Surface Soil Moisture from Sentinel-2 Imagery: A Systematic Review Complemented by a Case Study in Sardinia, Italy

Abstract

This study combines a systematic review with a case study to address the following question: can Sentinel-2 data yield accurate Surface Soil Moisture (SSM) estimates? Among the 1158 papers identified through the review, 66 met the eligibility criteria based on the exclusive use of Sentinel-2 data to estimate SSM and on reference measurements acquired simultaneously. Analysis of the eligible papers reveals six interconnected critical issues, the foremost being the insufficient spatial and temporal density of point-based reference measurements and their limited availability. The case study is aimed at addressing these issues. Specifically, Theia SSM products—freely available at the plot scale with an accuracy of approximately 5 vol%—are used as reference measurements, enabling a three-year multi-temporal comparison with Sentinel-2 bands. The comparison across different land and vegetation cover types, including a burned area, shows that SSM retrieval accuracy from Sentinel-2 can be strongly modulated by vegetation status and soil moisture magnitude. Specifically, the maximum R2 (Sentinel 2-bands against Theia SSM) increases by 0.44 (from 0.01 to 0.45) where NDVI is less than 0.35, and SSM is less than 15%. Moreover, extending the analysis to a multi-year time series improves R2 relative to single-date results (i.e., by up to 0.43). Across eligible papers, different methodologies (21 machine-learning algorithms, 18 optical trapezoidal models, and 7 statistical methodologies), different image processing products (4 algorithms, 77 indices, principal component, and albedo), and all bands were used to identify the best methodology and/or optimal image processing products and/or optimal bands. The median R2 for these outputs against the SSM reference data ranged from 0.44 (statistical methods) to 0.63 (optical trapezoidal models), indicating weak to moderate overall accuracy. Unlike other approaches, the optical trapezoidal models rely primarily on SWIR bands. Results show that SWIR bands yield moderate fit (R2 up to 0.65) during the dry season, but negligible fit during the wet season. These values align with those of the case study, suggesting that the use of the Sentinel-2 imagery for SSM estimation is critical and requires well planned approaches. Findings reported in this paper provide concrete guidance on image selection, reference measurement design, and time-series length for researchers seeking reliable SSM estimation from Sentinel-2 data.