Tea Tree Classification And Bio-chemical Parameters Estimation Based On Hyperspectral Remote Sensing At Near Ground Scale

Remote sensing technology can achieve target detection in a remote and non-contacting way.The development of remote sensing went through three stages:panchromatic,colored and multispectral imaging.Optical remote sensing has entered a new stage with the emergence of hyperspectral remote sensing.Hyperspectral remote sensing refers to the technology with continuous spectral information in a specific spectrum domain(visible,near-infrared and short-wave infrared wavelengths).It has the advantages of numerous bands and narrow band interval.Using hyperspectral remote sensing method makes it possible to distinguish diagnostic spectral characteristics of the surface information.In recent years,hyperspectral remote sensing plays an increasingly important role in vegetation,soil and water monitoring.Researches in the field includes:Chlorophyll concentration and nitrogen content determination in vegetation leaves,as well as,other chemical elements such as phosphorus,potassium,calcium,magnesium and phenolic determination;Soil organic content,ion content,moisture,soil loss and erosion and degradation,etc researches;Water quality monitoring in water environment.However,the further application of using hyperspectral imaging technology need to be explored in spatial distribution field and biochemical parameters inversion.Such as applying high spectral analysis technology to vegetation cultivar classification,detection of various chemical parameters and so on.High spectral resolution,however,is often accompanied with strong correlation,high redundancy and large noise within bands.As dimension reduction and noise reduction,choosing suitable pre-treatment methods for hyperspectral data and choosing fine classification,modeling method also become the key and difficult point for research in effective utilization of hyperspectral image.Tea trees(camellia sinensis)are wildly seen in mountainous area in the south of the Yangtze river,especially in Zhejiang,Anhui and Jiangxi province.With the growing demand for tea,concerning about tea quality also gradually improve.Thus tea quality becomes an important factors affecting the tea market prices.Sensory evaluation,however,the traditional tea evaluation method is very subjective,not unable to provide quantitative information of tea quality,and can only be used after the tea picking.This approach does not meet the requirements of monitoring tea quality through tea tree growth periods,which is unfavorable for the modernization and scientific fine management of large area of tea garden.The cultivars of tea,tea polyphenols,free amino acid content are significant factors that influences the quality of tea.The traditional use of handheld spectrometer and the method based on high resolution remote sensing image both can not meet the requirement for tea garden monitoring.In this paper,we select 8 kinds of representative tea varieties in the Huazhong Agricultural University as the experimental subjects,such as Taicha#12,Fu’an Dabai,Fuding Dabai,Wu Niuzao,Ying Shuang,Tie Guanyin,Huang Dan and Mei Zhan.We use the Airborne Imaging Spectrometer to get their canopy hyperspectral image.And we studyed:(1)the classification of tea varieties based on their canopy spectral feature.(2)compare different classification results from dimension reduction or not,different combination of dimension reduction and classification methods,trying to find the most appropriate combination methods for tea canopy hyperspectral classification,we estimate their bit-chemical content by canopy spectral image.(3)use deep learning to distinguish tea cultivars at near ground scale.Results as follows:(1)The results of different classification methods is that Support Vector Machine(SVM)and Artificial Neural Network(ANN)classification obtain the best classification results(OA>93%).But because of the large volume of hyperspectral image data,the classification efficiency is low.After noise reduction and dimension reduction methods such as Principal Component Analysis(PCA)transformation,although the classification accuracy decreased a little,but still is satisfied,and can improve the efficiency of classification(2)It can significantly improve the classification accuracy on traditional classification methods,such as Minimum Distance Classification(MDC)and Maximum Likelihood Classification(MLC)by dimension reduction than no dimension reduction,so dimension reduction has greater impact on traditional methods.The greatest enhancement can be 45.2%.(3)Based on tea canopy spectral information,tea polyphenols,amino acid and phenol ammonia are predicted.The prediction accuracy of phenol ammonia is the highest(Rcv=0.66,RMSEcv=13.27),followed with amino acids(Rcv=0.62,RMSEcv=1.16),prediction of tea polyphenol content is slightly lower than the first two(Rcv=0.58,RMSEcv=10.01).(4)In this study,using Partial Least Squares(PLS)to predict tea bio-chemical parameters achieved higher accuracy than using Artificial Neural Network(ANN).(5)Caffe,a feedforward convolution framework,is a neural network system which is able to learn and sum up.It has the strong ability of pattern recognition,and can realize tea tree cultivars identification with its accuracy of 0.76.This study used a variety of classification,inversion,recognition algorithm,based on the tea canopy spectra,and achieved its goal.Also it proposed a new method of quantitative monitoring the tea quality base on hyperspectral remote sensing images efficiently.