Spatiotemporal fusion for crop model assimilation: Linfen Basin case study, Chinese Loess Plateau

Spatiotemporal fusion addresses challenges in crop growth monitoring: namely, spatial discreteness of high‐resolution imagery and spectral mixing in high‐temporal‐frequency data. These can hinder accurate yield and phenology estimates, especially in fragmented semi‐arid landscapes. To improve winter wheat monitoring, we quantitatively and qualitatively evaluated two normalized differential vegetation index (NDVI) fusion methods: the spatial and temporal nonlocal filter‐based fusion model (STNLFFM) and enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM). Our primary focus was to address the limitations of MODIS data in crop phenology monitoring, specifically to resolve the challenges of spatial and temporal resolution. STNLFFM, incorporating inter‐image coefficients and temporal variation, outperformed ESTARFM by eliminating stripe artifacts from prolonged high‐resolution data gaps. Additionally, assimilating NDVI data via the four‐dimensional variational (4DVAR) method resulted in improved accuracy, with a 6.001% mean absolute percentage error in leaf area index (LAI) estimation, compared with 6.285% using the CERES‐Wheat model. Yield estimation accuracy was enhanced by 2.734% through the 4DVAR‐assimilated LAI, particularly in addressing inaccuracies in mountain‐cropland transition zones. This study addresses this gap by providing an integrated spatiotemporal fusion framework that mitigates data quality and scale mismatches, thereby improving the accuracy of crop growth monitoring and yield predictions. This study highlights the following: (1) the operational advantages of STNLFFM in fragmented landscapes and (2) the potential of variational assimilation to reduce model uncertainties. The proposed approach is applicable to precision agriculture in topographically complex regions and provides a scalable solution for addressing mixed‐pixel challenges in the Earth's observational data.