LAND COVER CLASSIFICATION METHOD ORIENTED TO GEOGRAPHIC NATIONAL CONDITIONS INVESTIGATION

Developing the project of geographic national conditions investigation is in order to obtain land cover change information which is caused by natural and human social and economic activities, and serve the information for government, enterprise and public. Land cover is an important method to describe the geographic national conditions information, which can truly reflect the land surface material type and its natural attribute. It has been contained in the content system preliminary scheme as an important portion. In this paper, it discusses and analyzes on the method and key technology, with according to the land cover content that geographic national conditions watches on. Then it evaluates the applicability of automatic classification method using in land cover information extraction, and comprehensively analyzes various extraction methods’ maximum effectiveness. Finally, it proposes a method that is based on high spatial resolution remote sensing imagery and can be used in engineering applications, which provides a reference method for geographic national conditions investigation. * Corresponding author.


INTRODUCTION
Land cover information obtaining is the primary mission of the project of geographic national conditions investigation; it's also the basis of the project.Utilizing some classification technology and methods, we can get the range, area, and spatial distribution information of land cover, such as plough, garden plot, woodland, lawn, building, road, construction, artificial stack, bare land, water, and so on, which constitute the classification system of the project and can provide important evidence for index information's statistic and proclamation.
There are five series of global land cover products issued so far, correlative research is also developed in China.However, the scale of global land cover research is too small, and it aims at satisfying the requirement of global change and earth's mode simulation, which uses lower spatial resolution images.
Secondly, the land cover classification systems in different areas are established only for the requirement of each area individually.
Oppositely, the classification system established in the project of geographic national conditions investigation is from the point of view of monitoring the changing information of geographic objects throughout the country, which has synthetically considered the requirements of multi-areas, and has referenced the national standards, technology norms and land cover classification systems that have been issued and applied broadly.Besides, in this project, it uses middle and high spatial resolution remote sensing images to extract information.
There are many methods to obtain land cover information based on remote sensing images, such as automatic classification, automatic classification by assistant thematic data, artificial interpretation and editing, outside investigation, and so on.In which the most efficient method is automatic classification.
However, mostly successful cases are oriented to thematic objects' extraction, or belong to researches in small areas, ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume II-4, 2014ISPRS Technical Commission IV Symposium, 14 -16 May 2014, Suzhou, China This contribution has been peer-reviewed.The double-blind peer-review was conducted on the basis of the full paper.
doi:10.5194/isprsannals-II-4-29-2014 which can't satisfy the requirements of wide range and synthetically objects' extraction in the project of geographic national conditions investigation.The paper proposes a feasible method based on analyzing the theory of classification methods, which can provide references for the project of geographic national conditions investigation.

INFORMATION EXTRACTION'S PRINCIPLES
The project of geographic national conditions investigation defines ten first-level classes according to certain principle and theory, including plough, garden plot, woodland, lawn, building, road, construction, artificial stack, bare land and water.Of course, it also defines some second-level and third-level classes.
As mentioned in the introduction, there are many methods for obtaining land cover information.However, each method has its respective process flow.Which method and process flow will be selected is mainly decided by the factors of image's quality, land cover's type, the abundance of fundamental geography information data and thematic data and so on.

Principle which is decided by image's quality
(1) If the image's quality is very good, which includes that the image is multispectral, clear and has no cloud, the temporal phase of the image is suitable, we can firstly use the method of automatic classification to obtain the preliminary results, and then use the method of artificial interpretation and editing to edit and perfect the results.
(2) Inversely, if the image's quality is not good, such as the image is not clear or has cloud, or the temporal phase is not suitable, we should use the method of artificial interpretation and editing; just rely on the experiential knowledge.

Principle which is decided by land cover's type
(1) For the types of natural land cover, such as water and vegetation, which can be classified automatically comparatively simple, we should use the method of automatic classification firstly; then, obtain the lower classes' types of plough, garden plot, woodland, lawn and so on based on the results.
(2) For the types of artificial land cover, such as building, construction, artificial stack, and some type of natural land cover, such as bare land, which cannot be classified automatically easily and accurately, we should firstly use the method of artificial interpretation and editing.
(3) Significantly, for some especial type of artificial land cover, such as road, which has especial geometry characters, for example, the width is uniform at each point in the same road, and the ratio of length to width is always very large.In this instance, we could use the method of automatic classification to obtain road's geometric information, and also we could use the method of artificial interpretation and editing.
The two methods can be complementary for each other.

Principle which is decided by the abundance of fundamental geography information data and thematic data
(1) If the fundamental geography information data or the thematic data is abundant and the quality is good, we can make full use of these data for classification.Especially for the land cover types which must be obtained by the method of artificial interpretation and editing, we can firstly obtain them from fundamental geography information data or thematic data; then, we can update it on the basis of images; based on these results, we could use the method of automatic classification to obtain other land cover types' information.
(2) Inversely, if the fundamental geography information data and the thematic data are not abundant, we can firstly use the method of automatic classification to obtain the preliminary results, and then use the method of artificial interpretation and editing to edit and perfect the results.

Results' integration
For the results of land cover classification, no matter which method is used to obtain the information, the final results must cover the total region, whatever we divide the unit into parts of image's scene, frame or administrative district.Besides, all the polygons must not overlap in the spatial coordinate system.So, for the same unit region, the result of land cover information obtained frontally should be as a basis for the following type.The results of land cover classification must be integrated to ensure the spatial relationships are reasonable and have no loopholes.
efficiency and precision.The guideline is furthest using automatic classification methods to obtain land cover information, and then using some other methods when the results' precision can't achieve the request.It's best to take on high efficiency and high precision.
Automatic classification method mainly uses the objects' characters of spectrum, texture and shape to build adaptive extraction regulations and automatically extract all levels' land cover information.The process flow of inside processing remote sensing images is: (1) Pre-process remote sensing image This process includes images' bands stacking, merge, ortho-rectification, and so on, which would construct fine foundation for the following processing.
(2) Segment image In order to obtain feature objects, firstly, we need to analyse pixels' spectral and shape characters and calculate eigenvalue for each band's spectral heterogeneity and shape heterogeneity; then, we need to define some parameters including scale, image layer weights, compactness, and segment image to feature objects by using multi-resolution segmentation arithmetic.
(   It's suitable to use multi-spectrum images of 10 meters resolution to extract first-level classes, 5 meters resolution to extract second-level classes, and 1 meter resolution to extract third-level classes.Furthermore, the higher spectrum resolution of remote sensing images, the better to improve the results' precision.Thus, if conditions permit, it's best to use high spectrum resolution images as far as possible. (2) Analyse objects' characters This method not only utilizes spectrum characters, but fully utilizes abundant spectrum, texture, shape, and spatial position characters to extract land cover information.The core of this method is images' segmentation.Segmentation arithmetic includes multi-resolution, grey-based, texture-based, and knowledge-based and so on, in which the multi-resolution arithmetic is used mostly.This arithmetic synthetically utilizes land covers' spectrum and texture characters, calculates the spectrum and shape differences between each bands, then sets certain threshold, when all the segmentation objects' characters' value extends the threshold, the segmentation process is completed.On the base of these processes, it could build proper decision regulations to extract land cover information.
Another method named decision-tree based on expert's knowledge is also a commonly used method.This method classifies pixels mainly according to spectrum characters, spatial relation, and other context relations between pixels.It's generated by analysing and concluding of mass samples' attributes.
A decision-tree consists of one root node, a series of inner nodes and offshoots, and some leaf nodes.Each inner node stands for an attribute will be tested in decision process.Each offshoot stands for a result in test.Different attribute forms different offshoot.And each leaf node stands for one class, which is the classification result.
In the project of geographic national conditions investigation, it's better to select proper classification method according to the images' spatial resolution and spectrum resolution, and improve the automatic classification's precision as far as possible, which could reduce the following workload and improve classification results' efficiency.
(5) Deal with special circumstances Firstly, if there is only panchromatic image in some regions, which would influence the information extraction of the types of natural land cover seriously.In this instance, we should collect some middle and lower resolution remote sensing images, such as 30 meters, 10 meters and 5 meters resolution and so on, whose spectral information would be very useful and can be assistant in land cover classification.
Secondly, if there are a lot of clouds in the high resolution optical image in some regions; furthermore, it is very difficult to obtain some good quality images in these regions, which are also influence the information extraction very seriously.In this instance, we should collect some high and middle resolution SAR remote sensing images whose resolution should be higher This contribution has been peer-reviewed.The double-blind peer-review was conducted on the basis of the full paper.doi:10.5194/isprsannals-II-4-29-2014 than 5 meters, and some middle and lower resolution remote sensing images, which can help classification.
From SAR remote sensing images, we could extract the types of artificial land cover; besides, water exhibits strong mirror reflex characters, based on this point, and we could extract water information from SAR remote sensing images.
Then, extract other types of land cover by using multi-source optical remote sensing images.

CONCLUSIONS
(1) At present, there are many automatic classification methods based on remote sensing images.However, because of the images' complexity, there isn't one method whose applicability is universal and has high efficiency.In practical application, it should find out a most proper classification method to achieve high precision as far as possible according to the type of images used in project.
(2) The classification efficiency and results' precision in the project of geographic national conditions investigation are very high.Therefore, it's inevitable to interpret and edit by artificial process when the automatic classification can't satisfy the requirement.
(3) Traditional Pixel-Based classification method is suitable for middle, low spatial resolution multi-spectrum images.Aim at extracting information from high spatial resolution images, it's better to utilize Object-Oriented method.This method's technology flow mainly includes pre-processing remote sensing images, obtaining objects' original characters, building indexes, inset assistant datum if necessary, selecting samples, extracting land cover information, analyse and evaluate the results' quality.

Quality checking
Land cover products  Based on this method, the first-level classes land cover information would be extracted by automatic classification, partial second-level classes should be added artificial interpretation and editing, and even outside investigation.This method is suitable for the project of geographic national conditions investigation and can provide references for this project.
Figure 1.R, G, B layer stacking images and spectral profile of land cover on each band

Figure 1 ( 6 )
Figure 1 exhibits that some indexes can be built on the basis of spectral characters, for example, NDWI (Normal Difference Water Index) = ( G -N ) / ( G + N ); Ratio B = B /

Figure 2 .
Figure 2. Work flow of automatic classification by computer

ISPRS
Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, VolumeII-4, 2014   ISPRS Technical IV Symposium, 14 -16 May 2014, Suzhou, China    This contribution has been peer-reviewed.The double-blind peer-review was conducted on the basis of the full paper.doi:10.5194/isprsannals-II-4-29-2014Objects on earth have many kinds of characters.The primary task of automatic classification is to analyse these characters.At present, the main characters used in automatic classification are spectrum characters, which are the basis of texture and shape characters and are the most direct information in images.To achieve automatic classification, it could directly build decision regulations or build some indexes based on objects' characters.(3)Build regulations for information extraction It's very important to build suitable decision regulations in information extraction, which would exhibit good capability in automatic classification.For example, NDVI index is always used to extract vegetation information, NDWI index is always used to extract water information.They are both built by ratio computing.The theory of this arithmetic is to find out the bands which the objects' exhibit strongest reflex and weakest reflect characters, and then put the strongest one on the numerator, the weakest one on the denominator, which can enhance the researching objects and bate the others.Influenced by the factor of hypsography, the spectrum reflex would change.And land covers' grown area is restricted by the height above sea level, slope and aspect.Therefore, it could improve the results' precision by using assistant DEM and DSM datum.According to the present research status, with the improvement of images' spatial resolution, the character difference between each land cover class is very tiny, which brings difficulty on building decision regulations for information extraction.The extraction result of single class always mixes the other land cover information.So, it should build detailed decision regulations as far as possible to obtain perfect results.(4) Select classification methods Traditional Pixel-Based classification method based on remote sensing images is on the conditions that the spectrum information is very abundant and the difference of spectrum between different land cover classes is obvious.In high resolution remote sensing images, the land covers' construction and texture are very outstanding, but spectrum character differences between different land covers are not that outstanding.Traditional Pixel-Based method would not satisfy the present requirement for information extraction.For satisfying the new application situation nowadays, the Object-Oriented classification method provides new solution in high resolution remote sensing images' information extraction.

Figure 3 .
Figure 3. Flow charts of classification based on high spatial resolution remote sensing imagery ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, VolumeII-4, 2014   ISPRS Technical Commission IV Symposium, 14 -16 May 2014, Suzhou, China