A novel agricultural drought index based on multi-source remote sensing data and interpretable machine learning

Drought is a frequent, destructive, and complex natural hazard, and seriously threatens eco-environment, socio-economy, and the health of human. Previous studies suggested that integrated multi-source remote sensing drought indices have the potential to comprehensively monitor drought conditions, however most existing integrated drought indices still have several limitations. Here, we used solar-induced chlorophyll fluorescence, water balance, soil moisture, and land surface temperature to develop a new integrated remote sensing drought index, namely interpretable machine learning drought index (IMLDI), based on the Bayesian optimized tree-based Light Gradient Boosting Machine and SHapley Additive exPlainations. The different land cover types were further considered, and the categories of drought severity were objectively determined by the iterative optimized method. The drought monitoring performance of IMLDI was validated in the eastern parts of China, and three integrated drought indies composited by PCA, multiple linear regression, and gradient boosting method were also included for comparison. The results show that IMLDI has a higher spatial and temporal consistency with standardized precipitation evapotranspiration index, can better reflect the real-world observed drought-affected cropland areas and gross primary production, and can also well describe the evolutions of 2009/2010 and 2019 drought events in the eastern parts of China, indicating higher drought monitoring performance of IMLDI. Besides, IMLDI-based agricultural drought risk analysis shows that the Huang-Hai Region and Yunnan, Guizhou, and Guangxi Provinces have a high risk to suffer from severe agricultural droughts. Overall, IMLDI has a great potential to use as a new integrated remote sensing drought index for agricultural drought monitoring.