Environmental Management

, Volume 36, Issue 4, pp 576–591

Modeling Change-Pattern-Value Dynamics on Land Use: An Integrated GIS and Artificial Neural Networks Approach

Authors

    • Institute of Geographical Sciences and Natural Resource Research (IGSNRR)Chinese Academy of Sciences
  • Shaohong Wu
    • Institute of Geographical Sciences and Natural Resource Research (IGSNRR)Chinese Academy of Sciences
  • Wenzhong Shi
    • Advanced Research Centre for Spatial Information TechnologyDepartment of Land Surveying and Geo-InfomaticsThe Hong Kong Polytechnic University
  • Chui-kwan Cheung
    • Advanced Research Centre for Spatial Information TechnologyDepartment of Land Surveying and Geo-InfomaticsThe Hong Kong Polytechnic University
  • Ahmed Shaker
    • Advanced Research Centre for Spatial Information TechnologyDepartment of Land Surveying and Geo-InfomaticsThe Hong Kong Polytechnic University
ENVIRONMENTAL ASSESSMENT

DOI: 10.1007/s00267-004-0165-z

Cite this article as:
Dai, E., Wu, S., Shi, W. et al. Environmental Management (2005) 36: 576. doi:10.1007/s00267-004-0165-z

Abstract

The use of spatial methods to detect and characterize changes in land use has been attracting increasing attention from researchers. The objectives of this article were to formulate the dynamics of land use on the temporal and spatial dimensions from the perspectives of the Change-Pattern-Value (CPV) and driving mechanism, based on multitemporal remote sensing data and socioeconomic data. The Artificial Neural Networks were used to identify the factors driving changes in land use. The Pearl River Delta Region of southeast China, which was experiencing rapid economic growth and widespread land conversion, has been selected as the study region. The results show that from 1985 to 2000 in the study region (1) the most prominent characteristics of change in land use were the expansion of the urban land at the expense of farmland, forests, and grasslands, (2) the land-use pattern was being optimized during this period, (3) in an analysis of value, built-up land can yield a return of more than 30 times that of farmland, water area, and forests lands, and (4) rapid economic development, growth in population, and the development of an infrastructure were major driving factors behind ecological land loss and the nonecological land expansion.

Keywords

Change in land useChange-Pattern-Value (CPV) analysisDriving factors for land-use changeRemote sensingGISArtificial Neural NetworksPearl River Delta Region

The studies of land-use change stand at the research frontier in global change (Turner 1990, Meyer and Turner 1994). The fundamental objectives of studying changes in land use are to investigate the social, economic, and spatial causes of changes (Batty and Longley 1994, de Koning and others 1999) and the process and trajectory of changes (Theobald and others 1997, Pijanowski and others 2002), so as to be able to make proposals on the suitable use of land and patterns of development. The Land-Use/Cover Change (LUCC) Science/Research Plan, a core project of the joint International Geosphere–Biosphere Programme (IGBP) and International Human Dimensions Programme (IHDP) proposed in 1991 and approved by the IGBP and IHDP in 1993, began by aiming to identify the driving factors behind LUCC (Turner and others 1995). A change in land use is the result of complex interactions among many factors, not only physical but also socioeconomic and environmental (Vesterby and Heimlich 1991, Dale and others 1993, Houghton 1994, Medley and others 1995, Pijanowski and others 2002).

Land-use change models are regarded as approaches to improve not only the ability to explore change dynamics in land use and to project future regimes on land use and patterns of development (Turner and others 1994, Bockstael and others 1995), but also helpful in the analyses of the factors driving land-use change and in the selection of appropriate development strategies. Realistic land-use change models need to integrate different spatial scales and their specific drivers to simulate changes in land use in response to biophysical and economic/human drivers (Turner and others 1993, Veldkamp and Fresco 1996). In the LUCC study, many models, such as the CENTURY model (Parton and others 1987), the AGE model (Fischer and others 1988), the Ehrlich model (Ehrlich and Ehrlich 1990), the FASOM model (Grainger 1990), the Adams model (Adams and others 1995), and the Riebsame model (Riebsame and others 1994), have been reported. In these models, the emphasis was on changes in land cover. Directly related to land-use change and the driving factors, de Koning and others (1999) addressed the CLUE model, which stressed the multiscale modeling of land-use dynamics, applied to regional- and global-scale drivers to determine the aggregate amount of change,and to geographic- and landscape-scale drivers to determine the pattern of change. The LTM (Land Transformation Model) has also been addressed by Pijanowski and others (2000, 2002). In the model, the Geographic Information System (GIS) was used to analyze the change process in land use, and a suite of complex factors, including policy, population change, culture, economy, and environmental factors, were selected as driving factors.

This article is an effort to model the dynamics of land-use change on the spatial and temporal dimensions from the perspectives of Change-Pattern-Value (CPV) and their driving factors. The major purpose of this study is to determine how to construct a series of indexes that reflect CPV and how to identify driving factors behind changes in land use. Multitemporal remote sensing images and socioeconomic data were the major sources of data. GIS was applied to demonstrate the process of CPV change. With the aid of artificial neural networks (ANNs), driving factors behind land-use change were also identified. In the following section, sketching and quantifying methods will be addressed to formulate the CPV model on changes in land use. This model will then be applied to the Pearl River Delta Region in southeast China.

CPV Analysis Model on Land Use

Being a critical dynamic element of ecosystems (Bockstael and others 1995, Pijanowski and others 2000), changes in land use result either from physical conditions or human behavior. Such changes often lead to changes in the pattern and value of the land-use system from the systematic perspective and in hydrology, vegetation, species diversity, the economy, and the social context. A land-use system is actually a comprehensive system composed of subsystems of nature, economy, and society. However complex a land-use system, its characteristics can be shown by indexes of change, pattern, and value. Following Turner and others (1995), a conceptual framework of a CPV function analysis model can be demonstrated by Figure 1.
https://static-content.springer.com/image/art%3A10.1007%2Fs00267-004-0165-z/MediaObjects/267_2004_165_f1.gif
Figure 1

Conceptual framework of the change, pattern and value analysis model in land use.

Data Processing

To analyze the dynamics of land use and their driving mechanism, data on changes in land use and related physical, social, and economic dimensions are required. Remote sensing is an important way of obtaining series data on the spatial and temporal dimensions of change (Saraf and others 2001), which can be used as a basis for describing past land-use processes and forecasting future land-use scenarios. At present, quantitative approaches can be used on global geo-sphere characteristics and their process of evolution, with the aid of integrated series data both on the temporal and spatial dimensions. In our study, a topographic map was used as the base map. A cube polynomial equation and the nearest-neighborhood interpolation were applied for geometric corrections of the remotely sensed images. Positional error adjustments were within one pixel. A scatter controlled regression (Christopher and Ding 1995) was used to normalize images from different periods in this study. The Thematic Mapper Images of 1985, 1995, and 2000 in the study region were used. Land-use types were classified into six types, namely farmland, forests, grasslands, water area, built-up land, and barren land. Assessments of accuracy were performed on the classification results, with simultaneous reference to ground data. A total of 106 sites were selected to calculate the confusion matrix and kappa coefficients. The results showed that the kappa coefficients were 0.71, 0.83, and 0.84 in 1985, 1995, and 2000, respectively, all of which exceed the requirement of 0.7 (Lucas and others 1994). After the supervised classification was completed, all of the data were transferred to a coverage format in Arc/Info software.

Change Analysis Module

To measure the process and degree of change in land use, series of diagnostic indexes are proposed (see Table 1). These indexes reflect characteristics change on land use from the aspects of spatial degree, temporal dimension, and degree of utilization.
Table 1

Series of diagnostic indexes on change analysis of land use

Diagnostic index

Algorithm

Land-use-change matrix \( (C_{i \times j}) \)

\( C_{i \times j}=A_{i \times j}^{k + 1} - A_{i \times j}^k \)

 

\( A_{i \times j}^k \)and \( A_{i \times j}^{k + 1} \): land-use map in time k andk + 1i and j: land use types at time k and k + 1

The Degree of change for a single type of land use (DDS)

\( DDS={{U_{i \times j}^{k + 1} - U_{i \times j}^k } \over {U_{i \times j}^k }} \times {1 \over T} \times 100 \)

 

\( U_{i \times j}^k \)and \( U_{i \times j}^{k + 1} \): amount of land use type U in timek and k + 1T: time period of change

The Total degree of change for various types of land use (DDC)

\( DDC={{\sum\limits_{i=1}^n {\Delta LU_{i - j} } } \over {\sum\limits_{i=1}^n {\Delta LU_i } }} \times {1 \over T} \times 100 \)

 

LUi: area of land use type i at the beginning ofthe period of detection

 

ΔLUi-j: area of change from land-use type ito land-use type j at the period of detection

T: period of detection

 

Degree of development in land use (LA)

\( LA=100 \times \sum\limits_{i=1}^n {A_i \times {AP}_{i} } ,{ LA} \in {\rm{[100 - 400]}} \)

 

Ai: land use index of type i

 

Api: percentage of area of land-use type i

 

A land use index is designed in the following way: barren landis 1, forests, grassland, and water area are 2, farmland is 3, andbuilt-up land is 4 (Zhuang and Liu 1997)

Pattern Analysis Module

To measure pattern variations in land-use change, diagnostic values in landscape ecology (Slater and Brown 2000, Griffiths and Mather 2000) were adopted and revised to formulate a pattern analysis module for an analysis of land-use change with the aid of the GIS and traditional statistical functions:
  • Area (A) and perimeter (P). The area and perimeter for each type of land use can be obtained through GIS software.

  • Fractal Dimension. The fractal dimension can be used to reflect the complexity of land-use patches, which can be quantified as:
    $$ D=2 log(P/4)/ logA. $$
    (1)
    It can be seen that D, which ranges from 1 to 2, is used to reflect the effect the shape of land-use patches to land-use pattern. If D equals 1, the land-use shape is square, whereas if D equals 2, it is the most complex land-use patch.
  • Dominance. Dominance can be used to measure the importance of any land-use patches to that of the whole landscape of change. A quantifying approach can be presented as
    $$ D_0=L/2 + R/2,{\rm{ }}L=N_i /N,{\rm{ }}R=A_i /A. $$
    (2)
    where L represents the number percentage of the land-use patch l (Nl) to that of landscape of change (N) and R represents the area percentage of the land-use patch l (Al) to that of landscape of change (A).
  • Diversity. A diversity index, which actually represents entropy measure, is used to reflect the complexity and variance of land-use patches to landscape of change. It can be quantified as
    $$ H=- \sum\limits_{i=1}^n {R_l \times log_2 R_l }, $$
    (3)
    where R has the same meaning as in Equation 2.
  • Fragmentation. A fragmentation index is used to reflect the fragmentation degree of the landscape of change. It is obvious that if there were more patches in the landscape system, the landscape structure would be more fragmented. Fragmentation can be measured as
    $$ F=N/A, $$
    (4)
    where N represents the total number of the patches in the landscape system and A represents the area of the landscape of change.

Value Analysis Module

It is essential to make a profit for any land-use system. The international framework for evaluating sustainable land management (FESLM) (FAO 1993) addressed the five pillars for assessing land-use sustainability. Four of them, namely productivity, stability, economic viability, and social acceptability, are related to economic output of land use. Economic output can be used to reflect the value of the land-use system.

According to the equilibrium principle of microeconomics, under the condition of full competition and economic and technological stability, marginal benefit (MB) will decrease with the development of the land-use system, whereas marginal cost (MC) will increase (Figure 2). The area under curve MB is the total benefit of the land-use system and that under curve MC is the total cost. E is the point where maximum profits can be made from a land use.
https://static-content.springer.com/image/art%3A10.1007%2Fs00267-004-0165-z/MediaObjects/267_2004_165_f2.gif
Figure 2

Economic dynamics of land-use system.

Transformations in the use of land are inevitable to maximize profits and to conserve limited resources, especially when there is a mismatch between suitability and target on land use. Such transformations include a single land-use type and multiple land-use types. The latter emphasize regional optimization of the land-use pattern. An important issue in the transformation of land-use types is to find an appropriate spatial criterion for the transformation of land use. In a purely market-oriented economy, a criterion for the transformation of land-use type (LTC) can be expressed as (Dai 2002)
$$ LTC=EB_i /EB_j, $$
(5)
where EBi represents the economic benefit of land-use type i and EBj represents the economic benefit of land-use type j. If LTC >1, type i will be transformed to type j on land use to make more profits.

Using ANNs to Identify Driving Factors Behind Changes in Land Use

Artificial neural networks are powerful tools that use the machine learning approaches to quantify and model complex behavior and patterns (Pijanowski and others 2002), especially on data pattern recognition. ANNs have the merit of being able to observe relationships in data by imitating the brain’s ability to sort patterns through the interconnected systems of many neurons. ANNs are composed of many parallel computing and simplified function units, which simulate biological neural networks. Because of the development of computing performance and powerful software, ANNs are used widely in many disciplines. ANNs are actually a nonlinear system with characteristics of distributed storage and parallel computing. Although one nerve cell has a limited structure and function, neural networks composed of many nerve cells can engage in complicated simulations.

The simplest ANNs, developed by Rosenblatt (1958), receive weighted inputs and thresholds according to a defined rule to classify linearly separable data through linear functions. Updated network types, like the Multi-layer Perceptron (MLP), Radial Basis Function (RBF), Kohonen Self-Organizing Map (Kohonen), and so on, are widely used on nonlinear phenomena. There will be input, hidden, and output layers in these networks in general. Through feed-forward algorithms, networks calculate weights for input values, input layer nodes, hidden layer nodes, and output layer nodes, which propagate through the hidden layer and output layers. Neural networks are applied in virtually every situation where relationships exist between independents (inputs) and dependents (outputs), especially when the relationship is too complex to be articulated by traditional approaches. A limitation in the use of ANNs is that ANNs provide a “black box” approach to the description of the relationship between two sets of variables. Even if ANNs are able to achieve a perfect performance, this would tell us nothing about the functional form of the relationship between the variables and the output layers.

Theobald and Hobbs (1998) described two basic types of spatially explicit model on change of land-use: regression-type models and spatial-transition-based models. The value of regression-type models is that the relative contribution of different variables in analyzing changes in land use is easily obtained (Pijanowski and others 2002). Correlation and logistic regression analyses are widely used to determine the factors driving changes in land use (Veldkamp and Fresco 1997, de Koning and others 1999, Schneider and Pontius 2001; Soares-Filho and others 2002, Almeida and others 2003). A statistical method, such as logistical regression, might have some limitations when variables interact in a complex way (Mas and others 2004); it might even be invalid when spatial variables interact with each other and have difficulties in handling poor and noisy data (Li and Yeh 2002). ANNs are able to directly take into account any nonlinear complex relationship between the driving variables and the change in land use. Fitted to regression-type models, Pijanowski and others (2002) used ANNs to establish functional relationships between a set of spatial predictor variables that was used to predict the locations of change on a landscape.

This article integrates the merits of regression-type models and ANNs to identify major driving factors behind changes in land use. The ANNs-based sensitivity analysis method that we used to identify the factors driving changes in land use also follows the regression-type category. Values of driving factors and land-use change are defined as the input data and output data of neural networks. After the networks are trained and tested, a sensitivity analysis of the input data can tell us which variables are considered the most important by that particular neural network. A sensitivity analysis can give important insights into the usefulness of individual variables. The unimportant variables can be safely pruned out and the key variables retained.

After the neural network is defined, inputs need to be divided into the training data subset, the testing data subset, and the verifying data subset. Through the training and testing stages when the neural network is worked normally, baseline errors (training and testing errors) are recorded for the purpose of making comparisons in the sensitivity analysis. Then, the network drops every variable out at a time to test the error performance. The network is tested repeatedly with the dataset, once for each variable. On each of these tests, all of the values for one variable are replaced with the missing value and the root mean square (RMS) error of the individual variable errors is calculated using the network’s training error function. Important variables have a high rate of error, indicating that the network performance deteriorates substantially if they are not present. The network then calculates the ratio between the error (if one variable is unavailable) and the baseline error (if all of the variables are available). If a ratio exceeds 1, this means that the dropping variable is important. If the ratio is lower than 1, this means that the dropping variable is not the important variable. In general, in selecting important variables, the ratio is set at 1.05 (Statistica 2001). A series of ratios of the variables can be used to reflect the order of importance of factors driving changes in land use.

Application of the Model to the Pearl River Delta Region of Southeast China

The Pearl River Delta Region (PRDR) in southeast China (Figure 3) is widely recognized as one of the most dynamic regions in the developing world. After two decades of the application of an open-door policy and market reforms, the PRDR has become one of the most vibrant and prosperous regions in China. The rapid pace of industrialization has transformed the PRDR from a predominately agricultural region to a “manufacturing workshop of the world” (The Economist, 12 October 2002). The development of the PRDR has successfully expedited the economic restructuring of Hong Kong into a world-class financial, service, and business center over the past two decades. For the economic success of this region, it is crucial to achieve a dynamic and cooperative integration between land use and economic development. The integration of remote sensing and GIS provides an efficient way of helping planners and decision-makers monitor changes in land use and propose sustainable development strategies to guide future development (Yeh and Li 1998).
https://static-content.springer.com/image/art%3A10.1007%2Fs00267-004-0165-z/MediaObjects/267_2004_165_f3.jpg
Figure 3

Location of the Pearl River Delta Region.

Analysis of Change in Land Use in the PRDR

The data processing mentioned earlier to classify land-use data and the spatial distribution are summarized in Table 2 and presented in Figure 4. To detect more details in the change of land use, the diagnostic index, such as the land-use-change matrix \( C_{i \times j} \)
Table 2

Land-use data of the Pearl River Delta Region in 1985, 1995, and 2000 (in ha)

 

Farmland

Forests

Grassland

Water area

Built-up land

Barren land

Total

Area

       

  1985

361,047

370,381

33,401

168,201

151,223

851

1,085,104

 

  33.27%

  34.13%

  3.08%

  15.50%

  13.94%

  0.08%

  100%

  1995

250,881

353,109

36,386

165,388

271,375

7,965

1,085,104

 

  23.12%

  32.54%

  3.35%

  15.24%

  25.01%

  0.73%

  100%

  2000

288,312

344,578

29,826

175,624

245,919

845

1,085,104

 

  26.57%

  31.76%

  2.75%

  16.18%

  22.66%

  0.08%

  100%

  Average percent

27.65%

32.81%

3.06%

15.64%

20.54%

0.30%

100%

Degree of change

       

  1985–1995

9.72

9.67

9.73

9.72

9.75

10

9.71a

  1995–2000

4.90

2.96

5.43

4.49

4.11

16.26

4.09a

Degree of land development

       

  1985

99.82

68.27

6.16

31.00

55.75

0.08

261.07b

  1995

69.31

64.80

6.67

30.34

101.04

0.73

272.89b

  2000

79.14

63.43

5.45

32.45

91.47

0.09

272.03b

a The total degree of change for various types of land use.

b The total degree of development in various types of land use.

https://static-content.springer.com/image/art%3A10.1007%2Fs00267-004-0165-z/MediaObjects/267_2004_165_f4.gif
Figure 4

Land use in the Pearl River Delta Region in 1985, 1995, and 2000.

, mentioned in the change analysis module (Table 1), was used to calculate the land-use-change matrix from 1985 to 1995 and from 1995 to 2000. The results are listed in Tables 3 and 4, where the rows are land-use data in time k and the columns are land-use data in time k+1, and intersected numbers are the amount of change from time k to time k+1.
From these results, the following conclusions can be drawn on land transformation in the PRDR over a period of nearly 20 years. First, in general, the amount of farmland, forests, and grasslands has been decreasing over that period; built-up land has been increasing, and there has been little change in the amount of barren land. Urban expansion was obvious in this region from the 1980s onward, at the expense of farmland, forests, and grasslands.
Table 3

Land-use-change matrix of the Pearl River Delta Region from 1985 to 1995 (in ha)

 

Farmland

Forests

Grasslands

Water area

Built-up land

Barren land

Total

Farmland

212,209

33,831

3,206

36,838

71,731

3,232

361,047

Forests

14,878

304,336

8,248

4,506

38,395

18

370,381

Grasslands

1,184

3,619

24,219

403

3,976

0

3,3401

Water area

13,766

5,142

335

119,983

25,124

3,850

168,201

Built-up land

8,843

6,182

378

3,658

132,137

25

151,223

Barren land

0

1

0

0

11

839

851

Total

250,881

353,109

36,386

165,388

271,375

7,965

1,085,104

Table 4

Land-use-change matrix of the Pearl River Delta Region from 1995 to 2000 (in ha)

 

Farmland

Forests

Grasslands

Water area

Built-up land

Barren land

Total

Farmland

201,214

13,503

1,052

17,876

17,235

0

250,881

Forests

32,313

300,712

3,278

4,241

12,565

1

353,109

Grasslands

3,099

8,193

23,940

322

832

0

36,386

Water area

17,568

4,578

326

133,317

9,596

3

165,388

Built-up land

29,985

17,574

1,230

18,701

203,873

12

271,375

Barren land

4,133

18

0

1,167

1,819

828

7,965

Total

288,312

344,578

29,826

175,624

245,919

845

1,085,104

Second, even under general situations of land transformation, there were differences for different land-use types. In the 10-year period from 1985 to 1995, there has been a serious loss of farmland and major transformations of built-up land, water area, and forests as calculated by area change rate per annum [degree of change for single type of land use (DDS) shown in Table 1]. From 1995 to 2000, however, less farmland, forests, and grasslands were transformed to built-up areas than in the preceding 10 years based on area change rate per annum. At the same time, more land was changed from urban area to farmland, forests, and grasslands compared to the preceding 10 years. This is coincident with the process of development in this region. Although great strides were made in the 1980s in economic development and in improving standards of living, the waste of resources and environmental pollution, especially the deterioration of the environment along the Pearl River has been severe, to the extent that it poses a threat to the long-term development of the economy, society, and the environment. The national and local governments have been taking action to restore the environment and to preserve natural resources and have also placed restrictions on urban expansion. The outside economic situation, especially the economy collapse of southeastern Asian, also sped up change of land use in this region. This economic collapse dampened urban development and real estate development.

Third, the degree of transformation of land use from 1985 to 1995 was greater than that from 1995 to 2000. Except for barren land, all types of land use from 1995 to 2000 were nearly half those in 1985 to 1995, which shows that the degree of change in land use was decreasing.

Finally, the development degree in land use in this region was comparatively high. As mentioned for the diagnostic index, degree of development in land use (LA), in land-use change analysis module earlier (Table 1), development degree index can be used to reflect the degree of land utilization. If all land-use types are to be barren land, the degree of development in land use value is 100, whereas if all land-use types are to be urban land, that value is 400. From 1985 to 2000, the total degree of development of various types in land use fluctuated around 270. With regard to different types of land use, the rate has been decreasing for farmland, forests, and grasslands, whereas that for built-up land has been increasing.

Pattern Analysis of Land Use in the PRDR

Some diagnostic values of pattern analysis, including area (A) and perimeter (P), can be obtained through GIS software. In our study, Arc/Info software was used. Other diagnostic values of pattern analysis were calculated using statistical package SPSS (Statistical Package for Social Science). The results are listed in Table 5.
Table 5

Pattern values of land use in 1985, 1995, and 2000 in the Pearl River Delta Region

 

Type

Farmland

Forests

Grasslands

Water area

Built-up land

Barren land

Total

1985

Area (ha)

361,047

370,381

33,401

168,201

151,223

851

1,085,104

 

Perimeter (m)

15,235,973

19,245,283

2,955,246

10,245,268

13,524,004

145,092

6,135,0865

 

D

1.4004

1.4001

1.3715

1.3908

1.3837

1.1547

1.4322

 

D0

0.2568

0.2991

0.0454

0.1759

0.2221

0.0007

1

 

F

0.0037

0.0051

0.0132

0.0086

0.0148

0.0047

0.0068

 

H

      

2.0334

1995

Area (ha)

250,881

353,109

36,386

165,388

271,375

7,965

1,085,104

 

Perimeter (m)

16,844,058

17,691,257

2,687,675

9,419,610

9,014,866

22,639

55,680,106

 

D

1.3863

1.3895

1.3679

1.3815

1.3844

1.0841

1.4239

 

D0

0.2019

0.2948

0.0471

0.1855

0.2664

0.0043

1

 

F

0.0057

0.0062

0.0138

0.0109

0.0084

0.0013

0.0076

 

H

      

2.1448

2000

Area (ha)

288,312

344,578

29,826

175,624

245,919

845

1,085,104

 

Perimeter (m)

4,451,750

19,621,296

1, 379, 361

5,289,777

6,998,546

5,154

37,745,883

 

D

1.3297

1.3771

1.2899

1.3177

1.3374

1.0736

1.3894

 

D0

0.2191

0.2827

0.0415

0.1944

0.2615

0.0008

1

 

F

0.0045

0.0054

0.0140

0.0096

0.0088

0.0052

0.0069

 

H

      

2.0957

For the fractal dimension, D increased from 1985 to 1995 and then decreased in 2000, which reflects the complexity of all the land-use patches in the PRDR first decreasing and then increasing. Regarding different land-use types, although in general all fractal indexes decreased from 1985 to 2000, built-up land had the lowest value.

For the dominance of land-use patches, the D0 for farmland, forests, and grasslands decreased, barren land changed little, and that of water area and built-up land increased. Changes in the D0 index demonstrated changes in the types of land use, which was from farmlands, forests, and grasslands to urban land in the past 20 years.

For the fragmentation index, C, it can be seen that from 1985 to 1995 fragmentation increased and from 1995 to 2000 it decreased in general, which meant that in the previous 10 years, all land-use patches became fragmented, whereas in the latter 5 years, land-use patches became consolidated. Different types of land use also showed different trends. Farmland, forests, and grasslands increased from 1985 to 1995, whereas urban land decreased. This is significant because, during this time, the urban area nibbled away at farmland, forests, and grasslands, so that urban land became more and more cohesive while farmland, forests, and grasslands became more and more fragmented. The large-scale economy collapse of southeastern Asian directly influenced the land-use pattern in the study region. At the same time, local governments also took actions to restore the environment and to preserve natural resources. These reversed the land-use pattern completely in the period from 1995 to 2000 comparing to the preceding 10 years.

The diversity index increased from 1985 to 1995, and then decreased from 1995 to 2000. This reflects that during the former period, human disturbance to land increased so that the heterogeneity of land-use patches increased, whereas from 1995 to 2000, human disturbance to land decreased and the homogeneity of land-use patches increased.

Value Analysis on Land Use in the PRDR

According to the value analysis on changes in land use mentioned earlier, the economic products of a land-use system can be used to measure the value of land use from the traditional land economy perspective. To quantify and compare different regions and various temporal periods, we used the Gross Domestic Product (GDP) of different land-use types to reflect the differences in the economic productivity of various land-use types. GDP data were collected from the Guangdong Statistical Yearbook for 1986, 1996, and 2001, published by the China Statistics Press, and data on Hong Kong and Macao were collected from the Statistical Yearbook for Asia and the Pacific in 1996 and 2001, published by the United Nations. The GDP per hectare for different land-use types in 1985, 1995, and 2000 are listed in Table 6.
Table 6

GDP per hectare of various land use types in 1985, 1995, and 2000 in the Pearl River Delta Region (HKD/ha)

 

Farmland

Forests

Grasslands

Water area

Built-up land

Average

1985

13,501

304

56,200

7,302

178,068

27,478

1995

28,711

402

131,534

31,369

1,127,338

299,616

2000

62,103

821

377,758

74,745

2,446,573

577,507

Increasing times from 1985 to 2000

4.60

2.70

6.72

10.24

13.74

21.02

Yearly increasing ratio

10.71

6.86

13.54

16.77

19.09

22.51

Great differences existed between different land-use types on the basis of economic product. In the study period from 1985 to 2000, the series on economic product from greatest to lowest was built-up land > grasslands > farmland > water area > forests. Built-up land has an economic product value of more than 30 times that of farmland and water area and thousands times that of forests. Another study done by Yeh and Li (1997) on sustainable land development in Dongguan showed that the encroachment of urban area on the best farmland was the greatest problem in land use. Our results on changes in land use were similar. To make maximum profits from land, farmland and ecological services land has been transformed to land yielding high economic profits.

Using ANNs to Identify Socio-economic Driving Factors Behind Changes in Land Use

With their advantages in handling nonlinear functions, performing model-free estimations of function, learning from data relationships that are not otherwise known, and generalizing to unseen situations, ANNs have been shown to be universal and highly flexible function approximators for any data (Mas and others 2004). Therefore, ANNs make powerful tools for models, especially when the underlying data relationships are unknown (Lek and others 1996, Lek and Guégan 1999). Although ANNs can be architectured as complicated as networks in the real world, the availability of data is often a problem. Moreover, the more complicated the network, the longer the time required for testing and training and the less stable the network will be. To identify the driving factors behind land-use change, the CPV analysis model and ANNs were combined in our study. Through CPV analysis, land-use data can be derived. These data were used in ANNs as output data, and the driving factors were used as input data. Through data elaboration, foregoing analysis on the proximate relationship of input data and output data, ANNs model calibration, and sensitivity analysis of the resultant explanatory variables, the CPV analysis model of land-use change, and ANN-based driving factors identification model were combined.

The model discussed in this article follows five sequential steps: (1) an elaboration of land-use data from 1985 to 2000, (2) an analysis of the relationships between land-use data and the proximate causes and selections of the driving variables, (3) the calibration of the models (training, testing, and verification of ANNs), (4) the selection of the model with the best performance, and (5) a sensitivity analysis of the resultant explanatory variables with the best performing ANNs (quantification of the relationships between land-use data and the “best” explanatory factors).

Elaboration of land-use data from 1985 to 2000

The amount of data required to determine changes in land use is rather high (de Koning and others 1999). Considering the fact that the landuse data were for 1985, 1995, and 2000, the average change in the amount of each type of land use from 1985 to 2000 was calculated. It can be found from these data that the major trends in changes in land use in the PRDR were a decrease in farmland, forests, and grasslands and an increase in built-up land. Therefore, land-use data on ecological service land, including farmland, forests, grasslands, water area, and barren land; and on nonecological service land, including built-up land, were also calculated. Land-use data from 1985 to 2000 are presented in Table 7.
Table 7

Land-use data for the Pearl River Delta Region from 1985 to 2000 (in ha)

Type/year

Farmland

Forests

Grasslands

Water area

Built-upland

Barren land

Ecological Service land

Nonecological service land

1985

361,047

370,381

33,401

168,201

151,223

851

933,881

151,223

1986

350,030

368,654

33,700

167,920

163,238

1,562

921,866

163,238

1987

339,014

366,927

33,998

167,638

175,253

2,274

909,851

175,253

1988

327,997

365,199

34,297

167,357

187,269

2,985

897,835

187,269

1989

316,981

363,472

34,595

167,076

199,284

3,697

885,820

199,284

1990

305,964

361,745

34,894

166,795

211,299

4,408

873,805

211,299

1991

294,947

360,018

35,192

166,513

223,314

5,119

861,790

223,314

1992

283,931

358,291

35,491

166,232

235,329

5,831

849,775

235,329

1993

272,914

356,563

35,789

165,951

247,345

6,542

837,759

247,345

1994

261,898

354,836

36,088

165,669

259,360

7,254

825,744

259,360

1995

250,881

353,109

36,386

165,388

271,375

7,965

813,729

271,375

1996

258,367

351,403

35,074

167,435

266,284

6,541

818,820

266,284

1997

265,853

349,697

33,762

169,482

261,193

5,117

823,911

261,193

1998

273,340

347,990

32,450

171,530

256,101

3,693

829,003

256,101

1999

280,826

346,284

31,138

173,577

251,010

2,269

834,094

251,010

2000

288,312

344,578

29,826

175,624

245,919

845

839,185

245,919

Selection of the socio-economic factors driving changes in land use

Changes in land use are driven by the interaction in space and time between biophysical and human dimensions (Turner and others 1993, 1995, Skole and others 1994). The need to better understand changes in land use as well as their causes and effects requires high-demand data, especially when a temporal analysis is performed. Although ANNs can be architectured to be as complicated as networks in the real world, the availability of data is often a problem. Biophysical data are more difficult to obtain on a temporal dimension. One alternative method is to regard some drivers as remaining constant over time (Pijanowski and others 2002). Socio-economic data are more easily obtained over time in comparison with biophysical data. To evaluate the influence of alternative policies and management regimes on land-use and development patterns is an important aspect of modeling changes in land use (Bockstael and others 1995). For pragmatic purposes, the influence of alternative policies and management regimes on socio-economic dimensions is more important for governments over both the short term and long term. Based on the above reasons, this article aims to identify the socio-economic factors driving changes in land use in the PRDR.

To make this study on the factors influencing changes in land use as comprehensive as possible, 8 categories and 49 factors have been selected, covering the aspects of economic quality, economic structure, economic development, infrastructure development, population, living standards, and environment. Descriptions of all the variables are listed in Table 8. The data to calculate these variables were from the Statistical Yearbook of Guangdong Province, published by the China Statistics Press from 1986 to 2001, for regions within Guangdong Province, and the Statistical Yearbook for Asia and the Pacific, published by the United Nations Economic and Social Commission for Asia and The Pacific in 1996 and 2001 for data of Hong Kong and Macao. All of these data were transformed to same base as their units shown in Table 8.
Table 8

Summary of factors driving changes in land use in the Pearl River Delta Region

Category of variables

Description of variables

I. Economic Quality

Var1:GDP (million HKD); Var2: Primary Industry Product (million HKD); Var3: Secondary Industry Product (million HKD); Var4:Tertiary Industry Product (million HKD); Var5: Economy Density (10 thousands HKD/km2); Var6: GDP per Capita (HKD); Var7: Primary Industry Product per Capita (HKD); Var8: Secondary Industry Product per Capita (HKD); Var9: Tertiary Industry Product per Capita (HKD); Var10: Fixed Capital Formation (million HKD)

II. Economic Structure

Var11: Primary Industry Percentage (%); Var12: Secondary Industry Percentage (%); Var13: Tertiary Industry Percentage (%)

III. Economic Development

Var14: GDP Increase (million HKD); Var15: Primary Industry Product Increase (million HKD); Var16: Secondary Industry Product Increase (million HKD); Var17: Tertiary Industry Product Increase (million HKD); Var18: GDP per Capita Increase (HKD); Var19: Primary Industry Product per Capita Increase (HKD); Var20: Secondary Industry Product per Capita Increase (HKD); Var21: Tertiary Industry Product per Capita Increase (HKD); Var22: Fixed Capital Formation Increase (million HKD)

IV. Infrastructure

Var23: Total Amount of Investment in Fixed Assets per Capita (HKD); Var24: Total Amount of Investment in Fixed Assets per GDP; Var25: Roads per Area (km/km2); Var26: Total Amount of Investment in Fixed Assets per Capita Increase (HKD); Var27: Motor Vehicles Owned by 10 Thousand Persons; Var28: Total Amount of Investment in Fixed Assets per Capita Increase (HKD)

V. Population

Var29: Estimates of Midyear Population (millions); Var30: Annual Growth Rate of Population (%); Var31: Population Density (persons/km2); Var32: Teaching Staff per Ten Thousand Persons; Var33: Students Enrolled per Ten Thousand Persons; Var34: Scientific and Technological Personnel per Ten Thousand Persons

VI. Living standard

Var35: Consumption Standard ((HKD); Var36: Central Government Expenditure per Capita (HKD); Var37: Central Government Revenue per Capita (HKD); Var38: Population per Physician; Var39: Population per Hospital Bed

VII. Environmental Quality

Var42: Volume of Waste Water Discharged per Ten Thousand GPD Products (tons per 10 thousand HKD); Var43: Rate of Industrial Waste Water Treated (%); Var44: Volume of Industrial Waste Gas Emission per Ten Thousand GPD Products (10 thousand m3 per 10 thousand HKD); Var45: Rate of Industrial Waste Gas Treated (%); Var46: Volume of Industrial Dust Emissions per Ten Thousand GPD Products (ton per 10 thousand HKD); Var47: Rate of Industrial Dust Recovered (%); Var48: Volume of Industrial Solid Wastes Discharged per Ten Thousands GPD Products (ton per 10 thousand HKD); Var49: Rate of Industrial Solid Wastes Utilized (%)

Training, testing, and verification of ANNs

In order to develop a network with adequate performance, it is necessary to train, test, and verify the ANN with different input data (Skapura 1996). Training involves presenting input values and adjusting the weights applied at each node according to the learning algorithm. In testing, a separate dataset is presented independently to the trained network to calculate the error rate. Verifying is used to track the network’s error performance and to stop training if overlearning occurs. A sensitivity analysis can lead to important insights on the usefulness of individual variables. However, such an analysis should be used carefully. In general, driving factors to land-use change are not independent, which means that interdependencies occur among driving factors. A sensitivity analysis involves assigning a rating value to each factor if it is no longer available to the model. The interdependence among driving factors might influence the rating of the driving factors to the subtle true situation in an absolute manner. However, it is extremely useful in practice. If a number of models are studied, it is often possible to identify key variables that are always of high sensitivity, other that are always of low sensitivity, and “ambiguous” variables that change ratings and probably carry mutually redundant information. In our study, we found that some of the data within the eight categories were interdependent. However, our objective was not to identify the subtle differences between variables within one category, but the driving factors with high sensitivity through the performance of a number of models. We could also see that the interdependence among those eight categories was very low.

The driving factors selected and land-use data calculated in the previous step were used as inputs and outputs to calibrate ANNs. The following equation was used to normalize all datasets:
$$ V(n)_{st}={{V_{st} - V(\min )_s } \over {V(\max )_s - V(\min )_s }} $$
(6)
where Vst represents input/output data s in year t, and V(min)st and V(max)st represent the minimum and maximum value of input/output data s from 1985 to 2000. The normalized data were stored in STASTICA software in the row with 49 driving factors and 8 land-use types, including ecological and nonecological service land, and the column with the data from 1985 to 2000 in the PRDR. The data were divided into three sections: the training subset for data from 1985 to 1994, the verification subset for data from 1994 to 1997, and the testing subset for data from 1997 to 2000. The ST Neural Networks in STATISTICA software developed by Statfoft was used to train, test, and verify ANNs.

Select the model with the best performance

The ANNs in this study were architectured as multiinput and single output. The selected driving factors were defined as inputs; and land-use types, including ecological service and nonecological service land, were defined as outputs in these respective ANNs at the PRDR spatial level. For each output, all network types, including Linear, Probabilistic Neural Networks, Generalized Regression Neural Networks, Radial Basis Function, and Multi-layer Perceptron, were selected for the Intelligent Problem Solver Tool in ST Neural Networks to find the best performing networks. In selecting important driving factor to land-use change, the ratio of the error (if one variable is unavailable) and the baseline error (if all of the variables are available) is set to be 1.05 (Statistica 2001). The driving factors with ratios lower than the criterion was pruned out in the networks, and those with higher ratios were kept for the best performance networks selection. Through an extensive search, the Intelligent Problem Solver kept the networks with the best performance. Three forms of results can be displayed for each network: a datasheet of the results for each case, an overall summary of the statistics, and a sensitivity analysis of best networks. Through the above steps, the best performing networks for each land-use type were selected and are listed in Table 9.
Table 9

Best performing networks for each land-use type

Land-use type

ANN architecture

Input

Hidden

Output

Farmland

MLP

37

13

1

Forests

MLP

35

16

1

Grasslands

MLP

45

18

1

Water area

MLP

40

17

1

Built-up land

MLP

39

15

1

Barren land

MLP

43

15

1

Ecological service land

MLP

39

12

1

Non-ecological service land

MLP

41

11

1

Sensitivity analysis on the resultant driving factors with the best performing ANNs

Three rows can be found on the sensitivity datasheet of the best performing ANNs: the Rank, Error, and Ratio for each input variable. Error indicates the performance of the network if this variable is “unavailable.” Important variables have a large error, indicating that the performance of the network will deteriorate substantially if these variables are not present. Ratio reports the ratio between the Error and Baseline error mentioned earlier. Rank simply lists the variables in order of importance (order of descending Errors). As mentioned earlier, there are a total of eight categories of factors driving changes in land use. Our objective in identifying driving factors was to identify the major drivers behind changes in land use at the variable category level, but not to identify subtle differences among different variables within each category. The ratio percentage and number percentage of the driving factors belonging to each category were calculated as shown in Table 10. Number percentage means the percentage of the sum of the number of driving factors belonging to each category to all of the driving factors in the network. Ratio percentage means the percentage of the sum of the ratios of the driving factors belonging to each category to the sum of the ratios of all of the driving factors in the network.
Table 10

Sensitivity analysis of the factors driving changes in land use

Land-use type

Driving category: number percentage/sensitivity percentage

Farmland

III: (18.52/21.30); I: (18.52/15.10); V: (14.81/15.60); VI: (14.81/15.27); IV: (11.11/10.71); VIII: (11.11/10.53); II: (7.41/8.06); VII: (3.70/3.43)

Forests

III: (20.00/22.55); VIII: (17.14/15.05); I: (14.29/16.23); IV: (14.29/15.28); VI: (14.29/10.93); V: (8.57/4.60); VII: (5.71/8.96); II: (5.71/6.39)

Grasslands

I: (22.22/22.81); III: (17.78/17.81); VIII: (17.78/15.76); IV: (13.33/16.49); V: (11.11/7.98); VI: (8.89/9.75); II: (6.67/8.18); VII: (2.22/1.22)

Water area

I: (30.00/ 30.57); VI: (30.00/ 29.81); VIII: (20.00/ 20.20); V: (10.00/ 9.71); VII: (10.00/ 9.71)

Built-up land

III: (34.78/31.14); I: (17.39/22.72); V: (13.04/11.93); VIII: (8.70/8.99); IV: (8.70/8.81); III: (8.70/6.53); VII: (4.35/5.21); II: (4.35/4.67)

Barren land

IV: (21.05/21.38); V: (15.79/15.56); V: (15.79/15.22); VI: (10.53/ 11.84); VII: (10.53/11.26); IV: (10.53/10.03); I: (10.53/ 9.87); VIII: (5.26/ 4.84)

Ecological service land

III: (19.44/14.43); V: (16.67/17.44); IV: (16.67/13.79); I: (13.89/18.68); VI: (11.11/13.56); VIII: (11.11/13.41); VII: (5.56/4.45); II: (5.56/4.24)

Nonecological service land

I: (28.57/27.62); III: (19.05/19.43); IV: (14.29/13.90); V: (9.52/11.45); VI: (9.52/8.04); II: (9.52/7.90); VIII: (4.76/7.55); VII: (4.76/4.12)

Although different driving factors and their related categories had a different importance for each land-use type, in general, from the results of the sensitivity analysis, it was found that the major driving factors behind the decrease in ecological land were rapid economic development, increases in population, and the development of infrastructure. Improvements in living standards, environmental pollution, external trade, and economic structure were of little influence. The major driving factors behind increases in nonagricultural land were rapid economic development, the development of infrastructure, and increases in population; improvements in living standards, economic structure, environmental pollution, and external trade have less influence.

Conclusion and Discussion

Land-use change issues are essential to the concerns of local, regional, and global resource management and community development. As an important approach to obtaining data both on spatial and temporal dimensions, remote sensing can be used as a way of obtaining spatial information on land use, as well as forecasting future scenarios on land use. GIS has been a useful tool for the formulation and implementation of analyses on changes in land use. Parameterized and quantitative studies can be performed based on integrated remote sensing data on geo-spatial characteristics. This artcile emphasizes formulation of a model based on changes in land use on the temporal and spatial dimensions using CPV and from the perspective of driving mechanisms.

Changes in land use involve change, pattern, and value. In a change analysis module, a land-use-change matrix, the degree of change for a single type of land use, the total degree of change for various types of land use, and the degree of development in land use were formulated and used as diagnostic indexes. For an analysis of pattern, a series of landscape ecology indexes, including area, perimeter, fractal dimension, dominance, diversity, and fragmentation, were used. Based on the principle of equilibrium in microeconomics under full competition and economic and technological stability, the economic dynamics of a land-use system was formulated for an analysis of value in land use. To examine the relationships of factors driving changes in land use, ANNs were used to identify the driving factors and sensitivity of each factor to land-use changes.

The land-use analysis model was applied to the Pearl River Delta Region of southeast China to investigate changes in land use from 1985 to 2000. Results showed that the most prominent characteristic of the changes in land use during this period was the expansion of urban land at the expense of farmland, forests, and grasslands. From the perspective of the landscape from 1985 to 2000, pattern characteristics became optimized on land use in this region. This can be shown as a decrease in the complexity index, a decrease in human disturbance to land, and an increase in the homogeneity of land-use patches. By calculating the benefits from different types of land use, an analysis of value showed that built-up land could bring returns of more than 30 times that of farmland and water area and thousands times that of forests, which might be an important reason behind changes in land use in this region. For the temporal dimension, CPVs were quite different in the study region for the time period from 1985 to 1995 and the period from 1995 to 2000. The large-scale economy collapse of southeastern Asian damped the development in this region and footprinted land use. Through the effects on economy quality, economic structure, economic development, infrastructure development, the economic situation outside the region changed land use. These reversed the land-use pattern completely in the period from 1995 to 2000 compared to the preceding 10 years.

A sensitivity analysis of those driving factors selected through ANNs showed that the major driving factors behind the decrease in ecological land were rapid economic development, an increase in population, and the development of infrastructure, whereas improvements in living standards, environmental pollution, external trade, and economic structure had little influence. For increases in nonecological land, the major driving factors were rapid economic development, the development of infrastructure, and increases in population; improvements in living standards, economic structure, environmental pollution, and external trade had less influence.

To analyze land-use change comprehensively, other data and techniques can also be considered if change, pattern, value, and driving factors identification analysis are modeled and integrated. This model demands a rather large amount of data to analyze the change, pattern, and value of changes in land use, as well as the driving factors behind such changes. This means that not only data on land use at different times but also social and economic data for the same time period are required. Remote sensing data have been widely recognized as an important way to obtain series data for land-use-change analysis. There exists potential error and uncertainty that might occur from remote sensing. Experience of the interpretater, quantity of image, ground data availability, and so on will influence the quality of the remote sensing data. In our study, although the kappa statistics are adequate, the use of remote sensing data can also derive some errors. Statistical yearbooks and censuses are major sources of economic and social data. However, the quality of such data is questionable. Error and uncertainty of statistical yearbooks and censuses might come from the data collection method, scale of the data collection and reporting, data classification, and data aggregation. Especially in the spatial analysis related to land-use issues, data collection units might have more influence on data quality. Although upscaling and downscaling methods on data aggregation and interpolating are widely used for these data, error and uncertainty will arise. To avoid such a disadvantage, we try to use not only quantity but also ratios, such as quantity of increase, ratio of increase, and per capita value. These approaches can help us obtain more valuable information for a land-use analysis than can be derived from remote sensing data, and the results are more scientific and useful.

The model formulated in this article can be used at different scales, from global, national, and regional to local, to detect different land-use changes with regard to change, pattern, and value, together with the factors driving these scale-related issues. Top-down and bottom-up analyses can also be performed to formulate complete series on land use on both the temporal and spatial dimensions. Even for a certain regions, land-use types can also be gridded so as to focus on different types of land use and land-use locations. However, how social and economic data should be gridded is a question that is still being debated.

The model can also be used to identify hot spots for changes in land use in any region. In this region, changes in land use are the most intensive in degree and have the greatest impact on natural resources. Based on an identification of hot spots and driving factors behind these changes, appropriate countermeasures can be taken and decisions can be made to achieve more a harmonious development of land resources.

Acknowledgments

The work described in this article was substantially supported by grants from the CRC scheme, the Research Grants Council of The Hong Kong SAR (Project No. 3-ZB40), The Hong Kong Polytechnic University (Project Nos. 1.32.37.87CK and 1.34.9709), Frontier Project on Knowledge Innovation in Institute of Geographical Sciences and Natural Resources Research (IGSNRR), CAS, China (Project No. CXIOG-A02-03), and Knowledge Innovation Project of CAS, China (Project No. KZCX3-SW-333). The authors also wish to express thanks to Professor D. Zheng and Professor Y. L. Cai for valuable discussion and support, and anonymous reviewers for their constructive comments.

Copyright information

© Springer Science+Business Media, Inc. 2005