CN106203523B - The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient - Google Patents
The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient Download PDFInfo
- Publication number
- CN106203523B CN106203523B CN201610561589.1A CN201610561589A CN106203523B CN 106203523 B CN106203523 B CN 106203523B CN 201610561589 A CN201610561589 A CN 201610561589A CN 106203523 B CN106203523 B CN 106203523B
- Authority
- CN
- China
- Prior art keywords
- sample point
- sample
- sample points
- semi
- class
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000003066 decision tree Methods 0.000 title claims abstract description 16
- 230000004927 fusion Effects 0.000 title claims abstract description 9
- 238000012549 training Methods 0.000 claims abstract description 47
- 238000012216 screening Methods 0.000 claims abstract description 7
- 239000013598 vector Substances 0.000 claims description 29
- 238000012360 testing method Methods 0.000 claims description 27
- 230000003595 spectral effect Effects 0.000 claims description 10
- 238000002372 labelling Methods 0.000 claims description 9
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000012545 processing Methods 0.000 abstract description 3
- 238000004088 simulation Methods 0.000 description 8
- 238000001228 spectrum Methods 0.000 description 5
- 238000003384 imaging method Methods 0.000 description 4
- 238000012706 support-vector machine Methods 0.000 description 4
- 235000008331 Pinus X rigitaeda Nutrition 0.000 description 3
- 235000011613 Pinus brutia Nutrition 0.000 description 3
- 241000018646 Pinus brutia Species 0.000 description 3
- 244000025254 Cannabis sativa Species 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000000701 chemical imaging Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000003331 infrared imaging Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000007786 learning performance Effects 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000005855 radiation Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
- G06F18/2113—Selection of the most significant subset of features by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention proposes a kind of hyperspectral image classification methods that the semi-supervised algorithm fusion of decision tree is promoted based on gradient, for solving the lower technical problem of nicety of grading present in the existing classification hyperspectral imagery combined based on Active Learning with semi-supervised learning, step includes: (1) input hyperspectral image data；(2) sample point feature is extracted；(3) training gradient promotes decision tree classifier parameter；(4) sample point classification is concentrated to study；(5) sample pixel confidence is assessed；(6) pass through rarefaction representation Screening Samples point；(7) updating has label training set；(8) output category result.The present invention assesses the confidence level of unmarked sample point using classifier prediction result and rarefaction representation, according to the height of unmarked sample pixel confidence, it is divided into two set and carries out different processing, the burden that handmarking is alleviated while improving nicety of grading, can be used for the fields such as geologic survey, atmosphere pollution.
Description
Technical Field
The invention belongs to the technical field of image processing, relates to a hyperspectral image classification method, and particularly relates to a hyperspectral image classification method based on gradient lifting decision tree semi-supervised algorithm fusion, which can be used in the fields of geological survey, atmospheric pollution, military target strike and the like.
Background
With the development of optical remote sensing technology, the remote sensing imaging process is from panchromatic (black and white) image, color photography, multispectral scanning imaging to today's hyperspectral remote sensing imaging and hyperspectral imaging. The hyperspectral remote sensing technology adopts 10-2The lambda continuous spectrum channel carries out continuous remote sensing imaging on the ground features, a large amount of ground feature image data with complete spectrum information are obtained, synchronous obtaining of ground feature space information, radiation information and spectrum information is achieved, the characteristic of 'map integration' is achieved, and convenience is brought to ground feature identification.
Commonly used hyperspectral image data include Indian pin dataset and Kennedy Space Center (KSC) dataset obtained by AVIRIS, an airborne visible light/infrared imaging spectrometer of NASA jet propulsion laboratory, NASA, and Botswana dataset obtained by EO-I hyper spectrometer of NASA, among others.
The hyperspectral image ground object classification problem is mainly that the ground object is classified by utilizing the spectral features of the ground object, the spectral form of each pixel content in the hyperspectral image is analyzed, and the category of the hyperspectral image ground object is judged according to the features of the content. The traditional hyperspectral image classification method mainly comprises a supervised classification method represented by a Support Vector Machine (SVM) and a neural network and an unsupervised classification method represented by a fuzzy clustering method. The supervised classification method needs a large number of labeled samples to train to obtain a classifier with good performance, the training data set of the hyperspectral remote sensing image classification problem is sample points labeled with class labels on the remote sensing images, and the class labels for labeling the sample points are all completed manually. However, requiring human experts to manually label hyperspectral images is a time-consuming, labor-intensive and costly task; for the unsupervised classification method, due to the lack of prior knowledge, the sample is divided into a plurality of classes only according to the spectral feature distribution rule of the ground features of the remote sensing image, the classification result only distinguishes different classes, the attribute of the class cannot be determined, and the correct correspondence between the clustered classes and the ground feature classes cannot be ensured.
Under the circumstances, a hyperspectral image classification method based on semi-supervised learning and active learning attracts wide attention of scholars at home and abroad. The semi-supervised learning trains the initial classifier by using a small amount of labeled data, and further improves the performance of the initial classifier by using a large amount of unlabelled data to achieve accurate learning, thereby making up the defects of supervised learning and unsupervised learning to a certain extent. Common semi-supervised classification methods include self-training methods, collaborative training, probabilistic model generation algorithms, semi-supervised Support Vector Machines (SVM), and graph-based methods. In these methods, a class label is assigned to unlabeled data, and the classifier is retrained using the data from which the class label is obtained, to obtain a final classification result. However, semi-supervised learning has the disadvantages that under the conditions of less samples and insufficient model training, the class label prediction of unlabeled data is often inaccurate, and adding the incorrectly labeled samples into the training set will result in the degradation of the learning performance of the classifier. Active learning aims to select samples which are valuable to a classification model through a certain query strategy and filter redundant sample information, so that the samples with abundant information are manually marked according to the knowledge and experience of field experts. The main task of active learning is to find an efficient sample query strategy, so that the quality of selectively marked samples is high and few, the classification performance can be ensured, and the workload of marking samples can be reduced. The current query strategies for active learning are as follows: 1) sampling based on sample uncertainty; 2) based on the sampling of the query expert committee, a committee is formed by a plurality of classifiers, and whether samples are selected or not is determined by voting. In active learning, the accuracy of the labeling can be guaranteed by experts labeling unlabeled samples, but manual labeling tends to be time consuming and labor intensive.
Active learning ensures one hundred percent accuracy by consulting human experts to introduce artificially labeled samples. The number of samples that can be manually marked is limited due to the time and effort involved in manual marking. The semi-supervised learning relies on a classifier to predict unmarked samples, and the quality cannot be guaranteed when the number of newly added samples is large. According to the characteristics of the two methods, scholars at home and abroad consider the combination of the two methods, a hyperspectral image classification method based on the combination of active learning and semi-supervised learning is provided, and the burden of manual marking is reduced while the number of newly added marked samples is ensured. For example, in a paper "a New Semi-supervised-assisted for Hyper-spectral image classification With differential Active Learning" (white pers, 2012), Inmaculada D' opino, Jun Li et al discloses a Semi-supervised Active Learning method for Hyper-spectral image classification, and a query strategy of Active Learning is used to screen unlabelled samples selected in the Semi-supervised Learning process, so as to select the sample With the most abundant information content. The method comprises the following specific steps: calculating the maximum posterior probability of unmarked samples in the neighborhood of the marked samples in a sparse polynomial logistic regression classifier; adding the given class mark with higher probability into a certain specific set; selecting samples in the set by using several common query strategies in active learning, and selecting the sample which contributes most to the performance improvement of the classifier; and adding the selected samples into the marked sample set, and retraining the classifier. The method saves time and labor, but due to the lack of a manual marking process, only the classifier is relied on for class marking prediction, and the classification precision needs to be improved.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides a hyperspectral image classification method based on gradient lifting decision tree semi-supervised algorithm fusion.
In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:
(1) inputting a hyperspectral image containing C-type and N sample points, taking a neighborhood window of each sample point, taking the maximum value of each dimensional feature of all the sample points in the window as the spatial feature of the central sample point, and connecting the spectral feature and the spatial feature of the sample points in series to obtain a spatial spectral feature vector of the sample points;
(2) selecting a marked training set, a learning set and a testing set from an input hyperspectral image, and realizing the steps as follows:
(2a) randomly selecting r sample points from each type of sample points of the input hyperspectral image to obtain a labeled training setIts corresponding class label set isWherein n is the total number of marked training sample points, and n is C × r, xiFor the ith labeled sample point of the labeled training set,/iFor the class to which the ith labeled training sample point belongs, liE.g., {1,2,. and C }, wherein R is a real number domain, and D is a characteristic dimension of the sample point;
(2b) randomly selecting sample points with per1 ratio from the selected sample points except n marked sample points to obtain a learning setWhere s is the total number of learning set sample points, s ═ N × per1, zqIs the q sample point in the learning set;
(2c) constructing a test set using the remaining sample pointsm is the total number of samples in the test set, and m is N-N-s, yjIs the jth test sample point of the test set;
(3) using labeled training setsTraining parameters of a GBDT classifier of the gradient lifting decision tree by using the feature vectors of the sample points and the corresponding class mark matrix, wherein every two classes of marked sample points can be trained to obtain a two-classifier model, and finally, the C class of marked sample points can obtain C x (C-1)/2 two-classifier models;
(4) study setThe sample points in the learning set Z are input into a plurality of obtained two-classifier models to obtain each sample point in the learning set ZThe prediction class mark k;
(5) according to the obtained learning setAt each sample point zqThe predicted class mark k of (2) judges each sample point z in each two-classifier modelqWhen the class mark is classified into the kth class, whether the winning times P of the class mark k is equal to C-1 or not is judged, if yes, the sample point is added into the empty set SsemiOtherwise, adding the sample point into the empty set SactPerforming the following steps; judging all sample points in the learning set Z one by one to obtain a setAnd collectionsWherein z isq1Is a set SsemiSample point of (1), zq2Is a set SactS' is the set SsemiThe total number of sample points in (1), S' is the set SactThe total number of sample points in (1), wherein s' + s ═ s;
(6) using sparse representation, the resulting set S issemiAnd set SactThe sample points in (1) are screened, and the implementation steps are as follows:
(6a) constructing dictionary A ═ X by using all sample points in labeled training set X1,x2,…,xn]And respectively aligning the sets S by using the constructed dictionary AsemiSample point z inq1And set SactSample point z inq2Performing sparse representation: z is a radical ofq1＝Aα1，zq2＝Aα2Wherein, α1And α2Is a sparse representation coefficient vector;
(6b) obtaining a sample point z by using an orthogonal matching pursuit algorithm OMPq1And sample point zq2Represents the coefficient vector:andwherein | · | purple2Is 12Norm, measure data reconstruction error; i | · | purple wind1Is 11Norm for guaranteeing vector α1Sum vector α2λ is a balance factor of the reconstruction error term and the sparse term;
(6c) representing coefficient vectors α from sparseness1And α2Class labels with labeled sample points corresponding to non-zero entries, i.e./, iniE {1, 2.., C }, and set SsemiMiddle prediction class mark k and class mark liSame sample point zq1Screening out and assigning the class labels of all the screened sample points to class labels li(ii) a At the same time set SactMiddle prediction class mark k and class mark liDifferent sample points zq2Screening out, and handing all the screened sample points to an expert for manual labeling;
(7) will gather SsemiMiddle-assigned class label liSample point z ofq1And set SactSample point z in which manual labeling is performedq2Adding the new classifier model into a labeled training set X, and retraining the classifier parameters to obtain a new classifier model;
(8) iterating the steps (3) to (7) until the set iteration times are met, and carrying out comparison on the test set by utilizing the finally obtained classifier modelThe sample points in the test set are classified to obtain the classification result of the test set
Compared with the prior art, the invention has the following advantages:
1. according to the invention, the confidence coefficient of the unmarked sample points is evaluated by adopting the classifier prediction result and sparse representation, and meanwhile, the two sets are divided according to the confidence coefficient of the unmarked sample points, and different processing is carried out according to the characteristics of the two sets.
2. The invention adopts the artificial mark and the unmarked sample points predicted by the classifier to update the marked training set, and simultaneously utilizes the marked sample points and the unmarked sample points to train the classifier, thereby effectively reducing the number of the required marked sample points, ensuring the classification accuracy and simultaneously lightening the burden of the artificial mark.
Drawings
FIG. 1 is a block diagram of an implementation flow of the present invention;
FIG. 2 is a simulation comparison diagram of classification accuracy when the number of marked training sample points is different between the present invention and the prior art.
Detailed Description
The invention is further illustrated below with reference to the accompanying drawings and examples.
Referring to fig. 1, the method of the present invention includes the following steps:
step 1, inputting hyperspectral image data:
and inputting a hyperspectral image, removing background sample points, wherein the number of the remaining sample points is N and comprises C categories.
Step 2, extracting the spatial spectrum characteristics of the sample points, wherein the implementation steps are as follows:
and 2a, using the spectral characteristic value of each wave band of each sample point as the spectral characteristic vector of the sample point, wherein the original characteristic dimension of the sample point is d.
And 2b, taking a neighborhood window of each sample point, wherein the window size is c multiplied by c, taking the maximum value of each dimensional feature of all the sample points in the window as the spatial feature of the central sample point, and the feature dimension is d.
And 2c, connecting the spectral features and the spatial features of the sample points in series to obtain a final feature vector, wherein the feature dimension is D, and D is 2 × D.
And 3, selecting a marked training set X, a test set Y and a learning set Z from the input hyperspectral image, and realizing the steps as follows:
step 3a, randomly selecting r sample points from each type of sample points of the input hyperspectral image to form a labeled training setIts corresponding class label set isWherein n is the total number of marked training sample points, and n is C × r, xiFor the ith labeled sample point of the labeled training set,/iFor the class to which the ith labeled training sample point belongs, liE.g. {1,2,. and C }, wherein R is a real number field;
step 3b, randomly selecting sample points with the proportion per1 from the selected sample points except the n marked sample points to form a learning setWherein s is the total number of the learning set sample points, and s is (N-N) x per1, zqIs the q sample point in the learning set;
step 3c, forming a test set by using the residual sample pointsm is the total number of samples in the test set, and m is N-N-s, yjIs the jth test sample point of the test set;
step 4, training parameters of a GBDT classifier, and performing class mark prediction on sample points in a learning set, wherein the implementation steps are as follows:
step 4a, inputting a marked training setThe feature vectors of the middle sample points and the corresponding class mark matrixes are input into a GBDT classifier, and classifier parameters are trained;
step 4b, inputting the learning setObtaining a sample point z by adding the feature vector of the middle sample point into the obtained classifier modelqA corresponding class label k;
step 5, dividing the sample points into two sets according to the confidence of the sample points in the learning set, and realizing the steps as follows:
step 5a, sample zqObtaining a prediction result value score (k) and score (t) through classification of the obtained two classifiers, wherein the two classifiers are obtained by training the kth class marked sample point and the tth class marked sample point, k belongs to {1, 2.. multidot., (C }, t belongs to {1, 2.. multidot., (C }, k is not equal to t, score (k) and score (t) are obtained through the classifiers on a sample zqThe predicted result values of the kth class and the t-th class of (1);
step 5b, sample zqThe winning times P of the class k obtained based on the two classifiers is
Wherein,
step 5C, if P ═ C-1, denotes sample zqThe confidence of the true class of (1) marked as k is higher; the main purpose of semi-supervised learning is to find out easy-to-mark high-confidence unmarked sample points, make class mark prediction by using classifier model, add them into marked training set, so that it uses z as referenceqPut into the empty set SsemiIn (1), get the setzq1Is a set SsemiS' is the set SsemiThe total number of sample points in (1);
step 5d, if P ≠ C-1, it means that sample zqThe confidence of the true class of (1) labeled k is lower; in active learning, samples which are difficult to be distinguished and rich in information content are screened out for manual marking, so that z is markedqPut into the set SactIn (1), get the setzq2Is a set SactSample points in (1), S' is the set SactThe total number of sample points in (1);
step 6, for set SsemiAnd set SactThe sample point sparse representation in (1) is realized by the following steps:
step 6a, constructing a dictionary A, wherein A is [ x ═ x1,x2,…,xn]，x1,x2,…,xnThe sample points in the marked training set are represented, n is the total number of the marked training sample points, and the characteristic dimension of the sample points is D, so that the size of the dictionary is Dxn;
step 6b, for set SsemiSample point z inq1And set SactSample point z inq2Respectively carrying out sparse representation to obtain sparse representation zq1＝Aα1And zq2＝Aα2；
Step 6c, obtaining a sample point z by utilizing an Orthogonal Matching Pursuit (OMP) algorithmq1And sample point zq2Represents the coefficient vector:andwherein | · | purple2Is 12Norm, measure data reconstruction error; i | · | purple wind1Is 11Norm for guaranteeing vector α1Sum vector α2The method is characterized in that the sparsity of the method is realized by the following steps that lambda is a balance factor of a reconstruction error term and a sparse term:
step 6c1, initializing residual term r(0)＝zqIndex setFor a K-dimensional zero vector, the initial iteration J is 1
Step 6c2, finding the residual r(J-1)And column j x in dictionary AjThe index lambda corresponding to the maximum inner product,
step 6c3, updating index set Λ(J)，Λ(J)(J) λ. According to the index set, selecting corresponding atom columns from the dictionary A to form a set A(J)＝A(:,Λ(J)(1:J))；
Step 6c4, obtaining J-order approximation by least squares
Step 6c5, updating residual r(J)＝zq-A(J)α(J),J＝J+1；
And 6c6, repeating the steps 6c 2-6 c5, judging whether J is larger than K, if so, finishing iteration, and otherwise, executing the step 6c 2.
zqIs a set SsemiAnd Sactα is a sparse representation coefficient vector;
step 7, representing the coefficient vector α according to the sparseness1And α2Class label l with labeled sample points corresponding to positions of medium and non-zero itemsiE {1, 2.., C }, set S of pairssemiAnd set SactSample point z inq1And zq2And (5) screening.
Step 7a, when the jth dictionary atom x in the dictionary AjAnd study collectionQ (th) sample point zqWhen belonging to the same class, α corresponds to position αjiThe value is 1, and when different, 0; if set SsemiMiddle sample point zq1Prediction class k and its sparse coefficient matrix α1Class label l with labeled sample points corresponding to positions of medium and non-zero itemsiSame, denotes the sample point zq1If the marked sample point belongs to the same class, the sample point z is setq1Is given asi。
Step 7b, if set SactSample point z inq2Prediction class k and its sparse coefficient matrix α2Class label l with labeled sample corresponding to middle non-zero item positioniDifferent, represents the sample point zq2The class mark predicted by the classifier is inconsistent with the class mark obtained by sparse representation, and the sample point zq2And screening out the sample points which belong to the sample points difficult to be classified, and submitting the sample points to an expert for manual marking.
Step 8, adding SsemiSample points z in the set assigned class labelsq1And SactSample points z in the set for manual labelingq2Adding the new labeled training set into the labeled training set X and inputting the new labeled training setRetraining classifier parameters by using the feature vectors of the middle sample points and the corresponding class mark matrix to obtain a new classifier model；
Step 9, outputting the classification result
Using a gradient boosting decision tree classifier, a new labeled training set is input in the first stepFeature vector and class label set of middle sample pointsFor training, second step input test setObtaining a class mark matrix of the test set by the feature vector of the middle test sample through a gradient boosting decision tree classifierWherein l'jIndicating the class label to which the jth test sample belongs.
Step 10, calculating classification accuracy
And comparing the real class mark matrixes to obtain a classification precision result.
The technical effects of the present invention will be further described below with reference to simulation experiments.
1. Simulation conditions are as follows:
the simulation experiment is carried out by MATLAB 2014a software on a WINDOWS 7 system with a CPU of Intel Core (TM) i3-3110M, a master frequency of 2.40GHz and a memory of 4G.
2. Simulation content and analysis:
simulation experiments adopt Ind Pine images obtained by AVIRIS of NASA jet propulsion laboratory in 1992 6 in North Indiana by 145, which have 220 wave bands, and remove noise and absorbed by atmosphere and water, which have 200 wave bands and 16 types of ground object information, wherein the number of data of part of types is very small, and in the simulation experiments, only 9 types of data shown in Table 1 are considered, and the whole image is divided into 9 types.
TABLE 1 type 9 data in Indian Pine images
| Categories | Category name | Number of |
| 1 | Corn-no till | 1434 |
| 2 | Corn-min | 834 |
| 3 | Grass/Pasture | 497 |
| 4 | Grass/Trees | 747 |
| 5 | Hay-windrowed | 489 |
| 6 | Soybeans-no till | 968 |
| 7 | Soybeans-min | 2468 |
| 8 | Soybean-clean | 614 |
| 9 | Woods | 1294 |
The invention is used for classifying the high-spectrum Image Indian Pine with the prior art, and the prior art for comparison is a Semi-supervised Active Learning method provided in a paper "A New Semi-supervised applied for Hyper-spectral Image Classification with differential Active Learning" (WHISPERS, 2012). The method takes a gradient lifting decision tree GBDT as a classifier, and is abbreviated as SSAc + GBDT on the basis of a hyperspectral image classification method combining active learning and semi-supervised learning.
The number of decision trees of the GBDT classifier is set to be 100, and the downsampling proportion is set to be 50%; the window size c × c is set to 15 × 15, and the selection ratio per1 of the learning set is set to 30%.
A fixed number of sample points are selected from 9 types of data shown in Table 1 as a marked training set, a certain proportion of sample points are selected as a learning set, the rest of sample points are used as a test set, the learning set and the test set are unmarked sample points, 10 classification experiments are carried out on the 9 types of data by using the method and the prior art, the average value of classification results is taken as the final classification accuracy, as shown in figure 2, the method is a classification accuracy simulation comparison graph when the number r of the marked training sample points of each type is respectively 5, 10 and 15, the abscissa represents the number of the marked training sample points of each type, and the ordinate represents the classification accuracy. It can be seen from fig. 2 that when the number of marked sample points in each category is different, the classification accuracy of the invention is obviously higher than that of the prior art.
In conclusion, the method classifies the hyperspectral images by combining the semi-supervised algorithm fusion on the basis of the gradient lifting decision tree, fully utilizes the structural information of the unmarked sample points, can reduce the calculated amount and obtain higher classification precision, and has certain advantages compared with the existing method.
Claims (3)
1. A hyperspectral image classification method based on gradient lifting decision tree semi-supervised algorithm fusion comprises the following steps:
(1) inputting a hyperspectral image containing C-type and N sample points, taking a neighborhood window of each sample point, taking the maximum value of each dimensional feature of all the sample points in the window as the spatial feature of each sample point, and connecting the spectral feature and the spatial feature of each sample point in series to obtain a spatial spectral feature vector of each sample point;
(2) selecting a marked training set, a learning set and a testing set from an input hyperspectral image, and realizing the steps as follows:
(2a) randomly selecting r sample points from each type of sample points of the input hyperspectral image to obtain a labeled training setIts corresponding class label set isWherein n is the total number of marked training sample points, and n is C × r, xiFor the ith labeled sample point of the labeled training set,/iFor the class to which the ith labeled training sample point belongs, liE is e to {1,2, …, C }, R is a real number domain, and D is a characteristic dimension of the sample point;
(2b) randomly selecting sample points with per1 ratio from the selected sample points except n marked sample points to obtain a learning setWherein s is the total number of the learning set sample points, and s is (N-N) x per1, zqIs the q sample point in the learning set;
(2c) constructing a test set using the remaining sample pointsm is the total number of samples in the test set, and m is N-N-s, yjIs the jth test sample point of the test set;
(3) using labeled training setsTraining parameters of a GBDT classifier of the gradient lifting decision tree by using the feature vectors of the sample points and the corresponding class mark matrix, wherein every two classes of marked sample points can be trained to obtain a two-classifier model, and finally, the C class of marked sample points can obtain C x (C-1)/2 two-classifier models;
(4) study setInputting the sample points into the obtained two-classifier models to obtain a prediction class mark k of each sample point in the learning set Z;
(5) according to the obtained learning setAt each sample point zqThe predicted class mark k of (2) judges each sample point z in each two-classifier modelqWhen the class mark is classified into the kth class, whether the winning times P of the class mark k is equal to C-1 or not is judged, if yes, the sample point is added into the empty set SsemiOtherwise, adding the sample point into the empty set SactPerforming the following steps; judging all sample points in the learning set Z one by one to obtain a setAnd collectionsWherein z isq1Is a set SsemiSample point of (1), zq2Is a set SactS' is the set SsemiThe total number of sample points in (1), S' is the set SactThe total number of sample points in (1), wherein s' + s ═ s;
(6) using sparse representation, the resulting set S issemiAnd set SactThe sample points in (1) are screened, and the implementation steps are as follows:
(6a) constructing dictionary A ═ X by using all sample points in labeled training set X1,x2,…,xn]And respectively aligning the sets S by using the constructed dictionary AsemiSample point z inq1And set SactSample point z inq2Performing sparse representation: z is a radical ofq1＝Aα1，zq2＝Aα2Wherein, α1And α2Is a sparse representation coefficient vector;
(6b) obtaining a sample point z by using an orthogonal matching pursuit algorithm OMPq1And sample point zq2Represents the coefficient vector:andwherein | · | purple2Is 12Norm, measure data reconstruction error; i | · | purple wind1Is 11Norm for guaranteeing vector α1Sum vector α2λ is a balance factor of the reconstruction error term and the sparse term;
(6c) representing coefficient vectors α from sparseness1And α2Class labels with labeled sample points corresponding to non-zero entries, i.e./, iniE {1, 2.., C }, and set SsemiMiddle prediction class mark k and class mark liSame sample point zq1Screening out and assigning the class labels of all the screened sample points to class labels li(ii) a At the same time set SactMiddle prediction class mark k and class mark liDifferent sample points zq2Screening out, and handing all the screened sample points to an expert for manual labeling;
(7) will gather SsemiMiddle-assigned class label liSample point z ofq1And set SactSample point z in which manual labeling is performedq2Adding the new classifier model into a labeled training set X, and retraining the classifier parameters to obtain a new classifier model;
(8) iterating the steps (3) to (7) until the set iteration times are met, and carrying out comparison on the test set by utilizing the finally obtained classifier modelThe sample points in the test set are classified to obtain the classification result of the test set
2. The hyperspectral image classification method based on gradient boosting decision tree semi-supervised algorithm fusion according to claim 1, wherein the winning times P of class mark k in the step (5) are realized according to the following steps:
(5a) using a classifier model obtained by training the kth class labeled sample point and the t th class labeled sample point to perform on the sample zqClassifying to obtain a prediction result value score (k) and score (t), wherein k belongs to {1, 2.., C }, t belongs to {1, 2.., C }, and k is not equal to t;
(5b) using the obtained prediction result values score (k) and score (t), each sample point z is obtainedqNumber of wins P for category k:
wherein the indication functionf＝score(k)＞score(t)。
3. The gradient boosting decision tree semi-supervised algorithm fusion based hyperspectral image classification method according to claim 1, wherein the orthogonal matching pursuit algorithm OMP is used to obtain the sample point z in the step (6b)q1And sample point zq2The sparse representation coefficient vector is realized by the following steps:
(6b1) initializing the residual term r(0)＝zqIndex setFor a K-dimensional zero vector, the initial iteration J is 1
(6b2) Finding the residual r(J-1)And column j x in dictionary AjThe index lambda corresponding to the maximum inner product,
(6b3) update index set Λ(J)，Λ(J)(J) λ; according to the indexSelecting corresponding atom columns from the dictionary A to form a set A(J)＝A(:,Λ(J)(1:J))；
(6b4) α for obtaining J-order approximation by least squares method(J)＝argmin||zq-A(J)α||2；
(6b5) Updating residual r(J)＝zq-A(J)α(J),J＝J+1；
(6b6) And (6b2) repeating the steps (6b5), judging whether J is larger than K, if so, finishing the iteration, and otherwise, executing the step (6b 2).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610561589.1A CN106203523B (en) | 2016-07-17 | 2016-07-17 | The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610561589.1A CN106203523B (en) | 2016-07-17 | 2016-07-17 | The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106203523A CN106203523A (en) | 2016-12-07 |
| CN106203523B true CN106203523B (en) | 2019-03-01 |
Family
ID=57474833
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610561589.1A Active CN106203523B (en) | 2016-07-17 | 2016-07-17 | The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106203523B (en) |
Families Citing this family (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108229271B (en) * | 2017-01-23 | 2020-10-13 | 北京市商汤科技开发有限公司 | Method and device for interpreting remote sensing image and electronic equipment |
| CN107316309B (en) * | 2017-06-29 | 2020-04-03 | 西北工业大学 | Hyperspectral image saliency target detection method based on matrix decomposition |
| CN107273938B (en) * | 2017-07-13 | 2020-05-29 | 西安电子科技大学 | Multi-source remote sensing image ground object classification method based on two-channel convolution ladder network |
| CN107367753B (en) * | 2017-07-14 | 2019-04-23 | 西南科技大学 | A kind of multicore element recognition methods based on sparse features and fuzzy decision-tree |
| CN108449342B (en) * | 2018-03-20 | 2020-11-27 | 北京云站科技有限公司 | Malicious request detection method and device |
| CN108509882A (en) * | 2018-03-22 | 2018-09-07 | 北京航空航天大学 | Track mud-rock flow detection method and device |
| CN108536938A (en) * | 2018-03-29 | 2018-09-14 | 上海交通大学 | A kind of machine tool life prediction system and prediction technique |
| CN108873829B (en) * | 2018-05-28 | 2020-09-15 | 上海新增鼎数据科技有限公司 | Phosphoric acid production parameter control method based on gradient lifting decision tree |
| CN108764212B (en) * | 2018-06-14 | 2021-04-20 | 内蒙古小草数字生态产业有限公司 | Remote sensing automatic identification method for surveying grass mowing field |
| CN108985365B (en) * | 2018-07-05 | 2021-10-01 | 重庆大学 | Multi-source heterogeneous data fusion method based on deep subspace switching ensemble learning |
| CN109242013B (en) * | 2018-08-28 | 2021-06-08 | 北京九狐时代智能科技有限公司 | Data labeling method and device, electronic equipment and storage medium |
| TWI692776B (en) * | 2018-10-29 | 2020-05-01 | 財團法人工業技術研究院 | Neural-network-based classification device and classification method |
| CN109614507B (en) * | 2018-11-22 | 2020-08-04 | 浙江大学 | Remote sensing image recommendation device based on frequent item mining |
| CN110321770B (en) * | 2019-03-25 | 2022-05-31 | 西安长城数字软件有限公司 | Pipeline monitoring method, device, equipment and storage medium |
| CN109978056A (en) * | 2019-03-26 | 2019-07-05 | 广东工业大学 | A kind of Metro Passenger classification method based on machine learning |
| CN110084318B (en) * | 2019-05-07 | 2020-10-02 | 哈尔滨理工大学 | Image identification method combining convolutional neural network and gradient lifting tree |
| TWI707137B (en) * | 2020-01-13 | 2020-10-11 | 憶象有限公司 | Intelligent production line monitoring system and implementation method thereof |
| CN111414942B (en) * | 2020-03-06 | 2022-05-03 | 重庆邮电大学 | Remote sensing image classification method based on active learning and convolutional neural network |
| CN113837209A (en) * | 2020-06-23 | 2021-12-24 | 乐达创意科技股份有限公司 | Method and system for improved machine learning using data for training |
| CN112465733B (en) * | 2020-08-31 | 2022-06-28 | 长沙理工大学 | Remote sensing image fusion method, device, medium and equipment based on semi-supervised learning |
| CN112070008B (en) * | 2020-09-09 | 2024-03-19 | 武汉轻工大学 | Hyperspectral image feature recognition method, hyperspectral image feature recognition device, hyperspectral image feature recognition equipment and storage medium |
| CN112699926B (en) * | 2020-12-25 | 2023-01-20 | 浙江中控技术股份有限公司 | Method for recognizing saturated grinding abnormity of cement raw material vertical mill based on artificial intelligence technology |
| CN112884050B (en) * | 2021-02-26 | 2024-04-12 | 江南大学 | Quality detection method based on unsupervised active learning |
| CN113111969B (en) * | 2021-05-03 | 2022-05-06 | 齐齐哈尔大学 | Hyperspectral image classification method based on mixed measurement |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
| CN103886342A (en) * | 2014-03-27 | 2014-06-25 | 西安电子科技大学 | Hyperspectral image classification method based on spectrums and neighbourhood information dictionary learning |
| CN104239902A (en) * | 2014-09-12 | 2014-12-24 | 西安电子科技大学 | Hyper-spectral image classification method based on non-local similarity and sparse coding |
| CN104281855A (en) * | 2014-09-30 | 2015-01-14 | 西安电子科技大学 | Hyperspectral image classification method based on multi-task low rank |
| CN104392251A (en) * | 2014-11-28 | 2015-03-04 | 西安电子科技大学 | Hyperspectral image classification method based on semi-supervised dictionary learning |
| CN104408478A (en) * | 2014-11-14 | 2015-03-11 | 西安电子科技大学 | Hyperspectral image classification method based on hierarchical sparse discriminant feature learning |
| CN105608433A (en) * | 2015-12-23 | 2016-05-25 | 北京化工大学 | Nuclear coordinated expression-based hyperspectral image classification method |
-
2016
- 2016-07-17 CN CN201610561589.1A patent/CN106203523B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
| CN103886342A (en) * | 2014-03-27 | 2014-06-25 | 西安电子科技大学 | Hyperspectral image classification method based on spectrums and neighbourhood information dictionary learning |
| CN104239902A (en) * | 2014-09-12 | 2014-12-24 | 西安电子科技大学 | Hyper-spectral image classification method based on non-local similarity and sparse coding |
| CN104281855A (en) * | 2014-09-30 | 2015-01-14 | 西安电子科技大学 | Hyperspectral image classification method based on multi-task low rank |
| CN104408478A (en) * | 2014-11-14 | 2015-03-11 | 西安电子科技大学 | Hyperspectral image classification method based on hierarchical sparse discriminant feature learning |
| CN104392251A (en) * | 2014-11-28 | 2015-03-04 | 西安电子科技大学 | Hyperspectral image classification method based on semi-supervised dictionary learning |
| CN105608433A (en) * | 2015-12-23 | 2016-05-25 | 北京化工大学 | Nuclear coordinated expression-based hyperspectral image classification method |
Non-Patent Citations (2)
| Title |
|---|
| "Class-Level Joint Sparse Representation for Multifeature-Based Hyperspectral Image Classification";Erlei Zhang等;《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》;20160225;第9卷(第9期);第4160-4175页 |
| "基于稀疏表示及光谱信息的高光谱遥感图像分类";宋相法等;《电子与信息学报》;20120215;第34卷(第2期);第268-271页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106203523A (en) | 2016-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106203523B (en) | The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient | |
| CN111860612B (en) | Unsupervised hyperspectral image hidden low-rank projection learning feature extraction method | |
| CN108537102B (en) | High-resolution SAR image classification method based on sparse features and conditional random field | |
| CN109766858A (en) | Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering | |
| CN106503727B (en) | A kind of method and device of classification hyperspectral imagery | |
| CN107145836B (en) | Hyperspectral image classification method based on stacked boundary identification self-encoder | |
| Qin et al. | Cross-domain collaborative learning via cluster canonical correlation analysis and random walker for hyperspectral image classification | |
| CN111401426B (en) | Small sample hyperspectral image classification method based on pseudo label learning | |
| CN108229551B (en) | Hyperspectral remote sensing image classification method based on compact dictionary sparse representation | |
| CN109543723B (en) | Robust image clustering method | |
| CN103489005A (en) | High-resolution remote sensing image classifying method based on fusion of multiple classifiers | |
| CN111222545B (en) | Image classification method based on linear programming incremental learning | |
| CN110414616B (en) | Remote sensing image dictionary learning and classifying method utilizing spatial relationship | |
| CN113936214B (en) | Karst wetland vegetation community classification method based on fusion of aerospace remote sensing images | |
| CN112613536A (en) | Near infrared spectrum diesel grade identification method based on SMOTE and deep learning | |
| CN104408731B (en) | Region graph and statistic similarity coding-based SAR (synthetic aperture radar) image segmentation method | |
| CN113723492A (en) | Hyperspectral image semi-supervised classification method and device for improving active deep learning | |
| CN109558803B (en) | SAR target identification method based on convolutional neural network and NP criterion | |
| CN107273919A (en) | A kind of EO-1 hyperion unsupervised segmentation method that generic dictionary is constructed based on confidence level | |
| CN110598753A (en) | Defect identification method based on active learning | |
| Treboux et al. | Decision tree ensemble vs. nn deep learning: efficiency comparison for a small image dataset | |
| CN114266961A (en) | Method for integrating, learning and classifying marsh vegetation stacks by integrating hyperspectral and multiband fully-polarized SAR images | |
| Guo et al. | Dual graph U-Nets for hyperspectral image classification | |
| Moliner et al. | Weakly supervised semantic segmentation for remote sensing hyperspectral imaging | |
| CN115496950A (en) | Neighborhood information embedded semi-supervised discrimination dictionary pair learning image classification method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |