Physically Based Feature Augmentation to Improve Classification Algorithm Performance

ABSTRACT A new method to improve the identification of gases or vapors in the environment is presented. We construct physically based nonlinear transformations of raw measurement variables from passive infrared remote sensing data and use them as inputs to classification algorithms such as linear discriminant analysis (LDA) and support vector machines (SVM). The physically based transformations are based on the concepts of differential absorption radiometry and contrast ratio from infrared remote sensing. The transformed data significantly improves the classification performance of the classifiers. We show that LDA with transformed input data outperforms an SVM trained on the original measurement variables (overall accuracy improved by up to 23%). The SVM also benefits from using the physically transformed data (up to 20% increase in overall accuracy). Both LDA and SVM give similar performance when the physically transformed data is used, hence the user can use the simpler and more interpretable LDA model rather than the more complex SVM model. These physically based transformations may also be effective in other problems for which one can utilize physical insight of the scenario to build meaningful physical transformations. This new approach has the promise of improving the accuracy of classification algorithms, which may allow using simpler, more interpretable models over more complicated black‐box approaches.