A large-scale spatiotemporal NDVI prediction method coupling multivariable constraints and structured modeling

Accurate prediction of vegetation dynamics is essential for understanding terrestrial ecosystems carbon and water cycles. However, existing Normalized Difference Vegetation Index (NDVI) prediction methods are limited in their ability to represent spatiotemporal heterogeneity and nonlinear interactions among multiple driving variables, which restrict predictive accuracy and extrapolation performance at large scales. To address these limitations, this study proposed a structured large-scale NDVI spatiotemporal prediction method incorporating multivariable constraints (MCS-LSTP) to characterize vegetation dynamics in complex and heterogeneous environments. MCS-LSTP incorporates multiple key driving variables and enhances the comprehensive representation of complex spatiotemporal dependencies and spatial heterogeneity through homogeneous spatial unit partitioning, unit-level temporal modeling, and continuous surface reconstruction. Results showed that MCS-LSTP model exhibited superior performance in terms of stability and prediction accuracy when compared to AutoRegressive Integrated Moving Average (ARIMA), Random Forest (RF), and Transformer models. The mean Population Stability Index (PSI) reached 0.0572, with 98.7% of prediction units exhibiting PSI values below 0.1, indicating no significant deviation between the observed and simulated NDVI values. The average R2 and RMSE on the test set were 0.8551 and 0.2171, respectively. Furthermore, MCS-LSTP accurately reproduced seasonal NDVI fluctuations and meridional spatial gradients, with particularly high agreement between simulated and reference values during the growing season. This study provides a robust and scalable solution for large-scale NDVI spatiotemporal prediction and offers critical technical support for ecological risk assessment in the Yellow River Basin and other ecologically vulnerable regions.