High-resolution Soil Moisture Prediction from SMOS using Machine Learning Models

Abstract

Soil moisture is essential for the land carbon cycle, surface and groundwater circulation, heat transport, energy exchange between these systems and other processes. SMOS's (Soil Moisture and Ocean Salinity) 36-kilometer spatial resolution and 3-day temporal resolution offer valuable insights into soil moisture dynamics. This research paper introduces an innovative approach to enhance our understanding and prediction of SMOS values by applying advanced machine learning models. Our research focuses on developing and implementing advanced downscaling techniques, leveraging advanced machine learning algorithms. The primary objective is to establish a robust framework for estimating soil moisture levels at multiple geographic locations within the study region of Oklahoma, USA. To achieve this, three years of SMOS (Soil Moisture and Ocean Salinity) data was integrated with remotely captured images spanning the full range of the electromagnetic spectrum, from visible to infrared wavelengths. The LSTM model performed significantly better in predicting soil moisture values with 0.041 RMSE (m3/m3) and 0.869 (R2) than the other models. © 2025 IEEE.