Research On The Development Of Hyperspectral Imaging-based Plant Phenotyping System And Methods Of Spectral Image Analysis

Plant phenotyping is critical for the analysis of phenotype-genotype-environment interaction and understanding of plant phycological life-cycle.The genome data have been explored adequately alone with the fast development of gene sequencing technologies.While phenotypic complexity can interact with genes and the environment which shows the spatiotemporal dynamic difference during the growth,which can be very challenging for phenotyping research,the severe scarcity of high-quality phenotyping data become the bottleneck problems for genetic analysis and crop breeding.The fast development of modern sensing technology,information technology,artificial intelligence as well computer science,make the effective and contactless optical imaging technique widely applied in phenotyping research,which is speeding up the development of plant phenomics,especially the hyperspectral imaging technology,which has become the crucial technique for plant phenotyping data collection and analysis,for its outstanding advantage of capturing the spectral and spatial information at the same time,it has become one of the most common techniques for capturing the plant growth,nutrient status,disease,and other phenotypic traits with high throughput.However,phenotyping facilities still mainly rely on imports,the system usually with poor scalability,high price,and is unable to upgrade iteratively when facing new application requirements.Besides,it also faces many problems such as phenotyping data analysis with high accuracy and data fusing.This research aims to design and develop a hyper-spectral imaging system for plant phenotyping,we proposed hyperspectral data pre-processing and analysis methods based on deep learning method.The system is applied in rice nitrogen status diagnosis and tomato resistance selection to evaluate its performance.The results and contents are as follows.（1）Phenotyping system mainly relied on import with poor expandability and update difficulty when facing special requirement,this research designed a hyperspectral imaging phenotyping system which can stably running,collaborative with plant transport conveyors efficiently as well as functionally expandable.As the core functional unit for hyperspectral imaging phenotyping platform,we implemented the extendibility of the imaging scheme to the maximum.We designed dual axis at the top of the dark chamber in order to positioning and pushing on X and Y axis for the hyperspecial camera.The bidirectional conveyor system can position sample flexibly.High-precision lifting on Z axis and multi-angle rotation of samples on ZR direction are realized by asynchronous motors drive,the light coupling of Z axis control and bidirectional conveyor control fully guarantee the continue update and extend operations of the platform.Limit switch control at the key positions and PLC emergency safety control are combined for the platform’s reliability and security concerns.For the software design and functional interaction,an automatic sequential image acquisition and system control mechanism are designed based on the plant phenotyping workflow,thus the auto phenotyping image acquisition demand are satisfied.（2）In order to evaluate the image quality of the system,the coefficient variation（CV）and structural similarity index（SSIM）between spectral are analyzed,in the visible area,the CV is higher especially in the 400～500 nm part,the most stable CV are located at 800～1000 nm around near-infrared area.SSIM is preferable in the middle area of the spectral range.On the issue of strip noise contamination during the data collecting,writing and processing errors,this research proposed de-striping convolutional neural network（DS-CNN）derived by deep learning method,the performance of the method are firstly explored on the 6 different strip noise scales datasets,The best performance of DS-CNN was on the lowest strip noise dataset（σ=0.02）,with the mean squared error（MSE）lower than 2×10^-4,highest structure similarity index metric（SSIM）of 0.99,and peak signal-to-noise ratio（PSNR）of around 36 d B.（3）In order to solve the problem of hyper-spectral data redundancy and the challenges of image-spectral data analysis,this research implemented rice leave nitrogen stress diagnosis base on the work had done ahead.A deep learning network for rice nitrogen diagnosis（ND-CNN）are proposed.DS-CNN can also remove the strip noise in hyperspectral image thus to solve the problem of diagnosis disorder and overfitting problem,enhanced the effectiveness of nitrogen diagnosis of ND-CNN.The result shows that the ND-CNN’s nitrogen diagnosis accuracy has increased to 95.56%from 84.44%after remove the strips in the hyperspectral image by DS-CNN,the t-SNE feature distribution map had also improved remarkably,the feature distribution space had also been reduced,strip removal had increased the nitrogen diagnosis performance confidently at the same time.（4）Concerning the high-throughput requirement of tomato wilt resistance selection,based on the platform constructed above,the potential of hyperspectral image technique for resistance screening has been explored.By collecting the HSI images of tomato samples inoculated bacterial wilt（Pseudomonas solanacearum E.F.Smith）and the control group,five resistance levels are designed based on the degree of wilting after sample infected the wilt,the screening potentials are explored among light-weighted convolutional neural network,transfer learning,neural architecture search（NAS）and auto machine learning（autoML）methods.The result showed that,for the full spectral dataset,light-weighted CNN with Tanh activation function reached an accuracy of 86%with the lowest time cost,but its robustness is not promising compared to others.For the featured band image dataset,light-weighted CNN reach an accuracy of 93%,which is little lower that of transfer leering methods（95%）and autoML（97%）.Besides,the transfer learning owns the best robustness.This research proposed the non-destructive screening technique for tomato wilt resistance,which can be a reference for the high-throughput phenotyping data collection and effective phenotype screening.