A time series deep learning algorithm for soil moisture retrieval from Sentinel-1 data using Transformer

Soil moisture (SM) retrieval from Sentinel-1 C-band Synthetic Aperture Radar (SAR) data has often applied change detection approaches under the assumption of minimal vegetation change over a short interval between retrievals: an assumption often violated, particularly for cropland and grassland due to phenology. Here we introduce a novel time series deep learning model that does not rely on this assumption to derive SM at 200 m resolution from a year of Sentinel-1 time series patches each with 20×20 pixels at 10 m resolution for cropland and grassland areas across the conterminous US (CONUS). The model leverages seasonal vegetation change patterns for retrieval instead of assuming minimal vegetation changes. The model used the Transformer architecture to handle the variations in the number of annual Sentinel-1 acquisitions across locations, and was trained using 34,711 high-quality SM daily measurements across 6-years (2017-2022) from 181 International Soil Moisture Network stations within the CONUS. A six-fold cross-validation shows that the Transformer SM retrieval achieved correlation coefficients (CC) of 0.581 and unbiased RMSE (ubRMSE) value of 0.093 m 3 /m 3 . It outperformed the NASA Soil Moisture Active Passive/Sentinel-1 1 km and 3 km SM products that exhibited CC of 0.416 and 0.496 and ubRMSE of 0.124 and 0.113 m 3 /m 3 , respectively. The Transformer model performed similarly to SMAP/Sentinel-1 product retrievals labelled as good-quality but significantly better than low-quality retrievals. The crop field boundaries were discernible in the 200 m agricultural area SM maps. The proposed method offers a complementary option for areas with extensive field measurements for model training. • A time series approach developed for Sentinel-1 200 m soil moisture mapping • A year of Sentinel-1 time series as input without assuming little vegetation change • Handle the variations in the number of annual Sentinel-1 observations across pixels • Trained and tested using 181 ISMN cropland and grassland sites across CONUS • Performed better than the 3 km and 1 km SMAP/Sentinel-1 products