Research Progress on Non-Destructive Detection of Crop Diseases Using Hyperspectral Imaging Technology

Abstract

Spotting plant ailments without harming the crop is a cornerstone of yield and quality protection: early warning curbs chemical overuse and shields farm income. Here we outline how hyperspectral cameras pick up pre-visual leaf stress, then survey global progress that couples these image cubes to support-vector classifiers, partial-least-squares regressors and deep convolutional networks for pinpoint disease identification. analyzes the principles and classification processes of hyperspectral image recognition algorithms for crop diseases, compares the advantages and disadvantages of three deep learning algorithms [Deep Belief Network (DBN), Stacked Autoencoder Network (SAE) based on Autoencoder (AE), and Convolutional Neural Network (CNN)] in the recognition of hyperspectral images of crop diseases and pests; explains the calculation process and principles of common deep learning classification metrics; points out the problems faced in hyperspectral detection and recognition of crop diseases: different objects with similar spectra, complicated data preprocessing and feature extraction processes, small data volume, and imbalanced training data, and suggests future research directions to address these issues.