Identification of tomato bacterial wilt severity based on hyperspectral imaging technology and spectrum Transformer network

Abstract

Bacterial wilt (BW) is an essential soil-borne disease in tomato production, which seriously decreases tomato yield and quality. Accurate monitoring of tomato BW status can prevent disease spread and reduce control costs. This study proposes a method for identifying tomato BW severity based on hyperspectral imaging (HSI) technology combined with spectrum Transformer network (STNet). The STNet utilizes Transformer modules to capture the associative information of spectral data in the global range, enhancing the ability of features to represent raw spectral data. The proposed network is trained by using the transfer learning strategy, that is, a huge amount of unlabeled pixel-level spectral data are used to pre-train STNet's backbone network to improve the generalization of extracted features, and then the labeled average spectral data are used to fine-tune STNet to identify tomato BW severity. Compared with the competitive random forest, support vector machine, one-dimensional convolution neural network (1DCNN), and self-attention 1DCNN, the STNet achieved the best identification performance with an average F1 of 0.9309, an overall accuracy of 91.93%, and a kappa coefficient of 0.8903. This study verifies the feasibility of using HSI technology combined with STNet to identify tomato BW severity.