Speaker
Description
Soil moisture (SM) is a critical variable in hydrological, ecological, and agricultural domains, playing a pivotal role in irrigation management, drought monitoring, and crop yield prediction. While in-situ measurements provide highly accurate SM data, they are costly and spatially limited. Remote sensing techniques enable large-area monitoring, but current satellite SM products, with spatial resolutions of 1 km or coarser, struggle to capture the rapid SM variations required at the agricultural field scale. Synthetic aperture radar (SAR) offers the spatial detail needed for such applications, but its backscatter is strongly influenced by surface roughness and vegetation, which complicates direct SM retrieval.
To address these challenges, this study proposes a high-resolution surface soil moisture (SSM) retrieval framework that integrates SAR and other data into a change detection-based downscaling (CD-DS) method. The framework integrates Sentinel-1 C-band SAR (VV and VH polarization), Sentinel-2 normalized difference vegetation indices (NDVI), and the 1 km Disaggregation based on Physical And Theoretical scale Change (DISPATCH) SSM product as background information, and is applied over Catalonia, a Mediterranean agricultural region with frequent droughts and a dense network of soil measurement stations. The key SAR variable is the backscatter difference derived from the CD approach: for each 30 m cell, and within discrete NDVI classes, the backscatter of each acquisition is referenced to the minimum backscatter observed under similar vegetation conditions, which represents the driest state. By comparing radar signals acquired over the same cell under approximately constant NDVI, the contributions of surface roughness and vegetation structure are largely removed and the remaining signal is dominated by SM changes. Sentinel-1 and Sentinel-2 data from 2017 to 2024 were used, and in-situ SSM measurements (5 cm depth) from 18 stations of the Soil Measurement Network of the Institut Car tografic i Geologic de Catalunya (ICGC) soil monitoring network served as labels: 13 stations for training (2017–2021) and all 18 stations for independent validation (2022–2024).
Four ML algorithms—Support Vector Regression, Random Forest, XGBoost, and LSTM—were evaluated. XGBoost outperformed the other models, achieving an R² of 0.933 and an RMSE of 0.023 cm³/cm³ on the test set. Validation across the 18 stations, which cover forest, vineyard, grassland, shrubland, and fruit-tree land cover, gave an average correlation coefficient (R) of 0.63, a ubRMSE of 0.045 cm³/cm³, and a bias of 0.024 cm³/cm³. A comparison with 54 sampling points from a field campaign carried out in the Urgell irrigated area in 2015 gave an R of 0.82 and an RMSE of 0.06 cm³/cm³. Feature importance analysis based on SHAP values showed that the backscatter difference is the second most influential variable after the background SSM, confirming the contribution of the CD variable to the retrieval accuracy. The retrieved SSM time series follow the in-situ observations across the years, and capture the abrupt increases in SM following precipitation and irrigation events. In terms of spatial distribution, the 30 m SSM maps closely match the SIGPAC irrigation type maps, with higher SSM values in irrigated parcels and lower values in rainfed farmland.
The results show that introducing a change detection variable into an ML-based downscaling scheme is an effective way to obtain field-scale SSM from Sentinel-1 SAR data. The proposed framework delivers high-resolution SSM estimates that can support precision irrigation, drought monitoring, and local water resource management.
Keywords: surface soil moisture; Sentinel-1; change detection; downscaling; machine learning; XGBoost; Catalonia