Neural networks for the automated detection of chemical plumes and marine oil spills in airborne infrared multispectral remote sensing images

The United States Environmental Protection Agency Airborne Spectral Photometric Environmental Collection Technology (ASPECT) program uses airborne remote sensing to support first responders in responding to incidents in which chemical or radiological materials are released into the environment. Two main tasks of the ASPECT program are the detection of chemical plumes in the atmosphere in accidental chemical releases and the detection of marine oil slicks on seawater in oil spill accidents. Detections are made in near real time by applying software-based mathematical classification models (classifiers) to the imagery data collected by a downward looking infrared (IR) multispectral sensor on the aircraft. The classifiers classify each image pixel as “plume” or “non-plume” or “oil” or “non-oil”, depending on the application. The research described in this dissertation focuses on the development of classifiers for future use in the above-mentioned emergency response applications.

Three classifiers have been developed for chemical plume detection and one classifier has been developed for oil spill detection. Methanol was used as the target compound to demonstrate the methodologies for the development of plume classifiers. To mimic an accidental chemical release at an industrial facility, the ASPECT program performed controlled methanol release field experiments. Classifiers 1 and 3 were developed on the same field data using two different algorithms. Both algorithms were branches of neural networks but had distinct architectures: one was “shallow” and one was “deep”. The shallow classifier required more effort in feature extraction and preprocessing and less effort on the model development, while the deep classifier was the opposite. Because the field experiments were time-consuming and expensive, strategies were also investigated to simulate plume radiances and to use those radiances in computing the classification model. Classifier 2 was developed by use of this procedure. Classifier 4 was built for the detection of oil spills. This classifier was developed using the shallow neural network, so the focus was the feature processing strategies needed to differentiate oil and water.