Research on comprehensive drought index prediction model based on CNN-LSTM

Ordos is a significant mineral-producing region and economically advanced in Inner Mongolia; yet, due to factors including as geography and climate, drought catastrophes frequently occur, which have become a critical constraint on regional growth. This study constructed the Regional Comprehensive Drought Index (RSDI) for the period 2001-2020 based on a random forest model, using disaster-causing factors such as temperature, precipitation, NDVI, surface soil moisture, land use type, and DEM as independent variables and SPEI as the dependent variable. The period from 2001 to 2016 was used for model training and validation, and the period from 2017 to 2020 was used for model testing. Three deep learning models, namely Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and CNN-LSTM hybrid model, were used to predict RSDI. The results showed that: (1) RSDI was consistent with the drought situation in the study area in 2005, and it showed a highly significant positive correlation with SPI at all four scales (r > 0.90, P < 0.01). (2) The predictive efficacy of the three models-CNN, LSTM, and CNN-LSTM-was compared and analysed over prediction durations of four years (2017-2020), two years (2019-2020), and one year (2020). Ultimately, it was determined that the hybrid CNN-LSTM model exhibited superior predictive and fitting performance compared to the individual models, with a reduced prediction error relative to the others. The two models experienced significant reductions, with the root-mean-square error decreasing by 0.30 and 0.22, and the mean absolute error diminishing by 0.24 and 0.13. The research results improved the prediction accuracy of the regional drought index and provided a technical reference for drought monitoring and prediction in the Ordos region.