Shadow Detection in Remote Sensing Images Based on Multibranch Feature Aggregation and Channel-Spatial Attention

Abstract

High-resolution remote sensing image shadow detection has wide applications in target recognition, land information retrieval, and other fields. However, current shadow detection technologies still face challenges, including shadow omission and difficulty in defining boundaries. To address these challenges, this article proposes a shadow detection method based on a multibranch channel-spatial attention network, which combines the multibranch feature aggregation module (MFAM) and the channel-spatial parallel attention feature fusion module (C-SPAFFM). The MFAM effectively integrates shadow information at different scales, reducing missed detections caused by changes in shadow size and shape. The C-SPAFFM enhances channel information to highlight boundary features and optimizes spatial information to more accurately capture spatial variations in shadows, thereby further reducing the possibility of missed detections. The effectiveness of the proposed method was validated on the public dataset AISD and the self-constructed satellite image dataset SISD. On the AISD dataset, the F1-score, OA, IOU, and BER metrics were 93.76%, 97.36%, 88.33%, and 4.19%, respectively. On the SISD dataset, these metrics reached 91.37%, 94.91%, 84.25%, and 6.87%. Experimental results show that the proposed method performs well in shadow detection tasks for high-resolution remote sensing images.