RSTNet: Rectified Spatial–Temporal Interaction Network for Remote Sensing Change Detection

Abstract

Currently, deep learning has become the mainstream technology for remote sensing change detection (RSCD) that aims to localize changes from bi-temporal observations. However, most existing prompt-guided methods generate prompts based on dual-phase pixel differences without fully perceiving the optimization goals of the subsequent detection task. In this letter, we propose a rectified spatial–temporal interaction network (RSTNet). Specifically, a forward mask temporal prompt generation block (FMB) generates forward prompts, while a reverse mask temporal rectification block (RMB) adaptively rectifies them by introducing explicit task-level supervision. Meanwhile, a high-level spatial–temporal interaction decoder (HLSTD) strengthens spatial–temporal interaction by fusing prompts and features via a high-level spatial–temporal guided fusion block (HLGB), and performs frequency-selective decoding via a dynamic selective frequency cross fusion block (DSFB). Extensive experiments on three public benchmarks demonstrate that RSTNet achieves superior performance over recent state-of-the-art methods by 2.83%, 0.84%, and 0.56% in $F1$ , with a favorable accuracy and efficiency trade-off. The code will be released at https://github.com/ZHANGJING7-BYTE/RSTNet