CN110390028A - A kind of method for building up in plant spectral library - Google Patents

A kind of method for building up in plant spectral library Download PDF

Info

Publication number
CN110390028A
CN110390028A CN201910304011.1A CN201910304011A CN110390028A CN 110390028 A CN110390028 A CN 110390028A CN 201910304011 A CN201910304011 A CN 201910304011A CN 110390028 A CN110390028 A CN 110390028A
Authority
CN
China
Prior art keywords
cluster
feature
class
wave band
cluster centre
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910304011.1A
Other languages
Chinese (zh)
Other versions
CN110390028B (en
Inventor
张竞成
刘鹏
王斌
陈冬梅
袁琳
吴开华
张垚
周贤锋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN201910304011.1A priority Critical patent/CN110390028B/en
Publication of CN110390028A publication Critical patent/CN110390028A/en
Application granted granted Critical
Publication of CN110390028B publication Critical patent/CN110390028B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/51—Indexing; Data structures therefor; Storage structures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/55—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23211—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with adaptive number of clusters
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Multimedia (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

The present invention discloses a kind of method for building up in plant spectral library.The present invention collects data using standardized collecting method, it proposes the complete Spectra feature extraction method of complete set, feature sensitivity analysis method, method for establishing model and carries out plant classification and status monitoring, wherein propose that wave band self-adaption cluster method can effectively remove the information redundancy in adjacent or close wave band, effectively improve data service efficiency.Based on the above method, the method established plant spectral library of spectra, and be provided complete set spectrum analysis and application.

Description

A kind of method for building up in plant spectral library
Technical field
The present invention relates to the foundation in plant spectral library and application, specifically includes the high-spectral data based on plant spectral library and adopt Set method, feature extracting method, feature sensitivity analysis method and sorting algorithm research based on high-spectral data, and establish reality Existing plant spectral and the storage of mating information and inquiry, the plant spectral library of the applications such as plant classification and status monitoring.
Background technique
Plant classification manages the farming such as agricultural planting section planning, nutrition regulation, the prevention and control of plant diseases, pest control particularly significant；In life Type, the distribution of plant are understood in state, Community composition indicates variation of ecology and environment, the floristics in Greening Design Selection, plantation distribution design etc. have very important application value.Utilize the spectrum in unmanned plane, satellite remote-sensing image Information, which carries out plant classification and status monitoring, can chart for large area plant classification and provide important means.
Library of spectra technology as remote sensing image automatic interpretation process important technology in terms of plant classification and status monitoring With potentiality.So far, based on different application purposes, domestic and foreign scholars construct a variety of different types of object spectrums Library relates generally to the types such as rock, vegetation, soil, ice and snow, construction material.And plant remote sensing monitoring is due to stronger spy Different property, the spectral reflectance or emission characteristics such as plant are common by its institutional framework, Physiological And Biochemical Parameters and own shape feature It determines, and it is these features and the development of plant, closely related by stress state and growing environment etc..Therefore, it is necessary to pass through Many experiments and the spectral signature for determining adaptation floristics and status monitoring feature, the core as plant spectral library technology The heart.Meanwhile the browsing and inquiry to spectral information are laid particular stress in some transmission spectra libraries, directly match or are based on often through spectrum Some general spectral signatures are inquired, and the application demand in this subdivision field of plant remote sensing monitoring is functionally unable to satisfy. For the missing of current special plant library of spectra technology, this method proposes a set of plant spectral base construction method, including standardization Collecting method, for plant classification and the spectral signature Sensitivity Analysis of status monitoring, Spectra feature extraction side Method, classification model construction method etc..Wherein, disappear in the adaptive wave band cluster that Spectra feature extraction link proposes with wave band redundancy Precision of the spectral signature on plant classification and status monitoring can be effectively improved by subtracting method.The Plant Light established according to this method Spectrum library can provide effective support for the analysis and application of plant remote sensing monitoring.
Summary of the invention
In view of the deficiencies of the prior art, it is an object of the present invention to provide a kind of method for building up in plant spectral library.The plant Library of spectra is shown for the spectrum of frequently seen plants and mating information, plant classification and status monitoring are studied.
Specific step is as follows by the present invention:
Step 1: data are obtained
Plant canopy high-spectral data and mating information are obtained according to the method for strict standard specification using the prior art, Wherein mating information includes floristics and status information, and status information refers to plant growth period, growing way and pest and disease damage situation etc..
In the present invention can use strict standard specification method, by ASD FieldSpec4 Pro FR (350nm~ 2500nm) instruments such as type spectrometer obtain plant canopy spectrum.Later, the mating information acquisition method of code requirement standard leads to Consultant expert is crossed, the modes such as special books and data is consulted and obtains floristics and status information etc. with sets of data.
Step 2: Spectra feature extraction
The plant canopy high-spectral data obtained according to step 1 and mating information extraction spectral signature, as plant classification Feature.Here spectral signature includes original wave band feature, first differential feature, continuum feature and vegetation index.
Original wave band feature, the original band spectrum data including all acquisitions；
First differential feature include wave band be 490-540nm blue side at maximum differential value (Maximum Differential value (BMV)), maximum differential value position (Position of the maximum differential Value (BPMV)), the sum of differential value (Sum of differential values (BSV))；Wave band is the Huang of 540-620nm Maximum differential value (Maximum differential value (YMV)) at side, maximum differential value position (Position of The maximum differential value (YPMV)), the sum of differential value (Sum of differential values (YSV))；And wave band is maximum differential value (the Maximum differential value at the red side of 660-780nm (RMV)), maximum differential value position (Position of the maximum differential value (RPMV)), differential The sum of value (Sum of differential values (RSV)).
The acquisition modes of the sum of above-mentioned maximum differential value, maximum differential value position, differential value belong to existing mature technology, therefore It does not explain in detail.
Continuum feature mainly includes depth (Depth), width at continuum near infrared band 530-770nm (Width) and area (Area)；The acquisition modes of above-mentioned depth (Depth), width (Width) and area (Area) belong to existing There is mature technology, therefore does not explain in detail.
Vegetation index is characterized in the spectral characteristic according to vegetation, and wave band is combined, and can construct a variety of vegetation indexs. Vegetation index can be used as simple, effective and experience measurement to earth's surface floristics, vegetative state and ambient conditions.This Specifically used vegetation index is shown in Table 1 in invention.
1 vegetation index of table
In upper table, R800Refer to the spectral reflectivity at 800nm wavelength, similarly, other RWavelengthRepresent the spectrum at the wavelength Reflectivity, X, a, b, S etc. are adjustment parameter.
Step 3: feature sensitivity analysis
For original wave band feature, in order to select the higher wave band feature of sensibility, JM distance can be used as each Sensibility index at wavelength, for JM apart from larger, it is bigger which is in the separability in plant classification and status monitoring, Then think that the wavelength is more sensitive.But during being selected using JM distance original wave band feature, being easy will be some Wave band at adjacent or close-spaced wavelength is chosen simultaneously, these wave bands are superfluous since mutual higher correlation results in a large amount of information It is remaining, the efficiency of feature selecting is reduced, therefore is using JM before as sensibility index first with adaptive wave band clustering procedure Original spectrum wave band is clustered, the higher wavelength of correlation can be clustered, recycle JM distance poly- for each at this time Class is analyzed, and chooses in each cluster JM apart from biggish wavelength as sensitive band feature, can effectively avoid correlation compared with High wave band feature is chosen simultaneously, removes the information redundancy in the original wave band data of EO-1 hyperion.
For index characteristic (referring to first differential feature, continuum feature, vegetation index feature), JM distance can be used and make For the sensibility index of each index characteristic, JM is obtained apart from maximum 10 and is used as Sensitivity Index feature.
3.1. original spectrum wave band is clustered using adaptive wave band clustering procedure, the higher wavelength of correlation is gathered Class；
In order to remove the high correlation in high-spectral data between adjacent or close wave band, redundancy letter in data is removed Breath is clustered using adaptive method wave band, adjacent and close wave band is clustered, main method and process are as follows:
Wave band self-adaption cluster is to be based on Kmeans clustering algorithm, but mainly have different at two.First is that and routine clustering with It is different that cluster is carried out centered on sample, adaptive wave band cluster is then centered on feature dimensions (i.e. this example medium wave band) in the present invention It is clustered, will effectively can be clustered similar wave band in this way, convenient for removing adjacent close wave band in the selection of subsequent sensitive band In information redundancy.Second is that in cluster process, increasing " merging " and " division " two behaviour to improve its adaptivity Make, and set the parameter of control algolithm operation, solves number of classifying in Kmeans clustering procedure and fix, it can not be according to Statistical The problem of matter adaptive classification number.By parameter designing, can effectively merge adjacent similar data category and division exist compared with The data category of big difference clusters wave band more flexible applied to spectrum practical application, better realization wave band cluster.Tool Steps are as follows for body:
(1) it determines predefined parameter, can be used for determining the clusters number in subsequent analysis.
C, it is expected that the class number of cluster；
Tn, every class allow minimum sample number
Ts is each component standard difference upper limit in class.
Td, the minimum range lower limit between two cluster centres, if being less than this number, two clusters need to be merged；
L, the most logarithms for the cluster centre that can merge in an iteration operation
It, the maximum number of iterations of permission
Nc, initial clustering number, may be the same or different with C.
(2) cluster centre is randomly generated.According to originally determined clusters number Nc, generate in the initial clustering of corresponding number The heart.It being concentrated on when to prevent cluster centre from obtaining at random near the wavelength of certain, classification center being obtained using layering, first by wavelength It is divided into Nc layers according to initial clustering number, then obtains cluster centre；According to the acquisition for being layered equally distributed stochastic clustering center Method is as follows:
Wherein, Rc is wave spectrum at the corresponding cluster centre of initial clustering number N c, and i is ith cluster center, and Ls is Wave band sum, Nc are initial setting clusters number, and rand is the random number between 0~1,For by all wavelengths as required The layering of initial clustering number is rounded,To be used as cluster centre at some random wavelength rounding in each layer；
(3) each wavelength data x is traversed according to range formula (2), x is categorized into nearest cluster centre Scj, obtain poly- Class center ScjThe cluster SC at placej；
Dj=min | | x-Sci| |, i=1,2 ..., j ..., Nc } formula (2)
X is wavelength spectroscopy data, SciFor ith cluster center, DjIndicate a certain wavelength data and all cluster centres away from From the distance value at middle minimum, which is jth class cluster centre, which is classified as the cluster.
(4) the removal cluster small numbers of classification of medium wave band.
When some in all cluster SC clusters interior wave band number sciWhen < Tn, cancel the cluster, and deletes in its cluster Heart Sci, while judging the number of iterations, if current iteration number reaches the maximum number of iterations It of permission, terminate, if not up to Then return to step 3.Until the number of samples of all clusters is all satisfied sci> Tn, then execute step 5, and Tn is that every class allows at least Sample number.
(5) cluster centre is corrected.
The sample mean of each cluster after step 4 is handled is calculated, and as new cluster centre.
Sci=NSciFormula (4)
Wherein, NSciFor new cluster centre, sciFor the wave band number in ith cluster, SCiFor ith cluster, x is Cluster SCiIn wavelength spectroscopy data, SciFor updated ith cluster center.
(6) average distance in each cluster SC between sample and each cluster centre is calculated.
Wherein,For the average distance between ith cluster medium wavelength data and each new cluster centre, sciFor ith cluster In wave band number, x be cluster SCiIn wavelength spectroscopy data, SciFor ith cluster center.
(7) whole wavelength datas and its overall average distance for corresponding to cluster centre are calculated.
Wherein,For whole wavelength datas and its overall average distance for corresponding to cluster centre, Nc is clusters number, sciIt is I cluster medium wave band number,For the average distance between ith cluster medium wavelength data and each cluster centre.
(8) if current iteration number reaches maximum number of iterations, terminate；Otherwise it then determines whether to meetIf It is to go to step 9, otherwise determines again；If current iteration number neither even number and meet Nc < 2C, then go to step 9, Division processing is carried out to having cluster, if current iteration number is even number or Nc >=2C, C is the class number of expectation cluster, then Step 10 is gone to, processing is merged.
(9) it divides.
Calculate the standard difference vector of sample distance in each cluster；
σi=(σ1iσ2i,…,σbi…,σni)TFormula (7)
Wherein T indicates that vector transposition, each component of vector are
Wherein, b=1,2 ... n indicate b-th of spectrum samples, i=1,2 ... k ..., Nc, σiFor in ith cluster Standard difference vector, σ1i,σ2i,…,σniFor standard deviation of each spectrum samples in ith cluster, sciFor ith cluster medium wave Number of segment mesh.Seek each standard difference vector { σi, i=1,2 ..., Nc } in largest component, with { σimax, i=1,2 ..., Nc } generation Table.In any one cluster set { σimax, i=1,2 ..., Nc } in, σ if it existsimax> Ts, while meeting two following conditions again One of:
1)And Ni> 2 (Tn+1)
2)
Ts is each component standard difference upper limit in T class, and Tn is that every class allows minimum sample number, sample distance in same Clustering Domain The standard deviation of distribution is then by SciTwo new cluster centres are split into, and Nc adds 1.One of them new cluster centre is SciIt is right The σ answeredimaxPunishment amount adds k times of σimax, the other is SciCorresponding σimaxPunishment amount subtracts k times of σimax, k is times of definition Number.
Otherwise third step is returned.
(10) union operation.
Calculate the distance of whole cluster centres between any two；
Dij=| | Sci-Scj| | formula (9)
Wherein, DijFor the distance between ith cluster center and j-th of cluster centre.Compare DijIt, will with the value of Td The value of Dij < Td is pressed apart from incremental arrangement, and Td is the minimum range lower limit between two cluster centres, i.e.,
Wherein
The above conditions (9)~(11) cluster centre will be met to merge, obtain new merging center, such as following formula (12):
Sc*For merge after new center,WithIt is the number in the i-th class and jth class cluster,WithIt is The cluster centre of i class and jth class cluster.The cluster centre logarithm merged every time is no more than L, each each iteration of cluster centre Can only at most it merge once, L is the most logarithms for the cluster centre that can merge in an iteration operation.
It is on the contrary then return to third step.
(11) if it is last time interative computation (i.e. I t times), then terminate；Otherwise, it if changing input parameter, goes to The first step；If inputting parameter constant, second step is gone to.
3.2 JM distances
Index by JM distance as sensitive band feature and Sensitivity Index feature selecting measures two due to JM distance The method of classification sample distance faces multi-class sample, traverses it all every two classes samples to each feature in the present invention JM distance, and average as measuring sensibility of some feature in all categories.In sensitive band selection, find out adaptive Answer in cluster result that for JM apart from biggish value as sensitive band, index characteristic then directly selects JM apart from maximum in every class 10 features are as sensitive features.JM distance represents the probability distribution distance of feature, and the value is bigger, illustrates this feature between two classes Separability it is stronger.
The JM distance between two classifications is sought, classification here refers to two kinds of floristics or the other classification of state area, and upper State the class declaration difference in wave band dimension cluster (i.e. clustering object is feature dimensions, and JM distance analysis object is sample dimension).Class Other ωiWith classification ωjBetween JM distance definition are as follows:
Wherein r is the feature vector of dimension k, and p (r | ωi) and p (r | ωj) be r two class conditional probability distributions.Work as p (r|ωi)) and p (r | ωj) when being Gaussian Profile, JM distance can simplify are as follows:
Wherein
It is wiAnd wjBetween Bhattacharyya distance.Herein, μiAnd μjIt is mean value in class, ∑IAnd ∑JIt is class association Variance matrix.
Step 4: appropriate model is determined
Using the KNN algorithm based on arest neighbors, is combined and supported based on the genetic algorithm of statistical mathematics method and optimisation technique Vector machine algorithm (GA-SVM), random forest (RF) scheduling algorithm based on probabilistic model, respectively to sensitive band, index characteristic (including differential characteristics, continuum feature and vegetation index) and whole feature (wave band adds index) three classes carry out modeling analysis, Obtain model accuracy；Overall classification accuracy (OAA) and Kappa coefficient are chosen as precision index, precision index maximum is regarded as It is suitble to model.Wherein,
In formula, rowFor the quantity of the row of confusion matrix crosstab；xiiFor the quantity of the type combination on diagonally； xi+For total observation number of row i；x+iFor total observation quantity of column；N is the total quantity of cell.
The present invention collects data using standardized collecting method, proposes the complete Spectra feature extraction of complete set Method, feature sensitivity analysis method, method for establishing model simultaneously carry out plant classification and status monitoring research, wherein proposing wave band Self-adaption cluster method can effectively remove the information redundancy in adjacent or close wave band, effectively improve data service efficiency.Based on upper Method is stated, plant spectral library of spectra is established, spectrum inquiry and Spectral matching, plant classification and status monitoring function containing plant Energy.
Detailed description of the invention
Fig. 1 is flow chart of the present invention；
Fig. 2 is wave band self-adaption cluster flow chart；
Fig. 3 is industrial crops canopy spectra curve and picture；
Fig. 4 is GA-SVM precision figure (confusion matrix is converted into accurate rate matrix), and (a) is corresponding " sensitive band feature ", (b) Corresponding " Sensitivity Index feature ", (c) corresponding " sensitive band+Sensitivity Index feature ").
Specific embodiment
Below in conjunction with attached drawing, the invention will be further described.
A kind of method for building up in plant spectral library, as shown in Figure 1, being specifically:
Step 1: data are obtained
The acquisition of plant spectral data.The measurement of plant high-spectral data is according to plant canopy hyperspectral measurement mark in this research Standard, in May, 2017 to September, the multiple testing sites of selection are carried out in Hangzhou, Zhejiang province city (Lon120.34 °, 30.31 ° of Lat), It include cowpea by ASD FieldSpec4 Pro FR (350nm~2500nm) type spectrometer collection farmland plant, by soybean, kind Potato, peanut, amaranth, eggplant, sesame, citrus, peach, Camellia yuhsienensis, ginkgo, shaddock, loquat, tea tree, rice, 16 kinds of plants such as wheat, Totally 450 spectrum, the plant canopy curve of spectrum and picture such as Fig. 3.
Step 2: Spectra feature extraction
It includes removing all original spectrum wave bands of steam band that ASD high-spectral data, which extracts and completes feature, 9 differential characteristics, 3 continuum features, 27 vegetation index features (being specifically shown in summary of the invention step 2).
Step 3: feature sensitivity analysis
Spectral band feature and Spectroscopy differential, continuum and vegetation index are selected using the methods of ISODATA, JM distance Feature (hereinafter referred to as this three category feature is index characteristic).Fig. 1 is wave band self-adaption cluster flow chart.
Wave band is clustered and obtains 9 classifications, takes JM wave band at maximum value in each classification to be characterized wave band, in order to go Except information redundancy between wave band and JM are apart from lesser wave band, wave of the JM at smaller value in adjacent or close wave band is removed Long feature obtains 394nm, 696nm, 731nm, 1037nm, 1505nm, 1585nm, 2020 nm, 2130nm, 2224nm etc. 9 altogether A wave band and Width, RMV, ATSAVI, PVIhyp, WI, SIWSI, NDII, NRI, sLAIDI, 10 index characteristics such as NDWI.
Step 4: suitable model is selected
According to total nicety of grading (OAA) and Kappa coefficient, can obtaining three kinds of classification method niceties of grading, it is as shown in the table (table).GA-SVM algorithm nicety of grading highest in industrial crops can be obtained by table, confusion matrix is as shown in Figure 4 (by confusion matrix Matrix is converted accurate rate matrix by total sample number ratio calculated in interior all classification samples frequencies and each prediction class categories), Therefore using the method as library of spectra farmland scene plant classification application method.
2 each method nicety of grading of table
Above-described embodiment is not for limitation of the invention, and the present invention is not limited only to above-described embodiment, as long as meeting The present invention claims all belong to the scope of protection of the present invention.

Claims (2)

1. a kind of method for building up in plant spectral library, it is characterised in that the following steps are included:
Step 1: data are obtained
Plant canopy high-spectral data and mating information are obtained, wherein mating information includes floristics and status information；
Step 2: Spectra feature extraction
The plant canopy high-spectral data obtained according to step 1 and mating information extraction spectral signature, as plant classification and shape State monitoring feature, wherein spectral signature includes original wave band feature, first differential feature, continuum feature and vegetation index；
Step 3: feature sensitivity analysis
For original wave band feature, original spectrum wave band is clustered first with adaptive wave band clustering procedure, by correlation compared with High wavelength cluster, recycles JM distance to be analyzed for each cluster, chooses in each cluster JM apart from biggish wave It is long to be used as sensitive band feature；
For index characteristic (referring to first differential feature, continuum feature, vegetation index feature), using JM distance as each The sensibility index of index characteristic obtains JM apart from maximum 10 and is used as Sensitivity Index feature；
It is above-mentioned to be clustered original spectrum wave band specifically using adaptive wave band clustering procedure:
(1) it determines predefined parameter, can be used for determining the clusters number in subsequent analysis；
C, it is expected that the class number of cluster；
Tn, every class allow minimum sample number
Ts is each component standard difference upper limit in class；
Td, the minimum range lower limit between two cluster centres, if being less than this number, two clusters need to be merged；
L, the most logarithms for the cluster centre that can merge in an iteration operation
It, the maximum number of iterations of permission
Nc, initial clustering number, may be the same or different with C；
(2) cluster centre is randomly generated；According to originally determined clusters number Nc, the initial cluster center of corresponding number is generated； It being concentrated on when to prevent cluster centre from obtaining at random near the wavelength of certain, classification center being obtained using layering, first presses wavelength It is divided into Nc layers according to initial clustering number, then obtains cluster centre；According to the acquisition side for being layered equally distributed stochastic clustering center Method is as follows:
Wherein, Rc is wave spectrum at the corresponding cluster centre of initial clustering number N c, and i is ith cluster center, and Ls is wave band Sum, Nc are initial setting clusters number, and rand is the random number between 0~1,For by all wavelengths as required initial Clusters number layering is rounded,To be used as cluster centre at some random wavelength rounding in each layer；
(3) each wavelength data x is traversed according to range formula (2), x is categorized into nearest cluster centre Scj, obtain in cluster Heart ScjThe cluster SC at placej；
Dj=min | | x-Sci| |, i=1,2 ..., j ..., Nc } formula (2)
X is wavelength spectroscopy data, SciFor ith cluster center, DjIt indicates in a certain wavelength data and all cluster centres distance Distance value at minimum, the center are jth class cluster centre, which is classified as the cluster；
(4) the removal cluster small numbers of classification of medium wave band；
When some in all cluster SC clusters interior wave band number sciWhen < Tn, cancel the cluster, and deletes its cluster centre Sci, while judging the number of iterations, if current iteration number reaches the maximum number of iterations It of permission, terminate, if not up to Return to step 3；Until the number of samples of all clusters is all satisfied sci> Tn, then execute step 5, and Tn is that every class allows minimum sample This number；
(5) cluster centre is corrected；
The sample mean of each cluster after step 4 is handled is calculated, and as new cluster centre；
Sci=NSciFormula (4)
Wherein, NSciFor new cluster centre, sciFor the wave band number in ith cluster, SCiFor ith cluster, x is cluster SCiIn wavelength spectroscopy data, SciFor updated ith cluster center；
(6) average distance in each cluster SC between sample and each cluster centre is calculated；
Wherein,For the average distance between ith cluster medium wavelength data and each new cluster centre, sciFor in ith cluster Wave band number, x are cluster SCiIn wavelength spectroscopy data, SciFor ith cluster center；
(7) whole wavelength datas and its overall average distance for corresponding to cluster centre are calculated；
Wherein,For whole wavelength datas and its overall average distance for corresponding to cluster centre, Nc is clusters number, sciIt is i-th Medium wave band number is clustered,For the average distance between ith cluster medium wavelength data and each cluster centre；
(8) if current iteration number reaches maximum number of iterations, terminate；Otherwise it then determines whether to meetIf then turning To step 9, otherwise determine again；If current iteration number neither even number and meet Nc < 2C, then go to step 9, to existing Cluster carries out division processing, if current iteration number is even number or Nc >=2C, C be the class number that expectation clusters, then goes to the 10 steps merge processing；
(9) it divides；
Calculate the standard difference vector of sample distance in each cluster；
σi=(σ1iσ2i..., σbi..., σni)TFormula (7)
Wherein T indicates that vector transposition, each component of vector are
Wherein, b=1,2 ... n, b-th of spectrum samples of expression, i=1,2 ... k ..., Nc, σiFor the mark in ith cluster Quasi- difference vector, σ1i, σ2i..., σniFor standard deviation of each spectrum samples in ith cluster, sciFor ith cluster medium wave Number of segment mesh；Seek each standard difference vector { σi, i=1,2 ..., Nc } in largest component, with { σimax, i=1,2 ..., Nc } and generation Table；In any one cluster set { σimax, i=1,2 ..., Nc } in, σ if it existsimax> Ts, while meeting following two items again One of part:
1)And Ni> 2 (Tn+1)
2)
Ts is each component standard difference upper limit in T class, and Tn is that every class allows minimum sample number, sample range distribution in same Clustering Domain Standard deviation then by SciTwo new cluster centres are split into, and Nc adds 1；One of them new cluster centre is SciIt is corresponding σimaxPunishment amount adds k times of σimax, the other is SciCorresponding σimaxPunishment amount subtracts k times of σimax, k is the multiple of definition；
Otherwise third step is returned；
(10) union operation；
Calculate the distance of whole cluster centres between any two；
Dij=| | Sci-Scj| | formula (9)
Wherein, DijFor the distance between ith cluster center and j-th of cluster centre；Compare DijWith the value of Td, Td is two poly- Apart from incremental arrangement, i.e., minimum range lower limit between class center presses the value of Dij < Td
Wherein
The above conditions (9)-(11) cluster centre will be met to merge, obtain new merging center, such as following formula (12):
Sc*For merge after new center,WithIt is the number in the i-th class and jth class cluster,WithFor the i-th class and The cluster centre of jth class cluster；For the cluster centre logarithm merged every time no more than L, each each iteration of cluster centre is most It can only merge once, L is the most logarithms for the cluster centre that can merge in an iteration operation；It is on the contrary then return to third Step；
(11) if it is last time interative computation (i.e. I t times), then terminate；Otherwise, if changing input parameter, first is gone to Step；If inputting parameter constant, second step is gone to；
Step 4: appropriate model is determined
Respectively to sensitive band, index characteristic (including differential characteristics, continuum feature and vegetation index) and whole feature (waves Section plus index) three classes progress modeling analysis, obtain model accuracy；Overall classification accuracy (OAA) and Kappa coefficient are chosen as essence Index is spent, precision index maximum is regarded as being suitble to model；
In formula, rowFor the quantity of the row of confusion matrix crosstab；xiiFor the quantity of the type combination on diagonally；xi+For Total observation number of row i；x+iFor total observation quantity of column；N is the total quantity of cell.
2. a kind of method for building up in plant spectral library as described in claim 1, it is characterised in that the calculating of JM distance is specifically:
Classification ωiWith classification ωjBetween JM distance definition are as follows:
Wherein r is the feature vector of dimension k, and p (r | ωi) and p (r | ωj) be r two class conditional probability distributions；When p (r | ωi)) and p (r | ωj) when being Gaussian Profile, JM distance can simplify are as follows:
Wherein
BI, jIt is wiAnd wjBetween Bhattacharyya distance, μiAnd μjIt is mean value in class, ∑IAnd ∑JIt is class covariance matrix.
CN201910304011.1A 2019-04-16 2019-04-16 Method for establishing plant spectrum library Active CN110390028B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910304011.1A CN110390028B (en) 2019-04-16 2019-04-16 Method for establishing plant spectrum library

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910304011.1A CN110390028B (en) 2019-04-16 2019-04-16 Method for establishing plant spectrum library

Publications (2)

Publication Number Publication Date
CN110390028A true CN110390028A (en) 2019-10-29
CN110390028B CN110390028B (en) 2021-08-10

Family

ID=68284943

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910304011.1A Active CN110390028B (en) 2019-04-16 2019-04-16 Method for establishing plant spectrum library

Country Status (1)

Country Link
CN (1) CN110390028B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111488822A (en) * 2020-04-09 2020-08-04 北华航天工业学院 Tree species information identification method based on full spectrum segment correlation analysis algorithm

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1209627A2 (en) * 2000-11-24 2002-05-29 Canadian Space Agency Vector quantization method and apparatus
CN106124049A (en) * 2016-06-20 2016-11-16 福州大学 A kind of implementation method of Vegetation canopy multi-optical spectrum imaging system
CN107729916A (en) * 2017-09-11 2018-02-23 湖南中森通信科技有限公司 A kind of interference source classification and identification algorithm and device based on ISODATA
CN108286962A (en) * 2018-01-31 2018-07-17 中国科学院遥感与数字地球研究所 A kind of method for building up and system of water environment library of spectra
CN108732133A (en) * 2018-04-12 2018-11-02 杭州电子科技大学 It is a kind of based on the plant disease of optical image technology in body nondestructive detection system

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1209627A2 (en) * 2000-11-24 2002-05-29 Canadian Space Agency Vector quantization method and apparatus
CN106124049A (en) * 2016-06-20 2016-11-16 福州大学 A kind of implementation method of Vegetation canopy multi-optical spectrum imaging system
CN107729916A (en) * 2017-09-11 2018-02-23 湖南中森通信科技有限公司 A kind of interference source classification and identification algorithm and device based on ISODATA
CN108286962A (en) * 2018-01-31 2018-07-17 中国科学院遥感与数字地球研究所 A kind of method for building up and system of water environment library of spectra
CN108732133A (en) * 2018-04-12 2018-11-02 杭州电子科技大学 It is a kind of based on the plant disease of optical image technology in body nondestructive detection system

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
PU HONG等: ""An improved hyperspectral image classification approach based on ISODATA and SKR method"", 《PROCEEDINGS OF THE SPIE 10030,INFRARED,MILIMETER-WAVE,AND TERAHERTZ TECHNOLOGIES IV》 *
赵春晖等: ""高光谱遥感图像最优波段选择方法的研究进展与分析"", 《黑龙江大学自然科学学报》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111488822A (en) * 2020-04-09 2020-08-04 北华航天工业学院 Tree species information identification method based on full spectrum segment correlation analysis algorithm

Also Published As

Publication number Publication date
CN110390028B (en) 2021-08-10

Similar Documents

Publication Publication Date Title
Elavarasan et al. Forecasting yield by integrating agrarian factors and machine learning models: A survey
Huang et al. Development of soft computing and applications in agricultural and biological engineering
da Silva Junior et al. Soybean varieties discrimination using non-imaging hyperspectral sensor
EP2430500B1 (en) Inverse modeling for characteristic prediction from multi-spectral and hyper-spectral remote sensed datasets
De Castro et al. Applying neural networks to hyperspectral and multispectral field data for discrimination of cruciferous weeds in winter crops
Ilic et al. Early cherry fruit pathogen disease detection based on data mining prediction
CN110991064B (en) Soil heavy metal content inversion model generation method, system and inversion method
Bendel et al. Evaluating the suitability of hyper-and multispectral imaging to detect foliar symptoms of the grapevine trunk disease Esca in vineyards
LÓPEZ‐GRANADOS et al. Multispectral classification of grass weeds and wheat (Triticum durum) using linear and nonparametric functional discriminant analysis and neural networks
Corrales et al. A surrogate model based on feature selection techniques and regression learners to improve soybean yield prediction in southern France
Chivasa et al. Phenology-based discrimination of maize (Zea mays L.) varieties using multitemporal hyperspectral data
Diniz et al. Forecasting frost risk in forest plantations by the combination of spatial data and machine learning algorithms
Tosin et al. Canopy VIS-NIR spectroscopy and self-learning artificial intelligence for a generalised model of predawn leaf water potential in Vitis vinifera
Meng et al. Fine hyperspectral classification of rice varieties based on self-attention mechanism
Dantas et al. Reduction of sampling intensity in forest inventories to estimate the total height of eucalyptus trees
Mangalraj et al. Recent trends and advances in hyperspectral imaging techniques to estimate solar induced fluorescence for plant phenotyping
CN110390028A (en) A kind of method for building up in plant spectral library
Khorrami et al. A fuzzy multi-criteria decision-making approach for the assessment of forest health applying hyper spectral imageries: A case study from Ramsar forest, North of Iran
Bagherian et al. Phenotyping agronomic and physiological traits in peanut under mid‐season drought stress using UAV‐based hyperspectral imaging and machine learning
Habibie et al. A Multicriteria Index using Neural Network to Evaluate the Potential Lands of Maize
Tussupov et al. Analysis of formal concepts for verification of pests and diseases of crops using machine learning methods
Khazaei et al. Yield estimation and clustering of chickpea genotypes using soft computing techniques
CN101726742A (en) Remote sensing measurement method for level of crops pollution stress
CN117077868A (en) Model-based grain yield prediction system and method
Santhosh et al. A Compendium Probabilistic Prospective for Predicting Coffee Crop Yield Based on Agronomical Factors

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant