Loading…
Oil spill segmentation in SAR images using convolutional neural networks. A comparative analysis with clustering and logistic regression algorithms
Synthetic aperture radar (SAR) images are a valuable source of information for the detection of marine oil spills. For their effective analysis, it is important to have segmentation algorithms that can delimit possible oil spill areas. This article addresses the application of clustering, logistic r...
Saved in:
Published in: | Applied soft computing 2019-11, Vol.84, p.105716, Article 105716 |
---|---|
Main Authors: | , , , |
Format: | Article |
Language: | English |
Subjects: | |
Citations: | Items that this one cites Items that cite this one |
Online Access: | Get full text |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Synthetic aperture radar (SAR) images are a valuable source of information for the detection of marine oil spills. For their effective analysis, it is important to have segmentation algorithms that can delimit possible oil spill areas. This article addresses the application of clustering, logistic regression and convolutional neural network algorithms for the detection of oil spills in Envisat and Sentinel-1 satellite images. Large oil spills do not occur frequently so that the identification of a pixel as oil is relatively uncommon. Metrics based on Precision–Recall curves have been employed because they are useful for problems with an imbalance in the number of samples from the classes. Although logistic regression and clustering algorithms can be considered useful for oil spill segmentation, the combination of convolutional techniques and neural networks achieves the best results with low computing time. A convolutional neural network has been integrated into a decision support system in order to facilitate decision-making and data analysis of possible oil spill events.
[Display omitted]
•The images studied show oil spills of diverse shapes and sizes.•To evaluate the results diverse empirical discrepancy methods have been used.•A fully convolutional neural networks achieves the best results.•The output can be directly interpreted as a classification for each pixel. |
---|---|
ISSN: | 1568-4946 1872-9681 |
DOI: | 10.1016/j.asoc.2019.105716 |