HCA-Net: An Instance Segmentation Network for High-Consequence Areas Identification From Remote Sensing Images

Abstract

The high-consequence area (HCA) is crucial for the safety management and operation of oil and gas pipelines. However, traditional models that rely on manual field investigations are costly, inefficient, and risky. Deep learning (DL)-based instance segmentation (IS) has the potential to enable automatic HCA identification. Unfortunately, the existing studies lack methods specifically designed to identify HCAs from remote sensing (RS) images. This letter proposes an IS network (HCA-Net) with spatial relation enhancement and mask decoupling refinement for HCA recognition is proposed. The proposed method first develops a spatial relation enhancement module (SREM) that queries the similarity of features at different spatial locations to represent spatial relations, further enhancing these features to promote completeness. Moreover, a unique decoupled mask refinement head (DMRH) is designed to refine the mask by decoupling boundary features from body features and optimally integrating them into the final features. Experiments on the constructed gas pipeline aerial dataset (GPAD) show that our method outperforms eight state-of-the-art (SOTA) methods. Compared to the baseline model mask R-CNN, HCA-Net improves the mAP of masks and the mIoU of HCA by 3.9% and 6.9%, respectively.