Abstract
This work presents a hierarchical semiautomated supervised machine learning oil spill classification algorithm for X-, C-, and L-band synthetic aperture radar (SAR) imagery using data acquired during the Deepwater Horizon (DWH) oil spill event. The approach consists of three stages: (a) a neural network pixel-by-pixel detector that creates candidate shapes; (b) a feature extraction stage where the detected shapes are analyzed and features most likely to separate spills from lookalikes are chosen; and (c) a 2nd neural network is used to classify the shapes into categories (namely oil and non-oil). The algorithm results were assessed using confusion matrix calculations (true positive, false positive, true negative, false negative) and accuracies were computed to evaluate algorithm performance. The influence of factors affecting performance of the hierarchical algorithm such as image quality issues and meteorological/oceanographic phenomena is discussed. Additionally, two analyses of the algorithm results were accomplished: 1) A recounting of the types of most common types of lookalikes found; and 2) a qualitative analysis and summary of the classification results for the periods of April-May, June, and July-August (complemented with ancillary data from the Global Forecast System). It was found that the algorithm produces efficient isolation of spills in SAR imagery in most cases, with the similarities of different types of lookalikes to oil spills and the presence of image pixel errors also influencing algorithm performance. Low wind areas, rain cells, natural seeps, and upwelling regions were the most misclassified areas across frequency band and imaging mode. The classification results showed that there was rapid expansion and dispersal of the oil spill above the wellhead, into the Mississippi River Delta and proximate coastal regions. These analyses illustrate some of the successes and limitations of the developed oil spill hierarchical classifier.