Research On Wheat Rust Identification Based On Deep Learning

Wheat is a dominant crop in food production,but the spread of crop diseases can lead to low yields and low quality,especially wheat rust,which is inherently fast spreading and has a large area of damage,so it is important to detect and treat the disease in a timely manner.However,wheat is widely grown and distributed,and identification using traditional methods is time-consuming and laborious,with a subjective bias that can seriously affect the identification results,so techniques such as deep learning can be used to complete the identification of crop diseases,which is currently a more cutting-edge research component.In this paper,we conduct a study on the identification of small-strip rust,combining optical data and UAV remote sensing data to achieve high accuracy identification of wheat rust using deep learning methods.Details of the research are as follows.The WR-EL model is proposed for the problem that it is difficult to distinguish between similar spots of wheat stem rust and leaf rust.In this paper,five convolutional neural network models,VGG,Res Net 101,Res Net 152,Dense Net 169 and Dense Net 201,are used as sub-models,while combined with a snapshot integration learning strategy to propose a WR-EL-based wheat rust recognition model.Due to the small size of the dataset and the fact that traditional data augmentation methods do not meet the needs of the task at hand,this study proposes to utilise nine different ways of data augmentation as an ensemble,one of which is randomly selected for data augmentation in each training session,which increases the diversity of the training samples and also improves the generalisation of the model.The number of iterations of the cosine annealing learning rate in the traditional SGDR algorithm is set to a fixed value or incremented by a multiplier,which can lead to the problem of missing the best model due to too large or too small a jump in the learning rate.Therefore,this study improves the SGDR algorithm by proposing the SGDR-S algorithm so that the number of iterations N for the next restart is determined by the product of the number of training epochs in the current training process and the number of cosine change cycles t.This approach ensures that N is in a reasonable range of change and also solves the problem of the magnitude of the learning rate change,which improves the efficiency of model training.In addition,the dataset used in this paper has category imbalance,so we finally determined the optimal weight proportion of the weight cross-entropy loss function by setting the weight proportion of the three categories several times.The experiments show that the recognition method proposed in this study achieves an accuracy of 92% on the test set and can perform the detection task better.A weak sample learning method is proposed for the time-consuming and laborious problem of acquiring labels from UAV remote sensing images.Based on the classification results obtained from traditional classification methods as sample labels,the deep learning model can not only solve the problem of labour and material consumption,but also obtain high recognition accuracy.In order to achieve the goal of large-scale crop disease identification,this paper investigates the spatio-temporal generalisation based on weak sample learning using field images from different regions at the same time,and the results show that this research method can achieve the identification of winter wheat stripe rust in large-scale scenarios.In addition,there are many causes of greening in winter wheat.To demonstrate whether this method can distinguish this condition,we cropped and added drone images of naturally mature winter wheat fields to an existing dataset of stripe rust wheat and experimentally verified that this method can distinguish different causes of greening in wheat with 98% accuracy.Therefore,this paper fills the gap in this research by making up for the high dependence of deep learning models on a large number of manually labelled samples,and the models obtained by training with weak samples have good generalisation ability,enabling the identification of large scale crop diseases to a certain extent,as well as the differentiation of different causes of wheat greening.