In a recent report a strategy was proposed for the classification and identification of toxic organic compounds observed in ambient air from mass spectra using computational pattern recognition based on SIMCA principal components modeling of the autocorrelation transformed mass spectra. With this technique very good classification and identification results (87% and 84%, respectively) were obtained with GC/MS from training and calibration data for the 78 toxic compounds targeted for routine monitoring in ambient air. However, when applied to GC/MS ambient air field data, a number of hydrocarbons were incorrectly classified as chlorocarbons indicating that the training sets were not optimal for discriminating between these classes. A new strategy for data reprocessing, variable selection and class model optimization has been developed to solve this problem. Only the sixteen most intense ions in each mass spectrum are retained. The MS data are scaled by taking the square root of the intensities and the autocorrelation transform is then taken. A training class has been introduced for hydrocarbons in addition to three other classes. The original SIMCA classification rule has been modified to give a more reasonable approximation of the training set pattern structure and object distances from the class models. (Copyright (c) 1989 American Chemical Society.)