Spatial-Semantic Alignment and Change-Aware Network for Remote Sensing Image Change Captioning

Remote sensing image change captioning aims to generate natural language descriptions of dynamic changes between bitemporal images. This task faces challenges of aligning high-level semantics with sparse change regions and modeling long-range dependencies in diverse change patterns. To alleviate these limitations, we propose the Spatial-Semantic Alignment and Change-aware Network, dubbed as SACNet. It introduces a semantic-spatial collaborative encoding module synergizing a CNN-based spatial encoder with a CLIP-based semantic encoder, using precise spatial cues to refine global semantic features. A change-driven adaptive feature calibration module emphasizes discriminative change signals via inter-temporal feature dissimilarity modeling, suppressing background noise and highlighting key changes. Additionally, a multi-path SSM-guided spatial representation module is designed to capture diverse complex spatial change topologies. Extensive experiments on the LEVIR-CC and WHU-CDC datasets demonstrate that SACNet achieves state-of-the-art performance, generating both high-precision textual descriptions and approximate change masks. Code will be available at https://github.com/CVer-Yang/SACNet.