Improving satellite microwave sensing of global soil moisture via radiative transfer process-guided machine learning

Accurately estimating top ~5 cm surface soil moisture (SM) is highly valuable for understanding the terrestrial water cycle. Based on the zero-order τ-ω radiative transfer model (RTM), the Soil Moisture Active Passive (SMAP) mission has provided daily global surface SM estimations at 9 km spatial resolution using L-band (1.41 GHz) radiometry since April 2015. As the parameterization of RTM for SMAP's official algorithm highly relies on in-situ measurements, SMAP SM has weaker performance in regions with few calibration sites. To improve the accuracy of global SM estimations, we developed a new radiative transfer Process-Guided Machine Learning (PGML) method, which integrates the mechanistic understanding of RTM and data-driven machine learning approaches to estimate global SM. We generated a synthetic dataset from RTM and developed a pre-trained PGML to quantify SM by using this synthetic dataset. Furthermore, we utilized SM measurements at 1131 in-situ sites collected from International Soil Moisture Network (ISMN) during April 2015 and December 2023 across the globe to fine-tune PGML. The validation result shows that the estimated  9-km daily PGML global SM has a good agreement with in-situ SM measurements from ISMN. Our model has significantly better performance on estimating global SM  than the SM retrievals from RTM (R from 0.413 to 0.636, RMSE from 0.132 to 0.100 m³/m³, bias from 0.042 to 0.001 m³/m³, ubRMSE from 0.125 to 0.100 m³/m³). This study highlights the potential of PGML to integrate machine learning and radiative transfer models for accurate remote sensing of SM at the global scale.