Hyperspectral image-based feature integration for insect-damaged hawthorn detection
Abstract
It was shown that hawthorn contains many substances that may benefit human health. However, this fruit is likely to be internally damaged by insects. Fruit pests have the potential to cause significant economic losses in its production, the damage of which is difficult to detect. The potential of a hyperspectral imaging technique for detecting insect-damaged hawthorn fruits is thus investigated. Three hundred and sixty samples, including normal and insect-damaged hawthorn fruits, are scanned in the 400 nm to 1000 nm wavelength range using a hyperspectral reflectance imaging system. The ability for insect-damaged hawthorn detection of fourteen types of image features extracted from hyperspectral images are evaluated using significant difference analysis. Reflectance sensitivity analysis of the hyperspectral data is then used for extracting the optimal wavelengths from full wavelengths. Partial least squares discriminant analysis (PLS-DA) is applied to develop the classification model. Among fourteen types of image features, three types of features, namely, mean, energy, and entropy feature, yielded relatively high classification accuracy for insect-damaged hawthorn. The classification accuracy is 98.0% based on full-spectrum PLS-DA model using integrated features of mean and energy and entropy, which is significantly better than that of PLS-DA models using mean, energy, and entropy features alone. The classification model with three-integrated features from the fifteen optimal wavelengths selected by reflectance sensitivity analysis, which is only 16.0% of full wavelengths, achieved 97.4% classification accuracy for test set. The paired t-test showed that there was no significant difference existed between the classification model using three-integrated features from the fifteen wavelengths and the full-spectrum model using three-integrated features. These results indicated the potential of detecting insect-damaged hawthorns using hyperspectral image-based feature integration.