Research On UAV Image Classification And Semantic Segmentation Of Maize Diseases And Pests Based On Deep Learning

Crop diseases and pests seriously threaten food security and sustainable and healthy development of agriculture in China.Currently,pest and disease monitoring mainly relies on traditional manual field sampling survey,which is time-consuming,inefficient and prone to misjudgment due to subjective influence.Therefore,it cannot meet the requirements of intensification under the construction of smart agriculture system.In recent years,the continuous development of UAV,sensor and deep learning technology provided a new idea for automatic monitoring of symptoms of diseases and pests.This study takes the precision monitoring of crop diseases and insect pests as the research key.In the study,we used small low cost unmanned aerial vehicle(UAV)to complete the RGB image acquisition.Based on convolutional neural network structure,we take group channel attention,semi-supervised learning methods and other methods to build and improve Image classification and semantic segmentation models respectively for different dimension of fine-grained monitoring requirements.The models achieve accurate extraction of information such as location and degree of maize disease and insect pest.The main research work is as follows:(1)The first part of study is indirect pest detection of Spodoptera frugiperda which used small UAV to collect field RGB images of damaged maize feeding by Spodoptera frugiperda.During the study,we constructed a dataset of multi-period Spodoptera frugiperda based on the pest characteristics.And Based on this data set and Res Net,we trained and optimized a multilevel classification model named SF-Res Net which including grouping channel attention and data enhancement method.At the same time,the model was trained and tested by images on different cycle.The validation accuracy reached 98.77% at jointing stage and the test accuracy reached 90.01% at heading stage,which verified the high-precision detection ability of convolutional neural network for damaged maize leaves after insect infestation and the robustness against cycle changes.(2)We taken the Northern leaf blight,a maize disease with a wide distribution,as the research object to construct semantic segmentation model based on an open source UAV RGB images.We combined the semi-supervised learning and deep learning methods aiming at decrease the difficulty and high cost of making mask labels for the monitoring of the blight of pixel level.Based on the semi-supervised learning method of cross pseudo-label verification,we designed the semi-supervised semantic segmentation model of maize Northern leaf blight disease.The semi-supervised method including pseudo-label generation and pseudo-label verification can improve the accuracy of the model and the ability of disease boundary extraction.We improved the effect of pseudo-label by introducing the low-level tag information fused with GradCAM and color threshold.Finally,we verified the ability of detecting and extracting the location and precise boundary of corn Northern leaf blight from UAV images.At the same time,Deep Labv3 and Unet network are introduced to semi-supervised and fully supervised comparative training experiments.The image ratio with and without labels is 1: 32 which only had 72 labeled images.The semi-supervised model obtained a DSC area index of 0.800 and a HD 95 boundary index of 7.62.The comprehensive effect is far better than that of the fully supervised model with the same labeled data,which verifies the effectiveness and universality of the proposed semi-supervised method in disease semantic segmentation.There are 42 figures,6 tables and 90 references in this thesis.