Automatic Fusion of Satellite Imagery and AIS data for Vessel Detection

Aristides Milios, Konstantina Bereta, Konstantinos Chatzikokolakis, Dimitris Zissis, Stan Matwin

Abstract: Being able to fuse information coming from different sources, such as AIS and satellite images is of major importance for maritime domain awareness, for example, to locate and identify vessels that may have purposely turned off their AIS transponder, preventing illegal activities. This paper presents a fully-automatic method for fusing AIS data and SAR satellite images for vessel detection. The proposed framework is based on the automatic annotation of satellite images by correlating them with AIS data producing train and test datasets which are provided as input to a convolutional neural network (CNN). The CNN was trained to detect the presence of ships in sectors of the image. Our automatic process allows the neural network to learn on a large amount of data, without the need for hand-labelled datasets. Our neural network, trained on our automatically-generated test set of images, achieved an accuracy of 88% at ship detection, and an area under ROC curve of 94,6%. Our estimate of real world accuracy is about 86-90%.