60-m Resolution Soil Moisture Estimation Based on a Multisensor Feedforward Neural Network Model

Understanding soil moisture (SM) at high spatio-temporal resolution provides crucial insights across various societal disciplines due to its direct impact on environmental and natural disaster monitoring, weather forecasting, agricultural productivity, and water resource management. In recent decades, a variety of algorithms have been developed to improve the spatial resolution of SM maps from passive sensors (∼40 km); however, the resulting maps, often with resolutions around 1 km or even hundreds of meters, still lack the necessary resolution for detailed local analysis. This study addresses this gap by presenting a machine learning methodology aimed at estimating SM at 60-m spatial resolution. A feedforward neural network is employed to capture the relationships among 14 different predictors, including several spectral bands and indices from Sentinel-2, land surface temperature from moderate-resolution spectroradiometer, elevation and slope from shuttle radar topography mission, precipitation from the fifth generation of the European Center for Medium-Range Weather Forecasts Reanalysis for Land, and the sand fraction from SoilGrids250m, with European Space Agency (ESA) Climate Change Initiative (CCI) SM serving as the target variable. The model is trained and applied over the central part of the Iberian Peninsula (38.9 °N–42.5 °N and 3.5 °W–7.2 °W) from 2019 to 2021. At 60-m resolution, the SM maps effectively capture the spatial heterogeneity of the terrain. The temporal analysis demonstrates that high-resolution SM maps preserve virtually the same sensitivity as those from the ESA CCI, with a correlation of 0.66, a bias of 0.095 m
³
/m
³
, and an unbiased root-mean-square error of 0.044 m
³
/m
³
on average.