The use of RADARSAT-2 Backscatter Coefficients (σo, βo, γo) to Discriminate Oil Seeps From Oil Spills: A Linear Discriminant Analysis in the Gulf of Mexico (Campeche Bay)

Using Synthetic Aperture Radar (SAR) measurements, we try to improve the discrimination accuracy of our previously published algorithm that categorizes oil slicks observed on the surface of the ocean as either oil seeps (natural source) or oil spills (operational origin). This empirical Linear Discriminant Analysis (LDA), that achieved an overall accuracy of about 70%, aims to contribute both to environmental assessments (public use), satellite oceanography research (academic use), and operational applications for the petroleum industry (commercial use). Since 2018, a multinational oil and gas company (Brazilian Petroleum Corporation: Petrobras) is exploring our proven seep-spill discrimination methodology, thus intending to ameliorate its routine operational search for new offshore oil grounds. Our database for this study is the same as our previous one: 244 satellite scenes of Campeche Bay (Gulf of Mexico) collected by the Canadian RADARSAT-2 platform – it contains 4,562 field verified slicks: 1,994 (44%) seeps and 2,568 (56%) spills. We now investigate the discrimination of the oil slicks’ category from this database using 61 different combinations of three types of SAR Backscatter Coefficients (sigma-naught (σ^o), beta-naught (β^o), and gamma-naught (γ^o)), four types of SAR Calibrated Products (received radar beam given in amplitude or decibels, with or without despeckle filtering), and the three Data Transformations (None, Cube Root, and Log₁₀). Although our more detailed data analyses have reached accuracies comparable to those of our previous LDA algorithms (~70%), we have simplified the problem using fewer attributes to start the analysis – this dimensionality reduction simplifies its applications to categorize oil (slicks) into oil (seeps) and oil (spills); be it for public, academic, or commercial use. We also observed that the effectiveness in discriminating seeps from spills is rather independent of the choice of Backscatter Coefficient or Calibrated Product, but is strongly influenced by the application of Data Transformations. Keeping up the investigative nature of our research, we suggest non-linear Multivariate Data Analysis (e.g. Cubist or Random Forest) to be considered as an aid to further improve our sound seep–spill discrimination power.