A wave-shaped CNN for marine-raft aquaculture-area extraction in SAR images

This paper introduces an innovative deep convolutional neural network, the wave-shaped CNN, tailored for the robust extraction of a marine-raft aquaculture area (MRAA) from synthetic-aperture-radar (SAR) images. Confronting the intricate challenges posed by coherent noise, environmental variability, and the distinction between rafting areas and their background within SAR images, the proposed wave-shaped CNN provides a novel solution for dynamic MRAA monitoring, which collaboratively incorporates the feature-attention subnetwork (FAS) and feature-refinement subnetwork (FRS) with residual connections to refine the feature-extraction process. Specifically, the FAS can adeptly extract both comprehensive global and nuanced local features from the multi-scale characteristics of SAR images. Furthermore, the FRS introduces a series of N-shaped subnetworks aimed at addressing the prevalent issue of edge adhesion observed within SAR images. In addition, a specialized SAR-MRAA dataset is developed, which allows for enriching the resource for the MRAA task in SAR images. Comprehensive experimental analyses conducted on the proposed wave-shaped CNN show its superior performance and effectiveness in MRAA extraction, demonstrating its advantages in establishing baselines within this domain. The code is available at https://github/gmy63000/Wave-shaped-CNN.