From Macro to Micro: A Lightweight Interleaved Network for Remote Sensing Image Change Detection

Remote sensing image change detection (RSICD) is a crucial technology for Earth monitoring, but it faces two major challenges in practical applications. First, the complex scenes in remote sensing (RS) images make it difficult to accurately locate and distinguish the small change targets. Second, most RSICD methods usually stack complex modules to improve model performance, which inevitably leads to high computational costs. To address these issues, this article proposes a lightweight interleaved network from macro to micro for RSICD, named M2M-LINet. First, a three-stage RSICD framework inspired by holistic registration theory in biological vision is designed to deal with complex and small change targets. Specifically, the first stage is the coarse location stage to obtain a coarse localization on change targets, which is similar to the macroscopic observation of human eyes. The second stage is the fine detail focusing stage for capturing fine-grained information by designing a CNN-Transformer interleaved model, which is akin to the microscopic observation of the human eye. The third stage is the decoding prediction stage for predicting the final change features by mimicking the human interpretation of change information. Second, a lightweight interleaved structure is designed in the fine detail focusing stage to reduce the model size for facilitating deployment. It consists of two critical lightweight components. One is the lightweight convolutional block (LCB) that is devised to enhance high-frequency and low-frequency information for fine-grained change localization. The other is the lightweight Transformer block (LTB) that is designed as a linear self-attention mechanism to integrate multiscale information into feature maps effectively. The experiments on three RSICD datasets demonstrate that the proposed method performs excellently for complex change scenes and small change targets and significantly reduces the consumption of computing resources.