Selection of optimal bands for developing multispectral system for inspecting apples for defects

Hyperspectral image technology is a powerful tool, but oftentimes the data dimension of hyperspectral images must be reduced for practical purposes, depending on the target and environment. For detecting defects on a variety of apple cultivars, this study used hyperspectral data spanning the visible (400 nm) to near-infrared (1000 nm). This paper presents the preliminary results from the selection of optimal spectral bands within that region, using a sequential feature selection method. The selected bands are used for multispectral detection of apple defects by a classification model developed using support vector machine (SVM). As a result, five optimal wavelengths were selected as key features. When using optimal wavelengths, the accuracy of the SVM and SVM with RBF kernel achieved accuracies over 90% for both the calibration and validation data set. However, the results of SVM with RBF kernel (>80%) based on image was more robust than SVM model (>50%). Moreover, SVM with RBF model classified between bruise and sound regions as well specular. The result from this study showed the feasibility of developing a rapid multispectral imaging system based on key wavelengths.