Research On The Identification Of Corn Leaf Diseases Based On 3D-2D Hybrid CNN Model

Corn diseases are one of the significant constraints to high-quality corn production,and accurate identification of corn diseases is of great importance for precise disease control.Corn anthracnose and brown spot are typical diseases of corn,and the early symptoms of the two diseases are similar,which can be easily misidentified by the naked eye.Widespread outbreaks of crop diseases can cause significant economic losses.Therefore,the detection,quantification,diagnosis and identification of crop diseases are particularly important for precision agriculture.Manual methods of crop disease detection were commonly used.Manual investigation is the most basic manual crop disease detection method.However,this approach requires expertise applied to relevant crop phenotypes and crop pathology.Another manual crop disease detection technique is biomolecular technique,which requires detailed sampling and complex processing methods.It is known that the process of manual method is expensive,time-consuming and laborious.These shortcomings limit the development and application of manual methods in large farms.In recent years,hyperspectral technique,as a typical non-invasive technique,has attracted more and more attention.Hyperspectral sensors can obtain spectral information from hundreds of spectral bands.These bands are highly sensitive to subtle changes in crops caused by diseases and make it possible to distinguish between different disease types and perform early asymptomatic detection.However,the digital information collected by hyperspectral instruments is always large,which contains a lot of redundant information,which brings trouble to data analysis.Therefore,feature selection becomes the key procedure of high-dimensional spectral data preprocessing.In this paper,we take corn brown spot and anthracnose as the research objects,target the similarity of the symptoms of the two diseases,extract the feature bands of corn disease hyperspectral data using the band selection module in a deep learning framework,and use a 3D-2D hybrid CNN model to identify the corn diseases.The following work was mainly accomplished:(1)Acquisition and pre-processing of hyperspectral data.Firstly,100 maize leaves for each of anthracnose and brown spot were collected and the hyperspectral data of each leaf were acquired using Hyperspec VNIR hyperspectral imaging system;then the Savitzky-Golay smoothing filter(SG)method was used to preprocess each hyperspectral image;finally,the region of interest was selected on each image to establish a sample set.(2)An exploration was conducted into the efficacy of classic hyperspectral data feature selection techniques and classifiers for recognizing corn illnesses.The continuous projection algorithm and partial least squares regression algorithm were used for feature selection of corn disease hyperspectral data,and then support vector machine was used for classification.The continuous projection algorithm’s chosen feature band was concentrated in the 786-901 nm and993-999 nm near infrared areas.The partial least squares regression algorithm selects the feature bands mainly in the 507-552 nm and 611-688 nm regions,in the NIR and red-edge regions.The continuous projection algorithm and the partial least squares regression algorithm selected feature bands with a classification accuracy of 57.42% and 61.14% in the support vector machine.(3)A 3D-2D hybrid CNN model combining a band selection module is proposed,which combines band selection,attention mechanism,spatial-spectral feature extraction and classification into a unified optimisation process.The model first inputs hyperspectral images to both the band selection module and the attention mechanism module,and then sums the outputs of the two modules as the input to the 3D-2D hybrid CNN.The spectral bands selected by the model’s band selection module achieve more reliable classification performance than traditional feature selection methods.The model achieved the best classification accuracy of 97.37% for Y-Net compared to support vector machines,1D CNNs and 2D CNNs.(4)The structure of the model was optimised using a network pruning method to remove the unimportant weights and obtain a lightweight model.The trained model was pruned to reduce its size to 1/3 of its original size,resulting in an accuracy rate of 98.34%.