Unsupervised Segmentation of Cage Aquaculture in SAR Images Based on Invariant Information

Abstract

Cage aquaculture is one of the important types of marine aquaculture, and reasonable monitoring can achieve sustainable and stable development. Using the Synthetic Aperture Radar (SAR) to realize the extraction of cage aquaculture is significant. The convolutional neural networks (CNN) extract cage aquaculture by learning semantic information from deep features. However, training CNN usually needs a large number of labeled samples. Unsupervised learning is difficult to discover the semantic information of aquaculture due to the speckle noise in SAR images. In this article, an invariant information differentiable feature clustering network (IIDFCN) is proposed to enhance spatial continuity and reduce the influence of speckle noise. The pseudo-labels are obtained by a differentiable function processing the deep features of network output. The network parameters are updated by back-propagation, and the deep features and pseudo-labels are alternately and jointly optimized. In addition, in order to obtain reasonable spatial continuity constraints, an invariant information loss is introduced into the global loss function. The IIDFCN solves the problem of needing a large number of labels in the extraction of SAR aquaculture and implements the unsupervised deep network learning of cage aquaculture semantic information. The experiments test the method on a cage aquaculture data set from the Sanduao area, which shows the approach to be effective.