Spatio-Temporal Modelling Of Field-Scale Soil Moisture Using Sentinel-1 SAR and Machine Learning In a Continental Agroecosystem

Abstract

This study focuses on the spatio-temporal modelling of field-scale soil moisture using Sentinel-1 SAR data and machine learning techniques in a continental agroecosystem in Nyírbátor, Hungary. It integrates multi-year satellite observations (2022–2025) with in-situ soil moisture measurements collected from a maize field to improve estimation accuracy. Machine learning models, particularly Random Forest and XGBoost, were applied to analyse the relationship between SAR backscatter (VV and VH) and soil moisture dynamics. The results show that XGBoost significantly outperformed Random Forest, achieving near-perfect accuracy in predicting soil moisture across all years. Seasonal and inter-annual analyses revealed consistent drying trends from June to August, influenced by precipitation and temperature variability. Overall, the study demonstrates the strong potential of Sentinel-1 SAR combined with machine learning for accurate, high-resolution soil moisture monitoring in agricultural systems.