Intelligent identification of heat stress in tomato seedlings based on chlorophyll fluorescence imaging technology

Accurate and timely detection of heat damage has been a high demand during different growth stages of crops. However, the traditional identification cannot clearly distinguish the specific characteristics under various heat damage levels, particularly with the high cost and labor-intensive operation. Taking the tomato seedling leaf of heat damage as the research object, a set of chlorophyll fluorescence imaging devices was designed to accurately classify the different heat damage using an improved AlexNet network model. Different levels of heat damage were selected in the tomato seedling leaf, including the health, the mild, moderate and severe heat damage. The original fluorescence images were captured for the single background of tomato seeding leaf at 5:00 each day. Threshold mask segmentation was then utilized to obtain the 12 sets of fluorescence parameters. Among them, the quantum yield of unregulated energy dissipation Y (NO) was the highest correlation in the Spearman rank correlation analysis. The data enhancement technique was used to expand the dataset, due to the small original number of samples. A total of 3 728 sample images were obtained, and then divided into a training and test dataset, according to the proportion of 7:3. Batch normalization, depth separable convolution, and global average pool layer were introduced into the AlexNet network for the improved one. As such, the improved AlexNet network model was achieved, including a traditional convolution layer, four depth separable convolution layers, a max-pooling layer, a global average pool layer, and a fully-connected layer. The Softmax classifier was also used in the last fully-connected layer. A self-regularized non-monotonic function was chosen as the activation function. The Adaptive Moment Estimation (Adam) was chosen as the optimizer. The improved AlexNet network model was used to identify the four-class heat damage samples. The optimal hyperparameters in the Convolutional Neural Network (CNN) models were determined via the performance under the different learning rates and optimizers. The test data showed that the AlexNet model performed best with the learning rate of 0.000 1 and an optimizer of Adam, fully meeting the requirements of fluorescence images. A visualization technique was used to analyze the features of convolutional layers, particularly on the feature extraction in a CNN model. It was found that the heat damage features were ever-increasing outstanding with the increase of the convolutional layers. The clearest features of heat damage were found in the captured images by the last convolutional layer. The optimal performance of the improved AlexNet model was achieved at the initial learning rate of 0.000 1, where the average accuracy was 98.8%, the model size was 1.13 MB, and the testing time per image was 11.6 ms. The accuracy of the improved AlexNet increased by 1.8 %, while the detection time was reduced by 23.2%, and the number of parameters decreased by 99.5%, compared with the original model. The comprehensive performance of the improved AlexNet network model was also better than that of Vgg16, ResNet50, MobileNet V2, and Shufflenet V2 models, indicating the outstanding advantages in the accuracy, weight space occupation, and testing time for the detection of tomato seeding leaf heat damages. A mobile and intelligent recognition system of heat damage in the tomato seedling was also developed to greatly reduce labor costs. The finding can provide technical support to the intelligent identification, timely prevention, and control of early heat stress of tomato seedlings.