Incorporating digital soil mapping-derived soil properties for enhanced soil moisture prediction
Solmaz Fathololoumi, Asim Biswas
Abstract
Accurate and up-to-date soil characteristic maps are essential for addressing global challenges. This study presents an innovative approach to enhance soil moisture (SM) modeling accuracy by incorporating key soil property layers derived from digital soil mapping (DSM) into a remote sensing-based machine learning framework. We investigated the impact of incorporating key soil properties as environmental covariates in SM modeling. Using multi-temporal satellite imagery, land use and geological data, and soil characteristics from 284 ground sampling points, we first implemented a Random Forest Regression algorithm to generate maps of seven key soil properties. We then compared two SM modeling strategies: a classical approach using common environmental covariates, and a proposed strategy incorporating the modeled key soil properties as additional environmental covariates. Results showed that land surface temperature, sand content, soil organic carbon, VV polarization, elevation, and clay content were among the most influential environmental covariates in SM modeling. The proposed strategy significantly improved modeling accuracy, reducing root mean square error by 17 %, 29 %, and 30 % for July, August, and September, respectively, compared to the classical approach. Additionally, average modeling uncertainty decreased from 11.3 %, 8.4 %, and 8.4–8.4 %, 6.3 %, and 4.5 % for the same months. This study demonstrates that integrating key soil properties derived from DSM can substantially enhance SM modeling accuracy and reduce uncertainty, offering a more comprehensive and reliable approach to mapping soil-water dynamics. • Key soil properties enhance soil moisture (SM) modeling accuracy. • Integrating digital soil mapping & remote sensing reduces uncertainty in SM predictions. • Land surface temperature is the most influential covariate in SM modeling. • Proposed approach decreases root mean square error in SM modeling by up to 27 %.