Identification of wheat classes using wavelet features from near infrared hyperspectral images of bulk samples

Wavelet texture analysis was used for classification of eight Western Canadian wheat classes using near infrared hyperspectral imaging of bulk samples. Hyperspectral images (slices) at 10 nm interval were acquired in the wavelength range 960–1700 nm. From each slice of hyperspectral data, central 256 × 256 pixels were analyzed using a wavelet transformation at five levels (resolutions) employing Daubechies-4 wavelets. Energy and entropy features were computed at each level in the horizontal, vertical, and diagonal orientations. Additionally, rotational invariant features were obtained by adding features from all three orientations. Based on a stepwise linear discriminant analysis, top 100 features were selected and used for classification of wheat classes. Linear and quadratic statistical classifiers and a standard back propagation neural network (BPNN) classifier were used for classification using top 10–100 features. In another approach, principal component (PC) score images obtained from hypercubes were used for wavelet analysis and classification.The wavelet energy features contributed more than the entropy features in class discrimination. The rotational invariant features were more important than the features at any individual orientation. The wavelet texture features at finer resolutions were more important than those at the coarser resolutions.The highest average classification accuracy of eight classes was 99.1% when top 90 features were used for classification in a linear discriminant classifier. The BPNN had the highest average classification accuracy of 92.1% using the top 70 features. Using wavelet features from score images, the PC2 features gave the highest classification accuracy (79.9%). The wavelet texture features of hyperspectral images can be used effectively for classification of wheat classes of Western Canada.