Coupling Machine Learning and Detrended Cumulative Drought Index for Lake Area Prediction

Lakes play a crucial role in supplying water resources, regulating regional climates, and supporting ecosystems. However, they are increasingly threatened by recurrent droughts. This study focuses on the Mu Us Sandy Land, the fourth largest desert in China, and presents an approach that combines machine-learning techniques with a newly constructed drought index—the detrended cumulative standardized precipitation-evapotranspiration index (DeCumSPEI)—to forecast monthly lake-area variations from 2001 to 2020. Remote sensing data, including lake-area measurements, were obtained using the Google Earth Engine platform. In addition, meteorological factors—such as precipitation, temperature, and actual evapotranspiration (ET_a) as well as anthropogenic variables, such as crop evapotranspiration, the normalized difference vegetation index, and land-use and land-cover change—were collected. The study assessed the performance of six machine-learning models using fivefold cross validation: gradient boosting decision tree, extra trees, random forest, adaptive boosting (AB), bootstrap aggregating (Bagging), and eXtreme gradient boosting. These models were evaluated for their ability to predict lake areas under both short-term (monthly) and long-term (annual) drought conditions. In addition, the influence of ET_a as an upper boundary condition was investigated. The results show that: First, all models, except for AB and Bagging, demonstrated strong predictive performance, achieving coefficients of determination (R²) as high as 0.833, and the lowest average root-mean-square error and standard deviation of 1.05 and 1.01 km². Second, incorporating the 12-month scale DeCumSPEI significantly enhanced model accuracy, with performance improvements of up to 32.01%. Third, comparing models with and without ET_a confirmed the critical role of ET_a in improving prediction accuracy. These findings offer valuable insights for future lake area forecasting in drought-affected regions and underscore the potential of machine-learning models in hydrological and drought response research.