Semantic segmentation of chemical plumes from airborne multispectral infrared images using U-Net

The United States Environmental Protection Agency Airborne Spectral Photometric Environmental Collection Technology program provides infrared (IR) remote sensing capabilities from an aircraft platform to assist first responders in managing chemical releases into the atmosphere. One of the instruments used is a downward-looking eight-band multispectral imaging system that receives the upwelling IR radiance from the ground and atmosphere below the aircraft. Volatile organic compounds absorb and emit IR radiation at characteristic wavelengths and produce unique signatures in the imaging data. To automate the detection of chemical plumes, this research applied a deep learning-based semantic segmentation model on multispectral images collected during controlled releases of methanol. A U-Net model was developed for this application, and multiple experiments were conducted to optimize the model. Issues studied included the use of a temperature and emissivity separation algorithm to suppress temperature effects, a custom normalization method to reduce scene composition variance, experiments to shrink the network architecture, and evaluation of the utility of data augmentation. The optimized U-Net model was able to detect the plume area in images collected across different temperatures and locations while achieving false detection rates < 0.02%. The U-Net model exhibited an improved ability to discriminate methanol plumes from other scene elements such as buildings and roads when compared to the performance of shallow neural networks studied in previous work.