Automatic neural classification of ocean colour reflectance spectra at the top of the atmosphere with introduction of expert knowledge

We propose an automatic neural classification method for ocean colour (OC) reflectance measurements taken at the top of the atmosphere (TOA) by satellite-borne sensors. The goal is to identify aerosol types and cloud contaminated pixels. This information is of importance when selecting appropriate atmospheric correction algorithms for retrieving ocean parameters such as phytoplankton concentrations. The methodology is based on the use of Topological Neural network Algorithms (TNA, so-called Kohonen maps). The pixels of the remotely sensed image are characterised by a vector whose components are the spectral TOA measurement and the standard deviation of a small spatial structure. The method is a three-step method. The first step is an unsupervised classification built from a learning data set; it clusters pixel vectors which are similar into a certain number of groups. Each group is characterised by a specific vector, the so-called reference vector (rv), which summarises the information contained in all the pixels belonging to that group. The second step of the method consists of labeling the reference vectors with the help of an expert in ocean optics. The groups are then clustered into classes corresponding to physical characteristics provided by the expert. The third step consists of analyzing full images and classifying them by using the classifier which has been determined during the first two steps. The method was applied to the Cape Verde region, which exhibits important seasonal variability in terms of aerosols, cloud coverage and ocean chlorophyll-a concentration. We processed POLDER data to test the algorithm. We considered four classes: pixels contaminated by clouds; two types of pixels containing mineral dusts; and pixels containing maritime aerosols only. The method was able to take into account the information given by the expert and apply it to unlabeled pixels. This methodology could easily be extended to a larger number of classes, the major problem being to find adequate expertise to label the classes.