DIBNet: Discrepancy-Invariant Boundary-Guided Network for Remote Sensing Change Detection

Abstract

Remote sensing change detection (RSCD) aims to conduct difference analysis on RS images obtained in different time phases of the same area. It plays a critical role in applications, such as disaster monitoring and forest cover analysis, and has evolved rapidly in recent years. However, how to suppress false changes while enhancing the response to minor real changes and maintaining fine boundaries under the interference of complex backgrounds and imaging differences remains a key challenge in high-resolution RSCD. To address this, this article proposes the Discrepancy-Invariant Boundary-Guided Network (DIBNet). Specifically, to address pseudochanges caused by imaging condition differences, this article proposes a discrepancy-invariant recalibration module, which explicitly utilizes invariant features between different time phases to calibrate the difference features and enhance the ability to suppress false changes. Second, a cross-granularity boundary modeling module is proposed, which uses the details in shallow difference features to provide precise boundary positioning, and introduces semantic context in deep difference features to calibrate the boundary response at the regional level. Through the mutual guidance between semantic granularity and detail granularity, clear and semantically consistent boundary features are generated. Subsequently, a boundary-guided image fusion module is constructed, which injects boundary features into the multiscale difference feature fusion process to improve the response completeness and boundary precision of tiny change regions. Experimental results on multiple public datasets show that the performance of DIBNet is superior to existing mainstream methods.