Study on Multi-Scale Window Determination for GLCM Texture ... - MDPI

International Journal of

Geo-Information Article

Study on Multi-Scale Window Determination for GLCM Texture Description in High-Resolution Remote Sensing Image Geo-Analysis Supported by GIS and Domain Knowledge Zeying Lan 1, * 1 2

*

ID

and Yang Liu 2

School of management, Guangdong University of Technology, 169 Yinglong Road, Guangzhou 510520, China Geographic Information Department, Guangzhou Urban Planning & Design Survey Research Institute, 10 Jianshe Road, Guangzhou 510060, China; [email protected] Correspondence: [email protected]; Tel.: +86-159-890-88556

Received: 18 March 2018; Accepted: 28 April 2018; Published: 5 May 2018

Abstract: Texture features based on the gray-level co-occurrence matrix (GLCM) can effectively improve classification accuracy in geographical analyses of optical remote sensing (RS) images, with the parameters of scale of the GLCM texture window greatly affecting the validity. By analyzing human visual attention characteristics for geo-texture cognition, it was found that there is a strong correlation between the texture scale parameters and the domain shape knowledge in a specified geo-scene. Therefore, a new approach for quickly determining the multi-scale parameters of the texture with the assistance of a geographic information system (GIS) and domain knowledge is proposed in this paper. First, the validity of domain knowledge from an existing GIS database is measured by spatial data mining algorithms, including spatial partitioning, image segmentation, and space-time system evaluation. Second, the general domain shape knowledge of each category is described by the GIS minimum enclosing rectangle indices and rectangular-degree indices. Then, the corresponding multi-scale texture windows can be quickly determined for each category by a correlation analysis with the shape indices. Finally, the Fisher function is used to evaluate the validity of the scale parameters. The experimental results show that the multi-scale value keeps a one-to-one relationship with the classified objects, and their value ranges are from a few to tens, instead of the smaller values of a traditional analysis; thus, effective texture features at such a scale can be built to identify categories in a geo-scene. In this way, the proposed method can increase the total number of categories for a certain geo-scene and reduce the classification uncertainty, as well as better meet the requirements of large-scale image geo-analysis. It also has as high a calculation efficiency and as good a performance as the traditional enumeration method. Keywords: window of GLCM texture descriptor; multi-scale; geographic information system (GIS); domain knowledge; spatial data mining; Fisher function

1. Introduction Texture is an important image spatial feature [1–3]. Many texture descriptors have been developed in the past, such as frequency domain analysis based on the Fourier transform, the wavelet transform approach, the hidden Markov approach, the local binary pattern (LBP), the local multiple patterns (LMP), the geostatistical-based approach, the watershed-based approach, the gray-level co-occurrence matrix (GLCM)-based approach, and other statistical approaches [2–18]. Most of these approaches are mainly designed for a simple and clear classification system or a specific object, similar to in the ISPRS Int. J. Geo-Inf. 2018, 7, 175; doi:10.3390/ijgi7050175

www.mdpi.com/journal/ijgi

ISPRS Int. J. Geo-Inf. 2018, 7, 175

2 of 24

fields of medicine, industry, the military, etc. [3–18]. However, when using high-resolution optical remote sensing (RS) images for geographical classification, the classification system will change with the needs of different geo-scenes, even in the same image, so the phenomena of “same object different spectrum”, “same spectrum different object”, and “mixed pixels” will become more prevalent, and the textures will be more complex and changeable [1,2,19–30]. Scholars have emphasized that such image analysis can be improved by employing geo-domain knowledge (especially various general statistical characteristics such as topology, location, spatial relationship, etc.) [19–30]. Thus, in this paper, a GLCM-based approach that can be easily associated with geo-knowledge is chosen to build texture features for geo applications. This is because the GLCM obtains statistics by using a certain scale window, where one can intuitively observe the actual objects and reflect the domain knowledge (mainly spatial shape attributes). Moreover, it provides information in image gray direction, interval and change amplitude, so that 14 kinds of texture features can be effectively defined based on it [3], and it meets the requirement for the classification of complex and variable geo-scenes [31]. Since 1990, scientists have created studies on the impact of various parameters on the validity of GLCM-based texture, and they found that it is very sensitive to the scale parameter, and the classification accuracy can even decrease if the parameter is unreasonable [31–36]. Marceau D.J. (1990) first used SOPT data to conduct a quantified study [31]. He used the factor analysis method to measure the impact on window size, quantization level, and four texture statistics (Energy, Contrast, Homogeneity, Entropy) for land cover types. The impacts are as follows: the scale of the window has the greatest effect on the classification, followed by the texture statistics, and the least significant is the level of quantization [31]. Later, other researchers also successively obtained similar study results [32–37]. In Cui L.’s Ph.D. dissertation (2005), this issue was described in depth [32]. She measured the impact on window scale, direction and level of quantization, and found that different directions had little effect on the classification reliability of most types, and the best solution was to use the average value of the omni direction. Different bands play a crucial role in the extraction of texture features for multispectral TM data, so the use of combined band information is important. In addition, she took the scale parameters from 3 to 50 separately for the TM image analysis and found the variation rule of 8 GLCM texture features of 14 types of geo-objects by using the enumerated window scale. She found that there was a correspondence between the different texture features and the different windows. When the window was small, the feature space composed of a combination of 2 texture features achieved good results. When the window was gradually increased, the combination dimension of the texture feature space was also increased, and the combination of 5 texture features achieved best results. Moreover, she found that the overall classification accuracy gradually increased with scale, reaching the highest value when the scale reached 35. Therefore, she concluded that the optimal identification windows of different objects were different, although only an optimal scale of 35 was used in the subsequent classification [32]. That is, multi-scale is an essential and necessary feature of spatial data, and it has a significant impact on GLCM-based texture features [31–43]. In many practical applications, the choice of scale parameters is still very arbitrary when using empirical method, and a small scale value is commonly chosen, either 3, 5, 7 or 9. It is easy to operate, but also extremely unreliable [32–34]. In addition, researchers have also developed various statistical methods for analyzing texture scales, including a histogram algorithm, a variogram algorithm, a Markov algorithm, a mean shift algorithm, and a fractal network evolutionary algorithm of Ecognition, etc. [1,8–18]. Additionally, Binaghi, E. et al. (2003) set the scale parameters by taking the statistics of the spatial characteristics of sampled data [43]. These methods can sometimes obtain multi-scale parameters. Therefore, they are able to achieve adaptive estimation to some extent and can improve the validity of texture features. However, the number of scales was random and few, and the values of the scales were usually low. In terms of the algorithms themselves, local optima or independent of actual categories will appear easily if they rely too much on the sample selection or the local image gray level variability analysis, or the errors will accumulate if they adopt an iterative calculation model. It can be seen that there are still many uncertainties in current studies for the selection of texture scale [31–44].


3 of 24

Some are only aimed at a certain application, some are only able to obtain a single scale parameter, and some are able to obtain multi-scales, but each time the results are commonly inconsistent and small for the number and value of scales. Therefore, it is necessary to find a quick and reliable setting method for multi-scale image texture in geo-analysis. To solve the problems above, this paper proposes a new method for determining multi-scale windows of image GLCM texture descriptors by means of the integrated use of a geographic information system (GIS) and geo-spatial domain knowledge. First, we explain the rules for the changes in texture scales under different geo-scenes and classification systems and construct a framework for texture scale recognition and extraction by simulating human visual attention and interpretation characteristics [1,2,19–29]. Second, a body of effective geo-spatial domain knowledge is mined from an existing GIS database. Then, the link between domain knowledge and texture scale is established, so that multi-scale windows can be quickly determined for each category. The approach uses geo domain knowledge that is extracted from different GIS informative layers (which is likely to be cheap to source from historical databases) to define the characteristic size and shape of the geo-scene categories of interest. These informative layers can be topographic maps, digital elevation models (DEM), land-use maps, etc. It must be noted that there is a possibility that these maps could cover a wide temporal range, and due to this potential temporal variability of the landscape, the accuracy of the domain knowledge should be taken into account. Finally, the Fisher function analysis based on the enumeration method is used to evaluate the validity of the above multi-scale parameters [44–47]. In this paper, experiments were designed to distinguish the categories of a certain geo-scene based on GLCM texture features with multi-scales by using high-resolution optical RS images and existing GIS data in the study area. The results show that the scale parameters of geo-textures can be quantified with a one-to-one model, making them easier to understand and express, and the proposed method shows a significant improvement in efficiency with similarly good performance when compared with the traditional enumeration method. Furthermore, GLCM texture descriptors under multi-scale windows can evidently represent different thematic categories; thereby, they will increase the total number of categories in the classification system, reduce the classification uncertainty, and better meet the requirements of large-scale image geo-analysis. Therefore, the integrated use of GIS and domain knowledge is proved feasible in image geo-texture analysis, and able avoid the disadvantages of local optimization, sample dependence, and initial error accumulation of iterative calculations. 2. Overall Framework GLCM is actually a statistical matrix of joint conditional probability distribution between the gray levels of image pixel pairs with some distance and direction within a certain window, and is used for texture descriptors [1–3]. Each geo-texture observation window based on GLCM must reflect the real spatial characteristics of the corresponding category. For example, in the classification system of land cover of I level, “woodlands” are taken as a single classified object. Thus, its texture unit is expanded, and a larger observation window and complex spatial structures are formed. In a classification system of land cover of II level, woodlands are subdivided into four sub-types: Dense woodland, Shrub woodland, Sparse land and other woodlands [31–33,48,49]. Obviously, for a texture unit of sub-types, its corresponding observation window should be relatively microscopic in the same place, and the spatial structure relatively simple. A case in which the classified geo-objects vary with the change of geo-classification system in an image can be seen in Figure 1. In the geo-analysis of an image, each category should have its own texture observation window, which is strongly correlated with the actual definition and shape of the classified objects [20–39]. In addition, the scale selection algorithm is essentially equivalent to setting the image pixel number as the optimal scale according to the size of the object shape and the image resolution. Therefore, this paper uses GIS data and technology to mine domain knowledge (mainly the statistical characteristics of objects’ shapes), in order to determine the multi-scale of GLCM texture observation windows. The overall framework is designed as shown in Figure 2; the steps will be described below.

expanded, and a larger observation window and complex spatial structures are formed. In a classification system of land cover of II level, woodlands are subdivided into four sub-types: Dense woodland, Shrub woodland, Sparse land and other woodlands [31–33,48,49]. Obviously, for a texture unit of sub-types, its corresponding observation window should be relatively microscopic in the same place, the spatial relatively simple. A case in which the classified geo-objects vary4 of with ISPRS Int.and J. Geo-Inf. 2018, 7, structure 175 24 the change of geo-classification system in an image can be seen in Figure 1.

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW

4 of 24

Figure 1. Schematic diagram of different classified objects under different classification systems in a region. (a) A RS images; (b) The I level land-cover system; (c) The II level land-cover system.

In the geo-analysis of an image, each category should have its own texture observation window, which is strongly correlated with the actual definition and shape of the classified objects [20–39]. In addition, the scale selection algorithm is essentially equivalent to setting the image pixel number as the optimal scale of the object shape and (c) the image resolution. Therefore, this (a) according to the size(b) paper uses GIS data and technology to mine domain knowledge (mainly the statistical characteristics Figureshapes), 1. Schematic diagram of different objects under different in The of objects’ in order to determine theclassified multi-scale of GLCM texture classification observation systems windows. a region. (a) A RS images; (b) The I level land-cover system; (c) The II level land-cover system. overall framework is designed as shown in Figure 2; the steps will be described below. Existing GIS data

Spatial partition Image segemation by GIS

High-resolution RS image

Evaluating the effectiveness of domain knowledge

Extracting GLCM texture features based on enumerated scales

Computing minimum enclosing rectangle indexes

Measuring features separability based on Fisher function

Measuring rectangularity of each polygon

Making multi-scale of enumerated method as evaluation criterion

Choosing effective polygons Setting multi-scale based on frequency histogram

Comparison and Analysis

Figure 2. Scheme of GLCM texture multi-scale selection and evaluation based on GIS and domain Figure 2. Scheme of GLCM texture multi-scale selection and evaluation based on GIS and knowledge. domain knowledge.

Evaluating the validity of domain knowledge by spatial partitioning and image segmentation Evaluating the validity of domain knowledge by spatial partitioning and image segmentation The validity of the domain knowledge must be taken into account when using it to analyze Thewindows. validity of the domain knowledge must taken intoan account when it to there analyze texture Because domain knowledge is be mined from existing GIS using database, is texture windows. Because domain knowledge is mined from an existing GIS database, there is potential knowledge variability in the two time-phases, i.e., from GIS data to RS data. To obtain more potential domain knowledge variability in reduce the twothe time-phases, from GIS data data. Toresults, obtain some more effective knowledge and impact ofi.e., such changes on to theRS analysis effective domain knowledge and reduce the impact of such changes on the analysis results, some spatial data mining algorithms are employed, including spatial partitioning and image segmentation spatial miningpartitioning algorithms are spatial partitioning image segmentation by GIS.data The spatial canemployed, ensure the including overall spatial consistency of aand geo-scene, and the GISby GIS. The spatial partitioning can ensure the overall spatial consistency of a geo-scene, and the the based image segmentation can further measure the stability of domain knowledge by analyzing GIS-based image segmentation can further measure the stability of domain knowledge by analyzing change rule of categories in segmentation blocks. the change rule of categories in segmentation blocks. • Domain knowledge mining by GIS • In Domain knowledge mining by GIS a stable time-space system, a set of effective polygons with representative shapes can be • •

obtained existing GIS database, and generalpolygons shape characteristics of each category can be In a from stableantime-space system, a set ofthe effective with representative shapes can be mined by employing GIS shape indices. obtained from an existing GIS database, and the general shape characteristics of each category can be by employingofGIS shape indices. •mined Determination multi-scale GLCM texture windows

to the principle ofGLCM determination, we set up a one-to-one model of the correlation • According Determination of multi-scale texture windows between domain shape knowledge and texture scale to extract the appropriate multi-scale window According to features. the principle of determination, we set up a one-to-one model of the correlation for GLCM texture between domain shape knowledge and texture scale to extract the appropriate multi-scale window for •GLCM Evaluation of multi-scale parameters texture features. To quantitatively evaluate the validity of texture scales extracted by GIS and domain knowledge, we employ the Fisher distance method, which can measure the separability of GLCM texture features


•

5 of 24

Evaluation of multi-scale parameters

To quantitatively evaluate the validity of texture scales extracted by GIS and domain knowledge, Int. J.the Geo-Inf. 2018,distance 7, x FOR PEER REVIEW of 24 weISPRS employ Fisher method, which can measure the separability of GLCM texture 5features under different scales; and the higher the separability, the better the setting of the corresponding scale under different scales; and the higher the separability, the better the setting of the corresponding scale parameters. First, a set of criteria for multi-scale parameters is obtained by measuring the separability parameters. First, a set of criteria for multi-scale parameters is obtained by measuring the separability of texture features based on enumerated scales. Then, the results deduced by the proposed method of texture features based on enumerated scales. Then, the results deduced by the proposed method can be evaluated by means of a comparison analysis. can be evaluated by means of a comparison analysis. Overall, the progressive visual attention and cognitive process for the geo-texture scale in a geoOverall, the progressive visual attention and cognitive process for the geo-texture scale in a geoscene can be seen and from from the thewindow windowtotothe thedetails” details” [21,45–51]. scene can be seenasas“from “fromthe thewhole whole to to the the local, local, and [21,45–51]. TheThe process is illustrated process is illustratedininFigure Figure3:3:

Visual attention pyramid

F1--Spatial partition of a geo-scene

F3--Multi-scales texture observation window

F2--Image segmentation

F4--Classified objects

Figure 3. Schematic diagram of the visual attention and cognitive process of the texture scale using Figure 3. Schematic diagram of the visual attention and cognitive of the texture scalelarge using GIS and domain knowledge. The cognitive pyramid summarizes theprocess visual attention space from GIStoand domain knowledge. The cognitive pyramid summarizes the visual attention space from small as F1 to F4. F1 shows the spatial partition result as “A, B, C” by using GIS and domain large to small as F1 to F4. F1 shows the spatial partition result as “A, B, C” by using GIS and domain knowledge. F2 shows the image segmentation blocks in the partition region C. F3 shows a set of knowledge. shows the image segmentation blocks inshows the partition region C. F3 shows a can set of observationF2 windows of image texture from “S1 to S8”. F4 the actual classified objects that observation windows of image texture fromthe “S1multi-scales to S8”. F4 shows theS8”. actual classified objects that can be identified by the texture features under of “S1 to be identified by the texture features under the multi-scales of “S1 to S8”.

3. Methods 3. Methods 3.1. Domain Shape Knowledge Mining by GIS 3.1. Domain Shape Knowledge Mining by GIS 3.1.1. Spatial Partition 3.1.1. Spatial Partition By considering various factors (including natural, social, and economic factors) that affect a certain geo-scene various (such asfactors the I level land-cover scene and and II level land-cover scene, thea rural By considering (including natural, social, economic factors) thatand affect certain land-use scene and urban land-use scene, etc.), the original GIS data can be divided into several geo-scene (such as the I level land-cover scene and II level land-cover scene, and the rural land-use homogeneous over a whole range. A partition the pattern rules of homogeneous a geo-scene, scene and urban pieces land-use scene, etc.), the original GISpiece datacan canretain be divided into several suchover as the consistency landscape units, the the consistency of ecological functions, the as pieces a whole range.ofAnatural partition piece can retain pattern rules of a geo-scene, such consistency of topology, or the integrity of administrative divisions [48,49,52]. The technology to the consistency of natural landscape units, the consistency of ecological functions, the consistency achieve this process is referred to as spatial partitioning, including the overlay analysis method, the dominant factor method, the multi-factor comprehensive evaluation method and the cluster analysis method [52]. Based on a comprehensive comparison, this paper adopts the dominant factor method


6 of 24

of topology, or the integrity of administrative divisions [48,49,52]. The technology to achieve this process is referred to as spatial partitioning, including the overlay analysis method, the dominant factor method, the multi-factor comprehensive evaluation method and the cluster analysis method [52]. Based on a comprehensive comparison, this paper adopts the dominant factor method combined with ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 6 of 24 the overlay analysis method, which considers the importance of various impact factors of a certain geo-scene andwith maintains the topology and continuity region. of various impact combined the overlay analysis integrity method, which considers of thethe importance Specifically, we extract various factor layers for a certain geo-scene from existing GIS databases, factors of a certain geo-scene and maintains the topology integrity and continuity of the region. such as digital elevation modelvarious (DEM)factor maps, administrative maps, main river maps, road maps, Specifically, we extract layers for a certain geo-scene from existing GISmain databases, digital elevation model (DEM) maps, administrative maps, main river maps, main roadimportance. maps, etc., such andasthen overlay polygons of these factor layers sequentially according to their etc., then overlay polygons of these factor layers sequentially according their importance. When theand overlapping layer is inconsistent with the boundary of basic units to (like a land-use polygon When the overlapping layer is inconsistent with the boundary of basic units (like a land-use polygon in a land-use scene), we adopt the maximum area principle to adjust the conflicting boundary (see in a land-use scene), we adopt the maximum area principle to adjust the conflicting boundary (see Figure 4). Finally, we obtain a series of homogeneous partitioned regions for which the mined domain Figure 4). Finally, we obtain a series of homogeneous partitioned regions for which the mined domain knowledge is basically consistent with that at the corresponding time of the GIS database. knowledge is basically consistent with that at the corresponding time of the GIS database.

DEM, administrative map,…

Basic units

Layer A2

Inconsistent boundaries

Layer A3

Boundary adjustment

Layer A4 Figure 4. Diagram of spatial partitioning using GIS overlapping analysis. Layer A1 shows an auxiliary Figure 4. Diagram of spatial partitioning using GIS overlapping analysis. Layer A1 shows an auxiliary map, such as a Digital Elevation Model (DEM), administrative map, etc. Layer A2 shows the basic map, such as a Digital Elevation Model (DEM), administrative map, etc. Layer A2 shows the basic polygon units in the GIS database. Layer A3 shows the situation in which the new boundary formed polygon units in the GIS database. Layer A3 shows the situation in which the new boundary formed by overlaying is inconsistent with the boundary of the basic units. Layer A4 shows the spatial partition by overlaying is inconsistent with the boundary of the basic units. Layer A4 shows the spatial partition map formed by adjusting the conflicting boundary based on the maximum area principle. map formed by adjusting the conflicting boundary based on the maximum area principle.

3.1.2. Image Segmentation

3.1.2. Image Segmentation In the partitioned region, above, the image segmentation blocks are obtained by cutting the RS

image the GIS basic unitabove, data. Then, the frequency histogram for are the mean grayby value of the In thewith partitioned region, the image segmentation blocks obtained cutting segmentation blocks of each category is drawn in order to analyze the unchanged rate of each RS image with the GIS basic unit data. Then, the frequency histogram for the mean gray value of category within the two time-phases from GIS data to RS data, with the results indicating the degree segmentation blocks of each category is drawn in order to analyze the unchanged rate of each category of domain knowledge validity in this partitioned region. within the two time-phases from GIS data to RS data, with the results indicating the degree of domain First, the single-band image is calculated, instead of the original multi-band image, by using a knowledge in this partitioned region. classical validity dimension-reduction algorithm (e.g., principal component analysis or color space First, the single-band image is calculated, instead of theisoriginal by using conversion), so that only one piece of gray value information stored formulti-band each pixel, image, for efficient a classical dimension-reduction algorithm (e.g., principal component analysis or color space conversion), calculation. For a rough estimate of the validity of domain knowledge, the description of the grayscale still reflects thevalue combination information of bands, such as the color After the so that only image one piece of gray information is stored for each pixel, for image. efficient calculation. the matrix dimension imageknowledge, is decreased,the butdescription the speed of is greatly For agrayscale, rough estimate of the validityofofthe domain of operation the grayscale image still

ω i , we improved, and the gradient information is still retained. Then, for image. categoryAfter the frequency reflects the combination information of bands, such as the color theuse grayscale, the matrix dimension of the image decreased, but of theimage speed of operation is greatly improved, and the gradient histogram to count theismean gray value segmentation blocks. information is still retained. Then, for category ωi , we use the frequency histogram to count the mean nωi − k gray value of image segmentation P blocks. (r ) = spectrum −ωi

In the equation above, abscissa.

Pspectrum−ωi (rk )

rk

and

k

Pspectrum−ωi (rk ) =

N spectrum −ωi n

(1)

ωi − k

Nspectrum−ωi

is the k-th gray level of the image, expressed as the histogram

nωi −k , respectively, indicate the frequency and numbers of the k-th

(1)


7 of 24

gray that equals the mean gray value of the segmentation blocks of class


expressed as the histogram ordinate, and

N spectrum−ωi

ω i . Pspectrum −ω (rk ) i

is

7 of 24

is the total number of segmentation blocks

In the equation ω i . above, rk is the k-th gray level of the image, expressed as the histogram abscissa. of category Pspectrum−ωi (rk ) and nωi −k , respectively, indicate the frequency and numbers of the k-th gray that equals the mean gray value the segmentation 3.1.3. Validity Evaluation ofofDomain Knowledgeblocks of class ω i . Pspectrum−ωi (rk ) is expressed as the histogram ordinate, and Nspectrum−ωi is the total number of segmentation blocks of category ω i . We analyze the distribution characteristics of the above histograms of segmentation blocks in order to describe the validity of the knowledge. 3.1.3. Validity Evaluation of Domain Knowledge In theory, if one of the above frequency histograms shows an obvious peak zone, then it indicates the distribution characteristics of the above histograms segmentation blocks in that We the analyze corresponding image segmentation blocks belong mainly to oneofcategory. Furthermore, if order to describe the validity of the knowledge. the category is the same as the historical one in the GIS database, then it indicates that these blocks if onechanges of the above histograms antime obvious peak zone, then it indicates haveIn notheory, significant in thefrequency land pattern betweenshows the two phases. Assuming that the peak that the corresponding image segmentation blocks belong mainly to one category. Furthermore, if the the  , we ，rωi −up then rωiGIS zone corresponds range can calculate the blocks number of − down category is the sametoasthe the gray historical oneasin  database, it indicates that these have no significant changes in the between time phases. Assuming that peak segmentation blocks that areland not pattern in the peak zonethe to two quantify this changed rate V. Asthe Yang C.zone and corresponds to the gray as rωi −down , rωithan , we can calculate the number of segmentation Zhou C. (2001) noted, if range the proportion is less the threshold value η, the time-space system is −up blocks thatstable, are not in itthe peak satisfy zone tothe quantify thisformula changed[25]: rate V. As Yang C. and Zhou C. (2001) relatively and should following noted, if the proportion is less than the threshold value η, the time-space system is relatively stable, rωi −up and it should satisfy the following formula [25]: Vωi = (1 − pspectrum −ωi ( rk )) × 100% ≤ η (2)

∑

−up rk = rωriω−idown

Vωi = (1 −

∑

rk =rω −down i

pspectrum−ωi (rk )) × 100% ≤ η

(2)

3.1.4. Shape Description 3.1.4. Shape Description In a stable time-space system, the domain shape knowledge of a geo-scene can be strongly In a stable the domain knowledge of a geo-scene can be strongly associated with time-space the texture system, scale parameters. Byshape comprehensively comparing various descriptive associated the texture parameters. By comprehensively comparing various descriptive operators ofwith the object shape,scale we adopt the minimum enclosing rectangle algorithm (MER). The MER operators of the object shape, we adopt the minimum enclosing rectangle algorithm (MER). The MER refers to the minimum range that can contain a polygon’s shape, expressed in two-dimensional refers to the The minimum range that can comprises contain a polygon’s shape,one expressed in two-dimensional coordinates. computation method two main steps, is to generate the enclosing coordinates. The computation method comprises two main steps, one is to generate the enclosing rectangle (ER)—that is, set a rectangle under the polygon borders of the maximum abscissa, the rectangle (ER)—that is, set a rectangle under the polygon borders of the maximum minimum abscissa, the maximum ordinate, and the minimum ordinate. Then, the MER of a abscissa, polygon the abscissa, the maximum ordinate, the at minimum ordinate. the MER a polygon canminimum be obtained by rotating the polygon to theand angle which ER has theThen, minimum areaof(see Figure can be obtained by rotating the polygon to the angle at which ER has the minimum area (see Figure 5). 5). The MER is often used in GIS to give the approximate location and size of a geographic element. The MER is often used in GIS to give the approximate location andindexing, size of a geographic Many operations in GIS, such as spatial queries and spatial require theelement. use of Many MER. operations GIS,operator such as has spatial queries and spatial when indexing, the use of MER. Meanwhile, Meanwhile,inthis excellent performance usedrequire to achieve RS image segmentation, this excellent used to achieve RS image andminimum it can be and operator it can behas a feature forperformance aiding directwhen classification [30], because the segmentation, MER adopts the arectangular feature for aiding direct classification [30], because MER the the minimum rectangular window window with edges parallel to the the axes to adopts describe polygon’s morphological with edges parallel to the axes to describe the polygon’s morphological characteristics, while GLCM characteristics, while GLCM also uses a rectangular window for texture feature description. also uses a rectangular window for texture feature description. Therefore, the MER window can directly Therefore, the MER window can directly describe the relationship between the polygon shape and describe the relationship between the polygon shape and texture scale. texture scale.

Figure 5. Diagram of the generation of the minimum enclosing rectangle (MER) for a polygon; (a) Figure 5. Diagram of the generation of the minimum enclosing rectangle (MER) for a polygon; Obtaining the enclosing rectangle (ER) of a polygon by coordinate direction measurement; (b) (a) Obtaining the enclosing rectangle (ER) of a polygon by coordinate direction measurement; Obtaining the minimum enclosing rectangle (MER) of a polygon by rotating the object. (b) Obtaining the minimum enclosing rectangle (MER) of a polygon by rotating the object.


8 of 24

First, we compute the MER of each polygon and record its parameters, including length and width, as {HER , WER }objID . Next, the polygon regularity must be measured in order to judge whether it is suitable for shape knowledge mining. If the polygon’s shape is too narrow or curved, then the simple enclosing rectangle cannot describe its shape characteristics. In this paper, the rectangular-degree index is employed to achieve this measurement. R = A/A MER = A/( H MER × WMER ) (3) where A represents the actual area of a polygon and AMER represents its MER area. This indicator is simple and effective for the shape regularity description, and can reflect the degree to which a polygon fills its enclosing rectangle, as well as the irregularities of the polygon. The value of R is in the range from 0 to 1. R has a maximum value of 1.0 for rectangular objects, a value of π/4 for circular objects, and becomes smaller for objects with narrow, irregular, or curved shapes. If the rectangular degree of a polygon is lower, the texture characteristics under the MER window will be not accurate. Therefore, the following equation must be created to select the polygons with square-like shape for the subsequent analysis: n n o o Eshape = objID objID ∈ 1, 2, · · · , Nobj &&RobjID > Tshape

(4)

In the equation above, Nobj represents the total number of polygons in the historical GIS database. RobjID represents the rectangular degree of polygon No.objID. Eshape represents the polygon sets with regular shapes, wherein each polygon’s squareness degree is larger than the threshold value Tshape , as determined by the histogram method. 3.2. Multi-Scale Texture Window Determination Based on Domain Knowledge and GIS 3.2.1. One-To-One Associated Rules Inspired by domain shape knowledge, texture scales can consequently be extracted; their associated rules are summarized as follows [31–39]:

•

•

•

A variety of categories should correspond to a variety of texture scales. This can be a one-to-one relationship or a many-to-one relationship. However, the number of texture scales should not be too large. Otherwise, it will cause information redundancy, increase algorithm consumption, and decrease the separability among the different categories. Therefore, the one-to-one model is chosen for analysis. Each scale parameter can determine an observation window, which range should be close to, but slightly larger than, the polygon size of the classified object. If the window is too small, it cannot reflect the object’s spatial features. If the window is too large, the number of mixed pixels will be increased, resulting in spectral ambiguity and enormously reducing the classification accuracy. If the classified object has a longitudinal structure (e.g., roads), it means that its width is relatively consistent, while its length and curvature are often uncertain. Therefore, we would choose an observation window related to the structure’s width.

3.2.2. Scale Determination According to the one-to-one model, we can obtain the scale parameter for each category from the regular polygon collection Eshape by analyzing the frequency histograms of the polygon shape. Assuming that Nωi is the total number of regular polygons of the category ωi , then the frequency histogram of MER length can be defined as the following discrete function: Pωi −length (S) =

nωi −length Nωi

(5)


9 of 24

In this frequency histogram, S represents the ranges of MER length of the category ωi , and ordinate nωi −length and Pωi −length (S), respectively, represent the number and probability of the polygons that are distributed in the range S. Additionally, the frequency histograms of the rectangular width Pωi −width (S) will be similarly drawn. In theory, if a category has similar shape for its polygons, then there should be a peak region distributed in its shape frequency histogram, above, whereby the corresponding shape size is able to cover the majority of polygons of this category and to achieve valid texture observation. According to the selection principles, among the peak data of the MER length and width, if the larger one is chosen, the texture observation window will contain too many mixed pixels, thus reducing the feature validity, so it is easier to choose the smaller of the two values as the observation window of GLCM texture for this category. In this paper, the scale parameter is defined as half the number of image pixels corresponding to the observation window in geo space. Therefore, the final scale value should be divided by double the pixel resolution of an image, that is, pResolution × 2, as in the following equation: Scalei = min(widthScalei , lengthScalei )/( pResolution × 2)

(6)

In this way, the best scale for the GLCM texture of each category will be obtained. 3.3. GLCM Texture Feature Evaluation Based on Fisher Function 3.3.1. GLCM Texture Features Based on Multi-Scales GLCM is a matrix that is defined by the following parameters: observation window scale, distance and direction of pixel pairs, band and gray level of distribution probability. Since this paper mainly discusses the influence and performance of texture scale parameters on the separability of categories, the other parameters for GLCM construction have been selected using conventional and common ways. In fact, 14 kinds of GLCM-based texture feature have been defined, and 8 kinds of classical feature are commonly used for RS image analysis. They are {mean, variance, homogeneity, contrast, dissimilarity, entropy, angular second moment, and correlation} [2,3,31,32]. Therefore, these 8 texture features are also used for our experiments. In this way, for a pixel k, its texture features set based on the scale scalei is denoted as xk (i) . Then, all the texture features can be expressed as a feature vector textureFeat with dimension of K × M (where K is the number of pixels, and M is the number of scales and categories), as shown in Equation (7).    textureFeat =   

x 1 (1) , x 1 (2) , · · · , x 1 ( M ) x 2 (1) , x 2 (2) , · · · , x 2 ( M ) .. . x k (i ) · · · ( 1 ) x K , x K (2) , · · · , x K ( M )

     

(7)

3.3.2. Separability Evaluation Extracting multi-scale windows for texture features is better able to distinguish different categories from an image. Therefore, a quantitative criterion for measuring the validity of these texture features for classification should be employed. In general, different categories can be distinguished because they are located in different regions of the feature space. Obviously, a smaller or no overlap of these regions indicates better separability of features, as expressed in Figure 6. The criteria for class separability evaluation commonly include Euclidean distance, Fisher distance, K-L distance, and others. In complex image classification situations, the Fisher distance adopts the determinant ratio of the between-class scatter matrix and the in-class scatter matrix as the separability criterion. Therefore, it can take the overall and local distribution characteristics into

ISPRS Int. J. Geo-Inf. 2018, 7, 175 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW

10 of 24 10 of 24

distinguished becausehas theyperformed are located in different regions of the featureThus, space. a smaller account. This method well in traditional studies [44–47]. it Obviously, has also been used to or no overlap of theseofregions separability features, evaluate the validity textureindicates features better in various scales inof this study. as expressed in Figure 6.

Figure 6. Schematic diagram of separability among categories. Figure 6. Schematic diagram of separability among categories.

The criteria for class separability evaluation commonly include Euclidean distance, Fisher ( j) denote the eigenvectors of categories ω and ω , respectively. Then, Specifically, xk (i) and i i the Fisher distance distance, K-L distance, andxkothers. In complex image classification situations, the average distance between eigenvectors of categories is defined as Equation (8): adopts the determinant ratio of the between-class scatter matrix and the in-class scatter matrix as the separability criterion. Therefore, it can take the overall and local distribution characteristics into Ni Nj M M 1 1 (i ) ( j) Thus, it has also been used account. This method has performed studies d( x ) = well Pj d( xk[44–47]. , xk ) (8) ∑ Piin∑traditional ∑ ∑ 2 N N i jscales to evaluate the validity of texture features i =1 in j=various 1 k =1 k =in 1 this study. Specifically, xk and xk denote the eigenvectors of categories ωi and ωi , respectively. In the above formula, Pi and Pj , respectively, represent the a priori probabilities of the corresponding Then, the average distance between eigenvectors of categories is defined as Equation (8): categories, which can usually be expressed on the basis of probabilistic statistics of the samples in each category. Assuming that the −number ofMsample ωi is N pi , then Pi can be obtained by N M pixels of category 1 1 Ni j Equation (9): d ( x) = Pi Pj d ( xk(i ) , xk(j) ) (8) 2=i 1 =j 1 NN k 1= k 1 i Npji = (9) Pi = M ∑ N pi represent the a priori probabilities of the In the above formula, P and , respectively, (i )

( j)

∑ ∑ P

i

∑∑

j

i =1

corresponding categories, which can usually be expressed on the basis of probabilistic statistics of the In addition, d( xk (i) , xk ( j) ) denotes the Euclidean distance between xk (i) and xk ( j) . This is defined samples in each as Equation (10):category. Assuming that the number of sample pixels of category ωi is Npi , then ( j ) T (i ) ( j) 21 (i ) ( j ) (i ) Pi can be obtained by Equation d( x(9): (10) k , xk ) = [( xk − xk ) ( xk − xk )] Then, the overall in-class scatter matrix of samples Npi can be defined as Equation (11):

Pi =

M

Sw =

∑ Pi

i =1

N pi

(9) (11)

M

∑

T 1 (i )Np (i ) (ix=k1 − imi )( xk − m j ) ∑ N p i k =1 (i )

In addition, d ( xk , xk ) denotes the Euclidean distance between xk and x The overall between-class scatter matrix of samples can be defined as Equation (12):k defined as Equation (10): (i )

( j)

M

S =(j)∑ Pi (m − m(j))(Tmi − m) T (j) ( i )i (i )

( i )b k

1

( j)

. This is

(12) (10)

2 d ( x , xk i=)=[( 1 xk − xk ) (xk − xk )]

The overall scatter matrixscatter of samples can definedcan as be Equation Then, the overall in-class matrix ofbe samples defined(13): as Equation (11): M

= S w S t∑ = Pi

11

M Np

Npi

∑ ∑ ((xxi −−mm)()(xix− m−)Tm ) M

=i 1 = i ki =11

(i ) k

i

(i ) k

T

j

∑ N pi The overall between-class scatter matrix of samples can be defined as Equation (12): i =1

(11) (13)


11 of 24

In the above formula, mi represents the average feature space vector of samples in category ωi and m represents the average feature space vector of samples of all categories. These are defined as Equation (14): Np

mi =

M i 1 (i ) xk , m = ∑ Pi mi ∑ N p i i =1 i =1

(14)

Then, the criteria for separability of categories can be defined as Equation (15): −2

a = 1,

J = d ( x ) = Tr [Sw + Sb ] = Tr [St ]

(15)

In the above formula, Tr {.} represents the trace of the matrix (the sum of the diagonal elements of the matrix). In this way, the trace of the overall scatter matrix of samples (the Fisher distance) can be used as the separability criterion. The larger the value, the more dispersed the samples and the better the separability. 4. Experiment and Discussion 4.1. Study Data In this paper, Changjiang County, Hainan Province was chosen as the study area. We aimed to obtain effective scale parameters for multi-scale GLCM texture features in distinguishing the land-use categories. Land-use classification is a common geo-scene in practical work, with a complex system and specific meaning. The collected study data include the following. (1) The 2008 SPOT5 RS image, which has a high spatial resolution of 2.5 m and a true color space (Red, Green, Blue). It is intended for land-use classification cartography of 1:10,000. (2) The existing GIS database in vector format, including 2005 and 2008 GIS land-use maps at the scale of 1:10,000, where the 2005 GIS data can be used to extract domain knowledge, and the 2008 GIS data can be taken as the current real geo-objects to verify the classification accuracy. The other auxiliary GIS data, which are mainly used for spatial partitioning, include the 2006–2020 land-use planning map and DEM graded map, etc. 4.2. Spatial Partition and Classification System of Land-Use First, by using the 2006–2020 GIS and land-use planning data, we extract the following partitioning factors: important resource distributions, administrative boundaries, and the functional division map. Meanwhile, the 2005 land-use polygon boundary is taken as the basic unit. Their cartographic priority is set according to the visual size in the local region as follows: DEM graded map > railway distribution map > provincial and municipal highway distribution map > important river distribution map > functional division map > village administrative boundary map. According to this order, each factor map should be superimposed on the basic land-use polygon. Finally, the study area is divided into five homogeneous regions, which have to maintain the integrity and continuity of land-use polygons and administrative boundary, as shown in Figure 7. According to land-use planning, the land-use pattern of the Zones 2–4 will not be extensively adjusted in 2005–2008, so the domain knowledge mined from the 2005 land-use map in these regions is seen to be usable. To better illustrate the adaptability of the proposed method, the 40 Zone has been chosen for the experiments because it has a more complex land-use pattern. According to the latest land-use classification system used in practical work, there should be 12 land-use classification categories (see in Table 1).

isis isseen seento tobe beusable. usable.To Tobetter betterillustrate illustratethe theadaptability adaptabilityof ofthe theproposed proposedmethod, method,the the4′ 4′Zone Zonehas hasbeen been seen to be usable. To better illustrate the adaptability of the proposed method, the 4′ Zone has been is seen to be usable. To better illustrate the adaptability of the proposed method, the 4′ Zone has been is seen to be usable. To better illustrate the adaptability of the proposed method, the 4′ Zone has been is seen to be usable. To better illustrate the adaptability of the proposed method, the 4′ Zone has been chosen for the experiments because it has a more complex land-use pattern. chosen for for the the experiments experiments because because it it has has aa more more complex complex land-use land-use pattern. pattern. chosen chosen for the experiments because has more complex land-use pattern. chosen for for the the experiments experiments because because itit it has has aaa more more complex complex land-use land-use pattern. pattern. chosen ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW

12 of 24

ISPRS Int. J. Geo-Inf. 7, 175 To better illustrate the adaptability of the proposed method, the 4′ Zone has been is seen to2018, be usable. chosen for the experiments because it has a more complex land-use pattern.

12 of 24

Figure Figure7.7. 7.Land-use Land-usespatial spatialpartition partitionmap mapand andclassification classificationsystem systemdetermination. determination. Figure Land-use spatial partition map and classification system determination. Figure 7. Land-use spatial partition map and classification system determination. Figure 7. 7. Land-use Land-use spatial spatial partition partition map map and and classification classification system system determination. determination. Figure

According Accordingto tothe thelatest latestland-use land-useclassification classificationsystem systemused usedin inpractical practicalwork, work,there thereshould shouldbe be12 12 According to the latest land-use classification system used in practical work, there should be 12 Figure 7. Land-use spatial partition map and classification system determination. According to the latest land-use classification system used in practical work, there should be 12 According to the latest land-use classification system used in practical work, there should be 12 land-use classification categories (see in Table 1). According to the latest land-use classification system used in practical work, there should be 12 land-use Figure classification categories (see in Table Tablemap 1). and classification system determination. 7. Land-use spatial partition land-use classification categories (see in 1). land-use classification categories (see in Table 1). land-use classification classification categories (see in in Table 1). 1). system used in practical work, there should be 12 land-use categories (see Table According to the latest land-use classification Table 1.1.Image classification system of land Table Imagecategories classification system of1). landuse useand andits itsinterpretative interpretativesigns signs(zone (zone4). 4). land-use classification (see in Table Table 1. Image classification system land use and signs (zone 4). Table 1. Image classification system ofof land use andits itsinterpretative interpretative signs (zone 4). Table 1. Image classification system of land use and its interpretative signs (zone 4). Table 1. 1. Image Image classification classification system system of of land land use use and and its its interpretative interpretativesigns signs (zone (zone 4). 4). Table

ClassIndex Categories Code GIS Database Interpretation Sign ClassIndex Table 1. Categories Codein inland GISuse Database Interpretation Sign ClassIndex Categories Code in GIS Database Image classification system of and its interpretativeInterpretation signs (zone 4). Sign ClassIndex Categories Code in GIS Database Interpretation Sign ClassIndex Categories Code in GIS Database Interpretation Sign ClassIndex Categories Code in GIS Database Interpretation ClassIndex Categories Code in GIS Database Interpretation Sign Sign Categories Code in GIS Database Interpretation Sign 1ClassIndex Paddy 011 1 Paddyland land 011 1 Paddy land 011 11 Paddy land 011 Paddy land 011 11 Paddy land 011 Paddy land 011 1

Paddy land

011

222 2 222

2

Dry Dryland land Dry land Dry land Dry land Dry land Dry land Dry land

013 013 013 013 013 013 013 013

333 3 333

3

Orchard Orchardland land Orchard land Orchard land Orchard land Orchard land Orchard land Orchard land

021 021 021 021 021 021 021 021

444 44 44

4

Other orchard land Other orchard land Other orchard land Other orchard land Other orchard land Other orchard land Other orchard orchard land land Other

023 023 023 023 023 023 023 023

5555 5 55

5

Forest land Forest land Forest land Forest land Forest land Forest land Forest land Forest land

031 031 031 031 031 031 031 031

6666 6 66

6

Other forest land Other forest Other forest land Other forestland land Other forest land Other forest land Other forest forest land land Other

033 033 033 033 033 033 033 033

ISPRS PEER 7 2018, ISPRS Int. Int. J. J. Geo-Inf. Geo-Inf. 2018, 7, 7, xxx FOR FORPond PEER REVIEW REVIEW ISPRS Int. J. Geo-Inf. 2018, 7, FOR PEER REVIEW ISPRS Int. Int. J. J.77 Geo-Inf. 2018, 7, 7, xx FOR FOR PEER REVIEW REVIEW Pond ISPRS Geo-Inf. 2018, PEER 7 Pond Pond

114

7 7 77 88 8888

Pond Pond Pond Pond River River River River River River

114 114 114 114 114 114 114 111 111 111 111 111 111

99 9999

Rural settlement Rural settlement Rural settlement Rural settlement Rural settlement Rural settlement

072 072 072 072 072 072

10 10 10 10 10 10

Urban settlement Urban settlement Urban settlement Urban settlement Urban settlement Urban settlement

071 071 071 071 071 071

11 11 11 11 11 11

Road Road Road Road Road Road

102 102 102 102 102 102

12 12 12 12 12 12

Rural road Rural road Rural road Rural road Rural road road Rural

104 104 104 104 104 104

13 13 of of 24 24 13 of 24 13 of of 24 24 13

4.3. Evaluation of Domain Knowledge 4.3. Evaluation of Domain Knowledge 4.3. Evaluation of Domain Knowledge 4.3. Evaluation Evaluation of of Domain Domain Knowledge Knowledge 4.3. In the ArcGIS software, we first split up the RS raster image using the GIS vector polygon In the ArcGIS software, we first split up the RS raster image using the GIS vector polygon In the ArcGIS software, we first split up the RS raster image using the GIS vector polygon In the the ArcGIS ArcGIS software, software, we we first first split split up up the the RS RS raster raster image image using using the the GIS GIS vector vector polygon polygon In boundary and calculate the mean gray value of each segmentation block. For this RS image with 3 boundary and calculate the mean gray value of each segmentation block. For this RS image with boundary and calculate the mean gray value of each segmentation block. For this RS image with boundary and and calculate calculate the the mean mean gray gray value value of of each each segmentation segmentation block. block. For For this this RS RS image image with with 3333 boundary bands, its gray image can be obtained by converting color space from RGB (red, green, blue) to HSI bands, its gray image can be obtained by converting color space from RGB (red, green, blue) to HSI bands, its gray image can be obtained by converting color space from RGB (red, green, blue) to HSI bands, its its gray gray image image can can be be obtained obtained by by converting converting color color space space from from RGB RGB (red, (red, green, green, blue) blue) to to HSI HSI bands, (hue, saturation, and luminance) and eliminating the hue and saturation information while retaining

11

Road

ISPRS Int. 12 J. Geo-Inf. 2018, 7,Rural 175

102

road

104

13 of 24

4.3. Evaluation Evaluation of of Domain Domain Knowledge Knowledge 4.3. In the the ArcGIS ArcGIS software, software, we we first first split split up up the the RS RS raster using the the GIS In raster image image using GIS vector vector polygon polygon boundary and calculate the mean gray value of each segmentation block. For this image with boundary and calculate the mean gray value of each segmentation block. For this RSRSimage with 3 3 bands, grayimage imagecan canbebeobtained obtainedby byconverting convertingcolor colorspace spacefrom fromRGB RGB(red, (red, green, green, blue) blue) to to HSI HSI bands, itsitsgray (hue, saturation, saturation, and and luminance) luminance) and and eliminating eliminating the the hue hue and and saturation saturation information information while while retaining retaining (hue, the luminance values. When we plot the frequency histograms of the blocks under the corresponding the luminance values. When we plot the frequency histograms of the blocks under the corresponding gray levels is is setset asas 26,26, which means thatthat there are are 26 breaks/bars in the gray levels of of each eachcategory, category,the thegray graylevel level which means there 26 breaks/bars in histogram, i.e., r = 1, 2, . . . , 26, and each level is 10. Since the category number is 12, we get 12 groups k the histogram, i.e., rk = 1, 2, …, 26, and each level is 10. Since the category number is 12, we get 12 of statistical graphs of the gray level of level segmentation blocks, as shown Figure groups of statistical graphs of the gray of segmentation blocks, asin shown in8:Figure 8:

C1

C5

C9

C2

C3

C4

C6

C7

C8

C10

C12

C11

Figure 8. The frequency histograms of segmentation blocks’ spectral means for each category. The xFigure 8. The frequency histograms of segmentation blocks’of spectral means for each The x-axis axis is the gray level of 26, and the y-axis is the frequency the blocks under eachcategory. gray level. is the gray level of 26, and the y-axis is the frequency of the blocks under each gray level.

Thus, we can easily obtain the peak zone of a histogram above because of its approximate normal Thus, weThen, can easily obtain peak zone to of count a histogram above because its number approximate normal distribution. Equation (2)the will be used the changed-rate V ofofthe of polygons distribution. Then, Equation (2) will be used to count the changed-rate V of the number of polygons for for each category. The statistics are shown in Table 2, where the total number of polygons of each each category. statistics aredata shown in Tableas 2, totNum, where the total polygons of for each category category in theThe historical GIS is denoted and thenumber numberofof polygons which the in the historical GIS data is denoted as totNum, and the number of polygons for which the category category has not been changed between the two-time phases is denoted as tipNum. Following related has not been changed between two-time phases denoted as tipNum. previous previous works (especially bythe Yang C., Zhou C. is 2001), in the classicalFollowing statisticalrelated theories of the works (especially C., Zhoumarking”, C. 2001), inthe thechanged classical rate statistical theories number of the “golden section” “golden section” by andYang “five-grade of polygons’ of a category and “five-grade marking”, the changed rate of polygons’ number of a category should be lower than the threshold η as 40%; that is, at least 60% of the polygons’ categories should not have been changed. In this case, the domain knowledge system can be characterized as stable and qualified [25,48,49]. Table 2. Statistics of the changed rate of the polygon category based on GIS data and image segmentation. ClassIndex

C1

C2

C3

C4

C5

C6

C7

C8

C9

C10

C11

C12

totNum tipNum V (%)

2240 1680 25.0

5610 3842 31.5

1124 753 33.0

2086 1451 30.4

4341 3537 18.5

287 256 10.8

1070 763 28.7

185 109 41.1

499 316 36.7

113 68 39.8

27 19 29.6

409 289 29.3

In the table above, a smaller polygon category’s changed-rate results in higher stability. Because most categories meet the pre-defined judgment of Equation (2), they are relatively stable and unified

ClassIndex C1 C2 totNum 2240 5610 tipNum 1680 3842 V J.(%) 25.07, 17531.5 ISPRS Int. Geo-Inf. 2018,

C3 1124 753 33.0

C4 2086 1451 30.4

C5 4341 3537 18.5

C6 287 256 10.8

C7 1070 763 28.7

C8 185 109 41.1

C9 499 316 36.7

C10 113 68 39.8

C11 27 19 29.6

C12 409 289 29.3 14 of 24

In the table above, a smaller polygon category’s changed-rate results in higher stability. Because under the selected scalejudgment in whichofthe corresponding domain knowledge is reliable. most categories meetspatial-time the pre-defined Equation (2), they are relatively stable and unified Meanwhile, the changed-rate of categories 8 and 10 are near 40%. Therefore, that corresponding under the selected spatial-time scale in which the corresponding domain knowledge is reliable. domain knowledge may be relatively ineffective in 10 theare subsequent analysis. Meanwhile, the changed-rate of categories 8 and near 40%. Therefore, that corresponding domain knowledge may be relatively ineffective in the subsequent analysis. 4.4. Scales Determined by GIS and Domain Knowledge 4.4. Scales by GIS and Domain Knowledge BasedDetermined on the secondary development program function of the ArcGIS platform, the following operators been obtained.development program function of the ArcGIS platform, the following Basedhave on the secondary operators have been obtained. (1) There are a total of 17,991 land-use polygons in the partition region, of which 3893 patches were are selected to their polygons regularityinas shapes (1) There a totalaccording of 17,991 land-use theeffective partitionpolygon region, of whichfor 3893morphological patches were knowledge statistics. selected according to their regularity as effective polygon shapes for morphological knowledge (2) statistics. According to the effective polygon sets, we plot out the frequency histograms of the length and (2) According to the effective polygon sets, we plotwe outare theinterested frequencyonly histograms of the length width of the MER for each category. Because in the shorter side ofand the width of the for each category. Because we are interested only in the shorter side of rectangle, asMER described in the discussion and Equation (6) above, frequency histograms of the the rectangle, described in the discussion Equation (6) above, frequency histograms of the widths of as MER should be plotted for eachand category, as shown in Figure 9. widths of MER should be plotted for each category, as shown in Figure 9.

C1

C5

C9

C2

C3

C4

C6

C7

C8

C10

C11

C12

Figure 9. The frequency histograms of the lengths and widths of the MER for each category. The xFigure 9. The frequency histograms of theand lengths and widths the MERof forthe each category. axis is the length with the unit of meters, the y-axis is the of frequency MER widths.The x-axis is the length with the unit of meters, and the y-axis is the frequency of the MER widths.

These figures show the following rules. (1) There are obvious peaks in most categories, which shows that figures a strongshow regularity lies in their morphology. Moreover, a more frequency These the following rules. (1) There are obvious peaks inconcentrated most categories, which distribution a more obvious morphological regularity with its category. (2) There is no shows that aindicates strong regularity lies in their morphology. Moreover, a more concentrated frequency obvious peak or peaks in obvious a few categories. Thisregularity includeswith category 8, which contains several distribution indicates a more morphological its category. (2) There is no obvious peak or peaks in a few categories. This includes category 8, which contains several morphological characteristics. Therefore, the validity of their domain knowledge has been further constrained. (3)

The texture scale parameter Scalei of each category is set by selecting the smaller value of the statistical width and length of MER, in accordance with Equation (6). Meanwhile, we indicate the variance of the length and width as STD_Width and STD_Length to measure the representativeness of the selected texture scale. Then, the scale value should be the nearest odd number obtained by the window size in geo-space divided by pixel resolution × 2 (2.5 × 2 = 5 m in this study). The results are shown in Table 3.


15 of 24

Table 3. The selection results of the best texture scale for each category. ClassIndex

Scale

widthScale

lengthScale

STD_Width

STD_Length

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12

81 79 63 83 59 63 27 83 39 71 13 9

80 79 64 83 62 63 30 87 39 73 30 34

88 82 63 88 59 70 27 83 42 71 13 9

47.00 41.00 33.00 50.00 39.00 44.00 13.00 102.00 29.00 41.00 10.00 9.00

49.00 48.00 36.00 49.00 41.00 40.00 14.00 77.00 26.00 37.00 11.00 9.00

According to the analysis in Table 3, some conclusions can be obtained. (1) The minimum texture scale corresponds to category 12 (rural road). It is a linear feature that takes the line width as a basis for analysis. The maximum texture scale corresponds to category 8 (river), which has the largest area and an irregular shape. At the same time, however, we should also note that there are still some short and small rivers in the study area. The texture features observed on this large a scale are not conducive to the extraction of these objects. For categories 1–6 (a series of cultivated lands), the selected texture scales are also larger, which indicates that there is a massive stretched distribution in the polygons. This conclusion can also be directly observed from the GIS historical database. Moreover, category 9 represents a rural residence, and its texture scale is much smaller than the urban residence in category 10. This result is also in accordance with the actual case. (2) The variances in the length and width of the enclosing rectangles for most categories are generally less than 50, except for category 8, which has a variance greater than 70. This result further illustrates that the differences in the morphological characteristics of category 8 will lead to a consequence of there beingno representative scale that can be selected. In conclusion, the texture scale sets that are selected with the aid of GIS historical knowledge are in accordance with the local actual situation and people’s general cognition. 4.5. Scales Determined by the Enumeration Method In this paper, the Fisher distance algorithm is employed to quantitatively measure the validity of the selected texture scales. The determinant ratio of the between-class scatter matrix and the within-class scatter matrix has been taken as the criterion of category separability in this index. (1)

First, the enumeration method is used to calculate texture features under a series of scales. On the basis of relevant research results [31–42], we select an enumeration scale of L = 3:101.

As for setting the other parameters of GLCM-based texture features, some conventional and efficient methods were used (as shown in Table 4), with the single-band image value also being obtained by converting the color space from RGB to HSI and retaining the luminance values of HSI, thereby obtaining the distance value that is most commonly set to 1, and obtaining the direction value by computing the average texture statistics in its four directions. Additionally, the dimension of GLCM is determined by the image gray level. Considering the calculation efficiency and the storage need of the matrix in practical applications, the gray level of the original image should be compressed, i.e., an 8-bit image with a gray level of 0–255 is compressed into a 3-bit image with a gray level of 0–8, which can effectively avoid the sparse matrices in GLCM. In this way, we can calculate 8 classic GLCM-based texture features as described in Section 3.3.1; meanwhile, we employ a principal component analysis algorithm to deduce 3 principal features, so as


16 of 24


16 of 24

thereby obtaining the distance value that is most commonly set to 1, and obtaining the direction value by computing the average texture statistics in its four directions. Additionally, the dimension of to reduce feature dimensions and ensure theirConsidering independence. These two groups of texture GLCM is determined by the image gray level. the calculation efficiency and thefeatures storage with number of 8 and 3 are denoted as I and II. need of the matrix in practical applications, the gray level of the original image should be compressed, i.e., an 8-bit image with a gray level of 0–255 is compressed into a 3-bit image with a gray level of 0– Table 4. The setting of parameters in constructing the GLCM-based texture features. 8, which can effectively avoid the sparse matrices in GLCM. In this way, we can calculate 8 classic texture as described Section The Number of Image Gray GLCM-based GLCM GLCM features GLCM TheinNumber of 3.4; Parameters Image Band Level Scale Distance Direction Texture Features meanwhile, we employ a principal component analysis algorithm to deduce 3 principal features, so Thereduce value feature dimensions 1 A set 1 TheseMean 8 or 3features as to and0–8 ensure their independence. two value groups of texture A common A common Principal component Selected Color space Gray level with number of 8 and 3 are denoted as I andGIS II. assisted method

converting

value

compress

value

analysis

Table 4. The setting of parameters in constructing the GLCM-based texture features.

(2)

Second, the Fisher distances of the two groups of texture features based on the enumeration The Number of Image Gray GLCM GLCM GLCM The Number of Texture Parameters method areImage computed. relational the category number Band In the following Level Scale diagram, Distancewe define Direction Features as the The value 1 0–8 y-axis. We A set 1 the value Meanof value 8 or 3 matrix x-axis and the texture scale as the also describe the overall scatter Selected Colordistance space Gray level GISblocks. A common A common Principalseparability, component and the Fisher using different color Blue represents the smallest method converting compress assisted value value analysis and red represents the largest separability. Therefore, the relationship between the category separability and the texture scales can be clearly expressed. (2) Second, the Fisher distances of the two groups of texture features based on the enumeration method are are computed. InFigure the following relational we define theand category The results shown in 10. Group I of the diagram, overall scatter matrix Fisher number distanceas is the under x-axis and the texture scale as the the group y-axis. IIWe alsooverall describe the value the Fisher overalldistance scatter derived 8 texture features, while of the scatter matrixofand matrixunder and 3the Fisher distance different blocks. from Blue these represents is derived independent textureusing features. It cancolor be observed graphsthe that:smallest (1) For separability, and red represents the largest separability. Therefore, the relationship between the the overall scatter matrix, the peak value is distributed in a more concentrated way and is more categoryin separability and in thegroup texture scalesincan be clearly expressed.most categories reach high inconsistent all categories I than group II. In addition, separability on some slightly smaller10. scale values group II than in matrix group and I. This illustrates that, The results are shown in Figure Group I ofin the overall scatter Fisher distance is when the texture features are more independent, the separability of all categories is more consistent derived under 8 texture features, while the group II of the overall scatter matrix and Fisher distance and can beunder obtained on a smallertexture scale value. (2) It For the distance, anygraphs category (column), is derived 3 independent features. can beFisher observed from in these that: (1) For there is always a peak zone in the separability bar chart, and the separability of such zone changes the overall scatter matrix, the peak value is distributed in a more concentrated way and is more mostly scalesinaccording to the in normal and it II. fluctuates. The peak of each category inconsistent all categories groupdistribution, I than in group In addition, mostvalues categories reach high in group I and are close. However, and noise occur occasionally group I, that, such separability ongroup some II slightly smaller scale fluctuations values in group II than in group I. This in illustrates as in category 6, while the peak region has a more obvious and concentrated distribution in group II. when the texture features are more independent, the separability of all categories is more consistent This illustrates that the separability is more prominent and clear when the texture features are more and can be obtained on a smaller scale value. (2) For the Fisher distance, in any category (column), independent, vice versa. there is alwaysand a peak zone in the separability bar chart, and the separability of such zone changes In addition, it should that most categories complex texture mixed pixels mostly scales according to be thenoted normal distribution, and itwith fluctuates. The peak features values oforeach category can be separated under a bigger texture scale value; thatand is to say, occur if onlyoccasionally one scale parameter be in group I and group II are close. However, fluctuations noise in group I,can such used to calculate the texture features in a particular case, its value should be as large as possible. as in category 6, while the peak region has a more obvious and concentrated distribution in group II. As a result,that thethe scales corresponding to prominent the highestand separability of each category were computed. This illustrates separability is more clear when the texture features are more The texture scale by the different methods can be seen in Table 5. independent, andsets vicedetermined versa.

(b)

(a) Figure 10. Cont.


17 of 24


17 of 24


(c)

17 of 24

(d)

Figure 10. The statistical results of the separability of texture features under different scales for each category. The number of categories is 12, the scale set is enumerated from 3:101, and the color from blue to red represents the changes in separability from low to high. (a) The overall scatter matrix of group I changes with the texture scales; (b) the Fisher distance of group I changes with the texture (c) (d) scales in the group I texture features; (c) the overall scatter matrix of group II changes with the texture scales; (d) distance of group II changes with the texture scales. Figure 10. the TheFisher statistical results of the separability of texture features under different scales for each Figure 10. The statistical results of the separability of texture features under different scales for each category. The number of categories is 12, the scale set is enumerated from 3:101, and the color from category. The number of categories is 12, the scale set is enumerated from 3:101, and the color from In addition, it should notedinthat most categories with complex texturescatter features or of mixed blue to red represents thebe changes separability from low to high. (a) The overall matrix blue to red represents the changes in separability from low to high. (a) The overall scatter matrix of pixels can be separated a bigger texture value; thatofisgroup to say,I changes if only one parameter group I changes withunder the texture scales; (b) thescale Fisher distance withscale the texture group I changes with the texture scales; (b) the Fisher distance of group I changes with the texture can be usedinto calculate the texture features a particular case, its be with as large as possible. scales the group I texture features; (c) thein overall scatter matrix of value group should II changes the texture scales in the group I texture features; (c) the overall scatter matrix of group II changes with the texture scales; (d) thethe Fisher distance of group II to changes with the texture scales. As a result, scales corresponding the highest separability of each category were computed. scales; (d) the Fisher distance of group II changes with the texture scales.

The texture scale sets determined by the different methods can be seen in Table 5. In addition, it should be noted that most categories with complex texture features or mixed Table 5. The scale sets that are extracted by the GIS-assisted method and the enumeration method. pixelsTable can 5. beThe separated bigger texture scale value; that and is tothe say, if only one scale parameter scale setsunder that areaextracted by the GIS-assisted method enumeration method. can be used to calculate the texture features in a particular case, its value should be as large as possible. ClassIndex ClassIndex Scale Sets Selected Method Scale Sets Selected Method As a result, the scales corresponding to the highest computed. 1 11 11 12 12 12 separability 23 3 4 4 5 of 5 6each 6 7 category 7 8 8 9 were 9 10 10 method 81 methods 79 79 63 can The texture scale setsGIS-assisted determined by the different Table 11 GIS-assisted method 81 6383 be 8359seen 5963 in 6327 2783 5. 83 49 49 71 71 13 13 9 9 Enumeration method 8 texture features 85 85 71 71 73 7387 8767 6763 6331 3169 69 27 27 83 83 9 9 7 7 22 Enumeration methodunder under 8 texture features 3 Enumeration method under 3 texture features 75 83 61 73 45 57 27 67 53 95 21 9 3Table 5.Enumeration method under 3 texture features 75 83 61 73 45 enumeration 57 27 67method. 53 95 21 9 The scale sets that are extracted by the GIS-assisted method and the

4.6. Comparison of the the Different Different Scale Sets 4.6. Comparison of Sets Scale Sets Selected Scale Method

1 2 3 For the the three three groups groups of texture scale sets shown in Table 1For GIS-assisted method 81 79 63 of texture scale sets shown in Table 2 Enumeration under 8 texture features 85 71 73 summarized as shown shownmethod in 11. summarized as in Figure Figure 11. 3 Enumeration method under 3 texture features 75 83 61

ClassIndex 4 5 6 7 8 9 10 11 12 5, the relationship among them can be be 83 59 63 27 83 49 71 13 9 5, the relationship among them can 87 67 63 31 69 27 83 9 7 73 45 57 27 67 53 95 21 9

4.6. Comparison of the Different Scale Sets For the three groups of texture scale sets shown in Table 5, the relationship among them can be summarized as shown in Figure 11.

Figure 11. The correlation of texture scale sets among different methods. Figure 11. The correlation of texture scale sets among different methods.

Figure 11. The correlation of texture scale sets among different methods.


18 of 24

Some conclusions can be reached. (1)

(2)

Intuitively, the scales that are selected by the different methods are very close and even have the same values in some categories. However, there is also a larger difference between categories 8 and 10. By analyzing the cause of the divergence, we find that category 8 represents river water, whose shape is larger in the GIS historical database. Therefore, the texture scale that is determined by its domain knowledge has a larger value of 83. However, significant changes have occurred in this category within two time-phases (see Table 1, with the change rate of 41.1%), because many rivers continue to diminish, and the standard scale parameter that is determined by the enumeration method has a smaller value of 69 or 66. Moreover, similar to category 10 (urban settlement), the scale that is determined with the aid of the GIS historical data is 71, but the standard texture scale that is determined by the enumeration method has increased to 83 or 95 with the continuous expansion of the town over several years. Therefore, under this unstable cognitive spatial-temporal system, the reliability of its domain knowledge is greatly reduced. In the scale sets 2 and 3, which are selected by the enumeration method, we can generally gain a smaller value in set 3 for most categories under 3 independent texture features. This shows that, under the same distinguishing performance, the independent texture features can slightly reduce the window scale value for texture observation in most instances. The correlation of the different texture scale sets was computed, as seen in Table 6. Table 6. Relevant indices of scale sets between the GIS-assisted method and the enumeration method. Relevance Indicators

Scale Sets 1 and Scale Sets 2

Scale Sets 1 and Scale Sets 3

Re p

0.93 0.000075

0.91 0.000039

The calculation results show that the relativity Re reaches 0.93, and the significance p reaches 0.00075 in the first group of scale sets, while the relativity Re reaches 0.91, and the significance p reaches 0.00039 in the second group of scale sets. The closer the Re value is to 1, the higher the relativity between the two sets of data is; and the further the p value is below 0.05, the more significant the correlation is. Although the relativity value in the second group is slightly worse than that of the first group, its significance value is better. It is reaffirmed that the increase in the independence of the texture features enables a more significant and clear value of the scale sets derived, while the scale sets also maintain a strong performance. 4.7. Classification Results Analysis To further evaluate the proposed method for setting texture scale sets, we carried out an experiment comparing classification ability. The SVM algorithm was employed. As a classic method, it has performed well in image analysis in many studies [1]. Specifically, we achieved image classification by using GLCM-based texture features under a series of typical single- or multi-scale sets. First, we choose any images cut from the RS data in the study area (as shown in Figure 7) for testing. The size of each cut image is 500 × 500 pixels, and its classification system is as shown in Table 7 and Figure 12. Then, the sample ratio parameter is set as 0.3, 0.6, or 0.8, while the other parameters remain unchanged from the SVM algorithm. Since the variation in classification accuracy is within 1%, its influence on the algorithm can be regarded as negligible. Therefore, we use the sample ratio of 0.3, in consideration of algorithm efficiency. The experiments results are as shown in Figure 13.


19 of 24

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW

19 of 24 19 of 24

(a1) (a1)

(b1) (b1)

(c1) (c1)

(d1) (d1)

(a2) (a2)

(b2) (b2)

(c2) (c2)

(d2) (d2)

Figure 12. The cut images and their true classification results in the GIS database. (a1) image 1, (a2) Figure 12. The cut images and their true classification results in the GIS database. (a1) image 1, (a2) true Figure 12. The cutimage2, images and true classification the result GIS database. (a1) image (a2) true result 1; (b1) (b2) their true result 2; (c1) imageresults 3, (c2)intrue 3; (d1) image 4, (d2)1, true result 1; (b1) image2, (b2) true result 2; (c1) image 3, (c2) true result 3; (d1) image 4, (d2) true result 4. true result 1; (b1) image2, (b2) true result 2; (c1) image 3, (c2) true result 3; (d1) image 4, (d2) true result 4. result 4.

(a) (a)

(b) (b)

(c) (c)

(d) (d) Figure 13. Cont.


20 of 24


20 of 24

(e) Figure 13. The classification results of the texture feature space under different scales. (a) Only Figure 13. The classification results of the texture feature space under different scales. (a) Only spectral spectral features are used; (b) scale = {89}, scale = {81}, scale = {79}, scale = {73}; (c) scale = scale set 1; features are used; (b) scale = {89}, scale = {81}, scale = {79}, scale = {73}; (c) scale = scale set 1; (d) scale = (d) scale = scale set 2; (e) scale = scale set 3. scale set 2; (e) scale = scale set 3.

The overall accuracy and kappa coefficient are calculated in Table 7. The overall accuracy and kappa coefficient are calculated in Table 7. Table 7. The classification results of the different features. Table 7. The classification results of the different features. Spectral Overall Image Data Categories Texture Features GLCM Scale Kappa Features Accuracy (%) Spectral Texture GLCM Overall / / 75.03 0.33 Kappa Image Data Categories Features Features Scale Accuracy (%) {89} 84.80 0.69 / 75.03 Image 1 {1,2,3,4,5,7,8,9,12} {R,G,B} scale/sets 1 90.58 0.860.33 scale sets 2 90.42 0.850.69 {89} 84.80 scale sets 3 90.25 0.850.86 Image 1 {1,2,3,4,5,7,8,9,12} {R,G,B} scale sets 1 90.58 / / 72.24 0.450.85 scale sets 2 90.42 {81} 85.60 0.730.85 scale sets 3 90.25 Image 2 {1,2,4,5,7,8,12} {R,G,B} scale sets 1 89.95 0.84 / / 72.24 0.45 scale sets 2 88.21 0.82 {81} 85.60 scale sets 3 87.65 0.810.73 Image 2 {1,2,4,5,7,8,12} {R,G,B} scale sets 1 89.95 / / 66.57 0.320.84 scale sets 2 88.21 {79} 79.72 0.630.82 scale sets 3 87.65 0.81 Image 3 {1,2,3,5,7,9,11,12} {R,G,B} 3 principal scale sets 1 87.95 0.83 / 66.57 scale/sets 2 86.83 0.810.32 component scale sets 3 88.69 0.840.63 {79} 79.72 / / 62.51 0.410.83 Image 3 {1,2,3,5,7,9,11,12} {R,G,B} 3 principal scale sets 1 87.95 {73} 84.35 0.780.81 component scale sets 2 86.83 Image 4 {1,2,3,4,5,6,7,12} {R,G,B} 3 principal scale sets 1 89.37 0.840.84 scale sets 3 88.69 scale sets 2 89.08 0.84 component / / 62.51 0.41 scale sets 3 90.55 0.86 {73} 84.35 0.78 Image 4 {1,2,3,4,5,6,7,12} {R,G,B} 3 principal scale setsfeature 1 89.37are as follows: 0.84 As shown in the results above, the classification abilities of different spaces component scale sets 2 89.08 0.84 the classification accuracy is lowest when only spectral features are used, followed by the integrated scale sets 3 90.55 0.86

usage of texture features under a single scale, and the highest accuracy, especially the highest kappa coefficient, is from the integrated usage of texture features under multi-scale sets. AsWhen shown in single-scale the results above, the classification abilities of 1, different feature spaces areaccuracy as follows: the value is gradually increased from the overall classification thesignificantly classification accuracy is reaches lowest when only spectral features arevalue, used, respectively, followed by is the89, integrated increases and the highest level when the scale 81, 79, usage texture under a single scale, and the highest accuracy, theclassification highest kappa 73 inof image 1–4.features When the multi-scale is respectively given as sets 1, sets 2, especially and sets 3, the accuracy is arefrom all high value and very close in images 1–4. However, the optimalsets. scale sets is different coefficient, the integrated usage of texture features under multi-scale forWhen these the different classification thatincreased is to say, no 1, a the kindoverall of scale sets has absolute single-scale value issystems, gradually from classification accuracy classification advantage. there are strong significant between the 3 groups significantly increases and Overall, reaches the highest leveland when the scalecorrelations value, respectively, is 89, 81, 79, 73 scale1–4. sets,When whichthe further proves the rationality and validity proposed method extracts in of image multi-scale is respectively given as setsof1,the sets 2, and sets 3, thethat classification the texture scale based on domain knowledge and GIS. accuracy are all high value and very close in images 1–4. However, the optimal scale sets is different for these different classification systems, that is to say, no a kind of scale sets has absolute classification advantage. Overall, there are strong and significant correlations between the 3 groups of scale sets, which further proves the rationality and validity of the proposed method that extracts the texture scale based on domain knowledge and GIS.


21 of 24


21 of 24

4.8. Algorithm Efficiency Analysis 4.8. Algorithm Efficiency Analysis (1) With the enumeration method, the best scale parameters can be gained by comparing texture (1) With theunder enumeration method, best parameters can be gained by comparing features numerous scales. the When thescale range of the enumerated scale was set from 3 totexture 111 in features under numerous scales. When the range of the enumerated scale was set from 3 to 111 this experiment, the algorithm’s time consumption changed, as shown in Figure 14a. in this experiment, the algorithm’s time consumption changed, as shown in Figure 14a.

(a)

(b)

Figure 14. 14. Algorithm Algorithm efficiency efficiency analysis. analysis. (a) (a) The The enumeration enumeration method method (2000 (2000 × × 2000 pixels); (b) (b) the the Figure 2000 pixels); GIS-assisted method. method. GIS-assisted

Clearly, this algorithm is extremely time-consuming. As the texture scale increases, the time Clearly, this algorithm is extremely time-consuming. As the texture scale increases, the time consumption for calculating texture features increases exponentially. When the scale is larger than consumption for calculating texture features increases exponentially. When the scale is larger than 15, 15, the time consumption exceeds 10 min, with the maximum even reaching 5 h (i.e., when scale = the time consumption exceeds 10 min, with the maximum even reaching 5 h (i.e., when scale = 111). 111). If the enumeration method is used to determine texture scale parameters, as the image data If the enumeration method is used to determine texture scale parameters, as the image data continues continues to increase, the accumulated time will reach an unbearable degree, such as in Equation (16): to increase, the accumulated time will reach an unbearable degree, such as in Equation (16): maxEN

y = maxEN ∑ k1eaxi y=

i =1 ∑

k1 eaxi

(16) (16)

i=1

The equation above is derived from the exponential regression line in this study, where k1 and a The equationdetermined above is derived the exponential regression line in this study, where k1 and a are the variables by thefrom computer’s performance, and maxEN represents the maximum are theofvariables determined by theIt computer’s performance, and maxEN represents the maximum value the enumeration number. shows that algorithm complexity is a relationship of cumulative value of the enumeration number. It shows that algorithm complexity is a relationship of cumulative exponential regression. exponential regression. (2) Based on the GIS and the domain knowledge-assisted method, the texture scale features can also effectively extracted. is realized through traversal, statistics, and the analysis of can patches (2) be Based on the GIS and theThis domain knowledge-assisted method, the texture scale features also from the existing GIS database using the ArcGIS platform. When the proposed method is widely be effectively extracted. This is realized through traversal, statistics, and the analysis of patches used, andexisting GIS data can be managed a standardized way,When then the its algorithmic complexity will from the GIS database usingin the ArcGIS platform. proposed method is widely be largely proportional to the number patches. The way, corresponding time consumption canwill be used, and GIS data can be managed in of a standardized then its algorithmic complexity reflected in the following statistical curve in Figure 14b. be largely proportional to the number of patches. The corresponding time consumption can be reflected in thethat following statistical curve in Figure 14b. It can be seen this algorithm has superior performance, especially in terms of efficiency. It only takes 180 s to perform the operation on the entire experimental area (region 4 of size 581 km2). It can be seen that this algorithm has superior performance, especially in terms of efficiency. Therefore, compared with the enumeration method, it substantially increases efficiency and is It only takes 180 s to perform the operation on the entire experimental area (region 4 of size 581 km2 ). therefore more suitable for high-resolution remote sensing image analysis of larger geo-data. We Therefore, compared with the enumeration method, it substantially increases efficiency and is therefore summarize the relationship between polygon numbers and time consumption in Equation (17): more suitable for high-resolution remote sensing image analysis of larger geo-data. We summarize the relationship between polygon numbers and y time in Equation (17): = bconsumption + k2 x (7)

The equation shows a simple linear relationship, and y = b + k2 xwhere b is the time of data management (17) data cleansing of GIS, and k2 is the variable determined by the computer’s performance. The algorithm has very low complexity; moreover, the calculation results are reasonable and valid, which is consistent with the scales selected by the enumeration method. It can be said that the texture scale


22 of 24

The equation shows a simple linear relationship, where b is the time of data management and data cleansing of GIS, and k2 is the variable determined by the computer’s performance. The algorithm has very low complexity; moreover, the calculation results are reasonable and valid, which is consistent with the scales selected by the enumeration method. It can be said that the texture scale extraction method based on the spatial knowledge of land-use proposed in this paper is scientific and reasonable and can meet the efficiency and precision requirements of classification. 5. Conclusions 1.

2.

3.

In a geo-scene classification of high-resolution RS images based on GLCM texture, the window scale of GLCM is an important parameter for determining its validity. Since the classification system changes with different needs, the observation windows of GLCM should also be changed correspondingly, so that the real texture features of classified objects can be intuitively reflected. Based on this, a one-to-one model that is strongly correlated with the actual definition and shape of classified objects was built to set the texture scales of each category. This is a new and practical method that is easy to understand and use. The proposed approach is based on the integrated computation of existing GIS data and current RS data, and the comprehensive use of a variety of domain shape knowledge, thereby avoiding the disadvantages of local optimum, sample dependence, and initial error accumulation when compared with other approaches relying only on the images themselves. Moreover, it shows a significant improvement in efficiency. In terms of the experimental results, the multi-scale value keeps the one-to-one relationship with the classified objects, and their value ranges are from a few to tens, rather than the smaller values of traditional analysis. Since most categories with complex texture features or mixed pixels are separable under bigger texture scales, the GLCM texture descriptors under these geo multi-scale windows can evidently represent the different geo categories. Meanwhile, the proposed approach has good performance that is the same as the classical enumeration approach and high calculation efficiency, so it can not only increase the total number of categories for a certain geo-scene, but also reduce the classification uncertainty. Obviously, the integrated use of RS image spatial features and GIS domain knowledge is proved necessary and effective in image geo-texture analysis.

Future research will construct a united platform in order to collectively handle and analyze the GIS and RS statistics and further contribute to the selection of the GLCM texture direction with the aid of GIS and domain knowledge. Author Contributions: Z.L. provided the initial idea for this study, conceived the experiments, and wrote the paper; Y.L. contributed the experimental data and supplied the infrastructure for the experiments; and Z.L. and Y.L. performed the experiments and analyzed the results. Funding: This research was funded by National Natural Science Foundation of China(NSFC) grant number 4130137. Acknowledgments: This paper was also received the help from B.J., M.Y., and Y.C., et al., in Guangzhou Urban Planning & Design Survey Research Institute. They provided important suggestions and reviews. Conflicts of Interest: The authors declare no conflicts of interest.

References 1. 2. 3. 4.

Lu, D.; Weng, Q. A survey of image classification methods and techniques for improving classification performance. Int. J. Remote Sens. 2007, 28, 823–870. [CrossRef] Liu, X. Summary of texture Research. Appl. Res. Comput. 2008, 25, 2284–2288. Haralick, R.M.; Shanmugam, K.; Dinstein, R. Textural features for image classification. IEEE Tran. Syst. Man Cybern. 1973, 3, 610–621. [CrossRef] Gong, P.; Howarth, P.J. Frequency-based contextual classification and gray-level vector reduction for land-use identification. Photogramm. Eng. Remote Sens. 1992, 58, 423–437.


5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31.

23 of 24

Randen, T.; Husoy, J. Filtering for Texture Classification: A Comparative Study. IEEE Trans. Pattern Anal. Mach. Intell. 1999, 21, 291–310. [CrossRef] Manjunath, B.S.; Ma, W.Y. Texture Features for Browsing and Retrieval of Image Data. IEEE Trans. Pattern Anal. Mach. Intell. 1996, 18, 837–842. [CrossRef] Reyes-Aldasoro, C.C.; Bhalerao, A. The Bhattacharyya space for feature selection and its application to texture segmentation. Pattern Recognit. 2006, 39, 812–826. [CrossRef] Malpica, N.; Ortuño, J.E.; Santos, A. A multi-channel watershed-based algorithm for supervised texture segmentation. Pattern Recognit. Lett. 2003, 24, 1545–1554. [CrossRef] Trevisani, S.; Rocca, M. MAD: Robust image texture analysis for applications in high resolution geomorphometry. Comput. Geosci. 2015, 81, 78–92. [CrossRef] Guo, Z.; Zhang, L.; Zhang, D. Rotation invariant texture classification using LBP variance (LBPV) with global matching. Pattern Recognit. 2010, 43, 706–719. [CrossRef] Atkinson, P.M.; Lewis, P. Geostatistical classification for remote sensing: An introduction. Comput. Geosci. 2000, 26, 361–371. [CrossRef] Herzfeld, U.C. Master of the obscure—Automated geostatistical classification in presence of complex geophysical processes. Math. Geosci. 2008, 40, 587–618. [CrossRef] Chica-Olmo, M.; Abarca-Hernandez, F. Computing geostatistical image texture for remotely sensed data classification. Comput. Geosci. 2000, 26, 373–383. [CrossRef] Zhang, C.; Franklin, S.E.; Wulder, M.A. Geostatistical and texture analysis of airborne-acquired images used in forest classification. Int. J. Remote Sens. 2004, 25, 859–865. [CrossRef] Lucieer, A.; Stein, A. Texture-based landform segmentation of LiDAR imagery. Int. J. Appl. Earth Obs. Geoinf. 2005, 6, 261–270. [CrossRef] Qian, X.; Hua, X.S.; Chen, P. An effective local binary patterns texture descriptor with pyramid representation. Pattern Recognit. 2011, 44, 2502–2515. [CrossRef] Ojala, T.; Pietikäinen, M.; Mäenpää, T. Multi-resolution gray-scale and rotation invariant texture classification with Local Binary Patterns. IEEE Trans. Pattern Anal. Mach. Intell. 2002, 24, 971–987. [CrossRef] Zhu, C.; Wang, R. Local multiple patterns based multi-resolution gray-scale and rotation invariant texture classification. Inf. Sci. 2012, 187, 93–108. [CrossRef] Qin, Q. Problems in Automatic Interpretation of Remote Sensing Images and Ways to Solve. Sci. Technol. Surv. Mapp. 2000, 25, 21–25. Chen, Z. Research on High-Resolution Remote Sensing Image Classification Technology. Ph.D. Thesis, Institute of Remote Sensing Applications, Chinese Academy of Sciences, Beijing, China, 2006. (In Chinese) Chen, S. Geo-spatial/temporal Analysis in Geo-processing. J. Remote Sens. 1997, 1, 161–169. (In Chinese) Huang, X.; Zhang, L.; Li, P. Remote sensing image classification based on high spatial resolution with shape and spectrum fusion. J. Remote Sens. 2007, 11, 193–199. Luo, H. Geo-Knowledge Assisted Soil Automatic Classification Method Based on Systematic Criteria by Remote Sensing in Mountainous Areas. Ph.D. Thesis, Wuhan University, Wuhan, China, 2005. (In Chinese) Shu, H. Knowledge based image classification approach supported by a GIS. Acta Geod. Cartogr. Sin. 1997, 26, 329–336. (In Chinese) Yang, C.; Zhou, C. Investigation on classification of remote sensing image on basis of knowledge. Geogr. Territ. Res. 2001, 17, 72–77. (In Chinese) Luo, J.; Zhou, C.; Yang, Y. Classification of Land Cover/Land Use Based on Remote Sensing Geo-intelligent Graphical Model. J. Nat. Resour. 2001, 16, 179–183. (In Chinese) Huang, H.; Yang, Q.; Wang, C. Relationship between land use types and topographic factors based on DEM. J. Southwest Univ. (Nat. Sci. Ed.) 2009, 31, 159–164. (In Chinese) Ou, L.; He, Z.; Ma, H.; Zheng, H. Application of knowledge-based stratified classification in remote sensing information extraction of land use/land cover. Sci. Technol. Surv. Mapp. 2008, 33, 173–175. (In Chinese) Wan, Y.; Huang, J. Influence of geometric and graph theoretical measures on land classification using high-Resolution remote sensing images. Geomat. Inf. Sci. Wuhan Univ. 2009, 34, 794–798. (In Chinese) Li, D.; Wang, S. Spatial Data Mining Theory and Application; Science Press: Beijing, China, 2006. (In Chinese) Marceau, D.J.; Howarth, P.J.; Dubois, J.M.; Gartton, D.J. Evaluation of the grey-level co-occurrence matrix method for land-cover classification using SPOT imagery. IEEE Trans. Geosci. Remote Sens. 1990, 28, 513–519. [CrossRef]


32.

33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52.

24 of 24

Cui, L. Integrative Analysis and Evaluation of the Interpretation Features in Remote Sensing Image. Institute of Remote Sensing Applications. Ph.D. Thesis, Chinese Academy of Sciences, Beijing, China, 2005. (In Chinese) Huang, H. Scale Issues in Object-Oriented Image Analysis. Ph.D. Thesis, Institute of Remote Sensing Applications, Chinese Academy of Sciences, Beijing, China, 2003. (In Chinese) Ma, H.; Li, D. Geographic Window Fixing and Its Application Based on Spatial Statistics. Geomat. Inf. Sci. Wuhan Univ. 2001, 26, 18–23. (In Chinese) Woodcock, C.E.; Strahler, A.H. The factor of scale in remote-sensing. Remote Sens. Environ. 1987, 3, 311–332. [CrossRef] Wang, L.; He, D. Texture unit, texture spectrum, and texture analysis. IEEE Trans. Geosci. Remote Sens. 1990, 4, 509–512. Goodchild, M.F.; Quattrochi, D.A. Scale in Remote Sensing and GIS; CRC Press LLC Lewis Pulishers: Boca Raton, FL, USA, 1997. Goodchild, M.F. Metrics of scale in remote sensing and GIS. Int. J. Appl. Earth Obs. Geoinf. 2001, 3, 114–120. [CrossRef] Cao, C.; Lam, N.S.N. Understanding the Scale and Resolution Effects in Remote Sensing and GIS; Scale in Remote Sensing and GIS; Lewis Publishers: Boca, Raton, FL, USA, 1997; pp. 57–72. Franklin, S.E.; Wulder, M.A.; Lavigne, M.B. Automated derivation of geographic window sizes for use in remote sensing digital image texture analysis. Comput. Geosci. 1996, 22, 665–673. [CrossRef] Cao, J.; Dong, Y. A Multi-scale Texture Segmentation Method. In Proceedings of the 2014 10th International Conference on Natural Computation, Xiamen, China, 19–21 August 2014; pp. 873–877. Siqueira, F.R.; Schwartz, W.R.; Pedrini, H. Multi-scale gray level co-occurrence matrices for texture description. Neurocomputing 2013, 120, 336–345. [CrossRef] Binaghi, E.; Gallo, I.; Pepe, M. A cognitive pyramid for contextual classification of remote sensing images. IEEE Trans. Geosci. Remote Sens. 2003, 12, 12–14. [CrossRef] Bo, H.; Ma, F.; Jiao, L. Analysis of computational complexity of gray level co-occurrence matrix for image texture. J. Electr. Inf. Technol. 2006, 34, 155–158. (In Chinese) Zhang, Y. Image Engineering; Qinghua University Press: Beijing, China, 1999. (In Chinese) Rafael, C.G. Digital Image Processing (MATLAB Version); Electronic Industry Press: Beijing, China, 2005. Luo, D. Pattern Recognition and Image Processing; Horwood Publishing: Chichester, UK, 1998. Wang, W.; Han, T. Land Use Planning; China Agriculture Press: Beijing, China, 2004. (In Chinese) Wu, J. Landscape Ecology—Concept and Theory. Chin. J. Ecol. 2000, 19, 42–52. (In Chinese) Yan, S.; Wang, T.; Liu, Y.; Xiao, C. Remote sensing image human-computer interaction interpretation system and its technical features. J. Remote Sens. 2002, 6, 198–204. (In Chinese) Tamura, H.; Shunji, M.; Yamawaki, T. Textural features corresponding to visual perception. IEEE Trans. Syst. Man Cybern. 1978, 8, 460–473. [CrossRef] Liu, Y. Study on Land Use Spatial Zoning Based on Multi-Objective Optimization Model. Ph.D. Thesis, Wuhan University, Wuhan, China, 2008. (In Chinese) © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Study on Multi-Scale Window Determination for GLCM Texture ... - MDPI

Study on Multi-Scale Window Determination for GLCM Texture ... - MDPI

Suggest Documents

Determination of the Optimum Harvest Window for Apples ... - MDPI

Determination of the Optimum Harvest Window for Apples ... - MDPI

Study on Components Determination and Performance ... - MDPI

Texture recognition by using GLCM and various ... - Google Sites

A Quantitative Study of Texture Features across Different Window ...

Extraction of Texture Features using GLCM and Shape Features using ...

A Study for Texture Feature Extraction of High-Resolution ... - MDPI

Multiscale Fractal Descriptors Applied to Texture Classification

Texture Descriptor Over Multiscale Edge ... - CiteSeerX

Texture Classification through Multiscale Orientation Histogram Analysis

a review on multiscale texture features using steerable pyramids

Multiscale Texture Characterization of Wastewater Sludges ... - ORBi

Multiscale Interface Effect on Homogeneous Dielectric ... - MDPI

3D-Guided Multiscale Sliding Window for Pedestrian Detection - Uab

Determining optimal window size for texture feature extraction methods

Window size selection for texture image generation ...

Experimental Study on the Precise Orbit Determination of the ... - MDPI

Multiscale Information Decomposition: Exact Computation for ... - MDPI

A study for multiscale information transfer measures based on ... - PLOS

Determination of Priority Study Areas for Coupling CO2 ... - MDPI

Texture Profile and Color Determination on Local and

preliminary study of print mottle analysis in flexography using glcm

A Comparative Study for Texture Classification ... - CiteSeerX

Determination of Window Size for Baseline Estimation ...