Abstract
This paper presents an empirical analysis of major drivers of land use change in China from 1988 to 2005. We compile a geographic information system database and develop a new method to estimate an econometric land use model that explicitly takes into account the spatial interactions between land use decisions. Results indicate that increasing urban land value was a major driver of farmland development, while rising rural income was a primary driver of conversion of farmland to forests and grassland. The growth of urban income and road density reduces grassland conversion but increases deforestation. (JEL Q15, R14)
I. Introduction
China has experienced rapid urbanization in the last 20 years. The fraction of population residing in urban areas increased from 26% to 47% during the period 1990–2009 (NBSC 2010).1 In particular, rapid urban expansion into fertile farmland in traditional agricultural regions such as the Huang-Huai-Hai Plain, the Yangtze River Delta, and the Sichuan Basin is believed by many to be a major threat to China's national food security. This concern is rooted in the fact that China is home to one-fifth of the world's population (World Bank 2010), but only 10% of the world's cultivatable cropland (FAO 2008). The Chinese government established the Basic Farmland Protection Regulation (State Council 1994) and revised the Land Administration Law (Standing Committee 1998) to impose a set of strict administrative controls over farmland conversion, but hundreds of thousands of acres of high-quality farmland are still being converted to development each year, particularly in rapidly urbanizing regions, such as coastal regions and areas near major cities.
According to the land use database provided by the Chinese Academy of Sciences (CAS), the total developed area increased by 3.3 million ha from 1988 to 2005. Urbanization not only caused an expansion of built-up areas but also has far-reaching indirect effects on land use change. Despite the loss of fertile farmland in traditional agricultural regions, the total acreage of farmland in China increased by 2.8 million ha from 1988 to 2005, due largely to the conversion of grassland to crop production in northern China. China had 303 million ha of grassland in 1988, accounting for approximately one-third of the total national land area. By 2005, the total acreage of grassland had declined to 291 million ha, a 3.9% reduction. Most of this reduction occurred in the farming-pasture zones of Northeast China, the North China Plain, the Loess Plateau, and Northwest China (Liu et al. 2003, 2010).
Understanding the drivers of land use change in China is useful for the design of agricultural, environmental, and land use policies. Many previous studies have made efforts to obtain such knowledge. For example, using land use inventory data and other statistics published by Chinese government agencies, Heilig (1997) identified five anthropogenic factors as major drivers of land use change in China, including population growth, urbanization, industrialization, changes in lifestyles and consumption, and shifts in political and economic arrangements and institutions.
With the increased concerns about the quality of land use data reported by China's statistical system, more and more studies relied on satellite remote sensing data. Most of the studies focused on relatively small geographic areas (Deng et al. 2002, 2010, 2011; Ostwald and Chen 2006; Seto and Kaufmann 2003). Several nationwide studies examined acreage changes for a particular land use but did not consider land use conversions between different uses (Deng et al. 2008). Two exceptions are studies by Liu et al. (2003; 2010), who documented the spatial pattern of land use change in China from 1995 to 2000 and from 2000 to 2005, respectively. A major finding of the previous studies is that both economic and geophysical variables affected land use change in China. Most previous studies, however, are not directly comparable because of differences in land use categories, geographic scopes, temporal dimensions, explanatory variables, and methods. We summarize some of the representative studies on China's land use change in Table 1, with respect to subject area, study region, time period, the type of land use data, and major findings.
Summary of Some Studies on Land Use Change in China
The purposes of this study are (1) to conduct a comprehensive, systematic analysis of land use conversion among six land use categories in China; and (2) to assess the relative importance of various economic drivers of land use changes. To achieve these objectives, we compiled a geographic information system (GIS) database on land use, weather conditions, land quality, topographic features, and economic variables and used it to estimate an econometric land use model that takes into account the spatial interaction between land use choices. We also developed a new method to estimate the model.
Discrete choice models (Carrión-Flores and Irwin 2004; Lewis and Plantinga 2007; Lubowski, Plantinga, and Stavins 2006; Nelson, Harris, and Stone 2001; Nelson and Hellerstein 1997; Wu et al. 2004; Wu and Cho 2007) and duration models (Irwin and Bockstael, 2002, 2004; Irwin, Bell, and Geoghegan 2003) are widely used to study land use changes in the economics literature. A common challenge facing these empirical studies is spatial dependence, which may arise when land uses in nearby parcels directly affect each other or are affected by the same unobserved factors. The former case is referred to as spatial lag dependence (or spatial interaction), and the latter case is referred to as spatial error dependence. Ignoring spatial dependence will lead to biased (or inconsistent) estimates if the dependence structure induces heteroskedasticity in a discrete dependent variable model (Yatchew and Griliches 1985).
Spatial dependence can occur for different reasons. Information, for instance, is a positive externality, in part because of its nonexcludable feature, that could induce industrial or agricultural clustering with neighboring manufactures or farmers adopting the same technology based on a shared but not necessarily symmetric or coordinated learning process. Such externalities cause changes in return structures for neighboring manufactures as a result of changes in transition costs, fixed costs, and infrastructure (Lewis, Barham, and Zimmerer 2008). Industrial agglomeration in China increased steadily from 1998 to 2005 (Lu and Tao 2009). Such agglomeration can shape urban land returns and hence land use decisions. In addition to industry and agriculture, spatial interactions between nearby forested land plots can also create additional values (Albers 1996).
Modeling spatial dependence in many cases means modeling externalities and spillovers.2 The difficulty of detecting and measuring spatial spillovers has been recognized in previous studies (Corrado and Fingleton 2012; Pinkse and Slade 2010). Gibbons and Overman (2012) argued that most applied spatial research achieves only weak identification when employing standard spatial econometric techniques, and suggested using the experimental paradigm to address identification and causality problems. McMillen (2010) was especially concerned about misspecified functional forms and suggested using nonparametric and semiparametric methods for spatial modeling. Despite these criticisms, spatial econometrics techniques are still being used to uncover the nature of spatial dependence.
The literature on spatial econometrics is extensive. Methods for estimation and diagnostic testing have been well developed in the context of linear regression models. It is technically challenging, however, to explicitly take spatial dependence into account in a discrete dependent variable model because of heavy computational burdens, particularly when the dataset is large. Some studies attempt to solve this problem by employing a spatial sampling technique (Carrio´ n-Flores and Irwin 2004; Nelson and Hellerstein 1997), others by constructing proxy variables based on neighbors and adding them to the right-hand side of the equation (Irwin and Bockstael 2002; Nelson, Harris, and Stone 2001; Wu, Fisher, and Pascual 2011). But most of the previous studies ignore the potential spatial dependence.
This paper departs from previous studies in three aspects. First, this national-level study is based on a highly accurate, disaggregated georeferenced land use database that includes data for four years (1988, 1995, 2000, and 2005). The database originally derived from the U.S. Landsat Thematic Mapper/Enhanced Thematic Mapper (TM/ETM) images. Landsat images were interpreted by CAS with a hierarchy land cover classification system and were validated by extensive ground-based surveys (Liu et al. 2003, 2010). The average interpretative accuracy is over 97% (Deng et al. 2006). Longtime coverage and repeated land use decisions in this panel dataset provide even richer sources of information on land use change relative to those observed for only one period. Second, the disaggregated GIS data are able to capture spatial interactions between land use choices and spatial heterogeneity in socioeconomic conditions, weather, land quality, and topographical features. This is particularly relevant because land use decisions are heterogeneous and interdependent among nearby parcels. Most previous studies on China's land use typically compiled data at an aggregated level such as county. As a consequence, some of the more complex interactions between land use decisions are not captured. Third, we propose a new method to estimate spatial interactions in a multinomial logit model. The method is computationally feasible even with a large dataset.
II. Modeling Land Use Change in China
Landownership and Land Allocation in China
To model land use change in China, one must fully understand China's land ownership. Unlike many western countries, China has no private land; land is either owned by the state or by a village collective, depending on land use types. For example, all urban land and most forests, grassland, water area, and unused land belong to the state; all farmland is collectively owned by villagers. The state retains the right to requisition farmland and other collectively owned land for urban and industrial development and other purposes. Land requisition is a unique type of land ownership transaction in China.
China's land markets are generally referred to as land use right markets, which was first introduced in Shenzhen in 1987, formally approved there on an experimental basis in 1988, and subsequently expanded to the rest of the country in 1992 (Lichtenberg and Ding 2009; Deng 2005; Zhu 2005). China's constitution specifies two types of land use right markets: conveyance markets and transfer markets. The conveyance market is the primary market in which local governments lease out use rights to private entities under long-term (40–70 years) contracts. Transactions in the conveyance market for use rights involve payment of an up-front conveyance fee, which was historically set mainly by negotiation but is increasingly set by auction or tenders subject to competitive bidding (Ding 2007; Lichtenberg and Ding 2009). The transfer market is the secondary market in which transactions occur between land users.
Land use decisions are made by two types of agents in China: government officials (county-level or higher) and village collectives.3 These two types of agents have different objectives when making land use decisions. Village collectives have authority to allocate land for agriculture, rural housing, village public works, and village enterprises. Because they benefit directly from land use, it seems reasonable to assume that they seek to maximize the present value of the stream of expected net benefits from the land use.
Government officials make decisions on land requisition and urban development. Many variables may affect their decisions. Lichtenberg and Ding (2009) investigated the role of economic incentives, including the value of urban and agricultural lands, in the primary land allocation in China. They argued that as part of its process of economic liberalization in the pursuit of higher economic growth rates, China implemented a number of fiscal and governance reforms that appear to have pushed local officials to take on the role of land developers.4 These reforms have given rise to land use conversion decisions that respond to economic incentives, even though the allocation of land between urban and rural uses is determined administratively. Deng et al. (2008) examined several fundamental hypotheses generated by the monocentric urban model for explaining the urban expansion in China from the late 1980s to 2000. Their investigation provides empirical evidence that China's urban area is increasing in urban income and decreasing in transportation costs, which is consistent with the classic urban economic theory.
The Econometric Model of Land Use
Consider land use in period t on a parcel of land of uniform quality within a land area indexed by n, where n = 1,..., N. Suppose land use decision makers (village collectives or local officials) choose land uses to maximize the present discounted value of the stream of expected net benefits from the land, and that decision makers base their expectations of future benefits on current and historic values of relevant variables. Under these simplifying assumptions, Lubowski, Plantinga, and Stavins (2006) showed that the decision rule that emerges from the related dynamic optimization problem is: Allocate the parcel to land use j in period t if
[1]where πnjt is the net benefit from land use j at time t.
The potential net benefits from alternative land uses on a parcel, πnjt, depend on attributes of the parcel, such as land quality, weather conditions, and locational characteristics, as well as economic conditions in the surrounding area. Although we have data about each land area, we do not have information about individual parcels within an area. A standard practice in the land use modeling literature is to decompose πnjt into a deterministic component and a random error term:
[2]where the deterministic component 
represents the expected average net benefit from allocating a parcel in area n to land use j at time t, and the random error term εnjt represents the deviation from the average net benefit and is often assumed to follow a normal or type-I extreme value distribution. For example, if εnjt is assumed to follow the type-I extreme value distribution, a standard multinomial logit model of land use is derived. This is the approach taken in many previous studies (see, e.g., Chomitz and Gray 1996; Hardie and Parks 1997; Nelson and Hellerstein 1997).5
Although a standard logit or probit model is easy to estimate, it ignores the spatial interactions between land uses on two neighboring parcels. Some of the interactions, such as habitat fragmentation and urban sprawl, have been integrated into land use decisions explicitly in the literature. For example, Albers (1996) developed an economic forest management model in which the pattern of forested land creates additional value when adjacent preserved plots form minimum habitat size. Irwin and Bockstael (2002) assumed that a parcel's value in a developed use is affected by the past land use choices on neighboring parcels. The cost of not correcting for spatial interactions is biased (or inconsistent) estimates if it induces heteroskedastic errors.
In this study, we take into account spatial interactions explicitly. Because of the forward-looking nature of land use decisions, we assume that net benefits from alternative land uses on a parcel in one area depend on net benefits from alternative land uses in neighboring areas. Having the same land use nearby may promote information spillovers, technology adoption, and labor market pooling and hence generates spatial externalities. Specifically, we rewrite equation [2] as
[3]where ρjt is a spatial autoregressive parameter (⎪ρjt⎪ < 1), representing the degree to which the propensity to have land use j in one area is affected by the propensity to have land use j on neighboring areas;6 wnm reflects the spatial relationship between land areas n and m. Such specification is often referred to in the literature as a spatial lag model (Brueckner 2003), commonly conceptualized as the empirical counterpart to the equilibrium solution of a spatial reaction function, which, in this context, represents the best response of land use decision makers in one location to land uses on other land locations.
The choice between a spatial lag model and a spatial error model is difficult in some settings. Algebraically, the spatial error model is a special case of a spatial lag model, but with additional nonlinear constraints on the parameters; the spatial lag model is a special case of a spatial error model that is nonlinear in the parameters (see Anselin 2006 p. 917 for more details). Although bearing these similarities, the two specifications are nonnested within each other, and the spatial error model is principally more appropriate in a situation where the dependent variable is affected by unobserved factors that are correlated over space. While a variety of identification strategies have been suggested in the literature, the problem in the land use model is complicated in ways that prevent ready adoption of these strategies (Irwin and Bockstael 2002). As a result, it is difficult to identify precisely the interaction parameter ρjt in practice. If the spatial correlation that exists in the unobserved factors is positive on net, the empirical estimate of the interaction effect is expected to be biased in the positive direction, and vice versa. Hence, the spatial autoregressive parameter must be interpreted with caution. In the case of a multinomial logit model, an additional attraction of a spatial lag model is that it leads to a particularly tractable estimation procedure (Klier and McMillen 2008).
Based on economic theories and previous studies (Hall 1966; Fujita 1989; Deng et al. 2008; Lichtenberg and Ding 2009), the average expected net benefits from allocating every parcel in area n to land use j at time t, 
, is specified as
[4]where snt− 1 is a vector of land use proportions in area n in period t − ;1; ynt−1 is a vector of variables describing land quality, topography, and weather conditions in area n between periods t− 1 and t; znt−1 is a vector of economic variables, including land values and household income in urban and rural areas, transportation costs, and public agricultural investments in area n in period t− 1. Note that we use variables in period t −1 to explain net benefits from land use in period t, so that statistical endogeneity seems less likely (Pfaff 1999). 
is a vector of coefficients on 
. These coefficients are specific to the land use converted because xnt− 1 does not vary across land use, hence separate coefficient vectors are estimated for each alternative. Besides, the magnitudes of these coefficients depend on the sensitivity of the net benefit to the initial land use, land quality, and the economic conditions in the surrounding area. For example, the net benefit from converting a parcel of grassland to crop production is likely more sensitive to the parcel's soil quality than the net benefit from converting the parcel to urban development, which is perhaps more sensitive to the county's level of urbanization.
Equation [3] can be written in a stacked form as follows:
[5]where 
. The reduced form of equation [5] is 
, where IN is an N-dimensional identity matrix. An important aspect of the spatial lag model is the spatial multiplier, which can be illustrated by expanding the inverse term in this reduced form: 
Hence the value of πnjt at area n depends not only on xnt−1, but also x values at other locations (−n), with locations further removed discounted by powers of ρ jt. This illustrates the global nature of the spatial multiplier effect in the spatial lag model. In particular, if a unit change were induced in a given explanatory variable 
at every location, the effect on πnjt would amount to 
(Kim, Phipps, and Anselin 2003).
Specification of the spatial weight matrix W is essentially an empirical question. We assume W is a row-standardized N ×N first order queen contiguity weight matrix, that is, 
if areas n and m share common borders or vertices, and wnm = 0 otherwise. This assumption is admittedly arbitrary, but the global nature of the spatial multiplier effect discussed above allows such specification capturing spatial reactions between any two locations through higher powers of W. It also facilitates the estimation of marginal effects as addressed below. Extension to alternative structures of a spatial weight matrix, such as the nearest−neighbor matrix, the binary block matrix, the distance inverse matrix, the negative exponential function, the Gaussian function, and the spherical function (see Dubin 1998 for detailed discussion), is straightforward but complicates the identification process without changing the overall conclusions of this paper.7
The variance-covariance matrix of 
is proportional to 
. Let 
be the diagonal elements of matrix 
and let 
. Under the assumption analogous to the standard multinomial logit model about εnjt, the share of area n allocated to use j at time t can be derived as follows (Klier and McMillen 2008; Train 2003):
[6]Changes in land use in area n between t − 1 and t are intrinsically captured by the lefthand side variable Pnjt (j = 1,..., J) and the right-hand side vector snt − 1.
The Estimation Method
Equation [6] defines a spatial multinomial logit regression model. When N is large, it is infeasible to estimate the model using a traditional maximum likelihood method, because the likelihood function involves an N-dimensional integration. To overcome this problem, we first linearize the generalized residuals around the starting point ρjt =0 and then apply the generalized method of moments (GMM). This approach was first used by Klier and McMillen (2008) to estimate a binary logit model with spatial interdependence. With the linearized model, the procedures reduce to a standard logit (that is, non-spatial) followed by two-stage least squares.
Specifically, from [6],
[7]where δ ( k = j ) is an indicator function, which equals 1 when k = j and zero otherwise;
[8]where 
and 
. Note that Kkt reduces to W if ρkt =0, implying 
since wnn = 0 by definition. Therefore, when 
, the gradient terms for βkt and ρkt reduce to
[9]
[10]Let 
, where 
and 
,8 and let the gradient terms 
. In a multinomial logit model, the generalized residuals are
[11]where snjt is the share of land use j in area n at time t. Linearizing the generalized residuals equation around the initial estimates 
, we have 
, which is equivalent to
[12]Based on equations [9]–[12], the spatial multinomial logit model [6] can be estimated using the following procedure: First, we estimate [6] by setting ρt = 0: βt is estimated consistently by a standard multinomial logit model using the maximum likelihood method. The log-likelihood function is 
. No matrixes need be inverted as 
.
Second, we estimate [12] using the linearized GMM approach. Estimation includes the following steps: (1) Given the estimated parameters 
from the standard multinomial logit model, the initial estimates for 
are 
. Based on 
, calculate the gradient term gnjt and the generalized residuals 
using [9]–[11]. Then calculate 
. (2) Regress 
on instruments 
. The predicted values are 
. Then regress 
on 
. The coefficient estimates are estimated values of θt , expressed as 
, which are estimates for the spatial multinomial model.
The model we estimate, equation [6], predicts the share of an area allocated to alternative land use. But the estimation procedure applies equally to a parcel level analysis, where snjt is a 0/1 indicator instead of a continuous variable ranging from 0 to 1.
We estimate separate coefficients for each conversion period. It would be desirable to conduct a panel data analysis by specifying a spatial fixed parameter that allows better addressing spatial heterogeneity and measurement errors such as the mismatch between the pixel-based imagery and the actual spatial boundaries of decision-making units. However, a fixed effects regression has at least two shortcomings in a multinomial logit model. Methodologically, the inconsistency arises because the number of incidental parameters (N) increases without bound while the amount of information about each incidental parameter (T) remains fixed (Neyman and Scott 1948). Practically, estimating parameters in a nonlinear fixed effects model involves a huge computational challenge. Furthermore, there is little evidence on the performance of a maximum likelihood estimator in a nonlinear fixed effects regression, and the relevant analysis focused almost exclusively on binary choice models (Greene 2004).
More important, land use conversion in China was affected by various policies over time. The period 1995–2000, for instance, witnessed a great concern about farmland protection, while the period 2000–2005 saw an increasing attention to ecological restoration. Limited by the data, these policies cannot be explicitly identified in the model but may induce changes in the structure (coefficients) of a land use model over time. Lacking appropriate policy variables, we estimate two versions of the model: one uses observations from each conversion period, and the other uses observations from the panel dataset by including a time-specific random effect to partly capture the temporal correlation of land use decision. If large differences exist between estimates from these two models, or if estimates are unstable over the three transition periods, then the estimated parameters based on the cross-sectional analysis are potentially superior in the context of this study.
Once the model is estimated, the marginal effects of the observed xnt−1 on land use choice can be calculated by
[13]III. Data
Our study covers the whole of mainland China. Most data used in this paper were provided by CAS, including land uses, topography, climate, and socioeconomic data. Based on the U.S. Landsat TM/ETM images (more than 500 TM scenes for each period), CAS generated the contiguous land use data. The data are available for four time periods—the late 1980s, the mid-1990s, the late 1990s, and the middle years of the 2000–2010 decade— denoted as 1988, 1995, 2000, and 2005, respectively (Liu et al. 2003, 2010; Deng et al. 2006, 2008). CAS made visual interpretations and digitized TM/ETM images to derive thematic maps of land use percentages with a spatial resolution of 1×1 km. There are more than 9.5 million contiguous 1 × 1 km cells to cover the entire country for each time period, which requires extremely intensive computation even with a standard logit regression. Therefore, we need to reduce the sample size to make our estimation computationally feasible.
There are two ways to reduce the sample size: one is to randomly draw a subsample, and the other is to aggregate the 1×1 km cells into a larger geographic unit. The advantage of the first approach is that when properly designed, it can be used to control spatial error autocorrelations (Carrio´n-Flores and Irwin 2004), but the approach could hinder the measurement of the substantive spatial effects, such as neighborhood externalities, when the sample size must be reduced substantially to make computation feasible. Therefore, instead of using a subsample of the 1 × 1km cells, we aggregate the 1 × 1 km cells into 10 ×10 km cells, each of which corresponds to a land area in the empirical model. The main limitation of this approach is that it cannot capture heterogeneity within a 10×10 km cell. But it facilitates modeling land use interactions.
CAS sorted the data with a hierarchical classification system of 25 land use classes, which were then further grouped into six aggregated classes: farmland, forests, grassland, water area, urban land, and unused land. In particular, water area is classified as land covered by natural water bodies and land with facilities for irrigation and water conservation; urban land includes land used for urban and rural settlements, industry, and transportation.9 A detailed explanation of the six aggregated land use classes is available upon request.
Table 2 depicts land use conversion by assigning each cell to the use with the highest proportion among these classes for 1988– 2005. Entries in a cell indicate the number of million hectares that were in the “land use” row in 1988 and “land use” column in 2005. The entries along the diagonal are areas where land use has not changed. Land use changes occurred mainly between farmland, forests, and grassland, and between grassland and unused land. All land uses except grassland increased. Urban area expanded by 56%, the largest change in percent among the six classes; farmland development accounted for 80% of that expansion.
Remotely Sensed Land Use Conversion in China, 1988-2005 (million hectares)
There are no official land value data available. Following Lichtenberg and Ding (2009), we use per hectare GDP in industry and service as a proxy for urban land value, and use per hectare GDP in agriculture as a proxy for agricultural land value. In addition, we include public agricultural investment. The idea is that such investment could lead to higher productivity in agriculture, hence increasing the opportunity cost of farmland conversion. We also include measures of per capita income in urban and rural areas. A higher value of urban income could increase the demand for housing in an urban area and consequently raise the urban land rent (Fujita 1989). Rural income, to a large extent, is determined by the marginal wage rate in urban areas. When the marginal wage rate increases in urban areas, more rural laborers will migrate to cities to work.
Data on county GDP for three years (1989, 1996, and 2000) are gathered from NBSC (2001). GDP consists of three sectors: primary, secondary, and tertiary industries. Primary industry is composed of farming, forestry, animal husbandry, and fishery. Secondary industry mainly includes activities of building, mining, manufacture, and electricity and gas production. Those not belonging to the former two sectors are classified into tertiary industry, for example, transportation, trade, finance, education, public service, and so forth. For convenience, we refer to GDP in primary sector as agricultural GDP and refer to GDP in industry and service sectors as urban GDP. We use urban GDP per unit of urban area (hereafter per hectare urban GDP) as a measure of urban land value. It would be desirable to separate agricultural GDP in the farming sector from that generated from forestry and other agricultural sectors, but such data are unavailable. Lacking better data, we use agricultural GDP per unit of farmland (hereafter per hectare agricultural GDP) as a measure of agricultural land value for each county and each year. Besides, per hectare agricultural GDP can be viewed as an indirect measure of proximity to urban areas. As suggested by Seto and Kaufmann (2003), people inhabit the most productive areas, and hence urban development occurs first there.
We collected data on public agricultural investment from province-and county-level statistical yearbooks available for four years (1994, 1995, 1999, and 2000). The investments came from state and local governments and were used mainly for developing agriculture infrastructure such as seeds, fertilizers, and irrigation projects. We use the investments in 1994, the average of investments in 1995 and 1999, and investments in 2000 when explaining land use change during the three periods (1988–1995, 1995–2000, and 2000– 2005), respectively.
Data on population including rural and nonrural residents are collected from MPSC (1996, 2001). Urban income is calculated as urban GDP per nonrural resident (hereafter per capita urban GDP). Rural income is calculated as agricultural GDP per rural resident (hereafter per capita rural GDP).
We use road density as a measure of transportation costs. Based on a digital map of transportation networks in the mid-1990s, road density is calculated as the total length of all highways, national expressways, provincial-level roads, and other minor roads in a county, divided by the land area of that county. County road density is available only for the mid-1990s. As a supplement, we collected provincial road length for three years (1988, 1995, and 2000) to calculate the province-level growth rate of road length. Lacking better data, we use the county road density in 1995 and the provincial growth rate to extrapolate the road density in 1988 and 2000 for each county.
A common suggestion is that the placement of a road network is endogenous. But in some applications, especially when roads are installed for political reasons, it is reasonable to assume road density is exogenous (Chomitz and Gray 1996). China, for instance, invested approximately $600 billion to upgrade its road system between 1990 and 2005. The road network was designed to eventually connect all cities of more than 200,000 people, and its construction aimed at trade facilitation and promoting faster development of China's poorer inland regions (Roberts et al. 2010). Therefore, the placement of roads was more likely influenced by the national development strategy rather than urban or agricultural returns.
Data on geophysical variables were generated from a GIS database, including timeinvariant data on land quality, terrain slope, and elevation. Land quality is an indicator of potential crop yield, originally measured at a 5 km cell at the equator. A research team from CAS using the stand-alone software of Estimation System for Land Productivity estimated the yield potential (Deng et al. 2006). Terrain slope and elevation were generated from China's digital elevation model as part of the basic CAS database.
Climate panel data were initially collected from over 600 weather stations and organized by the China Administration. The dataset includes annual precipitation and mean annual temperature from 1991 to 2005; CAS interpolated the point climate data into surface data with the method of thin-plate smoothing spline (Hartkamp et al. 1999) to get more disaggregated information for each cell. We calculate the averages and standard deviations of annual precipitation and mean annual temperature for each conversion period. The standard deviations measure temporal variations in weather. We assume these estimated means and standard derivations are constant through every short transition period (1988–1995, 1995–2000, and 2000–2005).
Because we could not determine exactly when during a time interval land use changes occurred, we use the 1988–1989, 1995–1996, and 2000 data (except public agricultural investment) as proxies for the expected returns to land use for the three respective transition periods. The lagged measures help to reduce endogeneity. All value variables are measured at the 2000 real Chinese yuan (¥) and at the county level. Missing data reduced the usable sample to 2,034 county-and 68,918 cell-level observations for each of three time intervals (1988–1995, 1995–2000, and 2000–2005).10 Table 3 provides summary statistics for these variables.
Summary Statistics of Explanatory Variables
IV. Results
The spatial multinomial logit model is estimated separately for each of the three conversion periods: 1988–1995, 1995–2000, and 2000–2005. It is also estimated using the three-period panel data. Variations are relatively large between estimates from the two versions of model and among estimates from the three cross-sectional analyses themselves, indicating possible structural changes of the land use model over time. Since these changes cannot be fully identified in the model due to limited data, we discuss the results below based on the cross-sectional analysis.
Model Performance
Table 4 reports spatial autoregressive parameter estimates by land use and time period. Almost all of them are statistically different from zero at the 1% level. This provides evidence that land use decisions interact with each other across space. The magnitudes of the autoregressive parameters indicate that over time, nonurban land uses became less dependent on land use in neighboring cells, while urban land use became more dependent on land use in neighboring cells. Specially, if a unit change were introduced in a given explanatory variable at every location in every time period, the effect on the expected net benefit from urban use would increase from 1.038 βurban1995 in the first period to 1.067 βurban2005 in the third period. This result is consistent with the observation that land use decisions in China have become more responsive to economic incentives because returns to urban land use are more likely to be affected by the surrounding land use due to agglomeration effects and other spatial externalities.11
Estimated Spatial Autoregressive Parameters
In a GMM regression, R2 does not have a statistical interpretation. Therefore, we evaluate the performance of the model through its prediction accuracy. First, following the winner-take-all principle, we assign each cell to the use with the highest predicted probability. We then compare the predicted major use with the observed major use and report two accuracy measures in Table 5. Borrowing the terms from the remote sensing literature, one measure is labeled as the “producer accuracy,” which equals the percentage of cells whose observed uses are correctly predicted. The other measure, labeled as the “user accuracy” or reliability, is the percentage of cells whose predicted uses are confirmed by observation. These two measures are commonly used in the remote sensing literature to assess the accuracy of classification (Congalton 1991).12
Prediction Accuracy and Predictive Power Assessment by Time Period and Land Use Category
Based on the two accuracy measures, the model performs quite well for the three periods. The overall accuracies are 80% or better for all land use categories except urban land, which has producer accuracies of 74%, 79%, and 74% and user accuracies of 68%, 65%, and 71%. The producer accuracy is higher than the user accuracy because the model overpredicts urban land use.
A parcel could be assigned to a use with a relatively low predicted probability, if the estimated probabilities for the other uses are even smaller. To assess the predictive power, we calculate the summary statistics for the predicted probabilities by time period and predicted land use. As shown in Table 5, the mean of predicted probability is greater than 0.55 in a sample of each use except urban land; the mean value is even higher in the unused land and forests samples. Besides, the maximum of predicted probability exceeds 0.94 in every predicted land use sample, implying that the model has the strength of identifying locations in that category. In summary, accuracy measures reported in Table 5 suggest that the model has strong in-sample predictive power at both aggregated and disaggregated levels.
Drivers of Land Use Change in China
Given that our main interest here is to identify the drivers of land use change, we report the estimated marginal effects. We group all observations into six subsamples by the initial land use in each transition period and evaluate the marginal effects at the sample means for each group using equation [13]. Tables 6–8 report the results for the initial land use that was farmland, forests, and grassland, respectively, for the period 2000–2005. The results for the other two periods, together with the coefficient estimates for the three models, are available upon request.
Marginal Effects on the Probabilities of Farmland Conversion, 2000-2005
Marginal Effects on the Probabilities of Forestland Conversion, 2000–2005
Marginal Effects on the Probabilities of Grassland Conversion, 2000–2005
Most of the marginal effects are statistically significant at the 1% level. Farmland with a higher yield potential, as measured by the land quality variable, was more likely to stay in agricultural use and less likely to be converted to forests and grassland in the first and second periods (1988–1995; 1995–2000). However, the land quality variable became less significant in the third period (2000– 2005). Farmland with higher elevation was more likely to be converted to forests and grassland and less likely to stay in farm use or be developed for urban use. Farmland with steeper slope was more likely to be converted to forestland.
There is evidence that land use change was affected by weather. Farmland was less likely to be converted to grassland and unused land in areas with higher precipitation. Farmland was more likely to stay in agricultural use in those places during the first two periods. This result is consistent with that of Wang et al. (2009), who ascribed it to farmers finding it profitable to switch from irrigation to rain-fed agriculture as precipitation increased and thus irrigation costs were saved. Farmland was more likely to stay in farm use and less likely to be converted to forests in areas with higher temperature. There is empirical evidence that warming was harmful to rain-fed crops but beneficial to irrigated agriculture in China (Tao et al. 2008; Wang et al. 2009; You et al. 2009). Wang et al. (2009) attributed such observation to irrigation farmers using water to offset the heat. The marginal effects of variations in precipitation and temperature on the probability of farmland conversion changed over time. For instance, farmland in areas with a higher variation in precipitation was less likely to stay in farm use and more likely to be converted to forests and grassland during 1988–2000, while the opposite was found during 2000–2005. This may reflect the increasing adaptabilities of farmers to climate change over time.13
The marginal effects of road density reported in Table 6 indicate that farmland in areas with a higher road density was more likely to stay in agricultural use or be developed to urban use and less likely to be converted to forests. This result may reflect that areas with higher road density have lower transportation costs and more convenient access to consumers of agricultural and industrial products and commercial services. Other things being equal, profits from urban and agricultural land uses would be larger.
The signs of the marginal effects of urban land value, urban and rural income, and public agricultural investments became increasingly consistent with economic theory over time. This provides additional evidence that land use decisions in China have been progressively responsive to economic incentives because of government reforms. During 1995–2005, farmland was less likely to stay in farm use and more likely to be converted to urban use in counties with higher urban land value. This result conforms to that of Lichtenberg and Ding (2009) and may reflect that if urban land values are much higher than farmland values, local governments are more likely to approve farmland requisition. In addition, local governmental officials in those places are more likely to engage in political games to increase their promotion opportunities, which often involve public investments in “image projects” designed to show off their “political achievements.” Those “image projects” often lead to conversion of large amounts of farmland into governmental, commercial, and industrial uses.
The marginal effects of agricultural land value on the probability of farmland conversion varied over time. From 1988 to 1995, farmland was less likely to stay in farm use and more likely to be converted to urban use as per hectare agricultural GDP increased. This result is consistent with the findings from a case study on China's Pearl River Delta for 1988–1996 by Seto and Kaufman (2003), who attributed this pattern in part to migration— migrants tended to move to the areas with high farmland productivity and therefore urban centers emerged from those regions. In contrast to the first period, agricultural land value has statistically insignificant effects on farmland conversion in the second and third periods. Similar results were found by Lichtenberg and Ding (2009), who studied urban development in 10 coastal provinces in China from 1996 to 2004. This may be viewed as the impacts of the first migrants versus later immigrants. China's pace of urbanization has accelerated since the early 1990s. First migrants were more likely to occupy areas with high farmland productivity than later immigrants. Consequently, urban development occurred first in the most productive regions in the beginning of urbanization (i.e., the first period), but the location of urban centers became less and less sensitive to land productivity.
Not surprisingly, a higher level of per capita urban GDP, as a proxy for urban income, increased the probability of farmland being converted to urban use and decreased the probability of farmland being converted to forests, grassland, and unused land. The marginal effects of this variable on the likelihood of farmland staying in its initial use changed over time. For example, the marginal values are estimated to be positive in the first two periods and negative in the last period, which reflects that land development is increasingly linked to market forces.
Rural income also affects the probability of farmland conversion. Farmland was less likely to stay in agricultural use and more likely to be converted to forests, grassland, and watered areas in counties with higher rural income. This may reflect that rural income is an indirect measure of the marginal wage rate of urban laborers as well. In China, returns to traditional farming are low. Farmers in counties with a higher marginal wage rate than urban laborers have more financial incentives to switch their livelihood strategies from encroaching into forests and grassland to off-farm employment, including rural-to-urban migration. Such changes may cause conversion of some farmland to forests and grassland.
Public agricultural investment is a government strategy for promoting agricultural development and increased agricultural productivity. The strategy is shown to reduce the conversion of farmland to forests and grassland. The results also show that areas that are attractive for agricultural investment were also attractive for industrial development.
Tables 7 and 8 present the marginal effects of various variables on forest and grassland conversions. Forests and grassland with higher yield potentials were more likely to be converted to farmland. Forests with steeper slopes were more likely to stay forested and less likely to be converted to farmland and grassland. Grassland with higher elevation was more likely to stay pastured and less likely to be converted to farmland and forests.
Forests and grassland were more likely to be converted to farmland in areas with higher road density. Forests were also more likely to be converted to grassland and less likely to stay forested in those places. There is a large body of literature that shows that building and upgrading roads led to deforestation (Chomitz and Gray 1996; Cropper, Griffiths, and Mani 1999; Pfaff 1999). One noticeable exception is a study by Deng et al. (2011), which found that road development in China's Jiangxi province had no impacts on the total forested acreage during the period 1995–2000.
Forests and grassland were more likely to be converted to unused land in counties with higher urban land value and to be developed for urban use in counties with higher urban income. Forests were less likely to be converted to farmland but more likely to be converted to grassland in counties with higher rural income; grassland was more likely to remain as pasture and less likely to be converted to farmland and forests in those places. Public agricultural investment is found to increase the conversion of forests and grassland to farmland, except during the first period.
Relative Importance of Alternative Drivers of Land Use Changes
Although the marginal effects reveal which variables are statistically significant in affecting land use change in China, they do not show their relative importance. To provide some sense of the relative importance of alternative forces driving land use change in China, we use the empirical models to estimate land use changes that would occur under a baseline and six counterfactual scenarios from 1988 to 2005. As described in Table 9, the baseline simulation uses historical observations. It provides a benchmark to measure land use changes under counterfactual scenarios; the counterfactual scenarios, respectively, hold urban land value, agricultural land value, urban income, rural income, road density, and public agricultural investment at a hypothetical level and keep the remaining variables at their historically observed values.
Description of Simulation Scenarios
Change in land use between 1988 and 2005 under each of the seven scenarios is estimated using the following steps: (1) predicting probabilities of land use for each cell based on the land use in 1988 and determining the land use shares on each cell in 1995; (2) repeating the procedure for the periods 1995–2000 and 2000–2005 to determine the land use shares on each cell in 2005; (3) comparing land use shares between 1988 and 2005 in each cell to determine the land use change under the scenario. The simulation results are reported in Table 10, where the odd columns report the simulated land use changes (million hectares), and the even columns report percentage changes from the baseline.14 A positive percentage change indicates that the increase is larger or the decrease is smaller relative to the baseline. A negative percentage change indicates the opposite. Comparing the magnitudes of these values across rows provides an estimate of the relative importance of different factors affecting national land use.
The direction of the acreage change in the baseline conforms to the direction of actual change, except for forest acreage. The baseline changes range from –11 million ha to 10 million ha; by comparison, the actual changes range from –11 million ha to 5 million ha. The discrepancies between the baseline change and the actual change are small (within 0.9 million ha) in farmland, grassland, water, and urban uses but are relatively large in forests and unused land. The model tends to under-estimate forests and overestimate unused land.
Relative Importance of Alternative Drivers of Land Use Changes. 1988-2005
The results suggest that farmland conversion was mostly driven by increasing urban land value and rising rural income. The average urban land value and rural income were, respectively, ¥56,000 per ha and ¥1,150 per capita in 1989. By 2000, the amounts had increased to ¥186,000 per ha and ¥1,870 per capita. Without the growth in urban land value, there would have been 1.6 million ha more farmland; without the increase in mral income, there would have been 1.8 million ha more farmland. In other words, the increases in urban land value and rural income accounted for more than 80% of the change in total farmland acreage in the baseline.
Increasing road density is found to be important in driving farmland expansion. From 1989 to 2000, the average county-level road density increased from 0.751 m/ha to 1.043 m/ha. With this growth, farmland increased by 0.7 million ha, accounting for 34% of the total farmland increase in the baseline. Public agricultural investment was another driver of farmland expansion. With an approximate annual average of ¥82,000 investment in agriculture per county, farmland increased by 0.2 million ha, accounting for 11% of the total farmland increase in the baseline.
The growths of urban income and urban land value are shown as primary drivers of urban development. From 1989 to 2000, the average county-level urban income increased from ¥8,870 per capita to ¥22,160 per capita; with this growth, urban area increased by 0.6 million ha, accounting for 17% of the increase in total urban area under the baseline. With the growth in urban land value, urban area increased by 0.4 million ha, accounting for 11% of the increase in total urban area under the baseline. Road network development also played relatively small but positive roles in urban expansion.
The results suggest that deforestation was primarily driven by increasing road density and rising urban income. Without these increases, there would have been 0.8 and 0.4 million ha more of forestland, accounting for 18% and 10% of the change in total forested area under the baseline. Surprisingly, both factors are found to help alleviate grassland loss. In particular, the growth in urban income abated grassland loss by 3.8 million ha, accounting for 34% of the change in total grassland under the baseline. This may be because road network development and rising urban income played a role of “pressure valve” in grassland conversion. In a region with faster economic growth or easier access to the outside world, rural households who previously relied on encroaching into grassland may have sought off-farm employment, which released pressure on grassland conversion.
Rising rural income also acted as a “pressure valve” for forestland and grassland conversions. With the growth in this factor, forestland increased by 0.5 million ha and grassland increased by 2 million ha, which accounted for 10% and 18% of the changes in forested and grass areas under the baseline. In contrast, public agricultural investment is shown to enhance the pressure on forested and grass areas. Without the investment, there would have been 0.3 million ha more forestland and 1.1 million ha more grassland, accounting for 7% and 10% of the changes in forested and grass areas under the baseline.
V. Conclusions
This paper presents an empirical analysis of major drivers of land use change in China from 1988 to 2005. We compile a GIS database on land use, weather conditions, land quality, topographic features, and economic variables and develop a new method to estimate an econometric land use model that explicitly takes into account the spatial interactions between land use decisions.
Results indicate that increasing urban land value was a major driver of farmland development, reflecting increasing responses of land use decisions to economic forces in China. Rising rural income was a primary driver of farmland conversion to forests and grassland. Rural income was, in part, determined by the marginal wage rate in urban areas. As wage rates in urban areas increased, more farmers migrated to cities to work. As a result, some farmland was converted to forests and grassland. The growth of urban income and road density reduced conversion of grassland to farmland, but increased deforestation. Public agricultural investment contributed to farmland increase, but increased conversion of forests and grassland to crop production.
Geophysical variables also affect land use change in China. Other things being equal, farmland with a higher yield potential and lower elevation was more likely to stay in farm use and less likely to be converted to forests and grassland. Forests with lesser slope and grassland with lower elevation were less likely to stay in their original uses and more likely to be converted to farmland. Farmland was less likely to be converted to grassland and unused land in areas with higher precipitation and to be converted to forests in areas with higher temperature. Farmland was also more likely stay in farm use in those places.
This big-picture study omits many details. For example, we do not take into account policy changes explicitly due to the lack of data, although our separate, cross-sectional analysis for each of the three periods takes into account policy changes implicitly. The six counterfactual simulations were conducted to assess the relative magnitudes of alternative forces affecting land use changes. These results must be interpreted with caution. Nonetheless, the results of this study can help policy makers make more informed decisions. Driven by fast economic growth, urban land value in many parts of China has been highly appreciated. Expanding facilities for industrial parks, commercial buildings, transportation infrastructure, and energy generation are consuming a large amount of arable land, which has raised concerns about China's food security. As Heilig (1997) pointed out, future land use in China depends heavily on how the country's leaders pursue reform. As land value is increasingly determined by market mechanisms, local government officials and village collective leaders will have more incentives to convert farmland to development, particularly in rapidly urbanizing regions. Incentivebased policies that encourage these decision makers to take on the role of conservationist, instead of land developer, will likely be more effective for farmland conservation than the traditional command and control approaches.
Public agricultural investment helps reduce farmland loss but has negative side effects on forests and grassland. As is well recognized, protecting forests and grassland in ecologically sensitive areas is of paramount importance to ecological preservation. Increasing rural income apparently helps achieve this purpose by providing farmers financial incentives to switch their livelihood strategies to off-farm employment. A popular conservation program in China is the Sloping Land Conversion Program (SLCP), initiated in 1999 and designed with dual goals of ecological restoration and poverty alleviation by paying farmers to increase forest cover on highly erodible cropland (typically sloped land). Although the goals of the SLCP are ambiguous, evidence indicates that participants shifted their labor endowment from on-farm work to the off-farm labor market (Kelly 2010; Uchida, Rozelle, and Xu 2009). This suggests that programs that promote rural development and encourage farmers to seek off-farm employment could have conservation benefits.
In the pursuit of a “new economic growth focus,” the central government has targeted construction and automotive sectors as its “pillar industries.” This, along with its national development strategy to facilitate domestic trade and to reduce regional disparities, led to fast development of road networks in China. China invested approximately $40 billion per year to upgrade its road system during 1988–2009. As a result, its total road length increased from approximately 1 million km in 1988 to 3.9 million km in 2009 (NBSC 2005, 2006–2010). Building roads through, near, or to forests and grassland may improve economic viability in the nearby areas, but increase deforestation. Therefore, appropriate measures must be taken to prevent deforestation when designing and upgrading road networks in areas with forest cover.
Acknowledgments
The authors gratefully acknowledge several anonymous reviewers for their constructive comments on earlier versions of this manuscript. The authors also greatly appreciate the land use/cover data support from the Data Center for Resources and Environmental Science, Chinese Academy of Sciences, and from the National Natural Science Funds for Distinguished Young Scholars (Grant No. 71225005) and the National Key Programme for Developing Basic Science of China (2010CB950900).
Footnotes
The authors are, respectively, research fellow, International Food Policy Research Institute, Washington, D.C.; professor and Emery N. Castle Chair, Department of Applied Economics, Oregon State University, Corvallis, and Chang Jiang Visiting Professor, Renmin University, Beijing, China; and professor, Center for Chinese Agricultural Policy, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China.
↵1 In addition to permanent migrants, urban agglomeration also attracted massive so-called floating populations (migrants without local household registration status) from rural areas, due largely to a huge demand for low-wage rural labor and a broad reservoir of the rural unemployed (Heilig 1997).
↵2 See Corrado and Fingleton (2012) for a review of the theoretical and empirical reasons for modeling spatial dependence.
↵3 China has five levels of government: central, provincial, prefectural, county, and township. Township officials rank lowest in China's government hierarchy; they have little power to make land use conversion decisions. Village is an informal subdivision below township, serving as a fundamental organizational unit for rural population.
↵4 In 1994, China implemented a comprehensive fiscal reform to decentralize its public finance: tax revenue was reallocated in favor of the central government at the sacrifice of local governments. This reform did not, however, change local governments' expenditure obligations. In 1998, the State Council launched the housing monetarization policy to replace the long-standing in-kind housing subsidy. This policy targeted the housing sector as “a new growth focus” through fostering the growth of related sectors, including construction, home furnishing, electricity, and home appliances, to offset the impact of the unfavorable economic environment caused by the Asian financial crisis of 1997 (Lee and Zhu 2006). Local governments, on behalf of the state, sell land use rights and retain all profits from land transactions.
↵5 Another econometric issue pertains to the independence from irrelevant alternatives (IIA) property of the standard multinomial logit model. Some applications to land use have demonstrated that IIA assumption is not a serious problem for empirical work (Lewis and Plantinga 2007; Lubowski, Plantinga, and Stavins 2006).
↵6 The net benefits from converting land use to j are a latent variable, explaining the propensity to have land use j.
↵7 In the initial version of this paper, we also tried to use the distance inverse to construct the weight matrix. The preliminary results indicate that the overall conclusions reached in this paper stay the same. With the simplification of the structure of the weight matrix to a first-order queen contiguity weight matrix and given ρt, it is easy to use a numerical approach to solve for (IN−ρtW)− 1 and the diagonal elements of matrix [(IN-ρsW)′(IN-ρtW)]−1, which are important coefficients to estimate marginal effects.
↵8 To avoid redundant parameters, we set the land use J as reference and normalize βJt = 0 and ρJt =0.
↵9 In this study, we allow land to be converted from urban to nonurban use because by definition, urban land includes rural settlements.
↵10 We eliminated the Tibet Autonomous Region, which is covered mainly by grassland and unused land from the dataset, due to severely incomplete economic observations. Nevertheless, it will not change the fundamental results derived from this study because Tibet is a plateau region characterized by sparse population, subsistence agriculture, and a less-developed economy. The land use change in Tibet is trivial in the sample period of this study.
↵11 Without taking into account the spatial dependencies, estimated coefficients for the economic variables would be less significant statistically. Specifically, for the period 1988–1995, 40% of economic variables in the spatial model are statistically significant within the 90% confidence interval, compared to only 10% in the nonspatial, standard multinomial logit model. For the period 1995–2000, 70% of economic variables in the spatial model are statistically significant, compared to 67% in the standard model. For the conversion period 2000–2005, 67% of economic variables in the spatial model are statistically significant, compared to 87% in the nonspatial model.
↵12 The producer accuracy informs the producer of a classification scheme how well a reference cell can be classified, whereas the user accuracy informs the users of a classification scheme how well a cell classified on the map actually represents that category on the ground.
↵13 Although the estimated effects of precipitation and temperature on farmland conversion in China are generally in line with previous studies, they must be interpreted with caution since the knowledge on those effects on crop productivity remains inconclusive in the literature.
↵14 Note that percentage changes under the six counterfactual scenarios cannot add up to 100%. This method follows Stavins and Jaffe (1990). Percentage changes can be interpreted as the share of land use change attributed to the factor. It would be desirable to incorporate uncertainty in the simulation, as suggested by Lewis and Plantinga (2007). However, such practice requires generating numerous land use maps, which imposes a heavy computational burden in the spatial multinomial logit. Hence, we follow Stavins and Jaffe (1990) and Lubowski, Plantinga, and Stavins (2006) and report only point estimates.
