Extraction of Raft Aquaculture in SDGSAT-1 Images via Shape Prior Segmentation Network

Abstract

Reliable extraction of raft aquaculture areas from high-resolution remote sensing data is vital for the sustainable development of coastal zones. Despite the success of semantic segmentation, challenges remain due to adhesion effects, weak and seasonal spectral signals against complex dynamic backgrounds, and limited labeled training data for robust and generalizable models. To overcome these challenges, this article proposes a shape prior segmentation network for the extraction of raft aquaculture areas from Sustainable Development Science Satellite-1 (SDGSAT-1) images. Based on the encoder-decoder framework of a U-shaped network, the method incorporates a shape prior module that flexibly integrates with the backbone network. This module combines global shape priors, offering coarse shape representations to model global contexts, and local shape priors, providing fine shape information to enhance segmentation accuracy while reducing dependency on learnable prototypes. By leveraging shape priors, the network can achieve satisfactory segmentation reliability, efficiency, and faster learning during training. Extensive experiments validate the proposed methodology, achieving an accuracy of 98.26%, a mean pixel accuracy of 88.26%, and a mean intersection over union of 85.16% .