Unsupervised Mutual Information and Superpixel Constraints in SAR Marine Aquaculture Extraction

The accurate extraction of floating raft aquaculture is significant to the scientific management and sustainable development of coastal zones. However, previous studies mainly focus on supervised detection technologies, which require a large number of high-quality training samples with labels. To solve this problem, an unsupervised mutual information differentiable feature clustering model with a superpixel algorithm (SMIDFM) is proposed to detect raft aquaculture using synthetic aperture radar (SAR) images. The core idea is to strengthen the aquaculture continuity based on the mutual information theory and the superpixel algorithm. Firstly, the network learns the global features to obtain the pseudo-labels. The pseudo-labels rely on the similarity of pixel values, and the number of classes tends to be unitary. Then, in order to make the pseudo-labels get the required number of classes, mutual information and superpixel algorithm are introduced into the global objective loss function. In addition, the superpixel spatial constraints are added to improve spatial consistency. The experiment demonstrated that the proposed model extracts effective features and overcomes the influence of speckle noise, and the proposed model can achieve effective extraction of SAR marine aquaculture information and does not require any labels.