Soil moisture downscaling in the state of Oklahoma: Employing advanced machine learning

This investigation employs a machine learning downscaling framework, explicitly utilizing the Random Forest (RF) algorithm, to enhance the spatial resolution of soil moisture data obtained from the Global Land Data Assimilation System (GLDAS) model, thereby transitioning from a resolution of 27 km–1 km, which significantly increases its utility for agricultural applications. By incorporating high-resolution predictors sourced from Moderate Resolution Imaging Spectroradiometer (MODIS) products, including Evapotranspiration (ET), Normalized Difference Vegetation Index (NDVI), Leaf Area Index (LAI), and Land Surface Temperature (LST), along with supplementary datasets such as Open Land Map Soil attributes (clay and sand fractions), NASA SRTM Digital Elevation Model (30 m), and MODIS land cover classifications (MCD12Q1), this research effectively executes the downscaling of GLDAS data. These methodological enhancements produce more precise soil moisture estimations for assessing agricultural droughts. To ensure alignment with empirical observations, this methodology incorporates Quantile Mapping (QM) bias correction, thereby enhancing the accuracy of the GLDAS data about in-situ measurements gathered over 20 years across six agricultural regions in Oklahoma. Following an evaluation of the significance of each predictor via p-value analysis, the Enhanced Vegetation Index (EVI) was omitted due to its high p-value, indicating a negligible impact on the accuracy of soil moisture estimations. The research validates the downscaled soil moisture estimations through performance metrics, including Root Mean Square Error (RMSE) Correlation Coefficients (CC), and coefficient of determination (R²), resulting in an average RMSE of 0.06 (m³/m³), which signifies concordance between the downscaled data and the observed measurements. Seasonal examination reveals that the highest correlation is observed during autumn, attributed to consistent precipitation patterns, with CC and R² values reaching a maximum of 0.74 and 0.85 in humid regions; conversely, correlations decrease in semi-humid and arid regions, likely due to the influence of irrigation practices and diminished rainfall variability. In the final phase, the application of QM bias correction further mitigates errors, particularly improving accuracy in humid areas. These findings underscore the effectiveness of downscaled GLDAS data, augmented with various ancillary datasets, for regional drought monitoring and agricultural management. This study is a reliable resource for addressing data limitations and supporting informed decision-making of farming sectors facing resource constraints.