Disease Spots Identification of Potato Leaves in Hyperspectral Based on Locally Adaptive 1D-CNN

Abstract

Early treatment of potato diseases can increase the yield in the later stage, so correct diseased areas identification of potato leaves is of great significance. Deep learning can effectively obtain invariant features and avoid the limitations of artificial feature extraction, it is gradually applied to hyperspectral image classification. Aiming at the local disease spots of potato leaves with different diseases, this paper used 1D-CNN to adaptively extract invariant features, so as to realize the identification of spots of different diseases. In order to verify the accuracy of the algorithm, the labels of the calibration data are needed, and the traditional calibration methods are cost in high. In this paper, a method of calibrating data for rough calibration followed by fine calibration is proposed. In the experiment, a total of 126 hyperspectral potato disease leaves were collected in Hohhot, there are three types of diseases, including 28 anthracnose, 49 leaf blight, 7 early blight and 42 mixed diseases of varying degrees. Among them, 9 of datasets were used to train and 117 tests. In the diseased area, the average accuracy of traditional SVM was 95.66%, and the number of misclassification pixels was 88,939. The average time of single data recognition was about 395s; the average accuracy of one-dimensional convolutional neural network was 97.72%, 39,684 pixels were misclassification, and the average time required to identify a single data was about 15s. The results showed that the one-dimensional convolutional neural network is faster and better in disease spots identifying of potato leaves in hyperspectral.