Identifying Oil Spills

Identifying Oil Spills using Self Organizing Maps and multi-spectral signatures. In our prototype studies we have found Neural Network (NN) Self Organizing Maps (SOM) to do a remarkable job of not just classifying an oil spill but also quantifying the severity. Neural networks are biologically inspired machine learning algorithms that have been trained to perform complex functions in various fields, including pattern recognition, identification, classification, speech, vision, and control systems.  Today neural networks can be trained to solve problems that are difficult for conventional computers or human beings. Neural networks are non-linear, multivariate, non-parametric learning algorithms inspired by biological nervous systems and are composed of simple elements operating in parallel (Bishop 1995; 1998; Haykin 1999; 2001). As in nature, the network function is determined largely by the connections between elements.

A self-organizing map (SOM) is a type of NN that is trained using unsupervised learning to produce a low dimensional discretized representation of the input space of the training samples, called a map. The map seeks to preserve the topological properties of the input space. This makes SOM useful for visualizing low-dimensional views of high-dimensional data. The model was first described as an artificial neural network by the Finnish professor Teuvo Kohonen (Kohonen 1984; 2001), and is sometimes called a Kohonen map. Like most artificial neural networks, SOMs operate in two modes: training and mapping. Training builds the map using input examples, in our case a suite of satellite products that will be used to characterize the ocean surface. Once the network is trained, mapping automatically classifies a new input vector. SOMs have already been used to great effect for a large variety of earth and space science applications. This includes the analyses of remote sensing spectral images, for example in the classification of different geological regions (Merenyi et al. 1990; Merenyi et al. 1996a; Merenyi et al. 1996b; Merenyi et al. 1997; Seiffert and Jain 2001; Villmann et al. 2003; Merenyi et al. 2007). Given this success by others, and our own recent success in using them for both the accurate delineation of dust sources for the Navy and the classification of ecosystems in the Gulf of Mexico (in press) it is timely to apply SOM for our need to delineate oil spill locations. The inputs to the SOM are a suite of satellite products from a suite of instruments including synthetic aperture radar (SAR) and the visible and near infrared water leaving radiances and absorbances, and their derived products. As an example, the figure above in panel (a) shows a picture of oil slick as seen from space by NASA’s Terra satellite on May 24, 2010. Panel (b) shows our prototype SOM classification for the same day. It can be seen that the yellow and red classes correspond to the oil slick, with the severest regions highlighted in red.