Edge-aware transformer for coastal raft aquaculture extraction in optical remote sensing imagery

Abstract

To scientifically plan and accurately manage the coastal aquaculture industry, it is especially critical to quickly and accurately extract raft aquaculture areas. In the study, the Raft-Former was designed to accurately extract coastal raft aquaculture in Sansha Bay using Sentinel-2 remote sensing imagery. Specifically, a Feature Enhancement Module (FEM) was designed to selectively learn the interest features for solving the omission and mis-extraction caused by changes in the coastal environment. For the boundary adhesion problems caused by the dense distribution of raft aquaculture areas, a Feature Alignment Module (FAM) was developed to enhance edge-aware ability. A Global-Local Fusion Module (GLFM) was introduced to effectively integrate the local features with multi-scale and global features to overcome significant scale differences in aquaculture areas. Numerous experiments show that our method is better than the state-of-the-art models. Specifically, Raft-Former respectively achieves 90.05% and 86.73% mIoU on the Sansha Bay dataset.

Keywords

• Raft extraction
• edge-aware
• multi-scale
• Sentinel-2 remote sensing imagery