Drought monitoring using downscaled soil moisture through machine learning approaches over North and South Korea

Abstract:

Soil moisture is one of the most important key variables for drought monitoring. It reflects hydrological and agricultural processes because soil moisture is a function of precipitation and energy flux and crop yield is highly related to soil moisture. Many satellites including Advanced Microwave Scanning Radiometer on the Earth Observing System (AMSR-E), Soil Moisture and Ocean Salinity sensor (SMOS), and Soil Moisture Active Passive (SMAP) provide global scale soil moisture products through microwave sensors. However, as the spatial resolution of soil moisture products is typically tens of kilometers, it is difficult to monitor drought using soil moisture at local or regional scale. In this study, AMSR-E and AMSR2 soil moisture were downscaled up to 1 km spatial resolution using Moderate Resolution Imaging Spectroradiometer (MODIS) data—Evapotranspiration, Land Surface Temperature, Leaf Area Index, Normalized Difference Vegetation Index, Enhanced Vegetation Index and Albedo—through machine learning approaches over Korean peninsula. To monitor drought from 2003 to 2014, each pixel of the downscaled soil moisture was scaled from 0 to 1 (1 is the wettest and 0 is the driest). The soil moisture based drought maps were validated using Standardized Precipitation Index (SPI) and crop yield data. Spatial distribution of drought status was also compared with other drought indices such as Scaled Drought Condition Index (SDCI). Machine learning approaches were performed well (R=0.905) for downscaling. Downscaled soil moisture was validated using in situ Asia flux data. The Root Mean Square Errors (RMSE) improved from 0.172 (25 km AMSR2) to 0.065 (downscaled soil moisture). The correlation coefficients improved from 0.201 (25 km AMSR2) to 0.341 (downscaled soil moisture). The soil moisture based drought maps and SDCI showed similar spatial distribution that caught both extreme drought and no drought. Since the proposed drought monitoring approach based on the downscaled soil moisture considered not only high resolution soil moisture but the other climatic and environmental characteristics such as precipitation, vegetation, and land surface fluxes, it is expected to provide useful drought information adaptive to different drought types.