A Dual-Perspective Spatiotemporal Fusion Model for Remote Sensing Images by Discriminative Learning of the Spatial and Temporal Mapping

Spatiotemporal fusion (STF) has received widespread attention as a cost-effective solution to the spatiotemporal conflicts in remote sensing images. Massive efforts have been made in the development of STF technology, and deep-learning methods have shown great potential in obtaining state-of-the-art results in recent years. However, it is still challenging to effectively fuse images with land-cover changes. The main problem is that the fine-resolution temporal changes are difficult to accurately model in the fusion stage due to the complex mapping relationships of the temporal features in different scale spaces. In this article, we propose a novel STF method with a dual-perspective framework, where the core idea is to predict the target information by discriminative learning of the spatial and temporal modeling for estimating heterogeneous temporal changes. Specifically, an encoder–decoder architecture based on a Swin transformer is designed to extract the global context information from the temporal change maps and predict the target image by learning the temporal mapping at a fine scale. A parallel subnetwork is further used to learn the spatial mapping across coarse-to-fine scales, considering the temporal changes. Channel-spatial attention is introduced to guide the model to focus on reconstructing the features delineating heterogeneous textures and temporal changes. The estimation from the dual perspectives is then fused to generate the final reconstructed image. Extensive experiments on three public datasets verified the superiority of the proposed method, compared with the mainstream STF algorithms, especially on image pairs with land-cover and phenological changes.