Feature-based classification of optical water types in the Northwest Atlantic based on satellite ocean color data

[1] We have developed an optical water type classification approach based on remotely sensed water leaving radiance, for application to the study of spatial and temporal dynamics of ecologically and biogeochemically important properties of the upper ocean. For CZCS and SeaWiFS imagery of the Northwest Atlantic region, pixels from several different locations projected into distinct clusters in water-leaving radiance feature space, suggesting that these waters can be distinguished using a few spectral bands of ocean color data. Based on these clusters, we constructed a Northwest Atlantic Training Set and developed two different classification techniques. The Euclidean Distance Classifier minimizes the raw distance between each pixel and the centroid of the class to which it is assigned, whereas the Eigenvector Classifier is based on scaling the raw distances by the variance of each class, thereby accounting for the shape of each class in feature space. We conducted an initial evaluation of these two classification techniques by constructing water type classes based on only half of the pixels of each water type (randomly selected) in the Northwest Atlantic Training Set; classification was then carried out on the remaining half of the training set data. Applying the Euclidean Distance Classifier resulted in an average of 97.4% correctly classified pixels over 20 trials; even higher success rates were achieved with the Eigenvector Classifier, which gave an average of 99.1% correctly classified pixels. The Euclidean Distance Classifier performed well with spherical classes, but with more ellipsoidal classes, classification success improved considerably using the Eigenvector Classifier. We then applied these classifiers to ocean color images of the Northwest Atlantic to elucidate the geographical location and extent of each water type. We interpreted classifier results based on our Classification Goodness of Fit measure, which indicates how closely a given pixel is associated with its assigned class. This revealed that sharp boundaries exist between water masses of different optical types, with pixels on either side of the boundaries being strongly associated with their water type class. We anticipate that our classification techniques will facilitate long-term time series studies by tracking optical water types through seasonal and interannual changes.