CLASSIFICATION OF PLANT DISEASES BASED ON HYPERSPECTRAL IMAGES OF LEAVES USING MACHINE LEARNING METHODS

Abstract

In this paper, we consider the problem of classifying cotton diseases based on hyperspectral images of leaves using deep learning methods. An approach based on the use of a three-dimensional convolutional neural network (3D CNN) with an attention mechanism is proposed, which makes it possible to efficiently extract spectral and spatial features. To increase the information content of the data, pseudo-hyperspectral image transformation is used, as well as preprocessing and dimensionality reduction methods. As part of the study, a dataset was formed that includes 6,500 images, divided into training (70%), validation (15%), and test (15%) samples. The experiments have shown that the proposed model provides high classification accuracy (up to 93-95%) and surpasses traditional machine learning methods. Additionally, the results were interpreted using visualization methods, including spectral signatures and Grad-CAM, which make it possible to identify informative areas and wavelength ranges. The results obtained confirm the effectiveness of the proposed approach and its prospects for application in the tasks of monitoring the state of agricultural crops.