SSMFusion: A Large Vision Mamba Model With Adaptive Feature Disentanglement for Remote Sensing Spatiotemporal Fusion

Abstract

Remote sensing spatiotemporal fusion aims to generate land surface observations with both high spatial and temporal resolution by integrating multisource satellite imagery, which is of great importance for applications, such as agricultural monitoring and environmental analysis. However, existing mainstream methods still face challenges in modeling long-range spatiotemporal dependencies and capturing high-frequency abrupt land-cover changes. Moreover, current studies lack explicit modeling of the relatively stable shared information and the dynamic differential information caused by land-cover changes during the temporal evolution of remote sensing imagery, which limits fusion accuracy. To tackle these challenges, this article introduces SSMFusion, a spatiotemporal feature fusion framework for remote sensing that explicitly models both “common” and “specific” characteristics via state-space equations. First, a hybrid hierarchical feature extraction module is designed. This module utilizes state-space equations to encode global-scale contextual correlations while maintaining linear computational efficiency. Meanwhile, spatial attention mechanisms and dynamic convolution are introduced to enhance local textures and edge details. Second, an adaptive spatiotemporal fusion module is developed to disentangle and effectively fuse stable shared features and dynamic change features under the guidance of spectral and temporal variation information, thereby improving spatiotemporal reconstruction in complex land-surface scenarios. Furthermore, to alleviate the lack of high-quality validation data in agricultural phenology applications, we construct the California Central Valley (CCV) dataset, which covers typical agricultural ecological zones in CCV, USA. Experimental results on two public datasets and the newly constructed CCV dataset demonstrate that the proposed method achieves superior quantitative performance over existing mainstream methods.