In this work some approaches to improvement of the chemical images classification performance due to preprocessing and dimensionality reduction of the multisensor array responses are studied. A criterion for linear separability of the analyte classes is proposed, allowing to choose an optimal method of responses processing for the particular classification task without the need of cross-validation involving multiple classifier retraining.