Master thesis and internship[BR]- Master's thesis : Ship detection from SAR imaging and machine learning techniques[BR]- Integration internship : Centre Spatial de Liège

Abstract

Human activities at sea have been not only impacting the maritime environment for years but also human development and society. Marine monitoring is one of the main thematic areas addressed by the Copernicus European Union’s Earth Obesrvation Programme. In particular, Sentinel-1 uses synthetic aperture radar (SAR) technology that provides global coverage and high-resolution images of the ocean surface, allowing monitoring of the position of ships at sea which may be useful to track illegal activities, such as fishing or oil spills, or for safety and security. Using machine learning algorithms allows to automatically analyze SAR images and supplies ship detection with better accuracy and efficiency. The present paper aims at evaluating the relevance of LS-SSDD-v1.0, i.e Large-Scale SAR Ship Detection Dataset-v1.0 Large, for training machine learning algorithms for ship detection from Sentinel-1 SAR images, and more particularly for small ship detection. To conduct this work effectively, the former dataset is used to train and evaluate RetinaNet and Faster R-CNN baseline models using common classification performance metrics. As a bonus, the Hungarian algorithm is implemented to solve the assignment problem between SAR ship predictions and AIS messages. The final selected model is applied on a SAR image of the Belgian coast and the detection is compared to the AIS messages of the region of interest at the same time of the Sentinel-1 acquisition. In conclusion, the comparison between both ship positions as well as the model performance results suggest that LS-SSDD-v1.0 is a relevant database for this master’s thesis subject but may still need to be augmented. A larger number of predicted ships than AIS messages also implies either illegal ships or a lack of received AIS signals.