ATOSPES：An Automated Organ Segmentation And Phenotypic Parameter Extraction System For Arabidopsis Thaliana

Plant phenotyping research aims to elucidate the influence of genome-environment interactions on plant traits.As image processing and artificial intelligence technologies advance swiftly,deep learning-based computer vision approaches are gaining prominence in high-throughput plant phenotyping research.Deep learning in image processing possesses the advantage of non-destructive detection and effectively reduces errors caused by human factors.Therefore,the use of deep learning techniques for non-destructive,rapid,low-cost,and automated measurement of plant phenotypic parameters has become an urgent need in the field of plant phenotyping research.Arabidopsis thaliana,as a model plant with a relatively short growth cycle and small genome,is suitable for genetics,genomics,and botany research,possessing significant value in plant phenotyping studies.This research focuses on Arabidopsis thaliana’s developmental leaves,mature pods,and root systems,improving deep learning algorithm models to achieve instance segmentation of leaf and pod images and semantic segmentation of root system images.Based on this,phenotypic parameters,including texture,morphology,and color features,are extracted and accuracy is assessed.In addition,this study relies on high-precision deep learning segmentation models to construct an Automated Arabidopsis Organ Segmentation and Phenotyping Parameter Extraction System(ATOSPES),covering open portal websites and applications,enabling non-destructive and high-precision extraction of Arabidopsis phenotypic parameters.The main contributions of this study are as follows:(1)Instance segmentation,phenotypic parameter measurement,and analysis of Arabidopsis thaliana leaves.To address the issues of leaf occlusion and under-detection of newly developed leaves during developmental stages,this study employs an improved Cascade Mask R-CNN model combining ResNeSt and GRoIE,achieving high-precision leaf segmentation.Experimental results indicate a segmentation accuracy of 96.52%,recall of 95.77%,and F1 score of 96.14%.Ablation experiments also demonstrate the effectiveness of the improved method.Based on image segmentation,52 relevant phenotypic features are extracted and accuracy is assessed.Compared to manual measurements,the average absolute error(DIC)for a single leaf is 0.7586,while the MAPE values for leaf number,leaf area,length,and width are 4.27%,1.30%,3.91%,and 4.64%,respectively.In addition,generalization experiments are performed on different Arabidopsis thaliana leaf datasets and other plant variety images,further demonstrating the model’s robustness and generalizability.(2)Instance segmentation,phenotypic parameter measurement,and analysis of mature Arabidopsis thaliana pods.Considering the characteristics of high overlap and intersection in pods,this study integrates GN+WS to improve the Detecto RS model structure,achieving automated instance segmentation of pods.The improved model is compared with mainstream instance segmentation algorithms such as Mask R-CNN,Cascade Mask R-CNN,and HTC,thus demonstrating its superiority.Experimental results show an accuracy of 95.42%,recall of 93.08%,and F1 score of 94.24%.On this basis,corresponding phenotypic traits are extracted and accuracy is assessed.Compared to manual measurements,the average absolute error(DIC)for a single pod is 0.8573,and the MAPE values for pod number,pod area,length,and width are 4.63%,7.57%,1.94%,and 2.10%,respectively,further validating the model’s accuracy.(3)Semantic segmentation,phenotypic parameter measurement,and analysis of Arabidopsis thaliana root systems.This research initially compares the performance of semantic segmentation networks such as PSPNet,DeepLabV3+,and SegFormer in root system segmentation tasks.Experimental results reveal that the SegFormer network structure demonstrates significant advantages in addressing Arabidopsis thaliana root system lateral root blur and multiple fractures,achieving a mean pixel accuracy(MAP)of 92.56% and mean intersection over union(MIo U)of 83.38%,far exceeding other semantic segmentation networks.This study automatically extracts 41 root system phenotypic features and performs regression analysis on root area,root depth,and width,obtaining regression coefficients of 0.986,0.997,and 0.995,respectively,thus proving the feasibility of the SegFormer model in crop root system segmentation tasks.(4)Design and development of ATOSPES.ATOSPES consists of two main components: an open portal website and corresponding applications.Based on high-precision segmentation and automatic extraction of phenotypic parameters for Arabidopsis pods,leaves,and root systems,an accessible portal website and a QT-based application are designed.These tools take user experience into account,providing convenient phenotypic parameter extraction services for researchers.The portal website facilitates online operations,allowing users to directly upload images for phenotypic parameter extraction,while the QT-based application offers offline usage possibilities,enabling researchers to extract Arabidopsis phenotypic parameters without internet access.The website is currently live,and the application is available for direct download.In summary,the methods introduced in this research attain high-accuracy segmentation for Arabidopsis thaliana images,and the derived phenotypic parameters display notable associations and extensive applicability potential.The open portal and application based on deep learning models and parameter extraction enable non-destructive,rapid,and automated assessment of Arabidopsis phenotypic traits.This research offers a robust practical system for Arabidopsis thaliana phenotyping research—ATOSPES,contributing to a reliable data foundation for Arabidopsis thaliana phenotyping and genetic breeding research.