Invasive weed optimization for optimizing one-agar-for-all classification of bacterial colonies based on hyperspectral imaging

Abstract Near-infrared hyperspectral imaging together with versatile chemometric algorithms including invasive weed optimization (IWO) were employed for optimizing fast classification of bacterial colonies on agar plates. Hyperspectral images of colonies from six strains of bacteria were collected, and classification models were established by applying partial least squares-discriminant analysis and support vector machine (SVM) on the original as well as difference spectra. The parameters of SVM models were optimized by comparing genetic algorithm, particle swarm optimization and the proposed IWO. The results showed that difference spectra amplified the variations among the spectra of the six strains thus potential for improving classification accuracy. The best full wavelength classification model was IWO-SVM model which produced overall correct classification rates (OCCRs) of 100.0% and 97.0% for calibration and prediction, respectively. Besides, competitive adaptive reweighted sampling (CARS), GA and successive projections algorithm (SPA) were utilized to select important wavelengths to establish simplified models. Among them, the simplified IWO-SVM model based on the feature wavelengths selected by CARS gave the best classification rates of 97.2% and 96.0% for calibration and prediction, respectively. The study demonstrated that IWO was a useful tool for optimizing calibration models thus potential for usage in many other applications.