The Construction and Migration of a Multi-source Integrated Drought Index Based on Different Machine Learning

Traditional methods of drought monitoring have high precision on the meteorological station scale, which needs to arrange numerous stations. Although the single-factor drought remote sensing index contains one or two parameters based on a single indicator can realize real-time and dynamic monitoring, which cannot accurately reflect drought information. Combining meteorological station data with remote sensing data and using machine learning to establish drought monitoring models has high data accuracy and spatiotemporal advantages. In this study, standardized indices of precipitation (P), land surface temperature (LST), evapotranspiration (ET), potential evapotranspiration (PET), normalized difference vegetation index (NDVI), soil moisture (SM), sun-induced chlorophyll fluorescence (SIF) and digital elevation model (DEM) are applied as independent variables, the one-month standardized precipitation index (SPI_1) as the dependent variable, a Multi-source Integrated Drought Index (MIDI) was constructed by Random Forest (RF), Back Propagation Neural Network (BP), and Support Vector Machine (SVM). MIDI was employed to monitor drought conditions in the North China Plain. Moreover, MIDI was utilized for monitoring the typical drought event in southwest China to verify its migration ability. The results showed that the correlation coefficients between each standardized indices and SPI_1 were all higher than the standardized evapotranspiration index (SPEI_1) except for surface temperature and potential evapotranspiration. Therefore, SPI_1 was selected as the dependent variable to construct MIDI. MIDI established by RF had higher accuracy in monitoring drought (R2 = 0.789, RMSE = 0.454, and MAE = 0.348) than BP and SVM. The correlation coefficient between MIDI and SPEI_1 was greater than 0.8 (P