Early Phenotypic Analysis And Drought Identification Of Maize Plants Based On Machine Vision

Plant growth is affected by biological or abiotic factors,such as pests,diseases and water stress.As an important factor,water stress seriously affects global agricultural production.Proper irrigation is necessary to maintain and improve crop yields.Maize plants are sensitive to water.As a result,identifying drought stress effectively can not only monitor crop normal growth,but also play an important role in water-saving irrigation.Traditional drought stress assessment is mainly based on the measurement of soil water content,which is not only time consuming,but also not measured completely.Besides,the soil water content does not always directly reflect the drought status of crops.The non-contact and automatic drought monitoring approaches mainly rely on remote sensing image analysis.In order to solve above problems,this thesis applied common digital image sensors to collect images and obtained samples in a simple and fast way.At the same time,combining with machine vision and machine learning techniques,it achieved rapid analysis of phenotypic characteristics of maize.Drought stress in the early stage of maize growth will slow down the process of crop development.In this thesis,the early growth of maize was regarded as the research object.With the collection of field population maize plants and laboratory pot plant maize images,this thesis extracted the phenotypic characteristics of maize from the overall and individual perspective.Besides,this thesis established the relationship between the phenotypic characteristics of maize and drought stress,which provided a scientific basis for the real-time monitoring of drought conditions.The main work of this thesis is summarized as follows:Firstly,in order to better extract the visual features of crops,a green crop segmentation algorithm based on multi-level features was proposed during image preprocessing.Considering the crop edge pixels that were easily misidentified,this thesis regarded crop segmentation as a multi-classification problem,and the probability value of each pixel in the image belonging to green vegetation was predicted with logistic regression model.At the same time,pixel-level and region-level features were extracted from multiple color spaces,including RGB,HSV and L*a*b*.As a result,the segmentation algorithm was robust.The proposed segmentation algorithm had high segmentation quality and fast calculation speed,which could be applied to various fields of agricultural engineering.Secondly,on the basis of image segmentation,the color and texture features of maize plants were extracted.Considering that the leaves of maize were generally green under suitable water conditions,while tending to curl and turning yellow under drought conditions,four color features based on green dominance were designed in this thesis.Using the extracted color features and texture features based on wavelet transform,two detection models were trained to judge whether the corn was subjected to drought stress,and then to identify the degree of drought stress.Specifically,two-stage strategy was applied to detect drought status of maize plants automatically.Thirdly,the phenotypic characteristics of a single leaf of maize were further extracted.Due to the high density of maize planting in the population,it was difficult to directly extract a single maize leaf.In this thesis,local leaves were manually cropped from the original image of maize population.Therefore,a data set of maize leaves under three different water stress conditions was constructed,containing 18040 leaf images.In order to automatically capture the bottom layer information such as the edge and color features of maize leaves,a deep convolutional neural network(CNN)model was designed to concatenate and fuse the information of the feature maps from different layers.By evaluating the importance of extracted feature maps,a small number of valuable feature maps were selected as input to the SVM classifier,which not only reduced the number of parameters saved in the model,but also had a high recognition accuracy.For potted plant corn,due to the interference of background,corn samples would produce local fracture when using conventional segmentation algorithm.This thesis proposed a semantic segmentation based on dense connected network architecture.The characteristics of the CNN learning ability was applied to realize automatic completion of the fracture blades.Since the labeling of the segmented samples was time-consuming and laborious,and the labeling samples were few,this thesis adopt the idea of transfer learning.Firstly,the segmentation model was trained using the open green vegetation segmentation data sets.Then the pre-training model was transferred to the enhanced maize plant segmentation data set with parameter fine-tuning strategy.The proposed segmentation model Dense U-Net in this thesis improved the information flow in the network and achieved the end-to-end segmentation of individual maize plant.Finally,on the basis of single maize plant extraction,the phenotypic characteristics of the whole plant and the local leaf were analyzed.Using segmented image,on the one hand,the shape,color and texture features of the whole maize plant were extracted.On the other hand,image processing techniques were used to separate the stems and leaves of maize,such as refinement and intersection detection.The separated stems and leaves were identified according to the morphological characteristics.Based on the algorithms of machine vision object detection and segmentation,the morphological characteristics such as plant height,elongation and leaf angle of a single leaf of maize were extracted.Combined with the color and texture features,this thesis provided technical support for future drought monitoring and selection of drought-resistant maize varieties.Facing the requirement of automatic maize drought detection,this thesis analyzed the phenotypic changes of maize plant under drought stress from multiple perspectives based on machine vision technology and machine vision theory.The proposed methods can be applied to the practical fields and extended to other crops,which provided a foundation for agricultural informatization,automated and intelligent production activities.