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

Abstract

[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.

Optical characterization of water masses within the Columbia River plume

“…Nonetheless, n‐dimensional analysis on 250 m MODIS data products is possible for the Columbia River plume and smaller plume systems of the U.S. west coast with the understanding that imagery may be scant due to atmospheric effects, and may thus be biased to those periods when clear satellite imagery is available. Water mass classification techniques have been applied to describe patterns in space [ Martin Traykovski and Sosik , 2003; Moore et al , 2001], and have been used to describe the migration of salinity fronts in time [ Oliver et al , 2004]. Daily satellite overpasses make this a possibility in many regions of the world, especially locations where cloud cover is less persistent, and could be applied to time‐averaged imagery for regions that evolve over longer time periods.…”

“…Classification of water types using ocean color imagery has far‐reaching applications for ecological and biogeochemical modeling [ Schofield et al , 2004]. Statistical, feature‐based classification techniques have been successful in characterizing river plumes [ Oliver et al , 2004; Thomas and Weatherbee , 2006] and phytoplankton blooms of the North Atlantic [ Martin Traykovski and Sosik , 2003; Moore et al , 2001]. Feature‐based classification techniques use the inherent characteristics of the input data, for example shipboard observations of ocean color or environmental data, to derive clusters within a multivariate statistical space.…”

An Inverse Model for Estimating the Optical Absorption and Backscattering Coefficients of Seawater From Remote‐Sensing Reflectance Over a Broad Range of Oceanic and Coastal Marine Environments

“…Note that the sum of a ph (k), a NAP (k), and a CDOM (k) can be referred to as the nonwater absorption coefficient, a nw (k), which can also be derived as a difference between a(k) and pure seawater absorption coefficient, a w (k). Satellite-derived IOP data products have been used in many studies addressing various characteristics of seawater constituents of biogeochemical significance (e.g., Balch et al, 2005;Behrenfeld et al, 2013;Kostadinov et al, 2009Kostadinov et al, , 2010Loisel et al, 2001aLoisel et al, , 2002Mannino et al, 2008;Siegel et al, 2002;Stramski et al, 1999;Vantrepotte et al, 2011), oceanic primary production (e.g., Behrenfeld et al, 2005;Lee et al, 2011), and characterization of water masses and other physical processes (e.g., Arnone et al, 2004;Traykovski & Sosik, 2003;Yang et al, 2015).…”

Spectra of a shallow sea—unmixing for class identification and monitoring of coastal waters

“…Another problem is that all endmembers should be available in the database. The study of Martin Traykovski and Sosik (2003) used expert knowledge to define their classes, although these were not used for endmember modelling. & Pure pixel method.…”