A cluster-based local modeling paradigm for high spatiotemporal resolution VPD prediction using multi-source data and machine learning

Vapor Pressure Deficit (VPD) is a critical environmental variable in terrestrial ecosystem and climate modeling, with broad applications in agriculture and hydrology. Currently, VPD reanalysis data suffers from relatively low resolution and insufficient accuracy validation. This study introduces a Cluster-Based Local Modeling (CBLM) paradigm, integrating meteorological data, surface characteristics, and in situ observations to achieve high-precision VPD prediction using machine learning. By comparing global and local modeling performances across various machine learning algorithms, the optimal model is selected for VPD inversion. The results show that the local modeling significantly enhances prediction accuracy, with the XGBoost model outperforming others across all clusters and maintaining high precision across seasons scales and different land use types. Spatiotemporal distribution statistics reveals higher prediction errors in southwestern mountainous and plateau regions, primarily due to heterogeneous local features driven by complex topography and climate. Furthermore, this study elucidates the dynamic influences of hydro-meteorological processes, climatic conditions, and terrain on VPD, highlighting their pronounced spatial heterogeneity. Given its high scalability, the proposed CBLM framework is applicable to large-scale, high-resolution VPD prediction globally, providing a reliable approach for refined VPD estimation.