A deep learning model for oil spill mapping

Global production, transportation, and consumption of oil can cause inevitable spills into the environment (typically from pipelines, oil wells, and storage facilities) contributing to surface and groundwater contamination. The hazardous and toxic contents of spilled petroleum can negatively impact flora and fauna, and human health, and accordingly, it should be immediately detected and remediated from the environment [1].

The advent of new satellite instruments with high spatial and spectral resolutions (e.g., WorldView-3 and EnMAP) has opened up new opportunities for terrestrial oil spill detection, mapping, and monitoring. The present challenge, however, is to adopt a robust processing algorithm capable of detecting the spectrally variable signal of oil with low reflectance properties and nonlinear spectral behavior in a heterogeneous background, which can also change over time due to degradation [2]. The supervised multi-class classifiers are not appropriate for this aim as they need all the classes (land cover types) in the scene to be known and incorporated in the training stage. However, one-class classification (OCC) techniques allow searching for the class of interest (target class) in a scene with the remaining classes being treated as unknowns (background clutter).