Corn seedling recognition algorithm based on hyperspectral image and lightweight-3D-CNN

Abstract

Weeds can severely harm corn seedlings. If weeds are not removed promptly, a series of problems arise, eventually resulting in a decrease in crop yield. In this study, corn seedlings and weeds were used as the research object to achieve rapid training and recognition of corn seedlings and weeds in hyperspectral images, a lightweight three-dimensional convolutional neural network (lightweight-3D-CNN) model was proposed, and two lightweight units were designed in the network. First, an improved band selection network based on fully connected networks (improved BS-Net-FC) was used to screen out characteristic bands to reduce the number of input channels of the model. Second, the training samples were rotated and flipped horizontally in the plane space, and the training samples were expanded six times. Then, the spectral data of the selected 15 optimal bands were tested on the lightweight-3D-CNN, lightweight-2D-CNN, 3D-CNN, and 2D-CNN models. Finally, the computational efficiency of the full band was compared with that of the optimal band. The test results showed that the average recognition accuracy of the model proposed in this study was 98.58%, which was improved by 1.30%, 3.23%, and 6.68% when compared with the lightweight-2D-CNN, 3D-CNN, and 2D-CNN models, respectively. In terms of computational efficiency, although the accuracy drops slightly in the optimal band, the trainable parameters are nearly 10 times less than those in the full band, and the training time, as well as the testing time, are significantly reduced. It provides a feasible technical approach for the fast training and identifying hyperspectral images of corn seedlings and weeds.