AMG-Net: A multitask network with adaptive mutual guidance for Semantic Change Detection

Semantic Change Detection (SCD), a pivotal task in Earth observation, aims to identify not only the locations of change but also the specific “from-to” semantic transitions between bi-temporal remote sensing images. The prevailing multitask learning paradigm, featuring a shared encoder and triple task-specific decoders, has emerged as the de facto standard. However, this architecture suffers from two key limitations. First, inter-task collaboration is typically unidirectional, which prevents mutual reinforcement between the semantic segmentation and change detection tasks. Second, this collaboration relies on generic features from a standard backbone, failing to provide the distinct, high-quality representations required by each sub-task. To address these limitations, we propose the Adaptive Mutual Guidance Network (AMG-Net). The AMG-Net features two main contributions. First, we introduce the Cross-Guidance Decoder (CG-Decoder), an architecture that establishes a mutual guidance mechanism to enable reciprocal refinement between the sub-tasks. Second, we develop two specialized modules, the Hierarchical Semantic Encoder (HSE) and the Preliminary Difference Module (PDM), to provide high-quality, task-specific features for the guidance process. Extensive experiments on the SECOND and Landsat-SCD benchmarks demonstrate that our proposed AMG-Net significantly outperforms existing methods.