New Fusion Network with Dual-Branch Encoder and Triple-Branch Decoder for Remote Sensing Image Change Detection

Deep learning plays a highly essential role in the domain of remote sensing change detection (CD) due to its high efficiency. From some existing methods, we can observe that fusion information at each scale is quite vital for accuracy CD results, especially common problems pseudo-change and difficult edges task. With this mind, propose New Fusion network with Dual-branch Encoder Triple-branch Decoder (DETDNet) follows codec structure as whole, where encoder adopts siamese Res2Net-50 extract local features bitemporal images. As decoder previous works, they usually employed single branch, approach only preserved encoder’s Distinguished from these approaches, adopt triple-branch architecture first time. The preserves not dual-branch left right branches, respectively, learn effective powerful individual temporal image but also middle branch. branch utilizes aggregation (TA) realize feature interaction three branches decoder, which enhances integrated provides abundant supplementary improve performance. ensures respective images well their fused are preserved, making extraction more integrated. In addition, employ multiscale module (MFE) per layer contextual enhance representation capability CD. We conducted comparison experiments on BCDD, LEVIR-CD, SYSU-CD datasets, were created Zealand, USA, Hong Kong, respectively. data preprocessed contain 7434, 10,192, 20,000 pairs, experimental results show DETDNet achieves F1 scores 92.7%, 90.99%, 81.13%, shows better compared recent means model robust. lower FP FN indicate error misdetection rates. Moreover, analysis problem pseudo-changes difficulty detecting small areas solved.