A Dual-Branch Temporal Deep-Supervised Framework for Remote Sensing Change Detection

In remote sensing applications, change detection serves as a pivotal technology to monitor changes on the land surface by identifying and extracting temporal discrepancies in bitemporal images captured over the same geographical area. Although existing deep learning methods can extract temporal difference features from bitemporal images, these methods often suffer from the attenuation of feature pair distances during long-range encoding and decoding, leading to weakened deep supervision signals. To address this core challenge, this study proposes a pioneering dual-branch temporal deep supervision network, termed DTDNet. The framework employs two independent branches—a front-end semantic branch and a back-end feature mining branch—to achieve multi-dimensional coupling of temporal change features. The former focuses on precise extraction of temporal difference semantics, while the latter is dedicated to effective mining of deep features from bitemporal images. Furthermore, a feature-enhancing Attention Module (FEAM) is innovatively introduced to capture long-range dependencies among token features and learn distance mappings between bitemporal feature pairs, significantly enhancing feature discriminability. Extensive experiments on the MACD, CDD, and SYSU-CD datasets demonstrate that DTDNet achieves breakthrough performance across multiple metrics, substantially advancing the state of the art in change detection technology.