Passive Airborne Fourier Transform Infrared Remote Detection of Methanol by Use of Wavelet Analysis as A Feature Extraction Method

Abstract

Wavelet analysis was evaluated as a data preprocessing tool in the construction of automated classifiers for the detection of volatile organic compounds from passive Fourier transform infrared remote sensing data collected in a downward-looking mode from an aircraft platform. The discrete wavelet transform was applied to single-beam spectra and patterns were formed with either the wavelet coefficients directly or with spectra reconstructed with selected resolution levels of the wavelet decomposition. Automated classifiers were constructed with support vector machines (SVM) and used to detect releases of methanol from an industrial site. A key issue in this work was the desire to use data collected during controlled experiments on the ground to train the SVM classifiers. Spectral backgrounds in these ground-collected data are different than those encountered as the aircraft flies, however, and the development of successful classification models requires spectral preprocessing to suppress background signatures. Biorthogonal wavelets were used to generate patterns and resulted in SVM models that produced no missed methanol detections and false detection rates of less than 0.1% when applied to prediction data not used in the development of the model. The SVM classifiers constructed with wavelet processing were compared to one based on unprocessed spectra and also to one computed with spectra preprocessed with Butterworth high-pass digital filters.

Keywords:

• Classification
• passive infrared
• remote sensing
• support vector machine
• wavelet

Acknowledgments

Robert Kroutil of Kalman and Company and Mark Thomas of the U.S. EPA are acknowledged for providing the infrared remote sensing data employed in this work.

Additional information

Funding

This work was funded by the U.S. EPA through Kalman and Company (contract 1501-001).