X-Change: An RGB-Inspired Spectral-Aware Framework for Multitask and Explainable Change Captioning

Abstract

Remote sensing change captioning describes land surface changes between bitemporal images. However, models trained on RGB inputs often underperform compared to multispectral (MS) counterparts due to limited spectral awareness. We present X-Change, an RGB-inspired spectral-aware framework that achieves MS-level descriptive quality while jointly performing multitask segmentation to predict change, NDVI, and NDWI masks-indicating where and how change occurs. Unlike prior RGB-based methods, X-Change employs rule-based spectral supervision from bitemporal Sentinel-2 data, enabling its shared encoder to internalize NDVI/NDWI-related cues for both captioning and segmentation tasks. Experiments on the MOSAIC-SEN2-CC dataset show that X-Change surpasses state-of-the-art RGB-based models and matches or slightly exceeds those trained on MS inputs, producing spatially consistent change and index maps. Overall, X-Change bridges the gap between RGB and MS modalities, offering an interpretable and practical framework for spectral-aware, multitask, and explainable change captioning.