Learn Geospatial Correlations for Remote Sensing Image Change Detection

Change detection plays a critical role in Earth’s surface surveillance via remote sensing images which usually have strong geospatial correlations in surface entities including specific global layout of different scenes and contextual dependencies of local regions or objects. To capture such geospatial correlations and facilitate recognition of changed regions, based on attention mechanism, we propose a geospatial correlation feature enhancement network (GCFENet) for remote sensing image change detection. In the network, to capture global scene context and change details, a dual feature interaction (DFI) module is firstly designed. Then, to model and fuse different correlations for each scale, a three branch multi-correlation attention (MCA) block is proposed which models local distributions, global scene dependencies and pixel-wise semantic prior respectively, and associates them via a self-attention module to capture the semantic distributions within the scene and between different regions. In the MCA network, a Gaussian filter based frequency separation (GFS) module is designed to generate the global scene dependencies and local distributions, which are used as the value and the query in self-attention respectively. Then a local semantic modeling (LSM) module is proposed to assign pixel-wise semantic prior, which is used as the key to generate correlations with the query and the value. This GCFENet enables the attention mechanism to compute similarity not only based on pixel-wise feature matching but also by perceiving the geospatial distribution correlations of ground objects and bi-temporal spatial correspondences. Experiments on three public datasets demonstrate the state-of-the- art performance of the GCFENet. The code will be opensourced in [https://github.com/4043-lab/Change-detection] once accepted.