A comparative study of different classification techniques for marine oil spill identification using RADARSAT-1 imagery

The discrimination of oil spills and look-alike phenomena (e.g., low wind area, wind front area and natural slicks) on Synthetic Aperture Radar (SAR) images is a crucial task in marine oil spill detection. Many classification techniques can be employed for this purpose. In order to make the best use of the large variety of statistical and machine learning classification methods, it is necessary to assess their performance differences and make recommendations for classifier selection and improvement. The objective of this paper is to compare different classification techniques for oil-spill detection in RADARSAT-1 imagery. The data of this study consists of 15 features of 192 oil spills and look-alikes identified by Canadian Ice Service between 2004 and 2008 off Canada's east and west coastal areas. The studied classifiers include the Support Vector Machine (SVM), Artificial Neural Network (ANN), tree-based ensemble classifiers (bagging, bundling and boosting), Generalized Additive Model (GAM) and Penalized Linear Discriminant Analysis (PLDA). Two performance measures, the specificity at fixed sensitivity (80%) and the area under the Receiver Operating Characteristic (ROC) curve (AUC), were estimated using cross-validation to evaluate the performance of classifiers at a high sensitivity. Overall, the bundling technique which achieved a median specificity of 90.7%, at sensitivity of 80%, significantly outperformed the second best (i.e. bagging) by 1.5 percentage points, and the worst (i.e. ANN) by 15 percentage points. The median values of AUC measure indicated consistent results. Bundling and bagging achieved comparable median AUC values of about 92%, followed by GAM and PLDA, with ANN yielding the smallest. Most classifiers (SVM, bundling and especially PLDA and ANN) performed significantly better on datasets pre-processed by log-transformation and standardization than on the original dataset. These results demonstrate the importance and benefit of selecting the optimal classifiers for oil spill classification, and configuring the classifiers by proper feature construction techniques.