Crop Classification and Soil Moisture Retrieval over Agricultural Fields Using Time-Series Synthetic Aperture Radar Images
Abstract
Crop monitoring is crucial for enhancing crop yields and ensuring food security. However, the diversity of climates and regional management policies present challenges for large-scale crop monitoring. The advent of high spatial-temporal resolution synthetic aperture radar (SAR) satellites offers a promising solution for extensive and long-term crop monitoring. This thesis explores the potential of time-series SAR data for monitoring two critical variables: crop type and soil moisture (SM). To leverage the statistical information in SAR images, new statistical descriptor, namely Object-based Generalized Gamma Distribution (OGΓD) features, were developed for time-series Sentinel-1 images. These features improved classification accuracy for corn, soybean, and wheat, achieving an overall accuracy of 96.66% and a Kappa coefficient of 95.34%, thereby facilitating further analyses such as crop rotation mapping. For early-season crop classification, the Local Window Attention Transformer (LWAT) was applied to RADARSAT Constellation Mission (RCM) and Sentinel-1 data. RCM with compact polarization outperformed Sentinel-1 with dual polarization during early crop stages, achieving comparable accuracy with the fusion of RCM and Sentinel-1 data. The optimal time for early-season crop mapping was determined to be six weeks after seeding, corresponding to the stem elongation stage of crops. To enhance the reliability for SM estimation in agricultural regions, a polarimetric decomposition-based change detection (PDCD) algorithm was proposed for compact polarization SAR data. It eliminates volume scattering from the backscattering, resulting in more accurate SM estimation. The PDCD method achieved a Root Mean Square Error (RMSE) of 5.058 Vol.% and 4.309 Vol.% in RCH and RCV polarizations, respectively. Consistent trends were observed when comparing time-series estimations with reference data, effectively reflecting rainfall events. Additionally, the retrieval depth of SM for multi-frequency and multi-polarization SAR was investigated using theoretical simulation and in-situ data evaluation. The theoretical penetration depth of L-band SAR ranges from 3 cm to 30 cm, which is deeper than C-band SAR, which penetrates only the top surface soil. Based on polarimetric features and in-situ data, the retrieval depth of L-band ALOS-2 was determined to be between 5 cm to 20 cm, while C-band RCM achieved lower accuracy for all depths due to its sensitivity to vegetation coverage.