (USDA). RF shows better performance than SSV for all scenarios. Non-Local Means filter delivers the best results. VH polarization performs better that VV in all scenarios. The best accuracy for SSV model is 0.73, and for RF model is 0.95 both using VH polarization.