Integrating drought indices and machine learning approaches to predict rainfed potato yield in North China

Although northern China has the largest potato cultivation area and highest total production nationwide, water scarcity is still the most important limiting factor on potato yield in this region. This situation necessitates accurate yield prediction for combating drought and formulating food allocation policy to safeguard food security. Based on simulated potato yields by well-validated APSIM-Potato model, we developed a prediction method of rainfed potato yield in northern China through integrating various drought indices (Standardized Precipitation Index, SPI; Standardized Precipitation Evapotranspiration Index, SPEI; Relative Moisture Index, RMI; Composite Meteorological Drought Index, CI) and machine learning approaches (Multiple Linear Regression, MLR; Support Vector Machine, SVM; Ridge Regression, RR; XGBoost; Random Forest, RF). The APSIM-Potato model simulated potato yield and water consumption with relative root mean square errors (coefficients of determination) of 21.8% (0.91) and 13.6% (0.87) compared to observed values, indicating the model accurately captures potato's response to drought. Simulated potential potato yields in northern China ranged from 16.3 to 72.4 t ha−1, with rainfed yields achieving 30–55% of the potential yields. The RMI demonstrated superior capability (R²=0.48) for predicting potato yield reduction rate compared to SPEI (R²=0.44), CI (R²=0.29), and SPI (R²=0.09). The RF model, coupled with drought index combinations during the tuber initiation phase (from tuberization to 30 days post tuberization), demonstrated the highest predictive capability for potato yield reduction, achieving a simulation accuracy of 88%. Under climate change emission scenarios SSP245/SSP585, future potato yield reduction rates in northern China remain relatively stable compared to current climate condition except a 20% decrease in yield reduction rate in the northern agro-pastoral ecotone under the SSP585 scenario during 20712100. This study provides an efficient method for predicting rainfed potato yields up to one month in advance, offering crucial support for crop yield forecasting and disaster early warning in rainfed agricultural system.