UniRSCD: A Unified Novel Architectural Paradigm for Remote Sensing Change Detection

Abstract

The Identification of changes within two different images from satellites or other sources of remote sensing is a basic issue in earth observation that seeks to recognize any semantically significant distinctions between two images. Methods using supervised learning approaches provide good performance on RSCD tasks, however, they suffer from generalization issues. In this paper, we present UniRSCD, a unified architectural paradigm for semantic change detection that integrates class-prior color statistics, dual-branch feature fusion, and lightweight decoder design into a single end-to-end trainable framework.
UniRSCD leverages an encoder network that is built on ResNet34 and executes simultaneously over the image pairs while using a special color signal stream to combine several difference features at different scales, along with an attention mechanism based on RGB statistics prototypes. Our extensive evaluations on the benchmark dataset SECOND show that the average class IoU of our approach reaches 38.7% and its F1 score is 53.5%, outperforming the previous SOTA open-vocabulary change detection method OmniOVCD by +11.6 IoU and +11.7 F1. In particular, the IoU scores of buildings and playgrounds are 61.3% (+16.1) and 64.4% (+37.4) compared to OmniOVCD, respectively, showing the effectiveness of prior-aware class disambiguation in spectral land-cover categories.