Soil Moisture Estimation with Geospatial Foundation Models and Satellite Imagery

Abstract

Soil moisture is a critical parameter influencing agricultural practices, climate conditions, hydrology, and various natural phenomena, including nature-based carbon sequestration. Estimating soil moisture at finer spatial resolutions poses significant challenges for traditional physics-based models due to their computational demands and dependence on extensive input variables at finer scales. Alternatively, data-driven approaches also require large amounts of ground truth data representing various climatic conditions to be generalizable across regions. To address these challenges, we propose adopting a Geospatial Foundation Model (GFM) pre-trained on a large corpus of satellite imagery for generic tasks and fine-tuning it with limited ground truth soil moisture data to map soil moisture at finer spatial scales. GFMs have garnered significant attention for their versatility in earth observation tasks such as classification and segmentation, but have seen limited study in estimating land surface variables like soil moisture. This study aims to showcase the efficacy of GFMs for soil moisture estimation at finer scales. To this end, we curated a comprehensive dataset, leveraging 12 different satellite bands from Sentinel-1 and Sentinel-2, along with multi-year observations from 40+ soil sensors densely placed around the Texas region (TxSON). The data curation process involved mosaicking, resampling, and co-registration to create 64x64 patches at a spatial resolution of 10m, with each patch containing at least one sensor observation. In total, we gathered 800 such patches from 2019 to 2021 to initiate our experiments. Before fine-tuning the GFM, we developed two baseline models: (a) linear regression and (b) U-Net-based regression. The U-Net model was selected for its proven efficacy in handling multi-resolution imagery and its encoder-decoder architecture. To address the challenge of missing values in the target soil moisture dataset, where only a few pixels have valid values, we implemented a customized loss function that only considers valid observation pixels and adjusted the finer layers of the U-Net model to perform regression tasks. This approach ensures that the model's performance is evaluated only on relevant data, improving the reliability of the results. We conducted numerous experiments with variations in data and models, reporting accuracy metrics of (a) the linear regression model with a test MAE of $0.062 {m^³/m^³}$ and RMSE of $0.081 {m^³/m^³}$, and (b) the U-Net, which achieved a test MAE of 0.0357 and RMSE of $0.055 {m^³/m^³}$. For GFM fine-tuning, we utilized the Prithvi-100M model, which was jointly developed by IBM and NASA. Fine-tuning experiments are ongoing, and results will be compared with the baseline models and presented. The fine-tuned experiments for soil moisture are expected to contribute new findings to the community.