Drought monitoring using Enhanced Soil Moisture Drought Index (ESMDI) downscaled with deep learning from multi-satellite data for achieving food and water security

Drought conditions are often assessed based on the available soil moisture from land surface models. However, these models operate at low resolution, rendering them suitable primarily for large-scale drought monitoring while constraining their ability to capture variability at the landscape level. The present study developed an Enhanced Standardized Soil Moisture Drought Index (ESMDI). The development of ESMDI was based on the 1 km downscaled soil moisture data from the Global Land Data Assimilation System (GLDAS_CLSM025) from 2004 to 2023 over the Gyeongsangbuk-do region. Three models- Deep Believe Network (DBN), Random forest (RF), and Extreme Gradient Boosting (XGB) were models used to downscale GLDAS soil moisture using six Moderate Resolution Imaging Spectroradiometers (MODIS), which include albedo, land surface temperature (LST), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Leaf Area Index (LAI), and evapotranspiration (ET)., precipitation data from Climate Hazards Group InfraRed Precipitation with Station Data (CHIRPS-V2.0), Digital Elevation Model (DEM) and bulk density. The downscaled soil moisture was validated against ground-based measurements. Among the three evaluated models, DBN outperformed in terms of in situ comparisons, achieving an average R-score of 0.93, a Root Mean Square Error (RMSE) of 0.0266 m³/m³, a Mean Absolute Error (MAE) of 0.0158 m³/m³, and a Bias of 0.0026 m³/m³ across the ten observation stations selected. ESMDI was developed by normalizing the downscaled soil moisture. The strong correlation of ESMDI with meteorological and hydrological drought indices, particularly in the spring and autumn seasons, and crop yield in July, indicates its effectiveness in drought management.