Passive Fourier transform infrared (FT-IR) spectrometry has been used for a variety of automated remote sensing applications. The automated detection of target analytes can be implemented by classification algorithms trained with short segments of spectra or interferograms. Measurement settings for remote sensing applications based on passive FT-IR spectrometry are quite flexible because of the compact and rugged instrumentation. With airborne platforms, the applications of passive FT-IR spectrometry can be extended to large area monitoring of gases released from the sites of chemical incidents or volcanoes.
For a successful airborne remote sensing application, it is impractical to perform costly and labor intensive airborne data collection to acquire large amounts of data required for training classification algorithms that implement the automated detection of the target analytes. It is highly desirable to collect training data under controlled conditions in the laboratory or during relatively low-cost experiments in the field and then apply the classifier trained with such data to airborne data classifications. The development of classification algorithms that are robust to differences between the training data and the data to which the classifier will ultimately be applied requires effective background suppression strategies.
In this dissertation, data analysis strategies are sought that allow us to train classification algorithms for the automated detection of analytes without requiring costly and labor intensive airborne data collection. Support vector machines (SVMs) are employed as a method to implement robust classifications with short segments of filtered interferogram points as input patterns. Classifiers based on SVMs compare favorably with those based on piecewise linear discriminant analysis that have been successfully applied to various remote sensing applications. Digital filtering implementations are explored to improve the performance of background suppression based on short interferogram segments, resulting in better classification performance of SVMs. The use of infinite impulse response (IIR) filters for interferogram processing is also investigated. Filtering with IIR filters significantly reduces the number of interferogram points involved in interferogram processing. Finally, high-pass filtering is compared with second-derivative computations as a method to suppress background contributions within spectral data. High-pass filtering compares favorably with second-derivative computations to implement effective background suppression.