Near real-time satellite soil moisture estimation via residual learning integrated with sensor networks

Abstract

Soil moisture (SM) is crucial for climate dynamics, hydrological processes, agricultural productivity, drought and flood management. However, real-time SM monitoring remains challenging due to sparse in-situ observations. This study presents a novel sensor driven residual learning framework that integrates multi-source data—including in-situ measurements (COSMOS-UK), satellite information (SMAP, AMSR2/GCOM-W1, SMOPS, and MODIS), and meteorological variables to generate high-precision, near real-time SM estimates across the United Kingdom (UK). The methodology employs a two-stage machine learning approach: the first stage utilizes an ensemble model to generate initial SM estimates, while the second stage applies residual learning informed by automated sensor networks to refine these estimates by correcting systematic deviations observed in the UK. Unlike conventional approaches that rely on historical time-series data, this framework demonstrates that reliable SM estimation can be achieved using single-time satellite observations with in-situ data, enabling near real-time monitoring. Initial SM estimates achieved an R² of 0.75 across 40 stations, with 37 stations achieving >70 % relative accuracy. Interestingly, residual analysis within the model revealed comparatively large residuals in central and southern UK regions, and the final refined SM estimations through residual learning improved the R² to 0.94. This computationally efficient, scalable framework offers a robust solution for data-sparse regions, advancing near real-time hydrological forecasting, drought assessment, and climate resilience strategies.