Validation-aware interpretable machine learning for sustainability-oriented monitoring of monthly NDVI anomalies in an alpine basin of Qinghai, China

This study evaluates the accuracy and spatiotemporal generalization of machine-learning estimates of baseline-referenced monthly NDVI anomalies in an alpine study area in Qinghai, China. Pixel-wise reference anomalies were calculated from a 2006 to 2015 monthly climatology, and predictors from 2016 to 2025 included contemporaneous and lagged hydrothermal variables, antecedent NDVI anomalies, and terrain and soil attributes on a common 30 m grid. Target-month NDVI was used to define the observed response, not as a direct model predictor. Five algorithms were assessed using random-split and spatial-block validation, together with a 2025 withheld-year assessment that also informed model selection. Under predefined configurations, gradient boosting decision trees achieved the most favorable overall withheld-year validation performance. A reduced configuration combining month, contemporaneous hydrothermal variables, and lagged NDVI anomalies achieved R 2 = 0.263 and RMSE = 1.145 in 2025, compared with R 2 = 0.254 and RMSE = 1.152 for the full configuration. Lagged NDVI supplied the strongest predictive signal, whereas terrain and soil covariates added little skill. The results characterize same-month anomaly estimation under the evaluated conditions; applications to missing-data reconstruction and management decision-making require separate validation.