CN108960276B - A Sample Expansion and Consistency Discrimination Method to Improve the Performance of Spectral Image Supervised Classification - Google Patents

Abstract

Translated fromChinese

本发明公开了一种提升光谱图像监督分类性能的样本扩充与一致性判别方法，包括以下步骤：构造形状匹配模板库；计算单个训练样本的邻域相似度矩阵；使用模板库对相似度矩阵进行匹配，并根据匹配度选取最佳匹配模板；根据最佳匹配模板扩充训练样本；使用扩充后的训练集训练监督分类器；计算测试样本的邻域预测矩阵；使用模板库对邻域预测矩阵进行匹配，并根据一致性度量计算最佳判别结果。本发明通过形状模板匹配有效扩充监督样本，利用模板库与分类预测矩阵一致性判别改善分类结果的局部聚集性，大幅提升监督分类器的精度，提升了小样本监督分类算法的鲁棒性，可适用于任何监督分类器。

The invention discloses a sample expansion and consistency discrimination method for improving the supervised classification performance of spectral images, comprising the following steps: constructing a shape matching template library; calculating a neighborhood similarity matrix of a single training sample; Match, and select the best matching template according to the matching degree; expand the training sample according to the best matching template; use the expanded training set to train the supervised classifier; calculate the neighborhood prediction matrix of the test sample; use the template library to perform the neighborhood prediction matrix match, and calculate the best discriminant result based on the consistency measure. The invention effectively expands the supervised samples through shape template matching, uses the template library and the classification prediction matrix consistency judgment to improve the local aggregation of the classification results, greatly improves the accuracy of the supervised classifier, and improves the robustness of the small sample supervised classification algorithm. Works with any supervised classifier.

Description

Translated fromChinese

提升光谱图像监督分类性能的样本扩充与一致性判别方法A Sample Expansion and Consistency Discrimination Method to Improve the Performance of Spectral Image Supervised Classification

技术领域technical field

本发明涉及高光谱图像监督分类算法中的预处理及后处理技术，具体涉及一种提升光谱图像监督分类性能的样本扩充与一致性判别方法。The invention relates to preprocessing and postprocessing technologies in a hyperspectral image supervised classification algorithm, in particular to a sample expansion and consistency discrimination method for improving the performance of spectral image supervised classification.

技术背景technical background

遥感高光谱图像是由几十甚至上百连续光谱波段组成的三维数据集，由于每个像元的通道数量远远超过普通RGB彩色图像的通道数量，所以在地物识别与分类应用中具有很强的分辨性能。因此，在诸如农业监测、污染监测、矿物识别等领域具有广泛的应用。而利用高光谱图像进行识别与分类是在图像获取后最重要的处理过程，其识别精度决定了高光谱图像对地物鉴别的最终效果。在高光谱图像分类研究中，机器学习相关方法受到了广泛关注。Remote sensing hyperspectral images are three-dimensional datasets composed of dozens or even hundreds of continuous spectral bands. Since the number of channels per pixel far exceeds the number of channels in ordinary RGB color images, it is very useful in ground object recognition and classification applications. Strong resolution performance. Therefore, it has a wide range of applications in fields such as agricultural monitoring, pollution monitoring, and mineral identification. The recognition and classification of hyperspectral images is the most important process after image acquisition, and its recognition accuracy determines the final effect of hyperspectral images on ground object identification. In the research of hyperspectral image classification, machine learning related methods have received extensive attention.

传统机器学习方法例如极限学习机、支持向量机、贝叶斯等方法在高光谱图像监督分类领域中都有一定的应用，但由于高光谱数据集样本间具有很强的相似性，以及同谱异物、同物异谱等现象的出现，导致分类器性能在一定程度上下降，使得最终分类精度不高。利用遥感图像的空间信息提升分类性能变得尤为重要。而常规的空-谱联合方法常将样本的空间信息与光谱信息进行融合并生成新的样本，例如增加样本属性维度的Gabor-SVM等方法会使得模型的输入维度增高，从而引入维度扩张带来的额外影响。目前，有许多针对在不改变数据维度的前提下利用预处理及后处理技术提升分类效果的方法，例如慕彩红等人在主动学习方法的基础上提出了利用测试样本邻域高相似像元判别结果来对目标样本的判别结果进行优化的方法(慕彩红,焦李成,王依萍,等.主动学习和邻域信息相结合的高光谱图像分类方法:,CN 104182767 A[P].2014.)，取得了较好的效果。然而，上述方法的判定规则较为复杂，且像元相似度参数固定，无法根据数据自适应调整，因此方法应用范围受限。Traditional machine learning methods such as extreme learning machine, support vector machine, Bayesian and other methods have certain applications in the field of hyperspectral image supervised classification. The appearance of foreign objects and different spectra of the same object leads to a decline in the performance of the classifier to a certain extent, which makes the final classification accuracy not high. Using the spatial information of remote sensing images to improve the classification performance becomes particularly important. The conventional space-spectral joint method often fuses the spatial information of the sample with the spectral information to generate new samples. For example, methods such as Gabor-SVM that increase the dimension of sample attributes will increase the input dimension of the model, thereby introducing dimension expansion. additional impact. At present, there are many methods for using preprocessing and postprocessing techniques to improve the classification effect without changing the data dimension. A method for optimizing the discriminative results of target samples (Mu Caihong, Jiao Licheng, Wang Yiping, et al. Hyperspectral Image Classification Method Combining Active Learning and Neighborhood Information:, CN 104182767 A[P].2014 .), and achieved better results. However, the judgment rules of the above methods are relatively complex, and the pixel similarity parameters are fixed, which cannot be adaptively adjusted according to the data, so the application scope of the method is limited.

发明内容SUMMARY OF THE INVENTION

本发明的目的在于提供一种提升光谱图像监督分类性能的样本扩充与一致性判别方法，通过形状模板匹配有效扩充监督样本，利用模板库与分类预测矩阵一致性判别改善分类结果的局部聚集性，大幅提升监督分类器的精度，提升了小样本监督分类算法的鲁棒性，可适用于任何监督分类器。The purpose of the present invention is to provide a sample expansion and consistency discrimination method for improving the performance of spectral image supervision and classification, effectively expanding supervision samples through shape template matching, and improving the local aggregation of classification results by using template library and classification prediction matrix consistency judgment, It greatly improves the accuracy of the supervised classifier and improves the robustness of the small sample supervised classification algorithm, which can be applied to any supervised classifier.

实现本发明目的的技术解决方案为：一种提升光谱图像监督分类性能的样本扩充与一致性判别方法，包括以下步骤：The technical solution for realizing the purpose of the present invention is: a method for sample expansion and consistency discrimination for improving the performance of spectral image supervision and classification, comprising the following steps:

第一步，构造形状匹配模板库，即构造若干大小为d×d的形状匹配模板，作为模板库；The first step is to construct a shape matching template library, that is, construct several shape matching templates with a size of d×d as a template library;

第二步，计算单个训练样本的邻域相似度矩阵，即截取以该训练样本为中心的d×d邻域像元块，并计算其邻域像元与中心像元的相似度，将相似度顺序排列成矩阵形式作为该训练样本的邻域相似度矩阵；The second step is to calculate the neighborhood similarity matrix of a single training sample, that is, to intercept the d × d neighborhood pixel block centered on the training sample, and calculate the similarity between its neighborhood pixels and the center pixel. The degree order is arranged in a matrix form as the neighborhood similarity matrix of the training sample;

第三步，使用模板库对相似度矩阵进行匹配，并根据匹配度选取最佳匹配模板；The third step is to use the template library to match the similarity matrix, and select the best matching template according to the matching degree;

第四步，根据最佳匹配模板扩充训练样本，即将最佳匹配模板所覆盖的像元划归到与中心像元同类别训练集中；The fourth step is to expand the training sample according to the best matching template, that is, classify the pixels covered by the best matching template into the training set of the same category as the central pixel;

第五步，使用扩充后的训练集训练监督分类器；The fifth step is to train a supervised classifier using the expanded training set;

第六步，计算测试样本的邻域预测矩阵，即截取以该测试样本为中心的d×d邻域像元块并分别预测其中每个像元的类别，将预测类别顺序排列成矩阵形式作为该样本的邻域预测矩阵；The sixth step is to calculate the neighborhood prediction matrix of the test sample, that is, intercept the d × d neighborhood pixel block centered on the test sample and predict the category of each pixel respectively, and arrange the predicted categories in a matrix form as The neighborhood prediction matrix of the sample;

第七步，使用模板库对邻域预测矩阵进行匹配，并根据一致性度量计算最佳判别结果，即依次计算邻域预测矩阵与模板库中各模板匹配所对应各类别的一致性度量，选取最大值所对应类别作为该样本的最终判别结果。The seventh step is to use the template library to match the neighborhood prediction matrix, and calculate the best discrimination result according to the consistency metric, that is, calculate the consistency metrics of each category corresponding to the matching of each template in the neighborhood prediction matrix and the template library in turn, and select The category corresponding to the maximum value is used as the final discrimination result of the sample.

本发明与现有技术相比，其显著特点在于：(1)通过形状模板匹配可以有效扩充监督样本，从而有效提升小样本监督学习的性能；(2)通过对邻域预测结果的最佳一致性判别，可以大幅度提升最终分类结果的精度；(3)本发明方法适用于任意监督分类器，具有广泛的应用性；(4)本发明对数据本身特性不敏感，因此拥有较好的自适应性和扩展性，并且在同物异谱和同谱异物问题上进行了改善，大幅度提升了算法的最终分类精度。Compared with the prior art, the present invention has the following notable features: (1) supervised samples can be effectively expanded through shape template matching, thereby effectively improving the performance of supervised learning with small samples; (2) the best consistency of neighborhood prediction results is achieved sex discrimination, which can greatly improve the accuracy of the final classification result; (3) the method of the present invention is suitable for any supervised classifier, and has wide applicability; (4) the present invention is not sensitive to the characteristics of the data itself, so it has better self-confidence Adaptability and expansibility, and improved in the problem of the same spectrum and the same spectrum foreign matter, greatly improving the final classification accuracy of the algorithm.

附图说明Description of drawings

图1为本发明提升光谱图像监督分类性能的样本扩充与一致性判别方法流程图。FIG. 1 is a flowchart of a method for sample expansion and consistency discrimination for improving the performance of spectral image supervised classification according to the present invention.

图2为大小为3×3的匹配模板库示例图。FIG. 2 is an example diagram of a matching template library with a size of 3×3.

图3为大小为5×5的匹配模板库示例图。FIG. 3 is an example diagram of a matching template library with a size of 5×5.

图4(a)为Indian Pines数据集真实地物分布图。Figure 4(a) shows the distribution of real objects in the Indian Pines dataset.

图4(b)为Indian Pines数据集采用ELM方法的分类效果图。Figure 4(b) shows the classification effect of the Indian Pines dataset using the ELM method.

图4(c)为Indian Pines数据集采用ELM方法配合本发明方法(使用大小为3×3的匹配模板)的分类效果图。Figure 4(c) is a classification effect diagram of the Indian Pines data set using the ELM method and the method of the present invention (using a matching template with a size of 3×3).

图4(d)为Indian Pines数据集采用ELM方法配合本发明方法(使用大小为5×5的匹配模板)的分类效果图。Figure 4(d) is a classification effect diagram of the Indian Pines data set using the ELM method and the method of the present invention (using a matching template with a size of 5×5).

图4(e)为Indian Pines数据集采用SVM方法的分类效果图。Figure 4(e) shows the classification effect of the Indian Pines dataset using the SVM method.

图4(f)为Indian Pines数据集采用SVM方法配合本发明方法(使用大小为3×3的匹配模板)的分类效果图。FIG. 4( f ) is a classification effect diagram of the Indian Pines data set using the SVM method and the method of the present invention (using a matching template with a size of 3×3).

图4(g)为Indian Pines数据集采用SVM方法配合本发明方法(使用大小为5×5的匹配模板)的分类效果图。Figure 4(g) is a classification effect diagram of the Indian Pines data set using the SVM method and the method of the present invention (using a matching template with a size of 5×5).

图5(a)为Pavia Center数据集真实地物分布图。Figure 5(a) shows the distribution of real objects in the Pavia Center dataset.

图5(b)为Pavia Center数据集采用ELM方法的分类效果图。Figure 5(b) shows the classification effect of the ELM method on the Pavia Center dataset.

图5(c)为Pavia Center数据集采用ELM方法配合本发明方法(使用大小为3×3的匹配模板)的分类效果图。Figure 5(c) is a classification effect diagram of the Pavia Center dataset using the ELM method and the method of the present invention (using a matching template with a size of 3×3).

图5(d)为Pavia Center数据集采用ELM方法配合本发明方法(使用大小为5×5的匹配模板)的分类效果图。Figure 5(d) is a classification effect diagram of the Pavia Center data set using the ELM method and the method of the present invention (using a matching template with a size of 5×5).

图5(e)为Pavia Center数据集采用SVM方法的分类效果图。Figure 5(e) shows the classification effect of the Pavia Center dataset using the SVM method.

图5(f)为Pavia Center数据集采用SVM方法配合本发明方法(使用大小为3×3的匹配模板)的分类效果图。Figure 5(f) is a classification effect diagram of the Pavia Center dataset using the SVM method and the method of the present invention (using a matching template with a size of 3×3).

图5(g)为Pavia Center数据集采用SVM方法配合本发明方法(使用大小为5×5的匹配模板)的分类效果图。Figure 5(g) is a classification effect diagram of the Pavia Center data set using the SVM method and the method of the present invention (using a matching template with a size of 5×5).

具体实施方式Detailed ways

结合图1，一种提升光谱图像监督分类性能的样本扩充与一致性判别方法，包括以下步骤：Combined with Figure 1, a method for sample expansion and consistency discrimination to improve the performance of spectral image supervised classification includes the following steps:

第一步，构造形状匹配模板库，即构造若干大小为d×d的形状匹配模板，作为模板库；The first step is to construct a shape matching template library, that is, construct several shape matching templates with a size of d×d as a template library;

第二步，计算单个训练样本的邻域相似度矩阵，即截取以该训练样本为中心的d×d邻域像元块，并计算其邻域像元与中心像元的相似度，将相似度顺序排列成矩阵形式作为该训练样本的邻域相似度矩阵；The second step is to calculate the neighborhood similarity matrix of a single training sample, that is, to intercept the d × d neighborhood pixel block centered on the training sample, and calculate the similarity between its neighborhood pixels and the center pixel. The degree order is arranged in a matrix form as the neighborhood similarity matrix of the training sample;

第三步，使用模板库对相似度矩阵进行匹配，并根据匹配度选取最佳匹配模板；The third step is to use the template library to match the similarity matrix, and select the best matching template according to the matching degree;

第四步，将最佳匹配模板所覆盖的像元划归到与中心像元同类别训练集中；The fourth step is to classify the pixels covered by the best matching template into the training set of the same category as the central pixel;

第五步，使用扩充后的训练集训练监督分类器；The fifth step is to train a supervised classifier using the expanded training set;

第六步，计算测试样本的邻域预测矩阵，即选取以该测试样本为中心的d×d邻域像元块并分别预测其中每个像元的类别，将预测类别顺序排列成矩阵形式作为该样本的邻域预测矩阵；The sixth step is to calculate the neighborhood prediction matrix of the test sample, that is, select the d × d neighborhood pixel block centered on the test sample and predict the category of each pixel respectively, and arrange the predicted categories in a matrix form as The neighborhood prediction matrix of the sample;

第七步，使用模板库对邻域预测矩阵进行匹配，并根据一致性度量计算最佳判别结果，即依次计算邻域预测矩阵与模板库中各模板匹配所对应各类别的一致性度量，选取最大值所对应类别作为该样本的最终判别结果。The seventh step is to use the template library to match the neighborhood prediction matrix, and calculate the best discrimination result according to the consistency metric, that is, calculate the consistency metrics of each category corresponding to the matching of each template in the neighborhood prediction matrix and the template library in turn, and select The category corresponding to the maximum value is used as the final discrimination result of the sample.

进一步的，第一步构造L个大小为d×d的形状匹配模板作为模板库，其中L≥8，d≥3且d为奇数；构造匹配模板的基本原则为：Further, the first step is to construct L shape matching templates with a size of d×d as a template library, where L≥8, d≥3 and d is an odd number; the basic principles for constructing matching templates are:

(1)在模板中以像素点为单位，将所有像素分为覆盖与未覆盖，覆盖取1，未覆盖取0；(1) In the template, take pixels as the unit, divide all pixels into covered and uncovered, take 1 for coverage, and take 0 for uncovered;

(2)在模板中所覆盖的像素区域须满足连通性，即模板中值为1的像素可组成满足4-邻接的连同区域，且连通区域须覆盖中心点；(2) The pixel area covered in the template must meet the connectivity, that is, the pixels with a value of 1 in the template can form a joint area that satisfies 4-adjacency, and the connected area must cover the center point;

(3)模板中的覆盖区域点个数n须满足n＝(d-1)²。(3) The number n of coverage area points in the template must satisfy n=(d-1)² .

进一步的，第二步计算单个训练样本的邻域相似度矩阵，具体过程为：Further, the second step calculates the neighborhood similarity matrix of a single training sample, and the specific process is as follows:

(1)对于第i个训练样本x_i，在高光谱图像中截取以像元x_i为中心的d×d邻域像元块，记为Dⁱ，表示如下：(1) For the i-th training sample x_i , intercept the d×d neighborhood pixel block centered on the pixel x_i in the hyperspectral image, denoted as Dⁱ , and expressed as follows:

1≤i≤N，N为训练样本数；1≤i≤N, N is the number of training samples;

(2)依次计算中心像元x_i与邻域像元块Dⁱ中的像元

的相似度,记为

其中1≤e≤d²，计算方法为：(2) Calculate the pixels in the center pixel x_i and the neighboring pixel block Dⁱ in turn

similarity, denoted as

Where 1≤e≤d² , the calculation method is:

(3)将中心像元x_i与邻域像元

的相似度

按照顺序摆放成矩阵形式，作为该训练样本的邻域相似度矩阵，记为Sⁱ，表示如下：(3) Connect the center pixel_xi with the neighboring pixels

similarity of

Arranged in matrix form in order, as the neighborhood similarity matrix of the training sample, denoted as Sⁱ , expressed as follows:

进一步的，第三步使用模板库对相似度矩阵进行匹配，并根据匹配度选取最佳匹配模板，计算方法为：Further, the third step uses the template library to match the similarity matrix, and selects the best matching template according to the matching degree. The calculation method is:

(1)依次计算模板库中第l个模板与训练样本x_i的邻域相似度矩阵Sⁱ的匹配度，记为

计算式为：(1) Calculate the matching degree of the lth template in the template library and the neighborhood similarity matrix Sⁱ of the training sample x_i in turn, denoted as

The calculation formula is:

其中，运算符

表示卷积运算，T_l为模板库中第l个模板，Sⁱ为样本x_i对应的邻域相似度矩阵，1≤l≤L。where the operator

Represents the convolution operation, T_l is the l-th template in the template library, Sⁱ is the neighborhood similarity matrix corresponding to the sample x_i , 1≤l≤L.

(2)在样本x_i与L个模板匹配得到的匹配度

中，选取最大值所对应的模板作为最佳匹配模板，记为

计算式为：(2) The matching degree obtained by matching the sample_xi with L templates

, select the template corresponding to the maximum value as the best matching template, denoted as

The calculation formula is:

进一步的，第四步根据最佳匹配模板扩充训练样本，即将训练样本x_i对应的最佳匹配模板

所覆盖的区域对应的邻域像元块中相同位置的像元划归到与中心像元x_i同类别训练集中。Further, the fourth step expands the training samples according to the best matching template, that is, the best matching template corresponding to the training sample_xi

The pixels at the same position in the neighborhood pixel block corresponding to the covered area are classified into the training set of the same category as the central pixel x_i .

进一步的，第五步使用扩充后的训练集训练监督分类器，即按照步骤二至步骤四依次对所有原始训练样本进行匹配扩充，并使用扩充后的训练样本训练监督分类器。Further, the fifth step uses the expanded training set to train the supervised classifier, that is, according to steps 2 to 4, all original training samples are sequentially matched and expanded, and the expanded training samples are used to train the supervised classifier.

进一步的，第六步计算测试样本的邻域预测矩阵，计算方法为：Further, the sixth step calculates the neighborhood prediction matrix of the test sample, and the calculation method is:

(1)对于第j个测试样本y_j，在高光谱图像中截取以像元y_j为中心的d×d邻域像元块，记为K^j，表示如下：(1) For the jth test sample y_j , intercept a d×d neighborhood pixel block centered on the pixel y_j in the hyperspectral image, denoted as K^j , and expressed as follows:

1≤j≤M，M为测试样本数量；1≤j≤M, M is the number of test samples;

(2)使用训练后的分类器依次对像元块K^j中的像元

进行判别，判别结果记为

其中1≤e≤d²，并将判别结果按照顺序摆放成矩阵形式，作为该测试样本y_j的邻域预测矩阵，记为Q^j，表示如下：(2) Use the trained classifier to sequentially classify the pixels in the pixel block K^j

discriminate, and the discriminant result is recorded as

Among them, 1≤e≤d² , and the discriminant results are arranged in matrix form in order, as the neighborhood prediction matrix of the test sample y_j , denoted as Q^j , and expressed as follows:

进一步的，第七步使用模板库对邻域预测矩阵进行匹配，并根据一致性度量计算最佳判别结果，具体过程为：Further, the seventh step uses the template library to match the neighborhood prediction matrix, and calculates the best discrimination result according to the consistency measure. The specific process is as follows:

(1)依次计算测试样本y_j与模板库中第l个模板T_l匹配下属于类别c的一致性度量，记为

计算方法为：(1) Calculate the consistency measure of the test sample y_j and the l-th template T_l in the template library belonging to the category c in turn, denoted as

The calculation method is:

其中，运算符

表示哈达玛积，Q^j为测试样本y_j的邻域预测矩阵，

表示矩阵

中元素等于c的个数，1≤c≤C，C为样本类别数。where the operator

represents the Hadamard product, Q^j is the neighborhood prediction matrix of the test sample y_j ,

representation matrix

The element in is equal to the number of c, 1≤c≤C, and C is the number of sample categories.

(2)在样本y_j与各模板匹配下分别属于各类别的一致性度量中，选取一致性度量最大值所对应类别为最终判别结果，记为

计算式为：(2) In the consistency measures belonging to each category under the matching of samples y_j and each template, select the category corresponding to the maximum value of the consistency metric as the final discrimination result, denoted as

The calculation formula is:

其中，

为样本y_j在各模板匹配下分别属于各类别的一致性度量集合，C为样本类别数，L为模板数量。in,

is the consistency measure set that the sample_yj belongs to each category under each template matching, C is the number of sample categories, and L is the number of templates.

下面结合实施例和附图对本发明进行详细说明。The present invention will be described in detail below with reference to the embodiments and accompanying drawings.

实施例Example

结合图1，一种提升光谱图像监督分类性能的样本扩充与一致性判别方法，步骤如下：Combined with Figure 1, a sample expansion and consistency discrimination method to improve the performance of spectral image supervised classification, the steps are as follows:

第一步，构造形状匹配模板库，即构造L个大小为d×d的形状匹配模板作为模板库，其中L≥8，d≥3且d为奇数。本实施例d＝3，L＝16，设计16个满足条件的不同匹配模板作为模板库，记为T₁,T₂,…,T₁₆，模板库如图2所示。每个模板实为一个矩阵，图2中描述的第一个模板T₁表示如下：The first step is to construct a shape matching template library, that is, construct L shape matching templates with a size of d×d as a template library, where L≥8, d≥3 and d is an odd number. In this embodiment, d=3, L=16, and 16 different matching templates that satisfy the conditions are designed as template libraries, which are denoted as T₁ , T₂ , . . . , T₁₆ , and the template library is shown in FIG. 2 . Each template is actually a matrix, and the first template T1 depicted in Figure₂ is represented as follows:

同理，其它模板也为矩阵表示形式。大小为5×5的匹配模板库形式如图3所示，模板设计方式按照前述规则进行设计。Similarly, other templates are also in matrix representation. The format of the matching template library with a size of 5×5 is shown in Figure 3, and the template design method is designed according to the aforementioned rules.

第二步，计算单个训练样本的邻域相似度矩阵，在d＝3，L＝16的条件下，具体过程为：The second step is to calculate the neighborhood similarity matrix of a single training sample. Under the conditions of d=3 and L=16, the specific process is:

(1)对于第i个训练样本x_i，在高光谱图像中截取以像元x_i为中心的3×3邻域像元块，记为Dⁱ，表示如下：(1) For the i-th training sample x_i , intercept a 3×3 neighborhood pixel block centered on the pixel x_i in the hyperspectral image, denoted as Dⁱ , and expressed as follows:

1≤i≤N，N为训练样本数；1≤i≤N, N is the number of training samples;

(2)依次计算中心像元x_i与邻域像元块Dⁱ中的像元

的相似度，记为

其中1≤e≤9。计算过程为：(2) Calculate the pixels in the center pixel x_i and the neighboring pixel block Dⁱ in turn

similarity, denoted as

where 1≤e≤9. The calculation process is:

(3)将中心像元x_i与邻域像元

的相似度

按照顺序摆放成矩阵形式，作为该训练样本的邻域相似度矩阵，记为Sⁱ，表示如下：(3) Connect the center pixel_xi with the neighboring pixels

similarity of

Arranged in matrix form in order, as the neighborhood similarity matrix of the training sample, denoted as Sⁱ , expressed as follows:

第三步，使用模板库对相似度矩阵进行匹配，并根据匹配度选取最佳匹配模板。在d＝3，L＝16的条件下，具体计算过程为：The third step is to use the template library to match the similarity matrix, and select the best matching template according to the matching degree. Under the conditions of d=3 and L=16, the specific calculation process is as follows:

(1)依次计算模板库中第l个模板与样本x_i的邻域相似度矩阵Sⁱ的匹配度，记为

计算过程为：(1) Calculate the matching degree of the l-th template in the template library and the neighborhood similarity matrix Sⁱ of the sample x_i in turn, denoted as

The calculation process is:

其中，运算符

表示卷积运算，T₁,T₂,…,T₁₆为模板库中的模板，Sⁱ为样本x_i对应的邻域相似度矩阵，1≤l≤16。where the operator

Represents the convolution operation, T₁ , T₂ ,..., T₁₆ are templates in the template library, Sⁱ is the neighborhood similarity matrix corresponding to the sample x_i , 1≤l≤16.

(2)在样本x_i对应的所有模板匹配度

中选取最大值所对应的模板作为最佳匹配模板，记为

计算式为：(2) All template matching degrees corresponding to sample_xi

The template corresponding to the maximum value is selected as the best matching template, denoted as

The calculation formula is:

假设

则样本x_i的最佳匹配模板即为T₁。Assumption

Then the best matching template of sample_xi is T₁ .

第四步，根据最佳匹配模板扩充训练样本，即将样本x_i对应的最佳匹配模板

所覆盖(值为1)的区域对应的邻域像元块中相同位置的像元划归到与中心像元x_i同类别训练集中。假设像元x_i的最佳匹配模板为T₁，则将

加入到与x_i同类训练集中。The fourth step is to expand the training sample according to the best matching template, that is, the best matching template corresponding to the sample_xi

The pixels in the same position in the neighborhood pixel block corresponding to the area covered (value 1) are classified into the training set of the same class as the central pixel x_i . Assuming that the best matching template for cell x_i is T₁ , then

Add to the same training set as_xi .

第五步，使用扩充后的训练集训练监督分类器，即按照步骤二至步骤四依次对所有训练样本进行匹配扩充，并使用扩充后的训练样本训练监督分类器，记分类器为q＝classify(y)，其中y为测试样本，q为测试样本y通过分类器得到的类别。The fifth step is to use the expanded training set to train the supervised classifier, that is, according to steps 2 to 4, all training samples are sequentially matched and expanded, and the expanded training samples are used to train the supervised classifier, and the classifier is recorded as q=classify (y), where y is the test sample, and q is the category obtained by the classifier for the test sample y.

第六步，计算测试样本的邻域预测矩阵，在d＝3，L＝16的假设条件下，具体过程为：The sixth step is to calculate the neighborhood prediction matrix of the test sample. Under the assumptions of d=3 and L=16, the specific process is:

(1)对于第j个测试样本y_j，在高光谱图像中截取以像元y_j为中心的3×3邻域像元块，记为K^j，表示如下：(1) For the jth test sample y_j , intercept a 3×3 neighborhood pixel block centered on the pixel y_j in the hyperspectral image, denoted as K^j , and expressed as follows:

1≤j≤M，M为测试样本数；1≤j≤M, M is the number of test samples;

(2)使用训练后的分类器依次对像元块K^j中的像元

进行判别，判别结果记为

其中1≤e≤9，并将判别结果按照顺序摆放成矩阵形式，作为该测试样本y_j的邻域预测矩阵，记为Q^j，计算方法为：(2) Use the trained classifier to sequentially classify the pixels in the pixel block K^j

discriminate, and the discriminant result is recorded as

Among them, 1≤e≤9, and the discriminant results are arranged in matrix form in order, as the neighborhood prediction matrix of the test sample y_j , denoted as Q^j , and the calculation method is:

第七步，使用模板库对邻域预测矩阵进行匹配，并根据一致性度量计算最佳判别结果，在d＝3，L＝16，样本类别数C＝9的条件下，具体过程为：The seventh step is to use the template library to match the neighborhood prediction matrix, and calculate the best discrimination result according to the consistency measure. Under the conditions of d=3, L=16, and the number of sample categories C=9, the specific process is:

(1)依次计算测试样本y_j与模板库中第l个模板T_l匹配下属于类别c的一致性度量，记为

计算过程为：(1) Calculate the consistency measure of the test sample y_j and the l-th template T_l in the template library belonging to the category c in turn, denoted as

The calculation process is:

其中，运算符°表示哈达玛积，

表示矩阵

中元素等于c的个数，1≤c≤9。where the operator ° represents the Hadamard product,

representation matrix

The element in is equal to the number of c, 1≤c≤9.

(2)在样本y_j对应的各模板匹配下分别属于各类别的一致性度量中，选取一致性度量最大值所对应类别为最终判别结果，记为

计算式为：(2) In the consistency measures belonging to each category under each template matching corresponding to the sample_yj , select the category corresponding to the maximum value of the consistency metric as the final discrimination result, denoted as

The calculation formula is:

其中

为样本y_j在16个模板匹配下分别属于9个类别的一致性度量集合。假设

则测试样本y_j的最终判别结果为属于类别7，同时最佳匹配模板为T₁。in

A set of consistency metrics that belong to 9 categories for samples y_j under 16 template matching. Assumption

Then the final discrimination result of the test sample y_j is that it belongs to category 7, and the best matching template is T₁ .

下面通过仿真实验说明本发明的效果：The effects of the present invention are described below through simulation experiments:

仿真实验采用两组真实高光谱数据：Indian Pines数据集和Pavia Center数据集。Indian Pines数据集为机载可见红外成像光谱仪(AVIRIS)在美国印第安纳州IndianPines实验区采集的高光谱遥感图像。该图像共包含220个波段，空间分辨率为20m，图像大小为145×145。去除20个水汽吸收和低信噪比波段后(波段号为104-108，150-163，220)，选择剩下的200个波段作为研究对象。该地区共包含16种已知地物共10366个样本。PaviaCenter数据集是由帕维亚的ROSIS传感器采集，共包含115个波段，图像大小为1096×490，在去除噪声波段之后，选择剩下的102个波段作为研究对象。每类样本均随机取10％作为训练样本，其余90％作为测试样本，并分别进行十次实验计算平均结果，并给出OA(OverallAccuracy)以及AA(Average Accuracy)评价指标。在实验前，两组数据集均采用MH预测方法进行平滑处理。仿真实验均在Windows 10操作系统下采用Python3.6与MATLAB R2016a完成。The simulation experiments use two sets of real hyperspectral data: the Indian Pines dataset and the Pavia Center dataset. The Indian Pines dataset is hyperspectral remote sensing images collected by the Airborne Visible Infrared Imaging Spectrometer (AVIRIS) in the Indiana Pines experimental area, Indiana, USA. The image contains a total of 220 bands, the spatial resolution is 20m, and the image size is 145×145. After removing 20 water vapor absorption and low signal-to-noise ratio bands (band numbers are 104-108, 150-163, 220), the remaining 200 bands were selected as research objects. The area contains a total of 10,366 samples of 16 known features. The PaviaCenter dataset was collected by the ROSIS sensor in Pavia, with a total of 115 bands and an image size of 1096×490. After removing the noise band, the remaining 102 bands were selected as the research object. 10% of each type of samples are randomly selected as training samples, and the remaining 90% are used as test samples. Ten experiments are performed to calculate the average results, and OA (Overall Accuracy) and AA (Average Accuracy) evaluation indicators are given. Before the experiment, both datasets were smoothed using the MH prediction method. The simulation experiments were completed under theWindows 10 operating system using Python 3.6 and MATLAB R2016a.

本发明采用的评价指标是常规精度的评价方法，包括平均精度(AA)和总体精度(OA)。由于本发明的用途是提升任意分类器的分类精度，在仿真实验中我们分别采用ELM和SVM作为基础分类器进行实验，为了显示该方法对分类器性能的提升，每组数据集均会使用ELM、TM-ELM(3x3)、TM-ELM(5x5)、SVM、TM-SVM(3x3)、TM-SVM(5x5)进行实验，并进行结果对比。其中TM-ELM(3x3)表示使用3x3大小的模板匹配(template matching)结合ELM的分类方法，TM-ELM(5x5)表示使用5x5大小的模板匹配结合ELM的分类方法，同理，TM-SVM(3x3)、TM-SVM(5x5)表示含义类似。大小为3x3的模板库样式如图2所示，大小为5x5的模板库样式如图3所示。The evaluation index adopted in the present invention is the evaluation method of conventional precision, including average precision (AA) and overall precision (OA). Since the purpose of the present invention is to improve the classification accuracy of any classifier, we use ELM and SVM as the basic classifiers in the simulation experiments respectively. , TM-ELM (3x3), TM-ELM (5x5), SVM, TM-SVM (3x3), TM-SVM (5x5) to conduct experiments and compare the results. Among them, TM-ELM (3x3) represents the classification method using 3x3 template matching (template matching) combined with ELM, TM-ELM (5x5) represents the classification method using 5x5 template matching combined with ELM, for the same reason, TM-SVM ( 3x3) and TM-SVM (5x5) have similar meanings. The template library style of size 3x3 is shown in Figure 2, and the template library style of size 5x5 is shown in Figure 3.

表1为本发明方法对Indian Pines与Pavia Center数据集进行仿真实验的分类精度(％)。Table 1 shows the classification accuracy (%) of the simulation experiments performed on the Indian Pines and Pavia Center datasets by the method of the present invention.

表1Table 1

从实验结果可以发现，在不改变原始ELM和SVM分类模型的前提下，仅通过使用本发明方法对数据进行训练样本扩充与预测结果最佳一致性判别，便可以大幅度提升分类精度。在Indian Pines数据集上，本发明方法可以将原始ELM模型96.27％的总体精度提升至98.85％(使用5x5大小模板匹配方法)，将原始SVM模型90.33％的总体精度提升至98.99％(使用5x5大小模板匹配方法)。在PaviaCenter数据集上，本发明方法可以将原始ELM模型97.46％的分类精度提升至98.08％(使用5x5大小模板匹配方法)，将原始SVM模型98.87％的总体精度提升至99.79％(使用5x5大小模板匹配方法)。显然，本发明方法在使用3x3模板和5x5模板进行匹配后对分类结果均有明显提升，并且随着模板大小的扩大，分类精度也会随之提升，充分表明了本发明方法具有较高的稳定性和扩展性。本发明方法在两组数据集上的结果效果图如图4、图5所示。以上两组真实数据集的仿真实验结果表明了本发明方法的有效性。It can be found from the experimental results that, without changing the original ELM and SVM classification models, the classification accuracy can be greatly improved only by using the method of the present invention to expand the data for training samples and determine the best consistency of the prediction results. On the Indian Pines dataset, the method of the present invention can improve the overall accuracy of the original ELM model from 96.27% to 98.85% (using the 5x5 size template matching method), and improve the overall accuracy of the original SVM model from 90.33% to 98.99% (using the 5x5 size). template matching method). On the PaviaCenter dataset, the method of the present invention can improve the classification accuracy of the original ELM model from 97.46% to 98.08% (using a 5x5 template matching method), and improve the overall accuracy of the original SVM model from 98.87% to 99.79% (using a 5x5 template). matching method). Obviously, the method of the present invention can significantly improve the classification results after using the 3x3 template and the 5x5 template for matching, and with the expansion of the template size, the classification accuracy will also be improved, which fully shows that the method of the present invention has high stability. flexibility and scalability. Figures 4 and 5 show the results of the method of the present invention on two sets of data sets. The simulation results of the above two sets of real data sets show the effectiveness of the method of the present invention.

Claims (7)

1. A sample expansion and consistency discrimination method for improving spectral image supervision classification performance is characterized by comprising the following steps:

firstly, constructing a shape matching template library, namely constructing a plurality of shape matching templates with the size of dxd as a template library;

secondly, calculating a neighborhood similarity matrix of a single training sample, namely intercepting a dxd neighborhood pixel block taking the training sample as a center, calculating the similarity between a neighborhood pixel and the center pixel, and arranging the similarity in sequence into a matrix form to be used as the neighborhood similarity matrix of the training sample;

thirdly, matching the similarity matrix by using a template library, and selecting an optimal matching template according to the matching degree; the calculation method comprises the following steps:

(1) calculating the first template and training sample x in the template library in sequence_iNeighborhood similarity matrix S ofⁱDegree of matching of (D) is recorded as

The calculation formula is as follows:

wherein the operator

Representing a convolution operation, T_lIs the first template in the template library, SⁱIs a sample x_iL is more than or equal to 1 and less than or equal to L of the corresponding neighborhood similarity matrix;

(2) at sample x_iMatching degree obtained by matching with L templates

In the method, the template corresponding to the maximum value is selected as the best matching template and recorded as the best matching template

The calculation formula is as follows:

fourthly, classifying the pixels covered by the best matching template into a training set of the same type as the central pixels;

fifthly, training a supervision classifier by using the expanded training set;

sixthly, calculating a neighborhood prediction matrix of the test sample, namely selecting a dxd neighborhood pixel block taking the test sample as a center, predicting the category of each pixel, and sequentially arranging the prediction categories into a matrix form to be used as the neighborhood prediction matrix of the sample;

and seventhly, matching the neighborhood prediction matrix by using the template base, calculating an optimal discrimination result according to the consistency measurement, namely sequentially calculating the consistency measurement of each category corresponding to the matching of the neighborhood prediction matrix and each template in the template base, and selecting the category corresponding to the maximum value as a final discrimination result of the sample.

2. The method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1 is characterized in that in the first step, L shape matching templates with the size of dxd are constructed as a template library, wherein L is more than or equal to 8, d is more than or equal to 3, and d is an odd number; the basic principle for constructing a matching template is as follows:

(1) dividing all pixels into covered pixels and uncovered pixels by taking pixel points as units in a template, wherein the covered pixels take 1 and the uncovered pixels take 0;

(2) the pixel regions covered in the template need to satisfy connectivity, i.e. pixels with a median value of 1 in the template can form a connected region satisfying 4-adjacency, and the connected region needs to cover a central point;

(3) the number n of coverage area points in the template must satisfy n ═ d-1²。

3. The method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1 is characterized in that a neighborhood similarity matrix of a single training sample is calculated in the second step, and the specific process is as follows:

(1) for the ith training sample x_iIntercepting pixel x in hyperspectral image_iThe D × D neighborhood block of pixels as the center is marked as DⁱExpressed as follows:

wherein i is more than or equal to 1 and less than or equal to N, and N is the number of training samples;

(2) sequentially calculating central pixel x_iAnd neighborhood pixel block DⁱIn (2)

The similarity of (2) is recorded as

Wherein e is more than or equal to 1 and less than or equal to d²The calculation method comprises the following steps:

(3) center pixel x_iAnd neighborhood pixel

Degree of similarity of

Putting the training samples into a matrix form according to the sequence, taking the matrix as a neighborhood similarity matrix of the training samples and recording the matrix as SⁱExpressed as follows:

4. the method of claim 1, wherein the fourth step of expanding the training samples according to the best matching template, i.e. the training sample x_iCorresponding best matching template

Pixels at the same position in the neighborhood pixel block corresponding to the covered area are classified to the central pixel x_iThe training of the same category is centralized.

5. The method for sample expansion and consistency discrimination to improve the performance of spectral image supervised classification as recited in claim 1, wherein the fifth step is to train the supervised classifier using the expanded training set, i.e. all original training samples are subjected to matching expansion in sequence from the second step to the fourth step, and the supervised classifier is trained using the expanded training samples.

6. The method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1, wherein the sixth step is to calculate a neighborhood prediction matrix of the test sample, and the calculation method is as follows:

(1) for the jth test sample y_jIntercepting pixel y in hyperspectral image_jA centered d × d neighborhood block, denoted as K^jExpressed as follows:

wherein j is more than or equal to 1 and less than or equal to M, and M is the number of test samples;

(2) sequentially aligning primitive blocks K using trained classifiers^jIn (2)

The result of the discrimination is recorded as

Wherein e is more than or equal to 1 and less than or equal to d²And putting the discrimination results in a matrix form in sequence as the test sample y_jIs the neighborhood prediction matrix of (1), denoted as Q^jExpressed as follows:

7. the method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1, wherein in the seventh step, a template library is used for matching the neighborhood prediction matrix, and the optimal discrimination result is calculated according to consistency measurement, and the specific process is as follows:

(1) calculating the test sample y in turn_jWith the first template T in the template library_lConsistency measure under match belonging to class c, denoted

The calculation method comprises the following steps:

wherein the operator ° represents the Hadamard product, Q^jFor testing a sample y_jThe neighborhood prediction matrix of (a) is,

representation matrix

C is more than or equal to 1 and less than or equal to C, and C is the number of sample categories;

(2) in the sample y_jSelecting the category corresponding to the maximum value of the consistency measurement as the final judgment result and recording the final judgment result as the final judgment result

The calculation formula is as follows:

wherein,

is a sample y_jAnd respectively belonging to the consistency measurement sets of the categories under the matching of the templates.

Images