CN107203790A - Utilize the Chinese land noctilucence Classification in Remote Sensing Image Accuracy Assessment of two stage sampling model - Google Patents

Abstract

The present invention relates to a method for evaluating the accuracy of China's land luminous remote sensing classification using a two-level sampling model. The method specifically includes the following steps: Step S1, extracting the built-up areas of China's land areas from the NPP/VIIRS series data, and obtaining unverified The extraction results of the luminous data; step S2, establish a secondary sampling model, complete the sampling of the evaluation area, and obtain a credible inspection data set; step S3, use the error matrix verification method to perform partition verification on the preliminary extraction of luminous data, and finally obtain Accuracy evaluation of the classification results of the whole luminous data. Using the classification accuracy evaluation method of luminous remote sensing using the two-stage sampling model, the system of luminous remote sensing data is completed, and the accuracy evaluation of large-area luminous remote sensing classification results is realized effectively. The quality inspection of the extraction results of luminous remote sensing towns can solve the previous verification set Too few, the evaluation parameters are not comprehensive, and a comprehensive and accurate verification result has been achieved.

Description

Evaluation method of China land night light remote sensing classification accuracy using two-level sampling model

technical field

The invention relates to the technical field of remote sensing image quality inspection, in particular to a method for evaluating classification accuracy of China's land luminous remote sensing classification using a two-level sampling model.

Background technique

Night light remote sensing data can detect low-intensity lights from urban lights and even small-scale residential areas, and distinguish them from dark rural backgrounds, indirectly reflecting the distribution of cities and towns, and providing new insights into large-scale urban detection research. train of thought. However, due to factors such as the spillover effect of night light, extracting cities and towns based on night light may lead to problems such as the urban area extraction result being larger than the actual value. Therefore, the quality of the extraction results must be verified before application.

Traditional verification methods, for example, in the study of the urbanization pattern of the Yangtze River Delta, only compared and verified the area attributes of the city extraction results with national statistical data; in the study of urban expansion in China, 15 large and medium-sized cities were selected as the verification set using Landsat data. The training set, using the error matrix to systematically verify the extraction results; using four quantitative indicators of landscape shape index, aggregation index, edge area ratio and connectivity index to verify the shape and area similarity of the night light remote sensing data extraction results; random sampling 1014 sample points were extracted in the same way to verify the night light data extraction results in Zhejiang Province.

In the past, there were some problems in the accuracy evaluation of the extraction results of luminous remote sensing lighting areas: (1) The selection of the verification set was too small, the distribution method was unreasonable, and it was not representative compared with the overall extraction area; (2) The selected verification evaluation indicators were too few and not enough. It is comprehensive and cannot really reflect the extraction results; (3) The evaluation method is not systematic, and the comparison with the national statistical values cannot explain the spatial distribution of the extraction results.

Chinese patent document CN201110230310.9, application date 20110812, the patent name is: a method for analyzing the impact of landscape features on the accuracy of remote sensing classification spots, including step 1, obtaining data, including performing data standardization on the original image; The data obtained in step 1 is subjected to image recognition, including classification and post-classification processing, wherein the classification process includes the determination of classification patterns; The correlation calculation among them is carried out by regression curve fitting to establish a regression model, which includes the definition of the landscape index expressing the landscape characteristics; step 4, the regression model established in step 3 is statistically tested, and at the same time, the classification error expressed by the landscape index is evaluated.

The above-mentioned patent documents use the spatial characteristics of remote sensing classification as the theoretical basis for accuracy evaluation and analysis, and use the landscape characteristics of remote sensing classification maps to describe and express the classification accuracy, providing basis and guidance for the application and research related to land cover thematic maps . However, regarding a method for evaluating the classification accuracy of luminous remote sensing using a two-stage sampling model, the systematic and effective implementation of the accuracy evaluation of the classification results of luminous remote sensing for large areas, and the quality inspection of the extraction results of luminous remote sensing cities and towns can be solved. In the past, the problems of too few verification sets and incomplete evaluation parameters were solved, and a comprehensive and accurate verification result was achieved. There is no corresponding disclosure of the technical scheme.

To sum up, there is an urgent need for a method for evaluating the classification accuracy of luminous remote sensing using a two-stage sampling model, to complete the system of luminous remote sensing data, and effectively realize the accuracy evaluation of large-area luminous remote sensing classification results, and to extract the results of luminous remote sensing towns. Quality inspection can solve the problem of too few verification sets and incomplete evaluation parameters in the past, and achieve a comprehensive and accurate verification result. But there is no report about this method at present.

Contents of the invention

The purpose of the present invention is to address the deficiencies in the prior art, to provide a method for evaluating the accuracy of luminous remote sensing classification using a two-stage sampling model, to complete the system for luminous remote sensing data, and to effectively realize the accuracy evaluation of large-area luminous remote sensing classification results , the quality inspection of night light remote sensing town extraction results can solve the problems of too few verification sets and incomplete evaluation parameters in the past, and achieve a comprehensive and accurate verification result.

For realizing above-mentioned object, the technical scheme that the present invention takes is:

A method for evaluating the classification accuracy of China's terrestrial luminous remote sensing classification using a two-stage sampling model, characterized in that the method specifically includes the following steps:

Step S1, realize the extraction of built-up areas in China's land area from the NPP/VIIRS series data, and obtain unverified extraction results;

Step S2, establishing a two-level sampling model, completing the sampling of the evaluation area, and obtaining a credible inspection data set;

Step S3, using the error matrix verification method to perform partition verification on the preliminary extraction of luminous data, and finally obtain the accuracy evaluation of the classification results of all luminous data.

As a preferred technical solution, the step S2 includes the following steps:

Step S21: Combining the division of the three major economic regions of China's eastern coast, central and western regions, extract some images from the Landsat 8 images covering China's land, respectively obtain the sampling sets in the three major regions, and complete the first-level sampling;

Step S22: Select the inspection pixels in the images corresponding to the three major regions of the eastern coast, the central region and the western region respectively to obtain the result set of the second-level sampling, and complete the second-level sampling;

Step S23: Extract the attribute value of the verification pixel from the high-precision Landsat 8 image, and determine whether the pixel belongs to a luminous area or a non-luminous area.

As a preferred technical solution, the first-level sampling method in step S21 is as follows: first, according to the sampling inspection model of this level, extract a picture from the Landsat 8 image covering mainland China as a sample, and use it in the eastern coastal, central, and western regions of China. Sampling is carried out according to different weights in the three major economic regions, and the quality level of the entire batch of atlases is inferred by checking the quality of each sample map until the verification set of n maps is obtained.

As a preferred technical solution, the step S22 also includes classifying the Landsat image according to NDVI and NDBI, and determining the urban area in the image in a large scale; at the same time, visual interpretation is used for the area at the junction of towns and non-towns The method is accurately determined, and finally the precise extraction result is achieved in the Landsat image, and the second-level sampling is obtained.

As a preferred technical solution, step S3 includes the following steps:

Step S31: Superimpose the vectorized luminous remote sensing image classification results and the verification set obtained by the two-stage sampling model; Step S32: Use the error matrix verification method to evaluate the accuracy of the classification results for the three major regions; Step S33: Combine the three major areas The verification results of the region are back-calculated, and finally the credible and reliable global classification accuracy evaluation verification results are obtained.

As a preferred technical solution, in step S3, the night light remote sensing extraction results are also systematically evaluated by selecting the parameters of overall accuracy, kappa coefficient, user accuracy, producer accuracy, misclassification error, and omission error.

As a preferred technical solution, in the step S3, the night light will be systematically evaluated by selecting the overall accuracy, kappa coefficient, user accuracy, producer accuracy, misclassification error, and omission error parameters in the three economic regions of China. The method of remote sensing extraction results obtains the extraction accuracy of night light remote sensing in each region, and then calculates the accuracy result of night light remote sensing extraction in China's land area by combining the weights of each area.

The present invention has the advantage that:

1, a kind of method of the present invention utilizes the Chinese land luminous remote sensing classification accuracy of two-level sampling model to evaluate, utilize the theory of two-level sampling for large area areas, from covering all Landsat 8 images of China's land area, select Appropriate and well-distributed verification data can effectively guarantee the reliability of subsequent verification results.

2. The evaluation method selects the error matrix for statistical analysis, and selects the overall accuracy (oa), kappa coefficient, user accuracy, producer accuracy, misclassification error, and omission error as evaluation parameters. From different angles, it does not require a comprehensive system The classification results were evaluated.

3. In view of the fact that there are few researches on the verification of night lights in the past, and the numerical comparison with national statistical data is often insufficient, this method starts from the perspective of geosciences and conducts sampling from the perspective of space to achieve a more accurate accuracy evaluation.

Description of drawings

Accompanying drawing 1 is the flowchart of two-stage sampling verification method of the present invention.

Accompanying drawing 2 is the classification result of luminous remote sensing image of the present invention.

Accompanying drawing 3 is the Landsat8 distribution map of the Chinese land area of the present invention.

Accompanying drawing 4 is the distribution diagram of the first-stage sampling verification set of the present invention.

Accompanying drawing 5 is the distribution chart of the second stage sampling verification set of the present invention.

detailed description

The specific embodiments provided by the present invention will be described in detail below in conjunction with the accompanying drawings.

Example 1

Compared with ordinary visible light remote sensing data, night light remote sensing data can only be used as an indirect remote sensing data when extracting cities and towns. Aiming at the phenomenon that in the past, the application of luminous remote sensing is more important than verification, this study extracts the luminous area in luminous remote sensing, and conducts a complete accuracy verification research on the extraction results, and finally proposes a complete set of extraction and research models.

The flow chart of this research method is shown in Figure 1, which mainly consists of three parts:

(1) Classification of luminous remote sensing images.

Using the support vector machine classifier to classify China's land NPP/VIIRS luminous remote sensing images, two classification results of luminous areas and non-luminous areas are obtained whose accuracy is to be verified;

(2) Establish a secondary sampling model.

①Combining the division of the three major economic regions of China's eastern coast, central and western regions, some images were extracted from the Landsat 8 images covering China's land, and the sampling sets in the three major regions were respectively obtained to complete the first-level sampling; among them, the three major economic regions The region is the eastern coastal region divided by the Chinese government: Liaoning, Beijing, Tianjin, Hebei, Jiangsu, Zhejiang, Fujian, Guangdong, Taiwan, Hong Kong, Macau, Guangxi; the central region: Heilongjiang, Jilin, Inner Mongolia, Shanxi, Henan, Hubei, Hunan , Jiangxi; Western Region: Shaanxi, Gansu, Ningxia, Sichuan, Guizhou, Yunnan, Qinghai, Xinjiang, Tibet;

Since the illuminated area in night-light remote sensing images reflects human activities, the model in the first-level sampling needs to consider both space and population density factors. In this study, a weight is set for each region in combination with spatial area and population density, and different proportions of samples are sampled.

The Landsat 8 images covering the land of China are distributed according to the World Wide Angle Reference System-2, covering about 536 scenes of the land of China.

② Select test pixels from the images of each scene corresponding to the three major regions of the eastern coast, the central region, and the west, and obtain the result set of the second-level sampling to complete the second-level sampling;

③ Extract the attribute value of the verification pixel from the high-precision Landsat 8 image, and judge whether the pixel belongs to the luminous area or the non-luminous area;

For the determination of the attributes of the inspection pixels, in the high-precision Landsat 8 images, the Landsat images are first classified according to NDVI and NDBI, and the urban areas in the images are determined in a large scale. At the same time, the area at the junction of urban and non-urban areas is accurately determined by visual interpretation, and finally an accurate extraction result is achieved in the Landsat 8 image, and it is judged that the pixel belongs to the luminous area or the non-luminous area.

(3) Evaluation of classification result accuracy.

① Overlay vectorized luminous remote sensing image classification results and the verification set obtained by the two-stage sampling model; ② Use the error matrix verification method to evaluate the accuracy of the classification results for the three major regions; ③ Combining the verification results of the three major regions, conduct The calculation is reversed, and finally a credible and reliable global classification accuracy evaluation verification result is obtained. The specific model design and decision-making rules are as follows:

(1) Secondary sampling model

In the sampling design of the second-level model in this paper, the first-level sampling inspection model takes the "map frame" covering Landsat 8 in mainland China as the sampling unit, and the second-level sampling inspection model takes the "light area and non-light area area" in the map frame as the sampling unit. sampling unit.

1) Selection of the first-level sampling model

The sampling unit of the first-level sampling inspection model is "Map". In the first-level sampling inspection, firstly, according to the sampling inspection model of this level, n images are selected from the Landsat 8 images covering mainland China as samples, and the quality of the entire batch of atlases is inferred by inspecting the quality of each sample image Level.

Since the illuminated area in night-light remote sensing images reflects human activities, the model in the first-level sampling needs to consider both space and population density factors. According to the principle of combining the level of economic and technological development and geographical location, the Chinese government divides the country into three major economic zones, namely: the eastern coastal area, the central area, and the western area. In this study, a weight λ _i is set for each region, and different proportions are sampled.

λ _i ＝θ(a _i )·θ(d _i ) (1)

Among them, θ(a _i ) is the proportion of this region in the national land area, θ(d _i ) is the proportion of the population density of this region, and the sample size of each region is

n _i =λ _i ·N (2)

Where N is all Landsat 8 images covering mainland China according to the WRS-2 system.

2) Selection of the second-level sampling model

In the second-level sampling inspection model, the pixel (Pixel) of the image is defined as the sampling unit. In the current-level sampling inspection model, all pixels in the sample frame are regarded as the inspection population. According to the second-level sampling inspection model , select an appropriate sample size n', and judge the quality of the map frame by checking the attribute value (luminous or non-luminous) of each inspection unit.

3) Acquisition of Landsat 8/OLI image verification set

The actual ground objects in the land light area of night light remote sensing images mainly include town lights, so Landsat 8 mainly extracts town areas for verification research. Landsat series remote sensing data is rich in band information, reasonable and effective use of band information can quickly and accurately extract the area of built-up areas. For a Landsat 8 remote sensing image, when extracting urban areas, the Landsat images are first classified according to NDVI and NDBI, and the urban areas in the image are determined in a large scale. At the same time, the area at the junction of urban and non-urban areas is accurately determined by visual interpretation, and finally an accurate extraction result is achieved in the Landsat 8/OLI image.

(2) Verification method

The error matrix can effectively evaluate the classification results of remote sensing images. This program first vectorizes the classification results, and establishes the error matrix with the help of the verification set obtained by secondary sampling through the overlapping analysis of the vector data, and selects the overall accuracy (oa), kappa coefficient, user accuracy, producer accuracy, and misclassification The error and omission error were used to systematically evaluate the results.

The distribution relationship of the error matrix is shown in Table 1.

Table 1: Error matrix and its element distribution

Among them, ω _i represents the classification category of the extraction result to be verified (producer), ω _j represents the classification category of the verification (user) set, and the two categories correspond to each other, and ρ _ij represents the number of pixels in which category ω _i is divided into ω _j , ρ _i+ and ρ _+i respectively represent the total number of pixels of producers and users in different categories, and m is the synthesis of all verification pixels.

Important verification results can be obtained from the error matrix:

The overall accuracy is the sum of the main diagonal elements in the error matrix (the number of correct classifications) divided by the total number of samples:

Producer accuracy and user accuracy can represent the accuracy of a single category. Producer accuracy is the number of correct classifications of a certain category divided by the total number of samples of this category (the sum of the columns of this category):

pa _i =ρ _ii /ρ _+i , (5)

While user accuracy is defined as the number of correctly classified classes divided by the number of samples classified into that class (row sum of that class):

ua _i =ρ _ii /ρ _i+ , (6)

In addition to the various descriptive accuracy measurements above, commission errors and omission errors are also included:

ce _i =1-ua _i , (7)

oe _i =1-pa _i , (8)

In addition, using various statistical analysis techniques on the basis of the error matrix can be used to compare different classification methods, the most commonly used of which is the Kappa analysis technique:

(3) Inversion of accuracy results

Using the method (2) to obtain the night light remote sensing extraction accuracy a _i of each area in the three economic regions of China, and then combine the weights of each area to calculate the accuracy result A of the night light remote sensing extraction in China's land area.

Through the global accuracy A, the user can intuitively obtain the extraction accuracy of the night light image.

Verify the effect

(1) Extraction results of luminous remote sensing

The verification experiment uses the night light remote sensing data mentioned above, and uses the SVM extraction method to classify the 2016 night light data. With the help of envi5.2 software, this paper uses its integrated SVM classification algorithm to supervise and classify luminous data, and obtain two types of classification information, luminous areas and non-luminous areas. The results are shown in Figure 2:

The purple area in Figure 2 is the extracted urban area (512284 pixels), the green area is the area without light at night (13339719 pixels), and the urban area accounts for 3.7% of the total land area of the country.

(2) Establishment of the secondary sampling model

According to the first-level sampling method, 10% of the 536 Landsat 8 images (Fig. 3) covering the land of China are selected as the verification set, and the samples are distributed according to China's three major economic regions. Table 2 is the statistical table of the area and population of the three major regions:

Table 2 Area and population statistics of the three major regions

According to the formula (1) (2), the sampling ratio of the eastern coastal region, the central region and the western region can be calculated as 8.94:6.13:4.03. Considering the integrity of the sampling map, the actual sampling ratio is 8:6: With a sampling ratio of 4, 24, 18, and 12 scenes of images were sampled in each area, and the sample distribution is shown in Figure 4: The yellow area in Figure 4 is the image position and covered area selected by the first-level sampling.

According to the second-level sampling inspection model, it is necessary to obtain the inspection pixel and the classification attribute of the pixel. In this paper, the method of spectral analysis and boundary visual recognition is used to extract the urban area of the 54 scene images selected by sampling. There are 7701 images in one scene. *7831 pixel points are all used as the verification set for verification. Landsat 8 image extraction results and luminous data extraction results were overlaid and analyzed. The overlay effect is shown in Figure 5. The shaded area in the figure is the luminous area extracted from the Landsat 8/OLI image. By comparing the classification results of the reference data and the luminous remote sensing The classification results can judge the pros and cons of night light image extraction results.

(3) Statistics of verification results

In this study, the method of error matrix is used to verify the extraction area of luminous data. According to the two types of extraction areas (Km ² ), the statistical results of the error matrix are obtained in Table 3. The proportion of non-luminous areas in the extraction results is much larger than that of luminous areas. Therefore, the follow-up The non-luminous areas in the evaluation have a greater impact on the results.

Table 3 Statistical table of error matrix (unit: Km ² )

According to the formulas (4) (9), the overall accuracy oa and kappa coefficients of the three regions can be calculated respectively, and the results are shown in Table 4

Table 4: Comparison table of overall accuracy and kappa coefficient (unit: %)

The results of user accuracy, production accuracy, misclassification error, and omission error are shown in Table 4:

Table 4: Evaluation parameter comparison table (unit: %)

From the verification results, it can be concluded that the overall accuracy of the luminous extraction results is relatively high, but the kappa coefficient is only 37.54%, and the values of the two comment parameters are quite different; in particular, the user accuracy of the luminous area in the classification results is only 25.43 %, the misclassification error reached 74.57%, indicating that the classification effect of the luminous area is poor; more precise.

Therefore, it can be judged that the classification accuracy of the noctilucent region in this luminous remote sensing extraction result is low, and the classification result of the dark night region is extremely high, because the area of the luminous region is much smaller than the area of the dark night region, and the proportion of the luminous region in the overall accuracy evaluation is relatively small. low, which makes the overall accuracy oa higher; from the kappa coefficient, the results of this luminous remote sensing classification result have a low similarity with the verification set, which is mainly reflected in the low classification accuracy of luminous areas, and most of the dark night areas are classified as It was mistakenly classified as a luminous area. It can be concluded that the overall accuracy of the classification results is relatively high, but the accuracy of the luminous area is lower than that of the non-luminous area.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the method of the present invention, some improvements and supplements can also be made, and these improvements and supplements should also be considered Be the protection scope of the present invention.

Claims (7)

1. A method for evaluating the classification accuracy of China's terrestrial luminous remote sensing classification utilizing a secondary sampling model, characterized in that the method specifically comprises the following steps: Step S1, realize the extraction of built-up areas in China's land area from the NPP/VIIRS series data, and obtain unverified extraction results; Step S2, establishing a two-level sampling model, completing the sampling of the evaluation area, and obtaining a credible inspection data set; Step S3, using the error matrix verification method to perform partition verification on the preliminary extraction of luminous data, and finally obtain the accuracy evaluation of the classification results of all luminous data. 2. the Chinese land luminous remote sensing classification accuracy evaluation method utilizing two-stage sampling model according to claim 1, it is characterized in that, described step S2 comprises the following steps: Step S21: Combining the division of the three major economic regions of China's eastern coast, central and western regions, extract some images from the Landsat 8 images covering China's land, respectively obtain the sampling sets in the three major regions, and complete the first-level sampling; Step S22: Select the inspection pixels in the images corresponding to the three major regions of the eastern coast, the central region and the western region respectively to obtain the result set of the second-level sampling, and complete the second-level sampling; Step S23: Extract the attribute value of the verification pixel from the high-precision Landsat 8 image, and determine whether the pixel belongs to a luminous area or a non-luminous area. 3. The Chinese land luminous remote sensing classification accuracy evaluation method utilizing two-level sampling model according to claim 1, characterized in that, the first-level sampling method of the step S21 is: at first according to the sampling inspection model of this level, from the coverage of China Take a map from the Landsat 8 image of the mainland as a sample, and sample according to different weights in the three major economic regions of China's eastern coastal, central, and western regions. Wait until the overall n-map verification set, and pass the inspection of each sample map to infer the quality level of the entire batch of atlases. 4. The Chinese land luminous remote sensing classification accuracy evaluation method utilizing two-level sampling model according to claim 1, characterized in that, in the step S22, it also includes classifying Landsat images according to NDVI and NDBI, and determining in a large range The urban area in the image; at the same time, the urban and non-urban border area is accurately determined by visual interpretation, and finally the accurate extraction result is achieved in the Landsat image, and the second-level sampling is obtained. 5. the Chinese land luminous remote sensing classification accuracy evaluation method utilizing two-stage sampling model according to claim 1, is characterized in that, comprises the following steps in the step S3: Step S31: superimposing the vectorized night light remote sensing image classification results and the verification set obtained by the secondary sampling model; Step S32: using the error matrix verification method to evaluate the accuracy of the classification results for the three major regions; Step S33: Combining the verification results of the three major areas, perform reverse calculation, and finally obtain a credible and reliable global classification accuracy evaluation verification result. 6. The Chinese land luminous remote sensing classification accuracy evaluation method utilizing two-stage sampling model according to claim 5, characterized in that, in step S3, by selecting overall accuracy, kappa coefficient, user accuracy, producer accuracy, misclassification error , Leakage error parameters, and systematically evaluate the night light remote sensing extraction results. 7. the Chinese land luminous remote sensing classification accuracy evaluation method utilizing two-stage sampling model according to claim 1, it is characterized in that, in described step S3, will be respectively in three economic regions of China by selecting overall accuracy, kappa coefficient, User accuracy, producer accuracy, misclassification error, omission error parameters, systematically evaluate night light remote sensing extraction results method to obtain the night light remote sensing extraction accuracy of each region, and then calculate the night light remote sensing of China's land area by combining the weights of each region Extracted precision results.