Ship and Oil Spill Detection using Convolutional Autoencoder
Author
Dhaka, Mohit Singh
Term
4. term
Publication year
2022
Submitted on
2022-06-02
Abstract
Maritim overvågning er blevet vigtigere i takt med stigende skibstrafik og behovet for at sikre maritime grænser. Denne afhandling fokuserer på to opgaver: at finde skibe og at finde oliespild i SAR-satellitbilleder (Synthetic Aperture Radar), en radarteknik der typisk kan fungere gennem skyer og i mørke. Vi behandler begge opgaver som anomali-detektion: skibe og olie er sjældne mønstre i et ellers ensartet hav. Metoden er en dyb konvolutionel autoencoder, en type neuralt netværk, der lærer, hvordan havet normalt ser ud ved at komprimere og gendanne billeder; afvigelser markeres som mulige anomalier. Vi sammensætter et anomali-datasæt ud fra xView3-konkurrencedatasættet og SOS oil spill-datasættet og bruger det til træning og test. I vores forsøg opnår modellen 96.64% nøjagtighed, med en præcision på 93.32%, recall på 98.80% og en F1-score på 97.96%.
Maritime surveillance is increasingly important as marine traffic grows and nations seek to secure their maritime boundaries. This thesis focuses on two tasks: detecting ships and detecting oil spills in SAR (Synthetic Aperture Radar) satellite imagery, a radar method that typically works through clouds and at night. We frame both tasks as anomaly detection: ships and oil are rare patterns against an otherwise uniform ocean background. Our method uses a deep convolutional autoencoder, a neural network that learns the sea’s normal appearance by compressing and reconstructing images; deviations are flagged as potential anomalies. We assemble an anomaly dataset from the xView3 competition dataset and the SOS oil spill dataset and use it for training and testing. In our experiments, the model achieves 96.64% accuracy, with 93.32% precision, 98.80% recall, and a 97.96% F1 score.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents