Research And Application Of Image-based Crop Phenotypic Feature Extraction Technology

The yield and quality of crops are closely related to crop phenotype information.The accurate extraction of phenotype information is of great significance for digital breeding,smart cultivation,and refined greenhouse management.Crop phenotypes refer to complex crop traits determined or influenced by genes and the environment.Due to the problems of low efficiency,large error,and high cost of large-scale phenotype equipment in traditional manual phenotype measurement methods,this paper proposes an image-based high-efficiency,low-cost and automated method for crop phenotype feature extraction.The main research contents and conclusions are as follows:(1)This project takes the lettuce and strawberry populations of different varieties planted in the greenhouse as the research object.Using a portable image acquisition device and an automated crop phenotyping platform,images of lettuce at maturity and strawberries at budding and flowering and fruiting stages are obtained.Image preprocessing and image enhancement were performed,and image datasets of lettuce and strawberry were constructed,which provided the basis for subsequent experimental data.(2)The image segmentation algorithm based on machine learning identifies the contours of 60 lettuce varieties and removes the noise in the contours to obtain the lettuce region,and then extracts 39 phenotypic features including shape,color and texture.Calculate the mean,maximum,minimum,standard deviation,maximum standard score,and minimum standard score for these features to observe the distribution of the index parameters.A regression analysis model was established based on the artificially measured fresh weight of lettuce leaves and the extracted leaf area index,and the regression coefficient R2 of the predicted result was 0.91,which verified the measurement accuracy of the data.Agglomerative hierarchical clustering was performed on 39 phenotypic indicators,and 60 lettuce varieties were divided into3 categories.According to the clustering results,a principal component analysis diagram and a clustering profile diagram were drawn to further analyze the differences of the 3 categories.Combined with the actual lettuce Images explain the rationality of classification and realize breed classification based on phenotypic characteristics.Based on this algorithm,lettuce leaf segmentation software was developed,and the video stream was segmented and information superimposed.(3)Based on the deep neural convolutional network YOLOX model,the leaf detection and identification of 35 varieties of strawberry at the budding stage were carried out,and the YOLOX model was compared with the typical target detection model.The results showed that the YOLOX model has higher detection pixels.Accuracy.The model was used to detect and identify flowers and fruits in the strawberry dataset during the flowering and fruiting period.The detection accuracy of the model in leaves,flowers and fruits reached 96.71%,92.32%,and 83.11%,indicating that the detection accuracy of the model meets the requirements.Output the trained model parameters,and count the leaves,flowers,and fruits of the strawberry growth within 7 days.Based on these three phenotypic indicators,35 strawberry varieties are clustered into 3 categories,and the principal component analysis diagram and heat map are drawn to show that The classification is reasonable,and the target detection of strawberry leaves is performed on the video stream.The average precision m AP is 81.61%,which meets the needs of high efficiency,low cost and automation,and provides reference for crop phenotype extraction.