Integrated meteorological drought monitoring framework using multi-sensor and multi-temporal earth observation datasets and machine learning algorithms: A case study of central India

Devising strategies for drought planning and risk mitigation in predominantly small holder agriculture regions like South Asia requires improved understanding of spatio-temporal characteristics of droughts at finer spatial scales. In this study, an integrated framework is developed to generate high-resolution gridded precipitation products from publicly available coarse scale data, and use the downscaled products for analysis of meteorological droughts. Data from a region in Central India is used to develop and test the framework. The Standard Precipitation Index (SPI) is used to characterize meteorological drought. The proposed framework incorporates: (i) a Random Forest machine learning algorithm to downscale CHIRPS (5 km) gridded long term rainfall data to a higher resolution gridded product (1 km) using high resolution multi-sensor and multi-temporal earth observation data of land surface characteristics (vegetation, temperature, and topography) as cofactors; (ii) Computation of 3-month SPI (SPI-3) for the downscaled gridded rainfall; (iii) Run theory applied to downscaled SPI grid to determine drought characteristics - duration, severity, intensity, frequency, and onset and cessation of droughts, and (iv) Principal Components Analysis (PCA) to combine effects of multiple drought characteristics into a single composite Effective Meteorological Drought Index (EMDI), to identify sub-regional drought patterns. The downscaled gridded precipitation product (1 km) was validated using: (i) original CHIRPS (5 km) data, and (ii) measured rainfall data from 65 rain gauge stations in the study area, with respect to conservation of both statistical properties and spatial structure. Comparisons were also made with drought characteristics obtained from Standardized Evapotranspiration Index (SPEI) derived from MODIS evapotranspiration data. The downscaled product captures spatio-temporal variability of droughts at finer village and sub-village scales, compared to original CHIRPS (5 km) data for which each pixel encompassed multiple villages. Spatial distribution of composite index EMDI delineates the study area into two characteristic drought risk regions. In one region, droughts are of longer duration, but are less frequent, with lower severity and intensity. In the second region droughts are of shorter duration but of higher severity, intensity and frequency. The integrated framework developed in this study for high resolution spatio-temporal analysis of droughts, starting from coarse scale precipitation data, is practical and sufficiently general to adopt in other regions to support local drought risk planning and targeting mitigation decisions and actions.