Identification of bluefin tuna cages in Mediterranean Sea fishing grounds from SAR images

Abstract

Synthetic aperture radar (SAR) images are widely used for ship detection to support fisheries control and enforcement. In particular, for bluefin tuna (BFT), a large number of SAR images have been used over the Mediterranean Sea, where floating cages are towed by vessels to transport live tuna towards near-shore farms. The aim of this article is to present the methodology used to detect the cages automatically and separate them from the vessels. Experience and comparison with ground truth data have proved that tuna cage SAR signatures can be distinguished from vessel signatures by their distinctive texture pattern and position with respect to the towing vessel. Different candidate features were extracted from the images: texture features and also some estimates based on the k-distribution of pixel intensities. The so-called local binary pattern (LBP) texture features were also used. Feature selection was performed by using the leave-one-out (LOO) procedure with the k-nearest neighbour (k-NN) classifier and the classification and regression tree (CART) methodology. Five features yielded the best results: the estimate of the k-distribution () of pixels' intensities, the standard deviation of the amplitude SAR image pixels in the neighbourhood of the possible cage or vessel, and the three principal component (PC) features of the LBP values. The classification performance of the features was estimated with both k-NN and multilayer perceptron (MLP) classifiers. A k-NN-based version of the cage/vessel detection algorithm was used successfully during the last BFT campaign in 2008.