Oil Spills or Look-Alikes? Classification Rank of Surface Ocean Slick Signatures in Satellite Data

Linear discriminant analysis (LDA) is a mathematically robust multivariate data approach that sometimes used for surface oil slick signature classification. Our goal to rank the effectiveness of LDAs differentiate spills from look-alike slicks. We explored multiple combinations (i) variables (size information, Meteorological-Oceanographic (metoc), geo-location parameters) and (ii) transformations (non-transformed, cube root, log10). Active passive satellite-based measurements RADARSAT, QuikSCAT, AVHRR, SeaWiFS, MODIS were used. Results two experiments are reported discussed: an investigation 60 several attributes subjected same transformation survey 54 other three selected different transformations. In Experiment 1, best discrimination was reached using ten cube-transformed attributes: ~85% overall accuracy six pieces size metoc variables, one parameter. 2, tied as most effective: ~81% area (log transformed), length-to-width ratio (log- or cube-transformed), number feature parts (non-transformed). After verifying classification 114 algorithms by comparing with expert interpretations, we concluded applying accounting optimizes accuracies binary classifiers (oil spill vs. slicks) simple LDA technique.