Ship and Oil Spill Detection using Convolutional Autoencoder

Student thesis: Master thesis (including HD thesis)

  • Mohit Singh Dhaka
Maritime surveillance has seen a growing interest in past few years owing to the fact that marine traffic has increased quite a lot, and securing a nation's maritime boundary is a crucial task. In this report we focus on two aspects of the maritime surveillance, namely ship detection and oil spill detection in SAR (Synthetic Aperture RADAR) satellite imagery. In this report the problem of ship and oil spill detection is formulated as that of an anomaly detection, where the ships and oil spills are the anomalies in the otherwise ocean background and use a deep Convolutional Autoencoder for the purpose of detecting these anomalies. Anomaly dataset was created from the xView3 competition dataset and SOS oil spill dataset for the training and testing of the autoencoder. The overall accuracy of the model is 96.64\%, and the precision, recall and F1 score being 93.32\%, 98.80\% and 97.96\%, respectively.
Publication date2 Jun 2022
ID: 472013745