Speaker
Description
Soil moisture is the amount of water contained in the unsaturated soil zone and is a key variable controlling plant water uptake, crop yields, groundwater recharge, and land–atmosphere exchanges of heat and carbon. Accurate knowledge of its spatial and temporal distribution is therefore important for applications such as precision agriculture, drought monitoring, and flood forecasting.
Remote sensing enables frequent, non-invasive soil moisture monitoring over large areas. Microwave radiometer missions such as NASA’s Soil Moisture Active Passive (SMAP) and ESA’s Soil Moisture and Ocean Salinity (SMOS) provide near-daily global observations at kilometer-scale resolution. More recently, Synthetic Aperture Radar (SAR), particularly the Sentinel-1 constellation, has attracted considerable interest because it provides observations at resolutions down to 10 m and offers the potential for higher-resolution soil moisture products.
In this study, we investigate a deep learning (DL) approach for time-tagged soil moisture retrieval by combining, for the first time, backscatter and repeat-pass interferometric information from multi-temporal Sentinel-1 data. Because high-quality reference measurements are scarce, we adopt a two-step training strategy. First, the model is weakly supervised using a large dataset of lower-accuracy soil moisture estimates. It is then fully supervised and fine-tuned using reliable in-situ measurements at 0–5 cm depth.
As initial DL architecture, we rely on a state-of-the-art fully convolutional neural network (CNN), chosen for its proven ability to extract multi-scale spatial features from radar data for regression tasks [1]. The architecture comprises approximately 760,000 trainable hyperparameters, ensuring sufficient model capacity to learn complex nonlinear relationships between radar observables and soil moisture dynamics while maintaining computational efficiency.
The experimental setup is based on distinct Regions of Interest (ROIs) selected to represent diverse environmental and climatic conditions across the study area in central and west Europe. For each ROI, a three-year time span (2018–2020) of Sentinel-1 acquisitions is considered, providing a consistent temporal coverage for soil moisture retrieval and model evaluation. Each ROI comprises approximately 180 time-series samples, capturing seasonal variations and a wide range of soil and vegetation states. This multi-regional, multi-temporal configuration ensures that the model is trained and validated under heterogeneous conditions, thereby enhancing its robustness and generalization capability across different land cover types and climatic regimes. A geographic subdivision of the ROIs is performed for splitting the dataset into independent training, validation, and testing regions.
For the weakly supervised stage, the Copernicus Global Land Operations Surface Soil Moisture (CGLS-SSM) product is used as reference [2]. The DL model receives several Sentinel-1-derived inputs: SAR backscatter intensity, InSAR coherence at different temporal baselines, interferometric phase triplets, and local incidence angle (LIA). Together, these features capture surface scattering properties, acquisition geometry, and structural changes related to soil moisture. The initial target resolution is 1 km, consistent with existing products, while the framework is designed to scale toward 100–200 m resolution.
As first results on the testing dataset, we have obtained peak performance by combining both SAR, InSAR and geometry-related features as input to the network, achieving a Mean Error (ME) of 0.96%, a Mean Absolute Error (MAE) of 7.25%, a Root Mean Squared Error (RMSE) of 9.65% and a coefficient of determination (R2) of 0.87. These preliminary results show in general a promising performance for the pretext, weakly-supervised task. The results indicate that InSAR information from short Sentinel-1 time series provides complementary information that enhances the model’s ability to characterize soil moisture dynamics.
Subsequently, for regressing absolute soil moisture values, we fine-tuned the pretrained model using in-situ measurements from the International Soil Moisture Network (ISMN) stations, located in Denmark, the Netherlands, the United Kingdom, and Spain at a depth of 5 cm. The selected ISMN stations are divided into training, validation, and test sets while maintaining a similar distribution across the datasets. Starting from the model weights obtained through the weakly supervised pre-training, the fine-tuning process adapts the proposed DL model to the characteristics of the in-situ observations. Specifically, it learns to map each short Sentinel-1 time series to the corresponding absolute soil moisture measurement, thereby refining the model to capture the finer-scale relationships present in the reference dataset. The best performing model evaluated against 700 ISMN stations measurements achieves a ME of 0.0243 m3/m3, a MAE of 0.0614 m3/m3, and a RMSE of 0.072 m3/m3.
The proposed DL approach demonstrates promising performance in exploiting both SAR and InSAR features from the Sentinel-1 mission to retrieve absolute soil moisture estimates that closely agree with in situ measurements. The scalability of the proposed model will be further exploited to improve retrieval performance and enable the generation of large-scale soil moisture maps across Europe at a spatial resolution of 100–200 m.
[1] D. Carcereri, P. Rizzoli, D. Ienco and L. Bruzzone, "A Deep Learning Framework for the Estimation of Forest Height from Bistatic TanDEM-X Data," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 8334-8352, 2023.
[2] B. Bauer-Marschallinger et al., "Toward Global Soil Moisture Monitoring with Sentinel-1: Harnessing Assets and Overcoming Obstacles," in IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 1, pp. 520-539, Jan. 2019.