Spatial Enhancement of Satellite Soil Moisture Using Machine Learning Approach
Harshita Rani Ahirwar
Department of Farm Engineering, Institute of Agricultural Sciences, Banaras Hindu University, Varanasi-221005, India.
Manish Kumar Nema
Water Resources Systems Division, National Institute of Hydrology, Roorkee, Haridwar, India.
Anupam Kumar Nema
Department of Farm Engineering, Institute of Agricultural Sciences, Banaras Hindu University, Varanasi-221005, India.
Vijay Shankar Yadav
National Institute of Hydrology, Central India Hydrology Regional Centre, Bhopal-462016, India.
Aaditya Jadhav *
Department of Farm Engineering, Institute of Agricultural Sciences, Banaras Hindu University, Varanasi-221005, India.
*Author to whom correspondence should be addressed.
Abstract
Background: Satellite-derived soil moisture supports agricultural and hydrological assessments, but the spatial resolution of the Soil Moisture Active Passive (SMAP) Level-3 radiometer product limits its representation of landscape-scale variability.
Aim: The current study aims to downscale the Soil Moisture Active Passive Level-3 radiometer soil moisture product from approximately 9 km to 1 km because the coarse resolution of satellite soil moisture may not capture field and landscape-scale variability that matters for agricultural monitoring, drought assessment, and hydrological applications.
Data and Study Area: The SMAP satellite soil-moisture product at 9 km resolution was selected for downscaling because of its widespread use and agreement with ground observations. Moderate Resolution Imaging Spectroradiometer (MODIS) predictors, namely daytime and night-time land-surface temperature (LST), diurnal LST, surface-reflectance bands (5, 6 and 7) and land-use/land-cover (LULC) were used at various resolutions. Ground soil-moisture observations from the Soil Moisture India Network were used to validate the coarse and fine resolution data. The study area was Pune district, which contains a SMIN station for validation of satellite-derived soil moisture against in-situ observations.
Methodology: A Random Forest (RF) machine learning framework was developed using Moderate Resolution Imaging Spectroradiometer (MODIS) predictor information (LST, surface reflectance and LULC). Downscaling was conducted for 2023. All datasets were spatially harmonised and summarised in 10-day periods. The model was trained on the native SMAP grid using predictors aggregated to the same scale, and then applied to the 1 km predictors to generate soil moisture estimates at the finer resolution. The downscaled product was assessed at the station level against ground observations for 2017–2023 wherever both satellite and ground observations were available for a 10-day composite.
Results: The station-based assessment showed that the RF-downscaled 1 km product provided greater spatial detail than the original SMAP 9 km product. The station-based assessment at Pune showed that the RF-downscaled 1 km product agreed more closely with in-situ observations than the original SMAP 9 km product. The correlation coefficient (r) increased from 0.753 for the 9 km product to 0.802 for the 1 km product, while the root mean square error (RMSE) decreased from 4.514 to 3.654 and the mean absolute error (MAE) decreased from 3.451 to 2.495.
Conclusion: The Random Forest workflow, driven by MODIS-derived predictors, can provide a more spatially detailed representation of soil moisture at 1 km resolution relative to the native SMAP product, making it a promising approach for applications that need finer-scale soil moisture information.
Keywords: SMAP, soil moisture, downscaling, Random Forest, SMIN, remote sensing