Identification of multiple plant leaf diseases using neural architecture search

Abstract: A plant disease is one of the key factors affecting the yield and safety of agricultural products. Traditional monitoring methods rely mainly on field sampling, and thereby to assess the species and severity of diseases, normally implemented by plant conservation experts. Nevertheless, the time-consuming and laborious method cannot meet the application requirements of rapid detection in actual large-scale production in modern agriculture. Alternatively, the image analysis method for automatic identification of plant diseases can provide an effective technical way for the real-time monitoring, particularly on deep learning algorithms with the advantages of high spatial resolution and high speed. However, most previous methods strongly depend on the experience of experts during the design of deep neural network architectures. It is inevitable to frequently adjust the network architecture and parameters, in order to obtain the optimal recognition model. In this study, a disease recognition method from plant leaf image was proposed based on the Neural Architecture Search (NAS). Bayesian Optimization (BO) algorithm was used to guide the network morphism, and thereby the network architecture can be selected as the optimal operation of network morphism every time. The proposed method can also automatically learn the appropriate deep neural network architecture according to the specific data set. A total of 54 306 plant disease images including 14 crops and 26 diseases of PlantVillage were used as experimental data. The balanced data after oversampling and subsampling, and the data after grayscale processing, where the 80% of images were used as the training dataset, whereas, the rest as the testing dataset. Firstly, the initial architecture of NAS was set as a three-layer convolutional neural network. Each layer was set as a convolution block, including a ReLU layer, a batch-normalization layer, a convolutional layer, and a pooling layer. Training data was used as the NAS input, while, the search history can be all generated network architecture, parameters learned from network architecture, and model loss values. Secondly, the acquisition function algorithm was optimized to generate the next network architecture for the observation. In the algorithm, the input data can be taken as the minimum temperature, cooling rate and search history of simulated annealing algorithm, whereas, the output data can be the new network architecture, and the required network morphing operation, in order to transform the existing architecture into a new one. After that, it needed to divide the data into multiple batches, and then to train each searched neural architecture. The optimal network architecture was automatically marked when the given search time reached. As such, the required network architecture was finally trained to obtain the disease classification model. Consequently, the disease identification can be gained using the test data as input to the model. Experimental results showed that the proposed method can search the appropriate network architecture in a short time. Furthermore, the method can also find out the optimal network structure, when the training sample data was in unbalanced and balanced conditions, where the accuracies of model recognition were 98.96% and 99.01%, respectively. Additionally, the color information of images related to plant disease has a positive promoting effect on the recognition of the diseases. Nevertheless, the accuracy of model recognition relatively decreased to 95.40%, when the gray image without balance processing was used as the training data. The proposed method can effectively simplify the workload of network architecture design, while identify accurately plant diseases, and thereby to provide a promising technical way for scientific formulation of disease control strategies.