| The root system is an important bridge for energy exchange between plants and soil.Accurately obtaining the complete configuration of the in-situ root system is of great significance for exploring its development rules and clarifying functional mechanisms.Affected by the limitation of observation methods,soil particle obscuration,and root incompleteness,it is difficult to obtain accurate root phenotypic information.With the rapid development of high-tech observation technology and deep learning,how to accurately segment the complete root configuration under the interference of complex soil background,and make the root phenotype analysis results as close as possible to the actual growth situation,is an important challenge in the field of automated phenotype analysis.In this thesis,based on the self-made digital equipment of in-situ root imaging device,a DeepLabv3+network based on subpixel convolution was proposed to realize non-destructive,real-time,complete and accurate identification of cotton in-situ root system by combining semantic segmentation and superpixel theory,and provides theoretical and technical support for high-throughput in-situ root phenotype research.This thesis is based on the logical main line of "in-situ root imaging device-cotton in-situ root system data set establishment-pixel-level semantic segmentation method-automated root system segmentation platform ".(1)A low-cost cotton in-situ root growth device based on digital imaging was designed.While meeting the needs of cotton plant growth,it can obtain high-resolution images of cotton in situ and establish a sample data set.(2)According to the characteristics of semantic segmentation network,the architecture,principle and parameter setting were analyzed and discussed respectively.Inspired by PixelShuffle algorithm,an automatic cotton root image segmentation method based on SP-DeepLabv3+was proposed.The sub-pixel convolutional is used instead of the traditional bilinear interpolation upsampling method,and the additional interpolation function is added and implicitly included in the convolutional layer,so that the network can gradually learn to adjust the interpolation function adapted to this task in the process of root edge feature resolution from low to high,and finally output the segmented image through periodic screening.After 80epochs of training,the final validation set F1-score,recall and precision are 0.9773,0.9847 and 0.9702,respectively,and the performance is better than the standard DeepLabv3+ and U-Net network.(3)WinRhizo was used to analyze the results of 161 root-line images which were separated by SP-DeepLabv3+and traditional WinRhizo manual segmentation,and the quantitative indexes of root phenotypes were obtained for comparative evaluation.The experimental results showed that the SP-DeepLabv3+network had a higher Spearman rank correlation of 0.9667(p<10-8),R2=0.9449,but there were still errors in the analysis results of the other three phenotypic parameters,root surface area,average diameter and volume.(4)A high-throughput automatic segmentation platform for cotton root images was constructed.Combining the imaging devices of digital equipment and root image segmentation methods,the end-to-end automatic acquisition,segmentation and storage of cotton root images were realized,which significantly reduced the complexity of traditional manual or semi-automatic root phenotype analysis,and improved the segmentation accuracy and efficiency.This thesis provides a new research idea and method for automatic analysis of plant phenotypes,which can replace the traditional manual segmentation to a certain extent.The results of this thesis will improve the efficiency of in-situ root image segmentation,solve the problems of traditional root phenotypic segmentation,such as low efficiency,complex operation and great interference from soil environment,and enrich the understanding of the development rules of cotton root system,which is conducive to the subsequent study on the morphology of cotton root system. |
PDF Full Text Request