Neural networks for the automated detection of methanol vapour from airborne passive infrared multispectral imaging data

ABSTRACT

Pattern recognition methodology was developed for the automated detection of methanol vapour plumes from passive multispectral infrared remote sensing data. The data employed in this work were collected with an infrared line scanner equipped with eight spectral bandpass filters mounted in a downward-looking mode on a fixed-wing aircraft. An automated classifier was developed by the application of a backpropagation neural network to the calibrated, registered, and preprocessed radiances. Preprocessing steps were performed to optimize the inputs for the neural network, which included: (1) contrast enhancement by calculating the ratios of band intensities; (2) assembly of training data by use of the $k$-means clustering algorithm; (3) removal of temperature information from the measured radiance data; (4) feature extraction for identifying the most representative ratios of band intensities; and (5) optimization of the starting assumed emissivity value for the temperature and emissivity separation algorithm. The best classifier achieved an overall classification accuracy of 98.07% on the testing set with a false detection rate of 0.90% and a missed detection rate of 13.50%. The prediction performance of the optimized neural network was demonstrated through its application to 16 images not included in the training procedure.

Acknowledgements

Funding for this project was provided by the U.S. Environmental Protection Agency through Kalman & Co., Inc. Mark Thomas and Robert Kroutil are acknowledged for providing the data used in this research.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was supported by the United States Environmental Protection agency through a sub-contract with Kalman & Co., Inc..