ICA based band selection for black walnut shell and meat classification in hyperspectral fluorescence imagery

There are approximately over 15.4 million acres of black walnut with each acre producing about 1000 to 1700 pounds of raw nuts in the U. S. However, only about 20 million pounds of the raw black walnuts are commercially processed every year. The reason that growers are not motivated to process the nuts is that there is not enough nut processing capacity available in the U.S. In the current walnut processing plant, small shell fragments are removed manually in order to meet the required marketable quality. This visual sorting work is a very labor intensive and difficult process because shell and meat fragments can be very similar in size and color. In this research, hyperspectral fluorescence imaging has been studied to analyze the difference type of walnut shell and meat. Although the hyperspectral fluorescence imaging has been found to be efficient for differentiating walnut shell from meat, the scanning speed of hyperspectral fluorescence imaging system is not satisfactory especially for the industry requirement of real-time online inspection. Furthermore, the cost of hyperspectral imaging system is still too expensive to be acceptable by the walnut processing plants. As a result, how to select the optimum wavelength for walnut shell and meat classification and keep the same classification performance simultaneously becomes a realistic issue. To solve aforementioned problem, the Independent Component Analysis (ICA) based hyperspectral band selection approach was proposed in this paper. Walnut samples used in this research included both and two-year old intact black walnuts provided by USDA AMS. Samples were scanned by a hyperspectral fluorescence imaging system. The images were taken at 79 different wavelengths ranging from 425 nm to 775 nm at the 4.5nm increments. The ICA ranking method was first applied to select the most optimal four wavelengths in discrimination of the walnut shell and meat. Then, the k-nearest neighbors (k-NN) classifier was used to do the classification. In order to evaluate the effectiveness of the proposed method, the classification results of ICA based band selection method with k-NN classifier were compared with that of direct k-NN classifier method. The experiment results showed that ICA based band selection with k-NN classifier had better performance than the direct kNN classifier, and ICA based band selection method were effective in classification of walnut shell and meat.