Application of partial least squares regression to the automatic detection of chemical vapors by passive infrared remotely sensed image data

Passive infrared (IR) remote sensors are gaining wide acceptance as an analytical tool for the remote detection of chemical vapor plumes. A common problem in plume detection for remotely sensed data is the ability to obtain a quality background signature. Many detection methods employ techniques to extract the signatures of the unknown components in order to determine the overall classification of a desired signature. However, this is often the most difficult step since no prior background knowledge is available. In this document, a novel implementation of partial least squares (PLS) regression is proposed for the automatic detection of dimethylmethylphosphonate (DMMP) vapors from remotely sensed hyperspectral image data. In this implementation, prior knowledge of the target signature is used to extract the analyte information directly from the scene. The various unknown and interfering signatures are implicitly modeled by the PLS algorithm as components that maximize a covariance criterion. This implicit modeling is beneficial since it allows for the detection of a single target chemical without the need for a separate background subtraction procedure.