Automatic recognition of coastal and oceanic environmental events with orbital radars

An automatic classification procedure was developed able to identify different oceanic events, detectable in orbital radar images. The procedure was customized to be used in the southeastern Brazilian coast, since the classification training and test used examples extracted from 402 RADARSAT-1 images acquired in this region. Different sets of spectral, geometric and contextual (meteo-oceanographic and location) features of selected low backscatter patches were evaluated. Machine learning procedures (neural networks, decision trees and support vector machines) were used to induce classifiers to differentiate between seven classes, belonging to two categories. The classification procedure involves two steps: first the features area classified in one of two categories - oil spill or meteooceanographic phenomena. In the second step, the identification of tree classes of oil spills and four classes of meteooceanographic phenomena is done. The oil spill related classes are associated to operational exploration and production spills, ship releases and others. The meteo-oceanographic phenomena include biogenic oils and/or upwellings, algae blooms, low wind areas and rain cells. The models induced by support vector machines and neural networks achieved good results, allowing the operational implementation of the proposed procedures.