Printer Friendly

Land use classification based on texture feature of multiple-source remote sensing image.

1. Introduction

The land use type has the characteristics of variability and diversity, and the land cover is scattered, which brings great uncertainty to the land use classification. A single remote sensing image cannot cover the comprehensive land use information; therefore, it is helpful to improve the accuracy of land use classification by acquiring multi-source remote sensing images and realizing the advantages of complementary on remote sensing data. But, at present of multi-source remote sensing land using classification research less, on land use classification is still focused on single source remote sensing data, recognition of the land use types, remote sensing data is not perfect and restrict land use classification accuracy improved(Chen, 2014; Liu, 2006). In addition, planting patterns of different texture characteristics of crops and different crops may be due to changes in cropping patterns with similar texture feature and the same crop may have the different texture features due to different planting patterns.

In view of the above problems, this paper proposes a multi-source remote sensing image land use classification based on texture. Use of multi-source remote sensing image acquisition is more abundant and more accurate land use of information, establishment of land use type texture library, and make up the texture characteristics may change due to planting mode, improves land use classification accuracy. Through the land use of the precise distinction between, each pixel in remote sensing image feature matching, can meet people from the mass remote sensing data fast recognition and need to retrieve different topics of land use types and land structural properties (Di, 2014), and some land use types of in-depth study provides a good technical foundation for the realization of the land resources dynamic monitoring and management, to learn about. Using remote sensing technology to classify the types of land use, the essence is to obtain the characteristic information of land type, and to match with the existing types of features, and then determine the type of land. The first classification technology is image visual interpretation classification, which is the most direct and basic method of land use classification. Its classification is based on the image of the surface of the main visual features, such as shape, size, color, location; shadow (Deng, 2008). The method is carried out by professional personnel to interpret the classification, classification accuracy is high, in some of the needs of the high accuracy of remote sensing image classification is still indispensable. Although this method can achieve higher classification accuracy, but on visual image interpretation personnel requirements are higher, dependence, in many cases will also need people personally conducted on-site investigation on-the-spot investigation, classification is improved and the cost and scope of application is small, strong subjectivity.

With the development of computer science and technology, using remote sensing image computer automatic classification become land use classification of the type of dominant. Its essence is using computer pattern recognition of remote sensing image information for the recognition and classification of property. Using computer system for remote sensing image classification, according to the different work principle is divided into statistical model method and syntactic mode method. Statistical method mainly includes the maximum likelihood method, spectral characteristics analysis method (Xu, 2010), texture analysis and other traditional remote sensing image classification method, syntactic pattern methods are mainly in the neural network method, fuzzy mathematics method and expert system method. Strahier in 1980, firstly using the maximum likelihood method to classify the image data, he in the acquisition of remote sensing data and assumes that all kinds of distribution function is based on the normal distribution, the calculation of every attribution probability of classification area, various regional data in the image are classified; similarly, Zhang using Landsat ETP images by maximum likelihood method of 7 kinds of land use types for the supervised classification, the classification accuracy is about 85%. Goldberg in 1983 by spectral characteristics of single multi band image classification, from obtain the forest resource information (Li, 2011), but as complex information contained in remote sensing image, with synonyms spectrum and foreign body with same spectrum there are a lot of restricts the spectral analysis of the improvement of classification accuracy. The expert classification method based on decision tree of 11 types of land for information extraction and classification. Through the extraction of remote sensing image in land use types of attribute parameters, combined with the knowledge base of the expert system in a similar form of binary tree gradually to identify the various types and the obtained classification accuracy was 84%. In addition, the current land types of remote sensing image classification is using a single source remote sensing data attribution probability calculation, a single data source, the land use information, classification and calculation is difficult to implement and need to determine in advance the relevant parameters.

Texture can be regarded as a pattern that is produced by the change of color in the image space, and it is one of the inherent properties of the image. So for remote sensing image, texture directly reflects the spatial distribution of image gray mode, contains the surface information of the remote sensing image and its relationship with the surrounding environment, can a good balance of remote sensing image of the macro structure and micro structure (Yang, 2008; Li, 2012). Common texture feature extraction methods include: gray level co-occurrence matrix, wavelet theory, fractal theory, local two value model, local three value model (Hu, 2007; Lan, 2010). Hardick proposed based on gray level co-occurrence matrix of 14 texture measure, and its application to the classification of remote sensing image, the image texture features to obtain more accurate expression; acqua and Gamba with gray level co-occurrence matrix, through the analysis of the different phase of SAR remote sensing image (Huang, 2003), it is concluded that the extraction can improve the accuracy of image recognition based on texture information. Then, with the rise of the statistics and the gradual improvement, the application of texture information extraction of remote sensing image based on the statistics of the ground, it is the use of random function to analyze the phenomenon of uncertainty. Huang use the logarithmic difference function application in texture image described, and separate spectral analysis for image classification and compared, finally came to a conclusion that the logarithmic deviation function and texture extraction combination can improve the conclusion of lithologic classification accuracy. In 1980s, Penthand et al proposed a fractal model. Ojala proposed a local two value model (Binary Pattern Local, LBP), which is used to describe the local texture features of the image (Liu, 2009; Wang, 2013). The algorithm has local characteristics of sound, to changes in illumination is not sensitive, algorithm complex degree is low, can be more effective to extract the essential characteristics of, become in recent years has drawn attention to texture extraction algorithm. The study for land use classification in the presence of the above problems is proposed based on the texture of multi-source remote sensing image classification; the land of multi-source remote sensing image using texture feature information fusion, texture classification database is established, in order to improve the land use classification accuracy.

2. Research methods

This study with multi-source remote sensing image as data source, based on the texture of the land use classification, first of all, the remote sensing image preprocessing, to convert the image as the feature vector information, texture feature library is established to achieve the fusion of multi-source remote sensing data. Its process can be described as:

1. The extraction of texture information from land use of multi-source remote sensing image. By using the LBP operator to get the texture feature vector of the land use type of multi-source remote sensing image, the scale invariant, gray scale invariant and rotation invariant.

2. Identification of land use in multi-source remote sensing image. Using LBP histogram relevance calculation of remote sensing images of land use types, a large number of tests based on the review of relevant literature and, get the two pieces of remote sensing image to the same or different land use types of correlation values in the range, the land use types of recognition.

3. Construct texture land use multi-source remote sensing image. Using LBP operator to get the texture characteristic information of land use, combining with BOW model, K-means clustering to establish classification texture database, realize the fusion of multi-source remote sensing data.

4. Scheme design and optimization of land use classification. Using support vector machine (SVM) to be in land classification and texture matching, classification of land use categories to be determined, the scheme contrast selected preferred scheme.

5. Verification of texture library, the use of a method of cross validation of the method of cross validation of the texture library for reasonable validation, to determine the feasibility of the texture database classification.

2.1. Acquisition of classified samples

This study using multi-source remote sensing images with different spatial resolution as the sample classification, resolution can be roughly divided into 0.5m, 1M, 3M, 5m, 10m several levels, for a total of 176 amplitude, the training sample 96 pieces (48), test samples of 80 (10*8). They respectively, include 8 kinds of land use types: woodland, grassland, terraces, residential areas, deserts, hills, mountains of the original and kiwi fruit planting area.

2.2. Texture feature extraction based on LBP algorithm

Local binary pattern LBP) is initially as a texture description algorithm is proposed, measure and describe the texture information of images of different and efficient characterization of the texture features of remote sensing image, in the image retrieval, robot vision, and face recognition and so on has been widely used. From the definition of LBP can be seen, there is gray scale invariance does not have rotation invariance, that is, when the original remote sensing image rotation, for each window will get a different LBP value. Cause changes in the whole remote sensing image LBP map to make the image rotation does not, Maenpaa presents with rotation invariance of the improved LBP Operator, in fact, this idea is: continuous rotating circular neighborhood get a series of initial definition of LBP values, take the minimum value as the neighborhood of LBP value.

[FIGURE 1 OMITTED]

As shown in Figure 1, Schematic diagram of the rotation invariant LBP, the original state under the remote sensing image of LBP values for 225, when occurred different degrees of rotation corresponding LBP values will change, figure said the 8 kinds of LBP patterns. After a rotation invariant processing, finally obtained with rotation invariant LBP value is 15. That is to say the figure in the 8 LBP modes corresponding to the rotation invariant LBP mode is 00001111.

2.3. Image classification based on BOW model

LBP texture feature extraction feature vector and cannot express the characteristic information of the whole remote sensing image to image complete and accurate expression, also need to continue to carry on processing to the texture data. Based on the BOW model, this paper establishes a visual word bag for remote sensing images, and uses a visual dictionary to represent the same land use types in remote sensing images. According to the matching of visual words and visual dictionary, land use types are identified, which can be used to distinguish the type of independent image blocks in the image.

BOW (of words Bag) model was originally used in information retrieval and Natural Language Processing. In the text information retrieval, all the statements as a combination of words, grammar, each word with its position in the sentence as a whole is not dependent on the in other words, in image retrieval, the bow by method of structure of the image of visual bag of words to express the image. This research using LBP histogram data as image feature descriptors using clustering the disorder data are clustered in order to realize the construction of the bow of the dictionary. After the construction of the dictionary, the relationship between each image feature descriptor and clustering to build the histogram, so each image can represented by statistical histogram features. In this study, all remote sensing images are processed on the basis of BOW model, and the texture database of land use type is obtained, which can realize the fusion of multi-source remote sensing data.

3. Research methods

3.1. Texture feature extraction and recognition of remote sensing image based on LBP algorithm

In this study, the original RGB remote sensing image gray scale transformation as the image input, the use of uniform and rotation invariant LBP operator to extract texture features, set P=8, R=1.

[TABLE 1 OMITTED]

Comparing the original remote sensing image histogram and LBP histogram can be seen distribution histogram of the original image, the majority of data clusters, not concentrated expression of the image characteristics; in contrast, LBP histogram except for some of the smaller data into state of distribution, so it is more representative. Therefore, the use of LBP histogram for land use classification, the use of a larger value as the texture feature vector of the remote sensing image, help to express the remote sensing image of land use types of texture feature information.

Identification and analysis of different land use types:

[TABLE 2 OMITTED]

A remote sensing image and remote sensing image B for two kinds of different land use types belong to woodland and grassland, using LBP Operator to extract the texture feature information drawn LBP histogram, to calculate the correlation of two images for 81.83%. Through multiple remote sensing image land use classification experiments, and access to relevant literature, can draw the conclusion: when the two pieces of remote sensing image correlation is less than 0.86, that the two types of land use is different.

Identification and analysis of the same land use types:

[TABLE 3 OMITTED]

The land use types of remote sensing images D and C belong to forest land, and their correlation is 97.73% through LBP histogram. After multiple remote sensing image land use classification experiments, and access to relevant literature, can come to the conclusion: when the two image correlation reached to 95% and above, that the two kinds of land use type is the same. In addition, synthesis of several experimental and literature data, when two remote sensing image correlation value between 86% and 95%, due to the remote sensing image resolution differences, shooting height difference etc., will cause land use type identification of uncertain phenomena, so 86%~95% can be said for multi-source remote sensing images of land use to identify the type of error band.

Through the LBP texture extraction algorithm multi-source remote sensing image texture feature vector, is used to identify the types of land use: two remote sensing image correlation reached 95% and, namely, two kinds of land use types are the same; correlation is less than 86%, namely two kinds of different land use type% of the 86%~95 as remote sensing images of land use type identification error. It can be seen that the recognition of the same or different land use types can be realized by using the texture feature information, which can lay the foundation for the further realization of the land use classification based on the texture of multi-source remote sensing image.

3.2. Classification based on BOW model

3.2.1. Classification of support vector machines (SVM)

In order to realize the LBP texture feature extraction based on the type of multi-source remote sensing images of land, while ensuring the integrity of texture extraction. First of all, the gray of the remote sensing image block dividing, does not increase the complexity of the K-means clustering, to avoid clustering efficiency is too low, the research choose the original remote sensing image was evenly divided as the size of a small patch of 1010, on every piece of LBP texture feature extraction, so as to ensure the land use types of texture information integrity and to improve recognition of the land use types. In order to facilitate the experimental test, the remote sensing image size used here is 100*100. How to use the BOW model and K-means clustering to describe the characteristics of the land use type of remote sensing image, after getting the LBP histogram of each small piece of remote sensing image, this paper proposes three schemes.

Scheme 1:

1. The remote sensing image is divided into a number of small pieces of patch 10*10;

2. LBP texture features extracted from each LBP and calculate the patch histogram;

3. Remote sensing image of all the K-means patch clustering, calculate the 65 point of mass;

4. Initialization frequency histogram, bin number is the number of points of the center of mass;

5. Patch to calculate all the Euclidean distance and the centroid distance, in accordance with the order from small to large;

6. Look at each patch to which the centroid of the nearest, then the center of mass corresponding to the bin plus 1;

7. In order to calculate the patch of each remote sensing image, and finally get the histogram which is represented by the 65 center of mass;

8. Histogram normalization, the land use type of each remote sensing image can be used to express the feature histogram;

9. The characteristic histogram of all the images constitute the classification of texture database, and realize the fusion of multi-source remote sensing information. So

So the method of test samples and training samples of all of the remote sensing image preprocessing, you can get all image centroid feature histogram representation, the feature histogram construction cost of land use texture database, realize the fusion of multi-source remote sensing image information, improve the texture in land use the accuracy of feature information.

To be sample classification and texture library existing land using pattern matching, using SVM classifier and libsvm toolkit for multi-source remote sensing image land use classification experiments can be obtained: when testing samples in remote sensing image a total of 30 pieces (3 * 10 pieces), income classification accuracy was 90% (27 / 30); when testing samples in remote sensing image a total of 40 (4 * 10 pieces), income classification accuracy was 80% (32 / 40). However under the premise of the remote sensing image of the same sample, repeatedly calculated accuracy results not only, for example, when the second test on the same 30 pieces of test samples for remote sensing image, the accuracy was 93% (28 / 30).

Known from the analysis of the experiment results, the classification accuracy decreases with the increase in the number of samples decreased, repeatedly test accuracy fluctuations. The reason is that LBP histogram data and more complex, and cluster centroid fluctuations, in further by calculating the center of mass of the patch with the centroid of the Euclidean distance and fluctuation further enlarged, resulting in the classification results of insufficient reliability problems.

Scheme 2:

1. The remote sensing image is divided into a number of small pieces of patch 10*10;

2. LBP texture features extracted from each LBP and calculate the patch histogram;

3. Remote sensing image of all the K-means patch clustering, and given the cluster centroid point K=65;

4. The average value of the centroid of all patches in a remote sensing image;

5. The centroid of all image points of the feature vector constitutes the classification of texture database; realize the fusion of multi-source remote sensing data.

Relative scheme through calculating Euclidean distance image centroid, the second program is the centroid averaging said various remote sensing images of land use types, which greatly reduces the computational complexity. After the SVM classifier tests can be obtained: when using test samples in remote sensing image of a total of 40 (4 * 10 pieces), income classification accuracy was 97.5% (39 / 40); when using test samples in remote sensing image of a total of 56 (7 8 pieces), income classification accuracy was 98.21% (55 / 56); when the test samples in remote sensing image with a total of 80 (8 * 10 pieces), income classification accuracy was 98.75% (79 / 80).

The experimental results show that the scheme two has higher accuracy, and in the same condition, the calculation results are unique, and the fluctuation of the calculation accuracy is not. Analysis of the scheme: LBP texture feature extraction of image block, improve the integrity and accuracy of texture feature information; use cluster centroid bow image model construction, image feature value extraction, avoid secondary calculation of frequency histogram in floating classification results. Thus, compared with other schemes, this scheme has high classification accuracy and reliable results. In order to compare with the above two schemes, the following two schemes were further tested in the experiment.

Scheme 3: due to LBP texture feature extraction of remote sensing image output vector and the comparison scheme does not divide the image block direct use of test samples for land use classification, results: using test samples of 50 pieces (5 * 10 pieces), the classification of income quasi accuracy was 24% (12 / 50), under the same conditions of the, accuracy was 28%.

Visible, this program is low accuracy and the results of fluctuations.

Scheme 4: compared to the scheme and scheme using image division block implementation of K-means clustering, this scheme directly different resolution remote sensing images in the same land use types of remote sensing image of K-means clustering, namely the clustering algorithm, the input matrix for the same land use types of LBP histogram. Experimental results are obtained: when using test samples in remote sensing image of a total of 10 (2*5), income classification accuracy was 70% (7/10); when using test samples in remote sensing image of a total of 18 (3*6 pieces), income classification accuracy was 55.56% (10/18); when using sample test of remote sensing image is 24 frames (4*6 pieces), income classification accuracy was 29.17% (7/24).

Obviously, the experimental results of this scheme are low, and with the increase of the number of samples, the result is more volatile.

Analysis of the contrast shows that: This study with multi-source remote sensing image as the source of data classification, on land use types, the image resolution and image size are quite different and without the image block classification and clustering, image in the land use texture information are unable to get effective expression.

Summary and comparison of the above schemes:
Table 4--Comparison of classification schemes based on SVM

number      training   Training   Test sample    Test sample
            samples    sample     remote         land use
            for        land use   sensing        type
            remote     type       image          number
            sensing               number
            images

            42         3          30             3

Scheme 1    48         4          30             3

            50         5          40             4

            48         4          40             4

Scheme 2    70         7          56             7

            96         8          80             8

            60         5          50             5

Scheme 3    60         5          50             5

            30         2          10             2

            30         3          18             3

Scheme 4    30         5          24             4

number      Classification   experimental result
            result

            90%(27/30)       The classification
                             accuracy is not stable,
Scheme 1    93%(28/30)       and the number of
                             samples varies
            80%(32/40)       significantly.

            97.5%(39/40)     Accuracy compared
                             to a larger upgrade,
Scheme 2    98.21%(55/56)    numerical stability,
                             reliability is
            98.75%(79/80)    significantly
                             improved compared
                             to other programs

            24%(12/50)       The accuracy is low,
                             the classification
Scheme 3    28%(14/50)       result is fluctuation,
                             and the reliability
                             is low.

            70%(7/10)        The accuracy is low,
                             the numerical value
            55.56(10/18)     is not stable, the
                             sample quantity
Scheme 4    29.17%(7/24)     change is obvious,
                             the precision is
                             not reliable


According to the construction method of BOW model of land use type of multi-source remote sensing image, this study proposes four schemes. Comparison on the scheme can be concluded: compared to other design scheme, scheme of classification results with high accuracy, reliable numerical values, and therefore the choice of scheme for the SVM classifier preferred scheme, select the entire remote sensing image centroid feature vector of a land use type texture database.

3.2.2. Parameter optimization of SVM

In this study, we use RFB (cross validation method) to optimize the parameters of C and G two in SVM multi classification function on the premise of using the CV kernel function. In the second scheme based on further experimental testing: before the good samples in remote sensing image replacement for edge samples in remote sensing image and poor samples in remote sensing image, the experimental test samples in remote sensing image of a total of 80, including 8 kinds of land use types, classification results was calculated as: 91.25% 73/80. Obviously, with the addition of the edge samples and the difference samples, the calculation accuracy has declined, but after several experiments under the same conditions, the scheme still maintains the uniqueness of the results. The calculated values of C and g were brought into the svm_train model, and the classification results were obtained under the same sample condition: 95% (76/80). Visible, compared to the optimization of the previous 91.25%, the classification accuracy is higher.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

3.3. Cross validation

Will this study all 176 (8*10; 8*12) remote sensing images of preprocessing, utilization scheme for the two ways set up image of the bow model and get the image feature vector representation, in order to establish eight land use texture database. In the leave one out cross validation, texture library in the entire remote sensing image in turn as a test sample and 21 training samples. Using the feature vector loop is used to test the classification, were used in the SVM classifier experiment: Using SVM classifier classification accuracy can be obtained from the average value is 96.73%.

The accuracy of the classification is more than 95%, which indicates that the texture database established in this study has good feasibility and rationality, and it shows that the research is reasonable for the fusion of multi-source remote sensing information. To sum up and combined with 3-3 can be obtained: the scheme for this study optimization plan. Using the SVM classification experiments on the preferred scheme, and to all remote sensing images as the experimental samples to improve the classification results of the scheme and persuasive, namely: at least five different side of multi-source remote sensing image with a total of 176 amplitude, the training sample 96 pieces (48), the test sample of 80 (10*8); total including 8 kinds of land use types, namely: woodland, grassland, terraces, residential areas, deserts, hills, mountains of the original, yangtao

From the table, we can see that based on scheme for the preferred scheme, selection of optimized parameters of the SVM classifier can realization based on texture of multisource remote sensing image land use classification and achieve good classification effect, also abandoned a cross validation show that the establishment of land use classification of texture is reasonable and feasible, thus proving of multi-source remote sensing data fusion for land use classification quality. In summary, scheme for the preferred scheme, SVM for the research project the preferred classifier, the entire remote sensing image centroid feature vector of constitutes the texture library; the texture library is a collection of the multi-source remote sensing data texture feature information.

4. Conclusion

Multi-source remote sensing image acquisition for this study provides the diversification of land use information and land use texture information, through the integration of multi-source remote sensing information, for texture is constructed to provide more abundant and accurate land use type characteristics to meet based on texture of multisource remote sensing image land use classification accuracy and applicability. This paper mainly according to the current land use classification of the status quo, combining texture extraction, classification method and the classification model, in order to improve the accuracy of classification of land use types for the purpose of research on this topic. The main features of this study are:

1. Using multi-source remote sensing images as the data source of classification, the classification results are more accurate and reliable;

2. put forward the idea of the land use classification method of remote sensing image based on texture;

3. Based on the BOW model theory, an algorithm for the above land use classification based on SVM is proposed;

4. By leave one out cross validation to validate the texture database. The results show that land use type texture creation can achieve good texture matching and classification function.

5. This study has good applicability, and can realize the identification and classification of various land use in the premise of ensuring the resolution of remote sensing image;

At the same time, there are some shortcomings in this research:

1. The lack of time for multi-source remote sensing images and the spectrum of multi-source;

2. Because of the complexity of remote sensing image data, remote sensing image acquisition height changes will affect land use classification accuracy. Therefore, it is necessary in the texture database with different heights of the remote sensing image to improve this problem.

3. The use of remote sensing image samples in this study is limited, and the establishment of texture library is lack of universality and integrity;

4. For a large number of complex remote sensing images, this research project may have the problem of low classification efficiency.

Outlook: the study currently only limited to remote sensing image land using types of sample test, and achieve a good identification effect, if can continue in the finite sample to derive a new sample, combined with neural network method based on remote sensing image texture of land use classification, the design method of this study will be further improvement and development.

Recebido/Submission: 10/04/2016

Aceitacao/Acceptance: 21/06/2016

Acknowledgments

This study was supported by the National Natural Science Foundation of China (No. 41401391), the Fundamental Research Funds for the Central Universities of China (No. 2014YB071), and the Exclusive Talent Funds of Northwest A&F University of China (No. 2013BSJJ017).

References

Chen, J. (2014). The land use classification method based on multi source remote sensing information fusion in Leizhou Province: a case study of Guangdong peninsula. Chinese Journal of ecology, (24), 7233-7242.

Di, X., Xi, Y. (2014). Chinese coast land use remote sensing classification system. Resources science, 3, 463-472.

Deng, K. (2008). TM remote sensing image classification based on multi source information. Northwest Agriculture and Forestry University, 50-55.

Hu, W. (2007). Remote sensing image texture information extraction method review. Yunnan geographic environment research, 3, 66-71.

Huang, Y., Li, P. (2003). Based on the application of geostatistical image texture in lithologic classification. Land resources remote sensing, 03, 45-49.

Liu, Y., Li, R. (2006). Research on image classification of land use dynamic monitoring based on multi source remote sensing: a case study of the Loess Hilly Gully Region of Northern Shaanxi Province. Bulletin of soil and water conservation, 26 (6), 63-66.

Li, N., Zhou, D. (2011). Community scale wetland classification and mapping of high resolution image support. Chinese Journal of ecology, 22, 6717-6726.

Li, S. (2012). Variety of land use classification method based on high precision remote sensing image. Northwest Agriculture and Forestry University, 40-45.

Liu, L., Kuang, G. (2009). Image texture feature extraction method. Chinese Journal of image and graphics, 04, 622-635.

Lan, Z. (2010). High resolution remote sensing image classification based on land use spatial knowledge mining. Wuhan University, 32-40.

Pereira, C., Ferreira, C. (2015). Identification of IT Value Management Practices and Resources in COBIT 5. RISTI--Revista Iberica de Sistemas e Tecnologias de Informacao, (15), 17-33.

Wang, B., Fan, B. (2013). An adaptive Mean shift tracking algorithm based on color texture association feature histogram. Journal of Nanjing University of Posts and Telecommunications, 03, 18-25.

Xu, Q. (2010). Semi supervised learning for remote sensing image classification research. Shaanxi Normal University, 60-65.

Yang, B. (2008). The different classification methods of land use and cover classification accuracy analysis. Inner Mongolia Agricultural University, 15-18.

Jinru Xue (1), Linya Huang (2), Baofeng Su (1), *

* bfs@nwsuaf.edu.cn

(1) College of Mechanical and Electronic Engineering, Northwest A&F University, Yangling 712100, Shaanxi, China

(2) College of Mechanical Engineering, Xi'an Jiaotong University, Xi'an 710049, Shaanxi, China
Table 5--Comparison of the results of different
classifiers based on the optimal scheme

SVM classifier

Before parameter    After parameter optimization
optimization

91.25%              95.0%
COPYRIGHT 2016 AISTI (Iberian Association for Information Systems and Technologies)
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2016 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Xue, Jinru; Huang, Linya; Su, Baofeng
Publication:RISTI (Revista Iberica de Sistemas e Tecnologias de Informacao)
Date:Aug 1, 2016
Words:5361
Previous Article:Nonlinear visual predictive control from invariant visual features.
Next Article:Adaptive sliding mode control based on RBF for slip ratio of electric vehicle.
Topics:

Terms of use | Privacy policy | Copyright © 2021 Farlex, Inc. | Feedback | For webmasters