Research On The Method For Plant Leaves Stomatal Based On Deep Learning

Plant stomata are channels for material and energy exchange between individual plants and the external environment.The stomatal characteristics and behavior of plant leaves significantly impact on plant photosynthesis and transpiration,water use efficiency,carbon sequestration capacity and drought resistance.The phenotypic traits of plant stomata and epidermal cells are important contents of plant phenomics research at the cell scale.Understanding and researching the characteristics and behavior of plant leaf stomata is crucial for comprehending vital aspects of plants,such as photosynthesis,transpiration,water use efficiency,carbon fixation capacity,and drought resistance.These studies provide a scientific foundation for practical applications in agriculture,forestry,ecology,and other related fields,ultimately promoting plant growth,increasing crop yields,and enhancing adaptability to diverse environmental conditions.At present,the stomatal analysis methods of plant leaves mainly use manual measurement or methods assisted by software.However,due to the significant differences in scale,shape,and texture of leaf microscopic images collected from various plant species using different imaging methods,the existing stomatal analysis methods have limitations in meeting the requirements of accurate and efficient processing.In this study,after building a living plant leaf stomata observation system,the automatic analysis method of plant leaf stomata and epidermal cells was studied.The main research contents are as follows:(1)Design of plant leaf surface stomata observation system and data set construction.First,an ultra-depth-of-field optical microscope was utilized to construct a live plant leaf stomata imaging platform.Secondly,a flexible and wearable humidity sensor was designed based on chitosan-doped graphene and carbon nanotubes to measure the humidity of the microenvironment on the leaf surface of plants.Next,the built plant leaf microscopic imaging system was utilized to capture live images of black poplar leaves.Additionally,microscopic images of leaves from various plant species and acquired through different imaging methods were gathered from publicly available sources,including Populus balsamifera,ginkgo,broad bean,Arabidopsis,etc.Lastly,LabelMe labeling software was used to annotate the stomata,pores,and epidermal cells in the microscopic leaf images,resulting in the creation of a dataset.(2)Research on stomatal segmentation method of plant leaf microscopic image based on instance segmentation Mask R-CNN.Firstly,the network model structure of Mask R-CNN is analyzed,the Swin Transformer backbone network with self-attention mechanism and the feature pyramid FPN are used to construct the feature extraction module,and the GRoIE module is introduced to make better use of the useful information output by the feature pyramid.At the same time,the PISA module is introduced into the sample sampling strategy to make the model pay more attention to important samples.Secondly,the constructed training set is used for model training,and the test set is used for leaf microscopic image stomata segmentation experiments.Finally,quantitative indicators such as precision rate,recall rate,and relative error of stomatal anatomical parameters were used to evaluate the performance of the model.The experimental results show that this research method has better stomatal detection and segmentation effects on leaf microscopic images obtained under different plants and different imaging methods and the measurement of stomatal anatomical parameters is more accurate.(3)Research on stomatal segmentation method of microscopic images of plant leaves based on self-supervised pre-training.Firstly,a self-supervised model based on the self-supervised contrastive learning network DenseCL was constructed,and the model was trained using an unlabeled plant leaf microscopic image dataset.Secondly,the weight of the model after selfsupervised pre-training is extracted,the Mask R-CNN instance segmentation model based on ResNet50 is constructed,and the model is transferred and learned by using the artificially labeled air hole dataset.Finally,the performance of the model is evaluated using visual and quantitative metrics.The results show that self-supervised learning can effectively utilize a large number of unlabeled microscopic images of plant leaves,extract effective image features,and improve the performance of the model for detection and segmentation of stomata.(4)Research on segmentation method of plant leaf stomata and epidermal cells based on semantic segmentation model.Based on the semantic segmentation DeepLabv3+model,the attention mechanism module is introduced into the residual module of the feature extraction network.Meanwhile,a generalized loss function based on the Tversky index is employed to address the data imbalance problem between stomata and epidermal cells.The improved semantic segmentation model is trained and tested using the produced dataset.The performance of the model is evaluated by quantitative indicators such as average accuracy and Intersection over Union(IoU).Experimental results show that the semantic segmentation model constructed in this study can simultaneously achieve accurate segmentation of leaf stomata and epidermal cells in microscopic images.