Research on maritime oil spill monitoring of multi-source remote sensing image based on deep semantic segmentation
Abstract
In recent years, with the increasing demand for crude oil in the world, the offshore oil transportation industry has developed rapidly. However, there are frequent oil spill accidents on the sea surface. For example, in 2010, the Gulf of Mexico rig are exploded, which led to the largest marine oil spill in history. The accident caused a large area of marine pollution. To sum up, after the oil spill accident, it is very important to detect the oil spill information accurately. Therefore, the development of oil spill monitoring is significant for marine environmental protection. The traditional method of oil spill monitoring is aerial or field investigation. But it needs a lot of manpower and material resources, which leads to high cost and difficult operation. Remote sensing satellites has the wide cover range. It can monitor the oil spill for day and night. And it becomes the best means of monitoring on the sea surface. At present, the Synthetic Aperture Radar (SAR) is usually used in the monitoring of oil spill based on remote sensing satellites. The advantages of SAR are wide coverage and all-weather operation. It can effectively monitor the position of oil spill. In addition, the research of oil spill monitoring in SAR image mainly focuses on the detection of oil spill. It spends less effort on calculation of oil pollution area. The optical image usually has high resolution and rich colors. The oil spill area can be effectively estimated, which is crucial for marine ecological damage assessment and oil spill control. For oil spill monitoring, it is mainly used image segmentation. The traditional methods of oil spill segmentation are: (1) Threshold segmentation method, it divides the image pixels into several classes. This method is simple and requires little computation. Nevertheless, it is easy to be affected by sea noise and uneven gray distribution of image. This will result in low segmentation accuracy. (2) Edge information detection, it combines the shape characteristics of the oil spill area and the edge information. The oil spill candidate area can be obtained. (3) Semantic segmentation, which clusters pixels belonging to the same category in an image into one region. The oil spill area and the sea surface can be clearly classified, and the semantic segmentation has a more detailed understanding of the image. The traditional classification methods used for semantic segmentation are: (1) Markov Random Fields, it is an undirected graph model, which defines the mark for each pixel. (2) Random Decision Forests, a classification method that uses multiple trees to train and predict samples. (3) Condition Random Field, it represents a Markov Random Field with a set of input random variables X and another set of output random variables Y. Among them, the Fully Connected Condition Random Field (fully connected CRF) overcomes the shortcomings of the traditional CRF that can miss fine structures. However, the classification effect of these traditional methods is still poor. Lately, deep learning has been widely used in the field of computer vision. It has achieved breakthrough success especially in image classification. Deeplab is proposed by the Google team for semantics segmentation. The network architecture is Resnet. It uses atrous convolution to adjust the resolution, this can expand the receptive field and reduce the amount of calculation. And it extracting features by deep convolutional neural networks(DCNN). However, there are still some problems in the Deeplab. For instance, first, it uses DCNN for rough segmentation. Then, the full connection condition random field is used for fine segmentation. The end-to-end connected cannot be realized, which will lead to low classification accuracy. Second, the fine details extraction of oil spill area is poor, and it is time-consuming. Aiming at the above problems and the characteristics of SAR and optical remote sensing images. We proposed a new semantic segmentation model based on Deeplab. It is a multi-source remote sensing image sea oil spill semantic segmentation model for monitoring oil spill area. Combining fully connected CRF with deep convolution neural network, it using Gaussian pairwise potential and mean field approximation theorem. The CRF is established as Recurrent Neural Networks (RNN). It is seen as the part of the neural network. These can obtain a deep end-to-end network with both DCNN and CRF. The model is used to monitor the oil spill area of SAR and optical sensing image. The oil spill area is estimated by optical sensing image.
- Publication:
-
43rd COSPAR Scientific Assembly. Held 28 January - 4 February
- Pub Date:
- January 2021
- Bibcode:
- 2021cosp...43E.105C