Temporal and Spatial Analyses of the Landscape Pattern of Wuhan City Based on Remote Sensing Images

International Journal of Geo-Information

Article Temporal and Spatial Analyses of the Landscape Pattern of Wuhan City Based on Remote Sensing Images

Jianjun Lv 1,2, Teng Ma 1, Zhiwen Dong 1, Yao Yao 1,* ID and Zehao Yuan 1

1 School of Information Engineering, China University of Geosciences, Wuhan 430074, China; [email protected] (J.L.); [email protected] (T.M.); [email protected] (Z.D.); [email protected] (Z.Y.) 2 National Engineering Research Center for Geographic Information System, Wuhan 430074, China * Correspondence: [email protected]; Tel.: +86-159-7210-3607

Received: 9 July 2018; Accepted: 20 August 2018; Published: 22 August 2018

Abstract: With the acceleration of the process of building a national-level central city in Wuhan, the landscape pattern of the city has undergone tremendous changes. In this paper, remote images are classiﬁed through the neural network classiﬁcation method, based on texture extraction, and the evolution of landscape patterns was quantitatively analyzed, based on the method of moving windows, landscape metrics and urban density calculation, in order to accurately extract landscape types and perform quantitative analyses. Wuhan City is taken as an example. The surface coverage of Wuhan City from 1989 to 2016 is divided into four types: agricultural landscape clusters, forest landscape clusters, water landscape clusters, and urban landscape clusters. It was concluded that, during the study period, the landscape heterogeneity of the entire area in Wuhan has increased, but the central urban area in Wuhan has decreased. The development of urban areas has compacted inwards but expanded outwards. In addition, the western part of Wuhan City developed better than the eastern part.

Keywords: landscape pattern; neural network; remote sensing application; urban land; land use

1. Introduction With the rapid development of global technology and economy, global problems, such as increased urbanization and reduced biodiversity, have surfaced [1,2]. China has been in a period of rapid development in recent years, and the speed of urbanization in Wuhan, a national central city in China, is accelerating. The proportion of urban land-use types has been constantly changing in the total area of urban land, thereby affecting the entire landscape pattern and ecological processes of the city [3,4]. Therefore, the study of landscape patterns is extremely important for the urban planning and management of Chinese cities with rapid development. Traditional landscape pattern analysis methods are mostly qualitative, analyzing the characteristics of the landscape pattern from the perspective of landscape ecology [5]. The most basic quantitative analysis method is spatial statistical analysis [6]. It analyses landscape changes by counting the area and proportion of various types of land. On this basis, several scholars have used the Markov transition matrix to analyze the loss direction of a certain type of land, and the source of a new type of land [7]. However, the accuracy of the traditional method was not high enough, and the results of the analysis greatly deviated from the actual situation [8]. With the development of remote sensing technology, some scholars have applied this technology to landscape pattern analysis [9]. Common methods for landscape classification are unsupervised classification and supervised classification [10]. Lillesand has experimentally pointed out that using

ISPRS Int. J. Geo-Inf. 2018, 7, 340; doi:10.3390/ijgi7090340 www.mdpi.com/journal/ijgi ISPRS Int. J. Geo-Inf. 2018, 7, 340 2 of 17 hybrid classification (meaning a combination of supervised with unsupervised classification) to classify images can achieve higher classification accuracy [11]. Moreover, with the rise of neural networks, some scholars have used Back Propagation (BP) neural networks to classify images and achieve better results [12]. Based on this finding, McGarigal et al. have proposed a calculation method of landscape metrics [13], which is widely used in the analysis of landscape patterns, such as the urban landscape pattern [14], vegetation pattern [15], urban growth [16], and other subjects of study. Landscape metrics can quantitatively describe and monitor changes in landscape spatial structures over time, but it cannot specifically reflect the direction and rationality of landscape changes [17]. Jiao has proposed an “inverse S-shaped function” to formulate urban land density for urban land [18], which can identify the extent of the built-up area of Chinese cities and measure the compactness and expansion intensity of urban land. However, he only discusses the spatial distribution of urban land use, and does not analyze the changes in land-use structure during the development of a city. There are two problems in the current studies on landscape pattern. On the one hand, the interpretative features of images of different landscape types have been unclear, which has led to the prevalence of the “same object with different spectra, different objects with the same spectrum” phenomena [19]. On the other hand, most of these analyses are based on landscape metrics; this research method is relatively simple, and cannot reflect the complexity of the landscape pattern change in the rapid development process of Chinese cities [20]. Based on the above problems, this study combines remote sensing technology and an “inverse S-shaped function” model to quantitatively analyze the landscape pattern evolution of Wuhan from 1989 to 2016 from multiple angles. First, with the aim of improving the accuracy of image classification, we selected the texture features of the remote sensing image using the gray level co-occurrence matrix method, and construed a texture-based BP neural network model. Then, with the aim of analyzing the spatial and temporal characteristics of the landscape pattern in Chinese cities during the rapid development process, from the angles of landscape spatial characteristics, urban compactness, and development direction, etc., we have combined the methods of moving the window and the “inverse S-shaped function”.

2. Materials and Methods

2.1. Study Area and Data Wuhan is located in the Jianghan Plain, the central part of China and the eastern part of the Hubei Province. The elevation is less than 50 m in most areas of Wuhan. It is a national, historical, and cultural city, and an important industrial, scientiﬁc, and educational base in China. From the statistical yearbook of Wuhan (http://www.whtj.gov.cn/), it is found that the built-up area of Wuhan has been expanding, and the population density has risen rapidly. As of 2017, the total built-up area had reached 719.85 km2. From 1995 to 2017, the total population increased from 7.10 million to 10.77 million, which is an increase of 3.67 million, and the urbanization rate of permanent residents reached 79.3% in 2017. Moreover, since 1995, Wuhan’s economy has continued to grow rapidly. The gross national product grew from 60.69 billion RMB Yuan in 1995 to 1191.22 billion RMB Yuan in 2017, which is an average annual increase of 53.83 billion RMB Yuan. Wuhan has played a leading role in the “1 + 8” urban circle, centered on Wuhan, and driven the economic and social development of the surrounding areas [21]. The advantages of its central position have brought policy and economic support to Wuhan. As a result, population, funds, and other resources gradually concentrate here under an agglomeration effect [22]. There are 13 jurisdictions in Wuhan, including seven central urban areas (Jiang’an District, Jianghan District, Qiaokou District, Qingshan District, Wuchang District, Hongshan District, and Hanyang District) and six suburbs (Dongxihu District and Hannan District are suburbs; Caidian District, Jiangxia District, Huangpi District, and Xinzhou District are far urban areas). This study focused on the 13 jurisdictions of Wuhan City. ISPRS Int. J. Geo-Inf. 2018, 7, 340 3 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 3 of 17

TheThe time period for this study was from 1989 to 2016. The Thematic Mapper (TM) images of Landsat-5Landsat-5 satellites satellites and and the the Operational Land Land Imag Imagerer (OLI) (OLI) images of Landsat-8 satellites, with with a resolutionresolution of of 30 30 m, m, were were used. The The image image data data of of th thee time time series series are are shown shown in in Table Table 11.. ThreeThree imagesimages were acquired at each time point, and the track numbers were 122/39, 123/38, 123/38, and and 123/39. 123/39.

TableTable 1. ImageImage data.

Image Type Acquisition Date (Month/Year) Image Type Acquisition Date (Month/Year) 2/1989 2/1989 4/19954/1995 4/20014/2001 Landsat-5Landsat-5 4/20044/2004 5/20075/2007 5/2010 5/2010 4/2013 Landsat-8 4/2013 Landsat-8 5/2016 5/2016

TheThe Environment for VisualizingVisualizing ImagesImages (ENVI) (ENVI) platform platform was was used used to to do do some some preprocessing preprocessing of ofimages images [23 [23].]. After After preprocessing, preprocessing, the the three three images images were were inlaid inlaid according according to the to required the required research research scope, scope,and the and administrative the administrative division divi datasion of data Wuhan of Wuhan were thenwere superimposed then superimposed on them. on them. Finally, Finally, the study the studyarea was area cut, was as cut, shown as shown in Figure in Figure1, and 1, and bands bands 6, 4, 6, and4, and 3 were3 were assigned assigned to to red, red, green, green, andand blue,blue, respectively,respectively, for synthesizing pseudo-color images.images.

Figure 1. Remote sensing image of Wuhan. Figure 1. Remote sensing image of Wuhan.

2.2. Study Method 2.2. Study Method Based on the texture features extracted from the images, this paper establishes a BP neural Based on the texture features extracted from the images, this paper establishes a BP neural network classification model to classify images in the study area. The temporal and spatial network classiﬁcation model to classify images in the study area. The temporal and spatial characteristics of landscape patterns were analyzed by selecting landscape indices, using moving characteristics of landscape patterns were analyzed by selecting landscape indices, using moving windows, and calculating density curve methods. A technical flow chart is shown in Figure 2. windows, and calculating density curve methods. A technical ﬂow chart is shown in Figure2.

ISPRS Int. J. Geo-Inf. 2018, 7, 340 4 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 4 of 17

Figure 2. Technical ﬂowflow chart.

2.2.1. Texture-Based BackBack PropagationPropagation (BP)(BP) NeuralNeural NetworkNetwork ClassificationClassification The elementelement with with the the lowest lowest level level in the in landscape the landscape pattern pattern system issystem the homogeneous is the homogeneous landscape unit,landscape and itsunit, upward and its classification upward classification complies with comp thelies relationship with the relationship between the between natural bodiesthe natural and landscapebodies and units landscape of similar units neighboring of similar neighboring natural landscapes natural [ 24landscapes,25]. Based [24,25]. on the Based principle on the of landscape principle classification,of landscape classification, and combined and with combined the research with data the andresearch purposes, data theand landscape purposes, ofthe the landscape overall Wuhan of the areaoverall is dividedWuhan into area the is following divided clustersinto the (landscape following clusters clusters are (landscape an organic cl combinationusters are an of multipleorganic similarcombination landscape of multiple functional similar areas): landscape functional areas): 1. Forest landscape cluster, including landscape units such as woodland, garden, and grassland. 1. Forest landscape cluster, including landscape units such as woodland, garden, and grassland. 2. Agricultural landscape cluster, including landscape units such as dry land and vegetable plots. 2. Agricultural landscape cluster, including landscape units such as dry land and vegetable plots. 3. Water landscape cluster, including landscape units such as the Yangtze River, lakes, and rivers. 3. Water landscape cluster, including landscape units such as the Yangtze River, lakes, and rivers. 4. Urban landscape cluster: including landscape units, such as houses, roads, and rural 4. settlements.Urban landscape cluster: including landscape units, such as houses, roads, and rural settlements. In thisthis paper,paper, the the gray gray level level co-occurrence co-occurrence matrix ma (GLCM)trix (GLCM) method method was usedwas toused extract to extract the texture the featurestexture features of the images of the ofimages Wuhan of CityWuhan [26]. City The [26]. GLCM The starts GLCM from starts a pixel from at a a pixel certain at positiona certainA position(x, y) in theA(x image,, y) in the and image, counts and the probabilitycounts the probabilityP(i, j, δ, θ) of P the(i, j occurrence, δ, θ) of the of occurrence pixels (x + ∆ofx, pixels y + ∆y ()x whose + ∆x, y image + ∆y) haswhose a gray image value has of a jgrayin a circularvalue of range,j in a circular with a distancerange, with of δ afrom distanceA. The of mathematicalδ from A. The formulamathematical was formula was P(i, j, δ, θ) = {[(x, y), (x + ∆x, y + ∆y)] f (x, y) = i (1) P(i,j ,f (x,+)∆x, y{[(+ ∆xy,)y =),j;(xx=0, 1,x 2,···y, NX−y1;)]yf=(x0,, 1,y ) 2 ··· ,iNy − 1 (1) f(x  x,y  y )  j;x  0,1,2  ,N X  1;y  0,1,2  ,N y  1} where x and y were the pixel coordinates in the image; and Nx and Ny were the number of rows and columnswhere x and of the y were images, the respectively.pixel coordinates in the image; and Nx and Ny were the number of rows and columnsHaralick of the has images, defined respectively. 14 kinds of texture feature parameters from four aspects: statistical characteristics,Haralick has characteristics defined 14 of kinds information of texture theory, feature characteristics parameters of correlation,from four andaspects: characteristics statistical ofcharacteristics, visual texture characteristics [26]. The gray of levelinformation co-occurrence theory, matrixcharacteristics can measure of correlation, various informationand characteristics in the image,of visual and texture the feature [26]. imageThe gray can level visually co-occurrence display the matrix thickness can of measure the texture various and the information characteristics in the of eachimage, direction. and the Commonfeature image statistical can visually features display are entropy, the thickness contrast, energy,of the texture and correlation. and the characteristics By extracting theof each texture direction. features, Common the phenomena statistical of thefeatures “same ar objecte entropy, with differentcontrast, spectra, energy, different and correlation. objects with By theextracting same spectrum” the texture can features, be better the solved, phenomena which helpsof the to “same improve object the classificationwith different accuracy spectra, of different images. objectsSince with there the wassame still spectrum” a certain can correlation be better between solved,the which bands, helps in orderto improve to reduce the theclassification time and amountaccuracy of of calculationimages. of neural network training, the fifth band (with the most abundant surface information)Since there and was the still red banda certain of visible correlation light werebetween selected the bands, after the in principalorder to reduce component the time analysis and (PCA).amount Finally, of calculation the images of withneural the network texture featurestraining of, the contrast fifth band (CON), (with angular the most second abundant moment surface (ASM), entropyinformation) (ENT), and and the correlation red band (COR)of visible were light extracted were selected (Figure 3after). the principal component analysis (PCA). Finally, the images with the texture features of contrast (CON), angular second moment (ASM), entropy (ENT), and correlation (COR) were extracted (Figure 3).

ISPRS Int. J. Geo-Inf. 2018, 7, 340 5 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 5 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 5 of 17

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 5 of 17

(a) (b) (c) (d) (a) (b) (c) (d) FigureFigure 3. Four 3. Four texture texture feature feature images images of Thematic of Thematic Mapper Mapper (TM) images(TM) images in the studyin the area:study( aarea:) correlation (a) Figure 3. Four texture feature images of Thematic Mapper (TM) images in the study area: (a) (COR)correlation texture feature;(COR) texture (b) angular feature; second (b) angular moment second (ASM) moment texture (ASM) feature; texture (c) feature; entropy (c (ENT)) entropy texture feature;(ENT)correlation and texture (d) (COR) contrast( afeature;) texture (CON) and (feature;d) texturecontrast (b()b feature. (CON))angular texture second feature. moment( c) (ASM) texture feature;(d) (c) entropy (ENT) texture feature; and (d) contrast (CON) texture feature. Figure 3. Four texture feature images of Thematic Mapper (TM) images in the study area: (a) Basedcorrelation on the extracted(COR) texture texture feature; features, (b) angular this second paper moment used (ASM) a BP texture networ feature;k model (c) entropy with only one Based on the extracted texture features, this paper used a BP network model with only one hiddenBased layer(ENT) on for thetexture classification. extracted feature; and texture The(d) contrast input features, (CON) layer thistexture of thepaper feature. neural used network a BP networ was sikx modelmulti-spectral with only bands one hiddenandhidden layer eight layer for texture classiﬁcation. for classification. images. The The Theoutput input input layer layer ofwas of the sixthe neural initial neural networkclassification network was was categories six six multi-spectral multi-spectral and one bands other and eightcategory.and texture eight images.BasedThe texture transfer on Theimages.the extractedfunction output The layer textureoutput(i.e., the was features, layer activati six was initialthison sixpaper function) classiﬁcationinitial used classification aof BP the networ categorieshiddenk categoriesmodel layer and with and oneand onlyoutput otherone one otherlayer category. hidden layer for classification. The input layer of the neural network was six multi-spectral bands The transferadoptedcategory. functiona The logarithmic transfer (i.e., functionlogic the activationto achieve(i.e., the a function)activati highlyon nonlinear offunction) the hidden mapping of the layer hidden from and the layer outputinput and to output layer the output adoptedlayer a and eight texture images. The output layer was six initial classification categories and one other layer.adopted The a basiclogarithmic principle logic diagram to achieve of neural a highly network nonlinear training mapping is show nfrom in Figure the input 4. For to the the training output logarithmiccategory. logic toThe achieve transfer afunction highly (i.e., nonlinear the activati mappingon function) from of thethe hidden input tolayer the and output output layer. layer The basic layer. The basic principle diagram of neural network training is shown in Figure 4. For the training principleof the diagram adoptedBP neural a oflogarithmic network, neural networkitlogic was to necessary achieve training a to highly dete is shown rminenonlinear several in Figuremapping important4 .from For the parameters input training to the in of theoutput the training BP neural of the BP neural network, it was necessary to determine several important parameters in the training network,model, itlayer. wasincluding The necessary basic training principle to determine contribution diagram of several neural threshold network important (c training), training parameters is show momentumn in in Figure the training (4.m For), training the model,training rate including (r), networkmodel, including global error training (e), and contribution number of training threshold iterations (c), training (n). momentum (m), training rate (r), training contributionof the BP neural threshold network, it ( cwas), training necessary momentum to determine several (m), training important rate parameters (r), network in the training global error (e), networkmodel, global including error (etraining), and number contribution of training threshold iterations (c), training (n). momentum (m), training rate (r), and numbernetwork of training global error iterations (e), and number (n). of training iterations (n). Target Target Target InputNeural Networks(Including Output InputNeuralconnection Networks(Including weights and Output Comparing InputNeural Networks(Including Output connectionconnectionthresholds) weights weights and ComparingComparing thresholds) thresholds) Adjust weights and Adjustthresholds weights weights and and thresholdsthresholds

Figure 4. The basic principle of neural network training. Figure 4. FigureFigure 4.The 4.The The basic basic basic principle principle principle of nene neuraluralural network network network training. training. training.

This studyThis study has selectedhas selected many many group group parameters parameters forfor comparison comparison experiments. experiments. The Theerror error curve curve of of This study has selected many group parameters for comparison experiments. The error curve of Thisneural studyneural network network has training, selected training, under many under variou group various sparameter parameters parameter settings, for comparison is is shown shown in inFigure experiments. Figure 5. Figure 5. Figure 5a The is 5aa set erroris ofa set curve of of neuralsettingsneural networksettings network with training, withgood training, good training under training under various results results variou parameterafter afters parameter multiple multiple settings, tesettings,sts, andisand shown isFigure Figure shown in5b Figure5b isin theisFigure the default5. Figuredefault 5. Figureset5 ofaset issettings, 5a of a setissettings, a of set settings of with goodrecommendedsettingsrecommended training with good resultsby the bytraining afterthesystem. system. multiple results It Itis isfoundafter tests,found multiple andthat that Figurethethe te trainingsts,5b and is timethe timeFigure defaultis isabout 5babout setis 4 the h of4 when h settings,default when the parametersetthe recommended ofparameter settings, by recommended by the system. It is found that the training time is about 4 h when the parameter the system.settingssettings It are is foundthose are those shown that shown the in trainingFigurein Figure 5a, 5a, time whereas, whereas, is about whenwhen 4 h thethe when parameter parameter the parameter settings settings are arethose settings those used areused in Figure those in Figure shown in settings5b, arethe thosetraining shown time is in about Figure 5 h. 5a, Moreover, whereas, the when final training the parameter error in thesettings group are illustrated those used in Figure in Figure Figure5b,5a, the whereas, training when time is the about parameter 5 h. Moreover, settings the are fina thosel training usedin error Figure in the5b, group the training illustrated time in is Figure about 5 h. 5a5b, is the lower5a training is lower than than timethat that inis aboutof in the of the group5 h. group Moreover, shown shown in inthe Figure Figure final 5b, training so so this this errorstudy study choosesin choosesthe group the theparameter illustrated parameter settings in settings Figure Moreover,demonstrated the final training in Figure error 5a. in the group illustrated in Figure5a is lower than that in of the group demonstrated5a is lower than in thatFigure in of 5a. the group shown in Figure 5b, so this study chooses the parameter settings showndemonstrated in Figure5b, in so Figure this study 5a. chooses the parameter settings demonstrated in Figure5a.

(a) (b)

(a) (b) (a) (b)

Figure 5. (a) is the neural network training error curve under a set of settings with good training results after multiple tests and (b) is the error curve under the default set of settings recommended by the system. ISPRS Int. J. Geo-Inf. 2018, 7, 340 6 of 17

2.2.2. Analytical Method of Landscape Metrics Landscape metrics is a measure that quantitatively represents the composition of the landscape pattern and the spatial form or distribution of different landscape types [27]. FRAGSTATS computes three levels of metrics, corresponding to (1) each patch in the mosaic, (2) each patch type (class) in the mosaic, and (3) the landscape mosaic as a whole [13]. Landscape metrics are large in number, although the repeatability and correlation of the landscape information contained in the metrics are high [28]. It is very important to select appropriate metrics for the study. In this paper, the patch density (PD) and Shannon diversity index (SHDI) of the landscape level were selected based on the regional features of Wuhan and the study objective.

1. Patch density (PD)

Patch density represents the patch number of one speciﬁc landscape type per one-hundred-hectare area. The formula is: N PD = i (2) A The PD value can reﬂect the fragmentation degree of the landscape in the study area [29]. The larger the patch number per unit area, the higher the fragmentation degree of the landscape in the area.

2. Shannon diversity index (SHDI)

Heterogeneity refers to the degree of confounding of different types of landscapes within the region [30]. A high degree of confounding indicates that the distribution of different types of landscapes is balanced and has strong heterogeneity [30]. If a certain type of landscape in a region is shown to be superior, the landscape pattern in the region is unbalanced and the heterogeneity is low, which is not conducive to the ecological development of the region [31]. The Shannon diversity index (SHDI) represents the heterogeneity of the landscape in the study area. The formula is:

m SHDI = − ∑ (Pi ln Pi) (3) i=1

The range of SHDI values is SHDI ≥ 0. The larger the SHDI, the greater the mixing degree and heterogeneity of patch types, so the distribution of patch types became more and more balanced in the landscape. Based on the calculation results of the selected landscape metrics, the landscape pattern was further analyzed using the moving window method. The moving window method is used to calculate the landscape metrics selected in the “calculation window”, and output the corresponding raster map to observe the spatial variability of the landscape pattern. In addition, this method could clarify the information from the landscape spatial pattern [32]. In comparison with traditional analyses, which are based on overall features, using the moving window method for performing spatial gradient analysis in landscape pattern could realize the quantiﬁcation and spatial visualization of the landscape metrics in the local area, and thereby more clearly exhibit the dynamic spatial change of the landscape pattern and reveal the differences in the internal structure of the landscape [33].

2.2.3. Urban Density Curve Urban density analyses have always been the focus in the study of urban expansion. This paper used the concentric circles division method to calculate the urban land use density of each ring [18]. Then, the spatial variation regularity of urban land use was analyzed, based on the relation between urban land use density variation and the distance from the urban center. The least square method is used to ﬁt the inverse S-shaped function of urban land use, based on Python. The formula is ISPRS Int. J. Geo-Inf. 2018, 7, 340 7 of 17

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 7 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 7 of 17 The least square method is used to fit the invers1e− S-shapedc function of urban land use, based on Python. The formula is f(r) = (4) The least square method is used to fit the inverseα S-shaped( 2r − ) function of urban land use, based on + D 1 Python. The formula is 1 e f(r) = ( ) (4) where f is urban density, r is the distance from the urban center, e is the Euler value, α, c, and D are f(r) = ( ) (4) constants,where and f is D urban is the density, maximum r is the radius distance of from the urban the urban center. center, e is the Euler value, α, c, and D are constants, and D is the maximum radius of the urban center. Thewhere function f is urban form, density, when r is theα = distance 4, c = 0.05,from the and urban D = center, 30, is showne is the Euler in Figure value,6 .α, The c, and urban D are land The function form, when α = 4, c = 0.05, and D = 30, is shown in Figure 6. The urban land density densityconstants, function and is D a continuouslyis the maximum monotonic radius of the decreasing urban center. differentiable function. It has two horizontal function is a continuously monotonic decreasing differentiable function. It has two horizontal asymptotes,The ffunction = 1 and form, f = when c. When α = 4, r c = = 00.05, (representing and D = 30, is the shown urban in Figure center), 6. The f(r) urban is close land to density 1; when r asymptotes, f = 1 and f = c. When r = 0 (representing the urban center), f(r) is close to 1; when r function is a continuously monotonic decreasing differentiable function. It has two horizontal approachesapproaches inﬁnity, infinity, f(r) is f(r) close is close toc. to c. asymptotes, f = 1 and f = c. When r = 0 (representing the urban center), f(r) is close to 1; when r approaches infinity, f(r) is close to c.

Figure 6. Graphs of the inverse S-shaped model. Figure 6. Graphs of the inverse S-shaped model.

3. Results 3. Results Figure 6. Graphs of the inverse S-shaped model. The Kappa coefficient of image classification results is shown in Table 2. The classification The3. Resultsresults Kappa meet coefficient the requirements of image of the classification accuracy of this results research. is shown The results in Table are 2shown. The in classification Figure 7. results meet the requirementsThe Kappa coefficient of the accuracy ofTable image 2. Calculation of cl thisassification research. results results of TheKappa is results coefficient. shown are in shown Table 2. in The Figure classification7. results meet the requirements of the accuracy of this research. The results are shown in Figure 7. Year Table1989 2. Calculation1995 2001 results 2004 of Kappa 2007 coefficient. 2010 2013 2016 Kappa coefficient 0.8854Table 2. 0.8063Calculation 0.7924 results 0.7928 of Kappa 0.8178 coefficient. 0.8056 0.8236 0.8389 Year 1989 1995 2001 2004 2007 2010 2013 2016 Year 1989 1995 2001 2004 2007 2010 2013 2016 Kappa coefficient 0.8854 0.8063 0.7924 0.7928 0.8178 0.8056 0.8236 0.8389 Kappa coefficient 0.8854 0.8063 0.7924 0.7928 0.8178 0.8056 0.8236 0.8389

Figure 7. Landscape classiﬁcation results of Wuhan from 1989 to 2016. ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 8 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, 340 8 of 17

Figure 7. Landscape classification results of Wuhan from 1989 to 2016. 3.1. Analyses of the Temporal and Spatial Features of Landscape Metrics 3.1. Analyses of the Temporal and Spatial Features of Landscape Metrics The patch density (PD) and Shannon diversity index (SHDI), at the landscape level, can reﬂect the The patch density (PD) and Shannon diversity index (SHDI), at the landscape level, can reflect basic spatial features of the landscape pattern, such as the fragmentation degree and heterogeneity [34]. the basic spatial features of the landscape pattern, such as the fragmentation degree and The calculation results of PD and SHDI of the entire area in Wuhan are shown in Table3. The variation heterogeneity [34]. The calculation results of PD and SHDI of the entire area in Wuhan are shown in trend of PD and SHDI is shown in Figure8. From 1989 to 2016, the PD of the landscapes of Wuhan Table 3. The variation trend of PD and SHDI is shown in Figure 8. From 1989 to 2016, the PD of the showed a decreasing trend, but SHDI showed an increasing trend. landscapes of Wuhan showed a decreasing trend, but SHDI showed an increasing trend. Table 3. Patch density (PD) and Shannon diversity index (SHDI) calculation results at the Tablelandscape 3. Patch level. density (PD) and Shannon diversity index (SHDI) calculation results at the landscape level. Year PD SHDI Year1989 3.6556 PD 0.904 SHDI 19891995 3.4896 3.6556 0.9064 0.904 19952001 3.801 3.4896 0.9257 0.9064 2001 3.801 0.9257 20042004 3.2578 3.2578 0.956 0.956 20072007 2.8566 2.8566 0.971 0.971 20102010 2.4352 2.4352 1.004 1.004 20132013 2.6703 2.6703 0.9998 0.9998 2016 2.2976 1.0166 2016 2.2976 1.0166

Patch densitydensity (PD) (PD) reflects reflects the the fragmentation fragmentation degree degree of the of landscape the landscape in the research in the research area. Based area. on Basedthe calculation on the calculation results of PD,results this of paper PD, firstthis analyzespaper first the analyzes fragmentation the fragmentation degree of the degree landscape of the in Wuhanlandscape City. in Wuhan In order City. to specificallyIn order to specifically analyze the analyze landscape the landscape changes inchanges different in different areas of areas Wuhan, of Wuhan,the moving the windowsmoving windows method was method adopted was fromadopted the upperfrom the left upper corner left of thecorner research of the area research at two area scales, at 1two km scales, and 3 km.1 km Then, and the3 km. calculated Then, valuesthe calculated were assigned values towere the centerassigned grid to of the the center window. grid The ofPDs, the window.on the landscape The PDs, level on ofthe the landscape entire city level (1 km of scale)the entire and city the city(1 km center scale) (3 kmand scale), the city were center calculated (3 km scale),separately were (Figure calculated9). separately (Figure 9).

(a) (b)

Figure 8. (a) Change in patch density (PD) and (b) change in Shannon diversity index (SHDI) in Figure 8. (a) Change in patch density (PD) and (b) change in Shannon diversity index (SHDI) in Wuhan Wuhanfrom 1989 from to 2016.1989 to 2016.

As shown in Figure9a, in 1995, the areas with a higher patch density were distributed in the concentrated downtown periphery: Southern Huanghua, Northern Xinzhou, and Eastern Hannan. Since the central urban area of Wuhan was covered by a large area of contiguous construction land, the patch density was relatively low (shown as large green), but the fragmentation degree of the city periphery, neighboring farmland, waters, and villages, had increased to a relevant degree. By 2013, the green area in the city center, with a low degree of fragmentation, had increased to a relevant degree, indicating that the land use of the city center had become more compact, and the fragmentation degree of the red area, between urban and rural areas, had been reduced to a relevant degree. By 2016,

ISPRS Int. J. Geo-Inf. 2018, 7, 340 9 of 17 the areas with a high degree of fragmentation still included the Zhangdu Lake periphery, Central Dongxihu District, and Eastern Hannan District. In addition, the degree of fragmentation of the Eastern Hannan District had increased to a relevant degree. That was because the Eastern Hannan District is at the junction of agricultural landscape clusters, urban landscape clusters, water landscape clusters, and forest landscape clusters, so landscape types quickly transformed. As shown in Figure9b, the patch density in the downtown area decreased by 53% from 1995 to 2013, because the smaller architectural patches were connected and merged into larger patches as a result of city development. However, the patch density of the Southern and Northern Hongshan District increased to a relevant degree, indicating a rapid landscape transformation. Due to the rapid development of the Guanggu area, the degree of fragmentation in 2016 became higher when compared toISPRS that Int. in J. 2013 Geo-Inf. (an 2018 overall, 7, x FOR increase PEER REVIEW of 42%). 9 of 17

High:100 High:50 High:100 Low:0 Low:0 Low:0

(a)

(b)

Figure 9. (a) Spatial distribution of patch density (PD) in the whole district and (b) spatial Figure 9. (a) Spatial distribution of patch density (PD) in the whole district and (b) spatial distribution distribution of patch density (PD) in the central city of Wuhan in 1995, 2013, and 2016. of patch density (PD) in the central city of Wuhan in 1995, 2013, and 2016.

As shown in Figure 9a, in 1995, the areas with a higher patch density were distributed in the concentratedThe SHDI downtown can be used periphery: to measure Southern landscape Huanghua, heterogeneity. Northern Based Xinzho on theu, and calculation Eastern resultsHannan. of theSince SHDI, the central this paper urban analyzes area of the Wuhan heterogeneity was covered of the by landscape a large area in Wuhan.of contiguous During construction the study period, land, thethe SHDIpatch ofdensity Wuhan was showed relatively an overalllow (shown upward as larg trend.e green), The distribution but the fragmentation of the landscape degree patterns of the city in Wuhanperiphery, had neighboring generally become farmland, more waters, balanced. and Thevillages, SHDI had calculation increased results, to a relevant at the landscape degree. By level 2013, of thethe wholegreen cityarea (1 in km the scale) city center, and the with city centera low (3degree km scale), of fragmentation, are shown in Figurehad increased 10. to a relevant degree,As shownindicating in Figure that 10thea, theland overall use landscapeof the city of Wuhancenter had became become more andmore more compact, balanced and as the cityfragmentation expandedand degree developed. of the red The area, original between large-scale urban agriculturaland rural areas, landscape had been clusters reduced were to broken a relevant into smalldegree. pieces By 2016, and transformedthe areas with into a otherhigh typesdegree of of landscape fragmentation clusters. still By included 2016, the the SHDI Zhangdu index ofLake the cityperiphery, periphery Central had increased Dongxihu by District, 21% compared and Easter to thatn inHannan 1995, and District. the area In withaddition, a high the index degree became of larger.fragmentation As shown of the in Figure Eastern 10 Hannanb, the SHDI District metrics had inincreased the city to center a relevant showed degree. a declining That was trend because from 1995the Eastern to 2016. Hannan This was District the inevitable is at the result junction of the developmentof agricultural of landscape the city center clusters, [35]. Theurban landscapes landscape of otherclusters, types water continuously landscape transformed clusters, and into urbanforest landscapelandscape clusters. clusters, The so heterogeneity landscape types of the quickly central urbantransformed. area of Wuhan decreased. As shown in Figure 9b, the patch density in the downtown area decreased by 53% from 1995 to 2013, because the smaller architectural patches were connected and merged into larger patches as a result of city development. However, the patch density of the Southern and Northern Hongshan District increased to a relevant degree, indicating a rapid landscape transformation. Due to the rapid development of the Guanggu area, the degree of fragmentation in 2016 became higher when compared to that in 2013 (an overall increase of 42%). The SHDI can be used to measure landscape heterogeneity. Based on the calculation results of the SHDI, this paper analyzes the heterogeneity of the landscape in Wuhan. During the study period, the SHDI of Wuhan showed an overall upward trend. The distribution of the landscape patterns in Wuhan had generally become more balanced. The SHDI calculation results, at the landscape level of the whole city (1 km scale) and the city center (3 km scale), are shown in Figure 10.

ISPRS Int. J. Geo-Inf. 2018, 7, 340 10 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 10 of 17

High:1.5 High:1.5 High:1.5 Low:0 Low:0 Low:0

(a)

(b)

Figure 10. (a) Spatial distribution of the Shannon diversity index (SHDI) in the whole district; (b) Figure 10. (a) Spatial distribution of the Shannon diversity index (SHDI) in the whole district; (b) Spatial Spatial distribution of the Shannon diversity index (SHDI) in the central city of Wuhan in 1995, 2013, distribution of the Shannon diversity index (SHDI) in the central city of Wuhan in 1995, 2013, and 2016. and 2016.

3.2. UrbanAs shown Land Density in Figure Distribution 10a, the overall Pattern landscape of Wuhan became more and more balanced as theWuhan city expanded is divided and into developed. three parts Theby original the Yangtze large-scale River agricultural and Han River.landscape Due clusters to the were special geographicalbroken into locationssmall pieces of and Wuhan, transformed in this into paper, othe ther types geometric of landscape center clusters. of the By three 2016, old the districts,SHDI whereindex commerce of the city has periphery been concentrated, had increased was by 21% selected comp asared the to circle that center,in 1995, and and the the main area with urban a high area of index became larger. As shown in Figure 10b, the SHDI metrics in the city center showed a declining Wuhan was divided into thirty rings, one every kilometer, as shown in Figure 11. In order to eliminate trend from 1995 to 2016. This was the inevitable result of the development of the city center [35]. The the impact of large water bodies, such as the Yangtze River and East Lake, the water portion of each landscapes of other types continuously transformed into urban landscape clusters. The ring was removed before the land use density in each ring was calculated. The results are shown in heterogeneity of the central urban area of Wuhan decreased. FigureISPRS 12 Int.a. J.The Geo-Inf. ﬁtting 2018, curve7, x FOR of PEER urban REVIEW land use density in 1995, 2013, and 2016 is shown in Figure 11 of 1712 b. 3.2. Urban Land Density Distribution Pattern Wuhan is divided into three parts by the Yangtze River and Han River. Due to the special geographical locations of Wuhan, in this paper, the geometric center of the three old districts, where commerce has been concentrated, was selected as the circle center, and the main urban area of Wuhan was divided into thirty rings, one every kilometer, as shown in Figure 11. In order to eliminate the impact of large water bodies, such as the Yangtze River and East Lake, the water portion of each ring was removed before the land use density in each ring was calculated. The results are shown in Figure 12a. The fitting curve of urban land use density in 1995, 2013, and 2016 is shown in Figure 12b.

Figure 11. Concentric ring division. Figure 11. Concentric ring division.

Urban landdensity Inverse S-shaped fitting S-shaped Inverse

Distance (km) Distance (km) (a) (b)

Figure 12. (a) Urban land density calculations and (b) inverse S-shaped fitting.

The density of urban land was high in the compact city, but low in the area around the city center. Therefore, the density curve was steep for the compact city, but gentle for the expanded city. The slope of the function curve of urban land use density could be used to reflect the total density of urban land use. The slope of the sharply descending part of the curve was used to reflect the compactness of the city. As shown in Figure 13, compactness was defined as the ratio between the difference value of the radius that corresponded to the minimum and maximum value of the second

derivative of the fitting curves, and the maximum radius. This difference value was k=(r −r)/D [18].

(a) (b)

Figure 13. (a) Urban land density original curve and (b) the second derivative curve.

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 11 of 17 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 11 of 17

ISPRS Int. J. Geo-Inf. 2018, 7, 340 Figure 11. Concentric ring division. 11 of 17 Figure 11. Concentric ring division. Urban landdensity Inverse S-shaped fitting S-shaped Inverse Urban landdensity Inverse S-shaped fitting S-shaped Inverse

Distance (km) Distance (km) Distance (km) Distance (km) (a) (b) (a) (b) Figure 12. (a) Urban land density calculations and (b) inverse S-shaped fitting. FigureFigure 12. 12. (a(a) )Urban Urban land land density density calculations calculations and and ( (bb)) inverse inverse S-shaped S-shaped fitting. fitting. The density of urban land was high in the compact city, but low in the area around the city The density of urban land was high in the compact city, but low in the area around the city center.The Therefore, density of the urban density land curv wase high was insteep the compactfor the compact city, but city, low inbut the gentle area for around the expanded the city center. city. center. Therefore, the density curve was steep for the compact city, but gentle for the expanded city. Therefore,The slope theof the density function curve curve was of steep urban for land the compact use density city, could but gentle be used for theto reflect expanded the total city. density The slope of The slope of the function curve of urban land use density could be used to reflect the total density of ofurban the functionland use. curve The of slope urban of land the usesharply density descending could be used part to of reflect the curve the total was density used ofto urbanreflect land the urban land use. The slope of the sharply descending part of the curve was used to reflect the use.compactness The slope of of the the city. sharply As shown descending in Figure part of13, the compactness curve was usedwas defined to reflect as the the compactness ratio between of thethe compactness of the city. As shown in Figure 13, compactness was defined as the ratio between the city.difference As shown value in of Figure the radius 13, compactness that corresponded was defined to the as minimum the ratio betweenand maximu the differencem value of value the second of the difference value of the radius that corresponded to the minimum and maximum value of the second radiusderivative that of corresponded the fitting curves, to the minimumand the maximum and maximum radius. value This ofdifference the second value derivative was k=(r of the −r fitting)/D derivative of the fitting curves, and the maximum radius. This difference value was k=(r −r)/D curves,[18]. and the maximum radius. This difference value was k = (r − r )/D[18]. [18]. 2 1

(a) (b) (a) (b) Figure 13. (a) Urban land density original curve and (b) the second derivative curve. FigureFigure 13. 13. (a(a) )Urban Urban land land density density original original curve curve and and ( (bb)) the the second second derivative derivative curve. curve.

Previous research points out that a compact city generally has a narrow area that includes the inner city and suburbs. The difference between r1 is r2 was small, so the k value is small [36]. On the other hand, a city with a large k value is a low-density or decentralized city [37]. The ﬁtting parameter values and calculated k values of Wuhan from 1989 to 2016 are shown in Table4. It can be seen that the value of k gradually decreases. By 2016, the value of k was reduced by 41.5%, and the development of the central urban area became more and more compact.

Table 4. Calculation results.

Year α c D k 1989 2.479 0.036 15.554 0.53 1995 2.493 0.041 18.528 0.48 2001 2.486 0.033 19.733 0.50 2004 2.481 0.029 22.672 0.44 2007 2.496 0.022 25.468 0.41 2010 2.507 0.021 26.732 0.39 2013 2.510 0.019 28.354 0.36 2016 2.894 0.006 30.366 0.31 ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 12 of 17

Previous research points out that a compact city generally has a narrow area that includes the inner city and suburbs. The difference between r1 is r2 was small, so the k value is small [36]. On the other hand, a city with a large k value is a low-density or decentralized city [37]. The fitting parameter values and calculated k values of Wuhan from 1989 to 2016 are shown in Table 4. It can be seen that the value of k gradually decreases. By 2016, the value of k was reduced by 41.5%, and the development of the central urban area became more and more compact.

Table 4. Calculation results.

Year α c D k 1989 2.479 0.036 15.554 0.53 1995 2.493 0.041 18.528 0.48 2001 2.486 0.033 19.733 0.50 2004 2.481 0.029 22.672 0.44 2007 2.496 0.022 25.468 0.41 2010 2.507 0.021 26.732 0.39 ISPRS Int. J. Geo-Inf. 2018, 7, 340 2013 2.510 0.019 28.354 0.36 12 of 17 2016 2.894 0.006 30.366 0.31

At thethe samesame time, time, for for the the ﬁtted fitted density density function function of urban of urban land, theland,D parameter the D parameter was an estimated was an estimatedvalue of the value radius of the ofthe radius main of urban the main area urban [18]. Thearea increase [18]. The of increaseD indicated of D urbanindicated growth. urban As growth. shown Asin Tableshown4, fromin Table 1989 4, to from 2016, 1989 the Dto value2016, forthe Wuhan D value increased for Wuhan from increased 15.554 to from 30.336, 15.554 indicating to 30.336, that indicatingthe distance that from the thedistance urban from boundary the urban to the bounda urbanry center to the increasedurban center from increased 15.554 km from to 30.33615.554 km.km toThe 30.336 variation km. The of the variationD value of of the Wuhan D value over of Wuhan the past over 20 years the past is shown 20 years in Figureis shown 14 a.in ItFigure can be 14a. seen It canthat be the seen urban that area the rapidlyurban area expanded, rapidly especially expanded, from especially 2000 to from 2016. 2000 In Figure to 2016. 14 b,In theFigure red 14b, represented the red representedthe construction the construction land in 1989 andland gray in 1989 represented and gray the represented construction the land construction in 2016. The land circle in represents2016. The circlethe expansion represents of the urbanexpansion boundary of the fromurban 1989 boundary (inner layer) from to1989 2016 (inner (outer layer) layer). to Clearly,2016 (outer during layer). the Clearly,study period, during the the urban study area period, of Wuhan the urba expandedn area of signiﬁcantly. Wuhan expanded significantly.

(a) (b)

Figure 14. (a) Results of the calculation of D values in Wuhan from 1989 to 2016, and (b) urban Figure 14. (a) Results of the calculation of D values in Wuhan from 1989 to 2016, and (b) urban boundary expansion. boundary expansion.

The results showed that, on the one hand, the commercial and political centers of Wuhan were still inThe the results three showed old urban that, areas—Wuchang, on the one hand, theHankou, commercial and Hanyang—during and political centers the of study Wuhan period, were whichstill in promoted the three the old development urban areas—Wuchang, of the entire Hankou, urban area. and As Hanyang—during a result, the development the study pattern period, tendedwhich promotedcompact inwardly. the development On the other of the hand, entire due urban to the area. rapid As economic a result, the development development of Wuhan pattern andtended the compactcontinuous inwardly. adjustment On the of otherthe industrial hand, due st toructure, the rapid the economicconstruction development land of old of urban Wuhan areas and couldthe continuous not meet adjustmentthe requirements of the of industrial industrial structure, and commercial the construction construction. land of The old central urban areasurban could area rapidlynot meet expanded the requirements to nearby of suburbs. industrial and commercial construction. The central urban area rapidly expanded to nearby suburbs.

3.3. Spatial and Temporal Evolution of the Development Direction Gradient In this study, samples were collected from the center of Wuhan to the East, West, South, and North, with an interval of 300 m, in order to explore the development direction of the central urban areas of Wuhan. There were a total of 142 sampling points in the eastern and western directions, and a total of 92 sampling points in the South and North directions. Samples were collected in the spatial distribution maps of 1995, 2013, and 2016, in order to acquire the distribution map of the PD and SHDI values, with the variation in the distance from the urban center (Figures 15 and 16). It could be concluded from Figure 15 that in 1995, the maximum PD value appeared in the West, about 19 km away from the urban center. The PD rapidly decreased, with the distance changing from 19 km to 6 km in the West (a decrease of 70.3%). This was the transition from the urban–rural border to the urban center. From 2013 to 2016, the PD value, ranging between 19 km and ISPRS14 km, Int. decreased J. Geo-Inf. 2018 significantly, 7, 340 (a reduction of 48% on average). This indicated that the urban–rural13 of 17 boundary expanded outward, and that the PD value of the original urban–rural boundary decreased. From the urban center to the East, the PD values for the three years had a similar fragmentationvariation trend. peaks There at were 10 km two and fragmentation 12 km away frompeaks the at urban10 km center, and 12 at km the junctionaway from of athe lake urban and constructioncenter, at the land, junction indicating of a lake the destructiveand construction effect of land, urban indicating expansion the on destructive the water landscape effect of cluster.urban Overall,expansion the on difference the water in landscape the West wascluster. greater Overall, than the that difference in the East, in indicatingthe West was that greater Wuhan than expanded that in morethe East, quickly indicating westward that thanWuhan it did expanded eastwards. more quickly westward than it did eastwards. It could be concluded from Figure 1616 thatthat thethe SHDISHDI values of 1995, 2013, and 2016 had the same variation trend in the East, West, South,South, andand North.North. The SHDI metrics represented a declining trend at the junction of of urban urban and and rural rural areas, areas, 16–19 16–19 km km west west of ofthe the urban urban center. center. The The small small patches patches on the on theurban urban margin margin continuously continuously merged merged into into large large clusters. clusters. The Theboundary boundary of Sand of Sand Lake Lake was was 1 km 1 kmto the to theeast east of the of theurban urban center, center, where where the theSHDI SHDI value value wa wass reduced reduced by by61%. 61%. This This was was because because the the area area of ofSand Sand Lake Lake was was continuously continuously shrinking shrinking and and transf transformingorming into into urban urban construction land. Fengshan forest reserve waswas easteast of the urban center. Due to proper protection measures, thethe SHDI was almost unchanged. In In the the north–south north–south direction, direction, the the SHDI SHDI in in 2013 2013 was was reduced reduced more more significantly signiﬁcantly than than it itwas was in in1995 1995 (a reduction (a reduction of 31.3%). of 31.3%). but was but wasslightly slightly lower lower than thanit was it in was 2016 in (a 2016 reduction (a reduction of 8.6%). of 8.6%).Agriculture, Agriculture, water, water,and forest and forestlandscape landscape clusters clusters transformed transformed into constr into constructionuction land landdue dueto the to theindustrial industrial and and commercial commercial development development in Qingshan in Qingshan District District and andHongshan Hongshan District. District. Therefore, Therefore, the theuniformity uniformity of the of thelandscape landscape was was reduced. reduced.

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 14 of 17 (a)

(b)

Figure 15. (a) The gradient change in patch density (PD) in the east–west directions; (b) the gradient Figure 15. (a) The gradient change in patch density (PD) in the east–west directions; (b) the gradient change in PD in the north–south directions. change in PD in the north–south directions.

(a)

(b)

Figure 16. (a) The gradient change in the Shannon diversity index (SHDI) in the east–west directions; (b) the gradient change in the Shannon diversity index (SHDI) in the north–south directions.

4. Discussion and Conclusions This paper proposes a landscape classification method based on texture, and a multi-angle landscape analysis method based on landscape metrics, urban land use density, etc. The challenges faced during urbanization, such as an increased degree of urban fragmentation and a decrease of connectivity, were effectively analyzed. Moreover, our study results can provide a scientific and effective technique, as well as data support for the construction of ecological cities and the

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 14 of 17

(b)

ISPRSFigure Int. J. Geo-Inf. 15. (a)2018 The, 7gradient, 340 change in patch density (PD) in the east–west directions; (b) the gradient14 of 17 change in PD in the north–south directions.

(a)

(b)

Figure 16. (a) The gradient change in the Shannon diversity index (SHDI) in the east–west directions; Figure 16. (a) The gradient change in the Shannon diversity index (SHDI) in the east–west directions; (b) the gradient change in the Shannon diversity index (SHDI) in the north–south directions. (b) the gradient change in the Shannon diversity index (SHDI) in the north–south directions.

4. Discussion and Conclusions 4. Discussion and Conclusions This paper proposes a landscape classification method based on texture, and a multi-angle This paper proposes a landscape classification method based on texture, and a multi-angle landscape analysis method based on landscape metrics, urban land use density, etc. The challenges landscape analysis method based on landscape metrics, urban land use density, etc. The challenges faced during urbanization, such as an increased degree of urban fragmentation and a decrease of faced during urbanization, such as an increased degree of urban fragmentation and a decrease of connectivity, were effectively analyzed. Moreover, our study results can provide a scientific and connectivity, were effectively analyzed. Moreover, our study results can provide a scientific and effective technique, as well as data support for the construction of ecological cities and the effective technique, as well as data support for the construction of ecological cities and the optimization of urban spatial layouts [38]. In this paper, a texture-based BP neural network classification model has been constructed by extracting texture features, which improved the accuracy of image classification. This paper also introduced similar function classification methods from landscape ecology, and the landscape of Wuhan was classified into multiple clusters, including forest landscape, water landscape, urban landscape, and agricultural landscape clusters. Finally, we analyzed three aspects of the landscape pattern: the temporal and spatial features of landscape metrics, the urban land density distribution pattern, and the spatial and temporal evolution of the development direction gradient. Moreover, two main conclusions can be extracted from this paper. Firstly, the heterogeneity of different landscape types in the entire area in Wuhan increased. However, the landscape types were unevenly distributed, and the heterogeneity of landscape patterns decreased as a result of the rapid development of the central urban area. The degree of fragmentation of the urban and rural boundary was higher than that of the central urban area. The landscape metrics of the western part was more intense than that of the eastern part. The development in the West was better than that in the East. Secondly, the compactness of Wuhan was quantitatively calculated by fitting the inverse S-curve based ISPRS Int. J. Geo-Inf. 2018, 7, 340 15 of 17 on the urban land use density. During the study period, the k value was reduced from 0.53 to 0.31, indicating that the development tended to compact inward. At the same time, the expansion status of the urban areas of Wuhan was quantitatively analyzed by calculating the urban radius D. Wuhan expanded outwards. Since many landscape metrics coincide in measuring landscape features, we have only selected two metrics that can reflect the changes in landscape characteristics. In future studies, more landscape metrics, such as the effective mesh size, Euclidean nearest neighbor, or interspersion index, should be used to identify changes in urban spatial structures more accurately. According to the results of analyses of the urban landscape pattern of Wuhan, the study methods and procedures in this paper had good applicability. In future planning and development of Wuhan, the protection of cultivated land and forest land should be strengthened to prevent the continuous reduction of the forest and farmland landscape. At the same time, durin the process of urban expansion, we should pay attention to the matching and connection of various landscapes. The contradictions between man and land that emerge during the rapid expansion of urban areas should be reduced, in order to ensure the rationality and ordering of the landscape structure. In this way, Wuhan can effectively exert the radiation effect of the central urban area, and promote the balanced development and harmonious progress of the Wuhan Urban Circle.

Author Contributions: Y.Y. and J.L. proposed the idea, J.L. wrote the paper, and T.M. and Z.D. built the model and performed the experiments; Z.Y analyzed the data. Funding: This research was funded by the National Key R&D Program of China (Grant No. 2017YFB0503804), the National Natural Science Foundation of China (Grant No. 41801306) and the Natural Science Fund of Hubei Province (Grant No. 2017CFA041). Acknowledgments: We really appreciate the two anonymous reviewers for their useful comments and suggestions. This study was supported by the National Key R&D Program of China (Grant No. 2017YFB0503804), the National Natural Science Foundation of China (Grant No. 41801306) and the Natural Science Fund of Hubei Province (Grant No. 2017CFA041). Conﬂicts of Interest: The authors declare no conﬂict of interest.

References

1. Roders, A.P. How can urbanization be sustainable? A reflection on the role of city resources in global sustainable development. BDC–Bollettino del Centro Calza Bini 2014, 13, 79–90. 2. Bhatta, B.; Saraswati, S.; Bandyopadhyay, D. Urban sprawl measurement from remote sensing data. Appl. Geogr. 2010, 30, 731–740. [CrossRef] 3. Xiao, D. On the Formation and Development of Modern Landscape Science. Sci. Geogr. Sin. 1999, 19, 379–384. 4. Wu, J. Landscape Ecology—Patterns, Processes, Scales, and Ranks, 2nd ed.; Higher Education Press: Bejing, China, 2007. 5. Xiao, D.; Zhong, L.S. Ecological principles of landscape classification and assessment. Chin. J. Appl. Ecol. 1998, 9, 217–221. 6. Brandt, J. Landscape ecological information through statistical analysis of the territorial structure of a sheep grazing system, Faroe Islands. In Proceedings of the First International Seminar on Methodology in Landscape Ecological Research and Planning of the International Association for Landscape Ecology (IALE), Roskilde, Denmark, 15–19 October 1984. 7. Pan, D.; Domon, G.; Blois, S.D.; Bouchard, A. Temporal (1958–1993) and spatial patterns of land use changes in Haut-Saint-Laurent (Quebec, Canada) and their relation to landscape physical attributes. Landsc. Ecol. 1999, 14, 35–52. [CrossRef] 8. Veldkamp, A.; Verburg, P.H. Modelling land use change and environmental impact. J. Environ. Manag. 2004, 72, 1. [CrossRef][PubMed] 9. Verburg, P.H.; Soepboer, W.; Veldkamp, A.; Limpiada, R.; Espaldon, V.; Mastura, S.S. Modeling the spatial dynamics of regional land use: The CLUE-S model. Environ. Manag. 2002, 30, 391. [CrossRef][PubMed] 10. Kitada, K.; Fukuyama, K. Land-Use and Land-Cover Mapping Using a Gradable Classification Method. Remote Sens. 2012, 4, 1544–1558. [CrossRef] ISPRS Int. J. Geo-Inf. 2018, 7, 340 16 of 17

11. Lillesand, T.M. Remote Sensing and Image Interpretation; Wiley: Hoboken, NJ, USA, 2006. 12. Ke, H.M.; Chen, C.Z.; Zhang, X.H.; Wang, J.L. Study on BP neural network classification with optimization of genetic algorithm for remote sensing imagery. J. Southwest Univ. 2010, 38, 157–166. 13. Mcgarigal, K.; Marks, B.J. FRAGSTATS: Spatial analysis program for quantifying landscape structure. Gen. Tech. Rep. PNW 1995, 122, 351. 14. Bai, X.; Shi, P.; Liu, Y. Society: Realizing China’s urban dream. Nature 2014, 509, 158–160. [CrossRef] [PubMed] 15. Wang, L.; Tian, B.; Koike, K.; Hong, B.; Ren, P. Integration of Landscape Metrics and Variograms to Characterize and Quantify the Spatial Heterogeneity Change of Vegetation Induced by the 2008 Wenchuan Earthquake. ISPRS Int. Geo-Inf. 2017, 6, 164. [CrossRef] 16. Dahal, K.R.; Benner, S.; Lindquist, E. Urban hypotheses and spatiotemporal characterization of urban growth in the Treasure Valley of Idaho, USA. Appl. Geogr. 2017, 79, 11–25. [CrossRef] 17. Lechner, A.M.; Reinke, K.J.; Wang, Y.; Bastin, L. Interactions between landcover pattern and geospatial processing methods: Effects on landscape metrics and classification accuracy. Ecol. Complex. 2013, 15, 71–82. [CrossRef] 18. Jiao, L. Urban land density function: A new method to characterize urban expansion. Landsc. Urban Plan. 2015, 139, 26–39. [CrossRef] 19. Wang, H.; Guo, J.; Ma, Z. Monitoring Wheat Stripe Rust Using Remote Sensing Technologies in China; Springer: Berlin, Germany, 2012. 20. Defries, R.S.; Asner, G.P.; Houghton, R.A. Landscape Level Analysis of the Spatial and Temporal Complexity of Land-Use Change; American Geophysical Union: Washington, DC, USA, 2013; pp. 217–230. 21. He, J.; Li, C.; Yu, Y.; Liu, Y.; Huang, J. Measuring urban spatial interaction in Wuhan Urban Agglomeration, Central China: A spatially explicit approach. Sustain. Cities Soc. 2017, 32, 569–583. [CrossRef] 22. Chen, Z.; Zhang, A.; Shan, X. Urbanization and administrative restructuring: A case study on the Wuhan urban agglomeration. Habitat Int. 2016, 55, 46–57. 23. Coulter, L.L.; Stow, D.A.; Tsai, Y.H.; Ibanez, N.; Shih, H.C.; Kerr, A.; Benza, M.; Weeks, J.R.; Mensah, F. Classification and assessment of land cover and land use change in southern Ghana using dense stacks of Landsat 7 ETM+ imagery. Remote Sens. Environ. 2016, 184, 396–409. [CrossRef] 24. Li, X.Q.; Zhou, L.; Sun, Z.; Song, Y. Analysis on changes of landscape pattern of Shenyang City assisted by TM images. Ecol. Sci. 2016, 35, 79–84. 25. Fan, Q.; Ding, S. Landscape pattern changes at a county scale: A case study in Fengqiu, Henan Province, China from 1990 to 2013. Catena 2016, 137, 152–160. [CrossRef] 26. Haralick, R.M. Statistical and structural approaches to texture. Proc. IEEE 2005, 67, 786–804. [CrossRef] 27. Yeh, C.K.; Liaw, S.C. Application of landscape metrics and a Markov chain model to assess land cover changes within a forested watershed, Taiwan. Hydrol. Process. 2016, 29, 5031–5043. [CrossRef] 28. Riitters, K.H.; O’Neill, R.V.; Hunsaker, C.T.; Wickham, J.D.; Yankee, D.H.; Timmins, S.P.; Jones, K.B.; Jackson, B.L. A factor analysis of landscape pattern and structure metrics. Landsc. Ecol. 1995, 10, 23–39. [CrossRef] 29. Chang, S.; Jiang, Q.; Wang, Z.; Xu, S.; Jia, M. Extraction and Spatial–Temporal Evolution of Urban Fringes: A Case Study of Changchun in Jilin Province, China. ISPRS Int. Geo-Inf. 2018, 7, 241. [CrossRef] 30. Barbaro, L.; Rusch, A.; Muiruri, E.W.; Gravellier, B.; Thiery, D.; Castagneyrol, B. Avian pest control in vineyards is driven by interactions between bird functional diversity and landscape heterogeneity. J. Appl. Ecol. 2016, 54, 500–508. [CrossRef] 31. Stephens, P.A.; Pretty, J.N.; Sutherland, W.J. Agriculture, transport policy and landscape heterogeneity. Trends Ecol. Evol. 2003, 18, 555–556. [CrossRef] 32. Dong, Y.; Liu, S.; An, N. Study on Dynamic Changes of Landscape Pattern in Daan City of Jilin Province Based on Landscape Index and Spatial Autocorrelation. J. Nat. Resour. 2015, 30, 1860–1871. 33. Hagen-Zanker, A. A computational framework for generalized moving windows and its application to landscape pattern analysis. Int. J. Appl. Earth Obs. Geoinf. 2016, 44, 205–216. [CrossRef] 34. Gao, B.; Huang, Q.; He, C.; Sun, Z.; Zhang, D. How does sprawl differ across cities in China? A multi-scale investigation using nighttime light and census data. Landsc. Urban Plan. 2016, 148, 89–98. [CrossRef] 35. Rahman, M. Detection of Land Use/Land Cover Changes and Urban Sprawl in Al-Khobar, Saudi Arabia: An Analysis of Multi-Temporal Remote Sensing Data. ISPRS Int. Geo-Inf. 2016, 5, 15. [CrossRef] ISPRS Int. J. Geo-Inf. 2018, 7, 340 17 of 17

36. Abdullahi, S.; Pradhan, B. Land use change modeling and the effect of compact city paradigms: Integration of GIS-based cellular automata and weights-of-evidence techniques. Environ. Earth Sci. 2018, 77, 251. [CrossRef] 37. Yokoyama, M.; Matsuo, K.; Tanaka, T.; Sadohara, S. Designing and evaluating land use scenario with effective sea breeze use: Study on land use scenarios of compact city with mitigating urban warming effect in Kanagawa prefecture Part 2. J. Environ. Eng. 2018, 83, 301–311. [CrossRef] 38. Peng, J.; Xie, P.; Liu, Y.; Ma, J. Urban thermal environment dynamics and associated landscape pattern factors: A case study in the Beijing metropolitan region. Remote Sens. Environ. 2016, 173, 145–155. [CrossRef]

© 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/).