Automatic Recognition of Oil Spills Using Neural Networks and Classic Image Processing
Oil spill events are one of the major risks to marine and coastal ecosystems and, therefore, early detection is crucial for minimizing environmental contamination. Oil spill events have a unique appearance in satellite images created by Synthetic Aperture Radar (SAR) technology, because they are byp...
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2022-04-01
|
| Series: | Water |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2073-4441/14/7/1127 |
| _version_ | 1797437447630487552 |
|---|---|
| author | Rotem Rousso Neta Katz Gull Sharon Yehuda Glizerin Eitan Kosman Assaf Shuster |
| author_facet | Rotem Rousso Neta Katz Gull Sharon Yehuda Glizerin Eitan Kosman Assaf Shuster |
| author_sort | Rotem Rousso |
| collection | DOAJ |
| description | Oil spill events are one of the major risks to marine and coastal ecosystems and, therefore, early detection is crucial for minimizing environmental contamination. Oil spill events have a unique appearance in satellite images created by Synthetic Aperture Radar (SAR) technology, because they are byproducts of the oil’s influence on the surface capillary, causing short gravity waves that change the radar’s backscatter intensity and result in unique dark formations in the SAR images. This signature’s appearance can be utilized to monitor and automatically detect oil spills in SAR images. Although SAR sensors capture these dark formations, which are likely connected to oil spills, it is hard to distinguish them from ships, ocean, land, and other oil-like formations. Most of the approaches for automatic detection and classification of oil spill events employ semantic segmentation with convolutional neural networks (CNNs), using a custom-made dataset. However, these approaches struggle to distinguish between oil spills and spots that resemble them. Therefore, developing a tailor-made sequence of methods for the oil spill recognition challenge is an essential need, and should include examination and choice of the most effective preprocessing tools, CNN models, and datasets that are specifically effective for the oil spill detection challenge. This paper suggests a new sequence of methods for accurate oil spill detection. First, a SAR image filtering technique was used for emphasizing the unique physical characteristics and appearance of oil spills. Each filter’s impact on leading CNN architectures performances was examined. Then, a method of a model ensemble was used, aiming to reduce the generalization error. All experiments demonstrated in this paper confirm that using the sequence suggested, in comparison to the common formula, leads to a 4.2% of improvement in the intersection over union score (IoU) for oil spill detection, and a 9.3% of improvement in the mean IoU among several relevant classes. |
| first_indexed | 2024-03-09T11:20:26Z |
| format | Article |
| id | doaj.art-a74e61efb25e4580b9bbb26a9976949c |
| institution | Directory Open Access Journal |
| issn | 2073-4441 |
| language | English |
| last_indexed | 2024-03-09T11:20:26Z |
| publishDate | 2022-04-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Water |
| spelling | doaj.art-a74e61efb25e4580b9bbb26a9976949c2023-12-01T00:20:34ZengMDPI AGWater2073-44412022-04-01147112710.3390/w14071127Automatic Recognition of Oil Spills Using Neural Networks and Classic Image ProcessingRotem Rousso0Neta Katz1Gull Sharon2Yehuda Glizerin3Eitan Kosman4Assaf Shuster5Electrical and Computers Engineering Department, Technion—Israel Institute of Technology, Haifa 32000, IsraelElectrical and Computers Engineering Department, Technion—Israel Institute of Technology, Haifa 32000, IsraelComputer Science Department, Technion—Israel Institute of Technology, Haifa 32000, IsraelComputer Science Department, Technion—Israel Institute of Technology, Haifa 32000, IsraelComputer Science Department, Technion—Israel Institute of Technology, Haifa 32000, IsraelComputer Science Department, Technion—Israel Institute of Technology, Haifa 32000, IsraelOil spill events are one of the major risks to marine and coastal ecosystems and, therefore, early detection is crucial for minimizing environmental contamination. Oil spill events have a unique appearance in satellite images created by Synthetic Aperture Radar (SAR) technology, because they are byproducts of the oil’s influence on the surface capillary, causing short gravity waves that change the radar’s backscatter intensity and result in unique dark formations in the SAR images. This signature’s appearance can be utilized to monitor and automatically detect oil spills in SAR images. Although SAR sensors capture these dark formations, which are likely connected to oil spills, it is hard to distinguish them from ships, ocean, land, and other oil-like formations. Most of the approaches for automatic detection and classification of oil spill events employ semantic segmentation with convolutional neural networks (CNNs), using a custom-made dataset. However, these approaches struggle to distinguish between oil spills and spots that resemble them. Therefore, developing a tailor-made sequence of methods for the oil spill recognition challenge is an essential need, and should include examination and choice of the most effective preprocessing tools, CNN models, and datasets that are specifically effective for the oil spill detection challenge. This paper suggests a new sequence of methods for accurate oil spill detection. First, a SAR image filtering technique was used for emphasizing the unique physical characteristics and appearance of oil spills. Each filter’s impact on leading CNN architectures performances was examined. Then, a method of a model ensemble was used, aiming to reduce the generalization error. All experiments demonstrated in this paper confirm that using the sequence suggested, in comparison to the common formula, leads to a 4.2% of improvement in the intersection over union score (IoU) for oil spill detection, and a 9.3% of improvement in the mean IoU among several relevant classes.https://www.mdpi.com/2073-4441/14/7/1127convolutional neural networkimage filteringpre-processingoil spill detectionSAR imagerysemantic image segmentation |
| spellingShingle | Rotem Rousso Neta Katz Gull Sharon Yehuda Glizerin Eitan Kosman Assaf Shuster Automatic Recognition of Oil Spills Using Neural Networks and Classic Image Processing Water convolutional neural network image filtering pre-processing oil spill detection SAR imagery semantic image segmentation |
| title | Automatic Recognition of Oil Spills Using Neural Networks and Classic Image Processing |
| title_full | Automatic Recognition of Oil Spills Using Neural Networks and Classic Image Processing |
| title_fullStr | Automatic Recognition of Oil Spills Using Neural Networks and Classic Image Processing |
| title_full_unstemmed | Automatic Recognition of Oil Spills Using Neural Networks and Classic Image Processing |
| title_short | Automatic Recognition of Oil Spills Using Neural Networks and Classic Image Processing |
| title_sort | automatic recognition of oil spills using neural networks and classic image processing |
| topic | convolutional neural network image filtering pre-processing oil spill detection SAR imagery semantic image segmentation |
| url | https://www.mdpi.com/2073-4441/14/7/1127 |
| work_keys_str_mv | AT rotemrousso automaticrecognitionofoilspillsusingneuralnetworksandclassicimageprocessing AT netakatz automaticrecognitionofoilspillsusingneuralnetworksandclassicimageprocessing AT gullsharon automaticrecognitionofoilspillsusingneuralnetworksandclassicimageprocessing AT yehudaglizerin automaticrecognitionofoilspillsusingneuralnetworksandclassicimageprocessing AT eitankosman automaticrecognitionofoilspillsusingneuralnetworksandclassicimageprocessing AT assafshuster automaticrecognitionofoilspillsusingneuralnetworksandclassicimageprocessing |