Development of a Monthly Meteorological Drought Forecasting Model for Yunnan Using Machine Learning and Increment Method

Accurate prediction of meteorological drought can effectively reduce drought disaster risks and provide scientific support for disaster prevention and mitigation. Based on the 1‐month scale Standardized Precipitation Evapotranspiration Index (SPEI‐1) and 114 atmospheric circulation and sea surface temperature indices, this study employed three machine learning methods, including the Classification and Regression Tree (CART), Random Forest (RF) and Neural Network (NN), combined with a prediction increment method, to construct monthly meteorological drought prediction models for 125 meteorological stations in Yunnan Province, Southwest China. The main conclusions are as follows: (1) Time series validation indicates that the CART and NN models demonstrate predictive capability for monthly variations at approximately half of the stations and for meteorological drought occurrence at over 60% of the stations. The RF model achieves accuracies exceeding 60% and 75% for monthly variations and drought occurrence, respectively. The correlation coefficients between the predicted and observed SPEI‐1 values for the CART, RF and NN models are 0.53, 0.77 and 0.59, respectively, all significant at the 0.01 level, which can provide a reference for drought disaster prevention in Yunnan. (2) All three models perform best in predicting the no‐drought category, with 10‐year average accuracies exceeding 60%. The extreme drought category ranks second, with an average accuracy above 45%. The prediction accuracies for mild, moderate and severe drought are only approximately 19%. This is partly related to the sample sizes of these categories. More importantly, however, there exist relatively large mean absolute errors between the predicted and observed SPEI‐1 values, which should be the priority for future model improvement. (3) Spatial distribution validation shows that the RF model achieves same‐sign rates of predicted increments and meteorological drought accuracies above 65% and 70%, respectively, while the CART and NN models achieve rates above 50% and 60%, respectively. Regions with higher accuracies for all three models are mainly distributed in the dry‐hot valleys of the Jinsha River in central‐northern Yunnan and in southwestern regions that are significantly influenced by the Indian monsoon. (4) The CART and NN models achieve a probability of detection (POD) of 50% for province‐wide meteorological drought events during 2014–2024, while the RF model achieves 57%. All three models achieve POD above 60% for the spatial distribution of drought events, with higher POD regions located in central‐western Yunnan where drought events occur frequently and persist for longer durations. (5) Leave‐one‐out cross‐validation demonstrates that all three models can reasonably reproduce typical meteorological drought processes since 1971. The temporal and spatial correlation coefficients between the predicted and observed SPEI‐1 values are close to or exceed 0.50, all significant at the 0.01 level. (6) Overall, the RF model outperforms the CART and NN models in predicting 1‐month scale meteorological drought in Yunnan.