International Journal of Applied Earth Observations and Geoinformation (Jul 2024)
CIBENet: A channel interaction and bridging-enhanced change detection network for optical and SAR remote sensing images
Abstract
Effectively utilizing the complementary characteristics between optical and SAR remote sensing images to accurately identify change information is of great practical significance. Direct pixel comparisons are challenging since they come from sensors with different imaging mechanisms. Therefore, in this paper, a novel domain transformation that incorporates channel interaction and a bridging-enhanced (CIBENet) heterogeneous change detection network is proposed, where twin-depthwise separable fused gated channel transformation is the convolutional block designed for channel information interaction. And samples without changed semantic information are used to train the deep transformation model, which effectively solves the interference of semantic information on NiceGAN. A different perspective is provided for the heterogeneous change detection task. The backend change detection network takes UNet++ with twin-depthwise separable convolution as the baseline, introduces the gated channel transformation and bridging-enhanced decoder, and models the feature relationship between channels to strengthen the channel information interaction while suppressing the expression of nonchanged information. In addition, the bridging-enhanced decoder can efficiently solve localized holes and discontinuities in binary maps by bridging identical pixels. CIBENet is supervised and experimented on three heterogeneous change detection datasets, Gloucester, Shuguang, and Italy, also compared with the classical unsupervised methods CGAN, SCCN, INLPG and advanced supervised methods DTCDN, DACDT. The network model proposed in this paper significantly improved F1 and recall, and the overall accuracies were 96.60 %, 98.23 %, and 96.29 %, respectively. The experiments validated the reliability of the proposed model.