Wheat Drought Stress Identification And Classification Based On Image Deep Learning

Drought stress is one of the major natural agricultural disasters,which seriously affects wheat growth and yield level.Rapid,accurate and non-destructive monitoring drought stress to improve the precision and timeliness irrigation was of great significance for saving water resources and ensuring high and stable wheat yield.Commonly used methods for drought diagnosis include soil moisture monitoring and agro-meteorological forcasting.These methods indirectly monitor wheat drought stress through soil moisture or ago-meteorological data with low precision and indirect monitoring with certain lag.However,under drought stress,wheat plants showed some phenotypic characteristics,such as withering,yellowing and rolling of leaves.At the same time,with the development of artificial intelligence,the deep learning model has achieved high accuracy in image recognition and the recognition accuracy has exceeded the human levels.Therefore,in this thesis,the deep learning and object detection model were used to identify and classify wheat drought stress based on the phenotypic characteristics of wheat under drought stress.It has important theoretical significance and application value to monitor drought stress directly through the phenotypic characteristics of wheat plant,namely drought disaster-bearing body.The main contents and contributions of this thesis are as follows:?1?In this experiment,a series of pot experiments were conducted using wheat plants.The experiments were designed with five different levels of drought conditions including:optimum moisture,light,moderate,severe,and extreme drought stress.Wheat digital images of different drought stress at three growth stages were obtained by single-mirror cameras.An extensive image dataset of total number of 130123 items was established,of which 47503 were in the stage of jointing,45352 were in the stage of heading-flowering and 37268 were at flowering-maturing stage.At the same time,the datasets of each stage were divided into three subsets including morning,noon and afternoon intervals?2?In view of the shortcomings of indirect drought monitoring,such as time-consuming,laborious and lagging,deep learning models based on wheat phenotypic characteristics was proposed to identify and classify wheat drought stress.The result showed that deep learning model achieved high accuracy and reliability in the identification and classification of wheat drought stress.The accuracy of model for identification and classification of wheat drought stress in the jointing stage were 99.47%and 97.40%,heading-flowering were 98.73%and 98.64%,flowering-maturing stage were 99.93%and 99.88%,respectively.The accuracy of deep learning model for wheat drought stress identification and classification at different time periods?morning,noon,afternoon?were above 99%and higher than 96%,respectively.In different drought stress treatments,the accuracy of deep learning model at optimum moisture and extreme drought stress treatments were higher than other treatments.?3?Aiming at the problem that traditional machine learning model needs to extract image features manually and the accuracy of image recognition is low,this thesis compared the accuracy of deep learning and traditional machine learning in wheat drought identification and classification.The result showed that the accuracy of wheat drought stress identification was significantly higher than that of traditional machine learning algorithms.The wheat drought stress identification and classification accuracy of deep learning at the jointing,heading-flowering and flowering-maturity stages was 0.24%-20.79%,0.33%-14.32%and 0.58%-12.84%higher than that of the machine learning algorithms,respectively.?4?Leaf rolling is one of the most typical phenotypic characteristics of wheat under drought stress.In this thesis,we proposed to identify wheat drought stress by rolled leaf detection,and we designed a pipeline using machine learning model to classify different wheat drought stress based the rolled leaf detection information.The results show that object detection model can accurately detect and locate wheat rolled leaves with high confidence.When the IOU threshold was 0.5,the mean average precision,precision of the model were 92.22%and 85.94%,respectively.The wheat drought stress classifcation accuracy of machine learning were more than 82%based on wheat rolled leaf detection information,The accuracy,precision,recall and F1_scorecore of KNN were the highest in all models,comprising of 95.83%,96%,96%and 95.67%,respectively.