Abstract: To address the challenges faced by existing remote sensing spatiotemporal fusion methods, particularly sensitivity to model parameters and inadequate utilization of prior information, this study proposed a nonparametric spatiotemporal fusion method for remote sensing images of complex topography. The proposed method employs a multi-stage progressive fusion strategy. Initially, an explicit spatial degradation model was established, which formulated a semi-coupled mapping relationship between coarse- and fine-resolution image patches using a cyclic downsampling matrix. Subsequently, based on a nonparametric Bayesian framework, dictionary learning and adaptive parameter inference were performed. Multi-temporal collaborative representation was achieved by jointly optimizing a shared mapping matrix and temporal-specific residuals. During the fusion process, the first stage reconstructed intermediate-resolution images from images with known temporal features to alleviate large-scale resolution discrepancies. The second stage integrated both images, accomplishing fine high-resolution reconstruction through cross-scale sparse coding. Applying to the Landsat7 Enhanced Thematic Mapper (ETM+) and Moderate Resolution Imaging Spectroradiometer (MODIS) datasets, the proposed method outperformed existing approaches across multiple quantitative metrics. It also achieved effective preservation of spectral characteristics and spatial details in heterogeneous regions. Using adaptive parameter-based inference and a phased fusion strategy were, the framework effectively mitigates fusion errors induced by resolution differences, thereby providing more reliable spatiotemporal fusion data for dynamically monitoring complex topography.