International Journal of Applied Earth Observation and Geoinformation
Oil spill detection using synthetic aperture radar images and feature selection in shape space
Introduction
An oil spill is the release of a liquid petroleum hydrocarbon into the environment, especially marine areas, due to human activity, and is a form of pollution. Oil spills are a major threat to marine ecosystems. Space-borne SAR can be used to gather information about maritime oil spills in any type of weather. Previous works have focused on the dark areas in SAR images, but oil spills are not the only things that can cause them. Darker areas may also be caused by other factors, such as currents, eddies, up- and down-welling, ice, wind fronts, sheltering provided by land, rain cells, and internal waves (Brekke and Solberg, 2005a, Brekke and Solberg, 2005b, Fiscella et al., 2000, Nirchio et al., 2005, Ferraro et al., 2007, Ferraro et al., 2009, Karathanassi et al., 2006). These objects (dark areas) resembling oil spills are called “lookalikes” (Brekke and Solberg, 2005a, Brekke and Solberg, 2005b, Topouzelis et al., 2008).
Oil films and lookalikes all have texture features, gray features, and frequency domain characteristics. Several researchers have used multi-features to detect oil spill from lookalikes. Fiscella et al. (2000) used 14 geometrical characteristics for both oil spill and natural features. Solberg and Theophilopoulos (1997) selected 11 features including surroundings and shape. Del Frate et al. (2000) considered 11 geometrical features in terms of its extension and its shape. A general description of the features calculated was published by Espedal and Johannessen (2000). They were the first to introduce texture features. Keramitsoglou et al. (2005) described 14 features. Karathanassi et al. (2006) described 13 features. Both these teams addressed physical, geometrical, and textural behavior. Several schemes have been designed for the purpose of unifying features with similar characteristics (Brekke and Solberg, 2005a, Brekke and Solberg, 2005b, Migliaccio and Trangaglia, 2004, Montali et al., 2006). Most of the published papers that describe attempts to detect oil spills discuss features such as shape, physical characteristics, and texture. However, none of the methods described in these studies addressed the issue of which features are useful and which are not.
Because oil films and lookalikes involve different materials, they have different physical properties and may take different shapes under certain wave and current conditions. Alessandro et al. (2007) simulate shift of oil-films shape. It indicates that shape features of oil films tend to be regular. In the present work, 9 space features were selected: marking ratio, solidity, rectangular saturation, circularity, narrowness, and edge density, interior angles based on bounding polygons (IABP), Hu moment invariance, and elliptical Fourier descriptors. These features were used to detect the oil spills from lookalikes. The results suggest that not all selected features have an effect on the classification process. The purpose of the present work was to identify the most important features (IEs) of the oil spills using differential evolution and a statistical repair mechanism (Rami et al., 2007, Rami et al., 2011). Then test samples were re-tested with IEs.
This paper is organized in the following manner. Section 2 introduces the method under discussion, and Section 3 shows the dataset. Results and the contributions to the detection of oil spill are discussed in Sections 4 Experimental results, 5 Evaluation of contributions, respectively. The paper is concluded in Section 6.
Section snippets
Methods
Generally, oil spill detection systems include two parts: image pre-processing and oil spill identification. Pre-processing is mainly responsible for filtering and segmenting SAR images. It provides samples (oil-spills and lookalikes) for identification (Yue and XiaoFeng, 2011). This paper mainly addresses identification (Fig. 1). First, the total eigenvectors from 9 shape features were computed. Next, a differential evolution feature selection (DEFS) algorithm was used in the total shape
Data
The present dataset comes from 20 full synthetic aperture radar (SAR) scenes of the South China Sea. The satellites include ers-1, ers-2, and Envisat. The SAR images are supplied as 16 bit system. The resolution is 20 m and the images are in GeoTIF Format. The raw satellite images are processed using semi-automated routines developed within ERMapper software. Fig. 2 shows all these 20 scenes in the ArcGIS bathymetry map. Every scene has a great deal of slick pollution, marked here by a colored
Experimental results
In order to find eigenvalues common to both oil films and lookalikes among a total of 95 eigenvalues, it was assumed that at least one eigenvalue belonged to a different subset. First, a DEFS algorithm was used to search for the similar eigenvalues with 1–94 eigenvalues. Removing these similar subsets, we get difference subsets from 94 to 1. This paper evaluates the accuracy of classification with difference subset in different numbers of iterations. Then the apparent frequencies of each
Evaluation of contributions
In order to evaluate the performance of the proposed detection scheme, different methods were tested, all involving 95 eigenvalues and 50 IEs in same data set. As illustrated in Table 5, the accuracy of the classification increased significantly, from 73.2% to 94.1%, and sensitivity increased from 62.6% to 79.3%. Accuracy is defined as the proportion of true results (both true positives and true negatives) in total population. Sensitivity measures the proportion of actual positives. It relates
Conclusions
The largest challenge in the detection of oil spills in SAR images is distinguishing oil spills from lookalike phenomena. However, most of the studies on this type of detection have only paid attention to the use of many different features and have only tested the accuracy of the classification. Few studies have analyzed the importance of individual features.
This study shows that shape features not only have a greater computational efficiency but also that they are easy to get. Most IEs are
Acknowledgments
We thank LetPub for its linguistic assistance during the preparation of this manuscript.
References (37)
- et al.
Oil spill detection by satellite remote sensing
Remote Sensing of Environment
(2005) - et al.
Towards an operational use of space imagery for oil pollution monitoring in the Mediterranean Basin: a demonstration in the Adriatic Sea
Marine Pollution Bulletin
(2007) - et al.
Accuracy of curve approximation by harmonically related vectors with elliptical loci
Computer Graphics and Image Processing
(1977) - et al.
Elliptic Fourier features of a closed contour
Computer Graphics and Image Processing
(1982) - et al.
Floating search methods in feature selection
Pattern Recognition Letters
(1994) - et al.
Feature subset selection and extraction using differential evolution and a statistical repair mechanism
Journal of Expert Systems with Applications
(2011) - et al.
Shape-based classification of spectrally identical objects
ISPRS Journal of Photogrammetry and Remote Sensing
(2008) Research on Visual Feature Analysis and Classification of Network Education Graphics Resources
(2009)- et al.
SAR simulation of ocean scenes covered by oil slicks with arbitrary shapes
IGARSS
(2007) - et al.
Feature extraction for oil spill detection based on SAR images
Lecture Notes in Computer Science
(2005)
Assessing the Accuracy of Remotely Sensed Data: Principles and Practices
Neural networks for oil spill detection using ERS–SAR data
IEEE Transactions on Geoscience and Remote Sensing
Detection of oil spills near offshore installations using synthetic aperture radar (SAR)
International Journal of Remote Sensing
Long term monitoring of oil spills in European seas
International Journal of Remote Sensing
Oil spill detection using marine SAR images
International Journal of Remote Sensing
Digital Image Processing Using MATLAB
A fast learning algorithm for deep belief nets
Neural Computation
Visual pattern recognition by moment invariants
IRE Transactions on Information Theory
Cited by (37)
Multi-source knowledge graph reasoning for ocean oil spill detection from satellite SAR images
2023, International Journal of Applied Earth Observation and GeoinformationMonitoring of oil spill in the offshore zone of the Nile Delta using Sentinel data
2022, Marine Pollution BulletinIntelligent computational techniques in marine oil spill management: A critical review
2021, Journal of Hazardous MaterialsCitation Excerpt :Four categories of features can be defined: geometrical features corresponding to the shape of oil slicks (e.g., area and perimeter of slicks); physical features defining wave backscattering in the physical space (e.g., mean and standard deviation of contrast ratio of slicks); contextual features linking objects with the surrounding area (e.g., the number of adjacent dark spots); and textural features presenting characteristics of the surface of the object (e.g., slicks continuity) (Mera et al., 2017; Topouzelis and Psyllos, 2012; Xu et al., 2014). Among them, shape features (a subset of geometrical category) have been introduced as one of the most practical features due to its extraction’s simplicity (Guo and Zhang, 2014). Accordingly, different indices can be defined based on extracted features to facilitate OSDCM.
Entropy-based matrix learning machine for imbalanced data sets
2017, Pattern Recognition Letters