As a modern geographic concept, the TGRR is the region directly or indirectly affected by the backwater of the reservoir, including 32 districts or counties of Chongqing and Hubei (Figure 1). Extending over the medium-high mountains and valleys in central Hubei and the ridge-and-valley region in eastern Sichuan, the TGRR’s topography consists primarily of mountains and hills. The region has a dense drainage network. The main course of the Yangtze River runs through the region, and the Jialing and Wujiang rivers are its largest tributaries within the region. The main types of soil include purple soil, yellow-brown earth, paddy soil and limestone soil. Subtropical evergreen/deciduous forest and mixed coniferous forest are the major vegetation types of the TGRR; evergreen/deciduous broad-leaved forest is the major type of the region’s zonal vegetation. This region has diverse types of land use, with agricultural land accounting for a large proportion of total land area. Farming is practiced primarily on sloping farmland to adapt to the topography. The TGRR has experienced rapid expansion of construction land as a result of the migration, development and construction, especially expansion of urban functional development zone of Chongqing, built-up area of Yichang, and regions centered on cities and towns along rivers. The project of storing clear water and discharging muddy water of the reservoir frequently causes significant short-term changes to the water bodies in the TGRR. In addition to storing reservoir water, the Three Gorges Project involves relocation, removal and reconstruction of cities and towns, supporting infrastructure construction as well as socio-economic development and post-resettlement support (environmental migration and industrial restructuring, etc.) after completion of construction. These activities all have prompted rapid and profound changes in the land-use structure and land functions of the TGRR, especially sudden, continuous and long-lasting changes that have never being found in previous projects. Moreover, environmental problems induced by LUC become increasingly prominent and serious. Soil erosion by water, soil degradation and non-point source pollution are becoming new bottlenecks in the operation of the reservoir and regional development (Mao et al., 2002). After the Three Gorges Project was nearly completed and came into full operation in 2010, the entire reservoir region entered a new historical period called the post-resettlement period. The Central People’s Government issued the Follow-up Work Plan for the Three Gorges Project (the Plan) in order to address potential problems that may surface in the post-resettlement period. The Plan covers issues such as the well-being and prosperity of the migrants, socio-economic development, ecological conservation, prevention and control of geological disasters, and the impacts of the reservoir operation on the regions in the middle and lower reaches of the Yangtze River. The execution of the Plan would necessarily bring about new structural and functional changes in the TGRR’s land, thereby affecting the spatial pattern, process, and path of evolution of land use.

This subsection presents the main data sources used in the study: 1) The land-use data was obtained by correction and interpretation of the land use/cover data (1:100,000) for 2000 and 2010, which were provided by the Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences. The data correction and interpretation were performed based on the results of a field survey and the outcomes were proved to meet related accuracy requirements. Application of the CLUE-S model requires each LUT involved to make up more than 1% of the total land area of the study area (Luo et al., 2010). Because the unused land in the TGRR accounted for far less than 1% of its total land area, this study only investigated the following six LUTs: paddy field, dryland, woodland, grassland, construction land and water area, which were measured in hectare. 2) The digital elevation model (DEM) data was downloaded from the website of the Environmental and Ecological Science Data Center for West China. Then the DEMs (in meter) of the TGRR were derived from the downloaded data by background subtraction and clipping. The ArcGIS Spatial Analyst was employed to extract data about slope gradient and aspect. 3) The data about administrative boundaries, roads, rivers, and residential areas came from the National Geomatics Center of China. 4) The precipitation and temperature data were offered by the National Meteorological Information Center and the meteorological bureaus of Chongqing and Hubei; they were sourced from 25 national weather stations and 42 local weather stations in the TGRR and its surrounding areas; the annual precipitations and annual average temperatures of different areas were extracted from the weather data and then the collaborative Kriging interpolation was performed on the data extracted in combination with the elevations of the stations. 5) The statistical data used in this study, including data on population (total population, urban population, and rural population), administrative area, GDP, food production, and fertilizer input, were obtained from the Statistical Yearbooks of Chongqing and Hubei, Rural Statistical Yearbooks of Chongqing and Hubei, and China County Statistical Yearbook for the period from 2000 to 2015.

1.3.1 Logistic regression analysis based on neighborhood enrichment

The traditional binary Logistic regression is often used to determine the probability of a certain LUT occurring in a certain cell. However, this method was rejected as a basis for simulation and prediction in future scenarios because it overlooks the spatial dependence of land-use data (Wu et al., 2008; Wu et al., 2010). In this study, a spatial autocorrelation factor was combined with land conversion probability and spatial autocorrelation between LUTs in neighboring cells to construct an auto-logistic model that is capable of reflecting the effects of spatial autocorrelation inherent in spatial statistical analysis (He et al., 2003). This model can be expressed as

${{P}_{i}}=\frac{exp({{\beta }_{0}}+{{\beta }_{1}}{{X}_{1i}}+\cdots \cdots +{{\beta }_{n}}Autoco{{v}_{i}})}{1+exp({{\beta }_{0}}+{{\beta }_{1}}{{X}_{1i}}+\cdots \cdots +{{\beta }_{n}}Autoco{{v}_{i}})}$ (1)

where P_i is the probability of the i-th LUT in a particular cell; X_1i, X_2i, …, X_ni, are driving factors behind the conversion of the i-th LUT; β₀ is a constant, and β₁, β₂, …, β_n are the regression coefficients for the explanatory variables X; and Autocov_i is the spatial autocorrelation factor.

The neighborhood enrichment in neighborhood analysis was introduced as the spatial autocorrelation factor Autocov_i, in order to reflect the neighborhood relation among LUTs (Verburg et al., 2004). Then a binary Logistic regression analysis was performed, with the influence of neighbors being regarded as a driver behind LUC. The introduction of neighborhood enrichment complements the simple Logistic regression analysis. The neighborhood enrichment is defined as:

${{F}_{i,k,d}}=\frac{{{n}_{k,d,i}}/{{n}_{d,i}}}{{{N}_{k}}/N}$ (2)

where F_i,k,d is the neighborhood enrichment, in which i denotes the location of a raster cell, k represents LUT, and d is neighbor radius; n_k,d,i is the number of cells containing the k-th LUT within a d radius of the i-th cell; n_d,i is the total number of cells within a d radius of the i-th cell; N_k is the total number of cells containing the k-th LUT within the TGRR; and N is the total number of cells within the TGRR. The Neighborhood Statistics tool provided by ArcGIS9.3 was employed to set the size of each neighbor at 10 m by 10 m, which means that the distance between neighbors is 10 m.

The regression results were then evaluated using a Receiver Operating Characteristic (ROC) curve, also known as the sensitivity curve. As a composite indicator of the sensitivity and specificity of continuous variables, the ROC curve offers a graphical representation of the relationship between sensitivity and specificity. The values of sensitivity and specificity can be calculated by giving multiple cut-off values to the continuous variable. Then the ROC curve can be generated by plotting the sensitivity on the ordinate versus the (1-specificity) on the abscissa. The larger area under the curve, the higher the diagnostic accuracy (Pontius and Laura, 2001; Chen et al., 2015). The selected driving factors are normally considered to have good explanatory power when the values of ROC are greater than 0.7.

1.3.2 Simulation using the CLUE-S model

The CLUE-S model comprises two modules: non-spatial land demand and space allocation (Verburg et al., 2008; Verburg et al., 2010). The former is responsible for calculating a region’s annual demand for each LUT through analysis of driving factors like population, socio-economy and relevant policies and regulations. The latter serves to generate the spatial probability distributions of LUTs using methods such as empirical analysis and spatial variability and then allocate the results calculated by the module of non-spatial land demand to particular spatial locations. Simulation of the spatio-temporal dynamics of land use can be achieved through a number of iterations (Verburg et al., 2009; Zheng et al., 2015). The specific parameters are presented below:

1) Selection and acquisition of driving factors. The spatial pattern and process of land use have complex relationships with site conditions. Human activities aimed at meeting any kinds of human needs ultimately rely on land, but all take place within existing biological, economical, institutional, and technical frameworks. The physiographical factors (topography, hydrology, climate, soil, etc.) have significant control over land-cover types and are restricted by LUC. They exert effect on human activities related to land use, and thus on land-use conversion. The probability and possibility of LUC are to some extent determined by these factors. According to the competition model in neoclassical economics and the principle of optimal utility and best use, maximizing socio-economic and ecological benefits are the highest goal people pursue in activities related to land use. Therefore, socio-economic development (involving factors like population, traffic, economy) as well as policies and regulations (e.g. government project plans) that provide it convenience and guidelines are the major factors driving LUC. To use the CLUE-S model, the selected driving factors are required to be relatively stable or change in an abrupt manner rather than a gradual manner during the study period. In this study, four groups of factors that meet this requirement were selected after considering their availability, data consistency, measurability and correlation (Verburg, 2008), including topographic factors (elevation, slope gradient and slope aspect), climatic factors (annual precipitation and annual average temperature), distance factors (the distances to the nearest urban area, rural residential area, water area, and main roads), and socio-economic factors (population density, per capita GDP, rate of urbanization, food production per unit land area, and fertilizer input). The main roads refer to railway, expressway, national road, and county road. The distances can be obtained using the Euclidean distance tool offered by ArcGIS. Statistical data about the socio-economic factors were collected at district (county) level and the data were then rasterized.

2) Calculation of land demand. Calculation of land demand is a relatively independent module in the CLUE-S model. The land-use data from 2000 and 2010 were used as the base data needed for the simulation. The yearly land-use data for the study period were derived from the base data by quadratic polynomial interpolation. After verification, an empirical interpolation formula was yielded. The time step was set at one year for the quantitative simulation of land demand.

3) Delineation of the restricted area. The restricted area was the whole TGRR. The designed maximum water level of the reservoir has been 175 m since 2010. Because muddy water is discharged from the reservoir during the flood season from July to September and clear water is stored in the relatively dry season months of October and November, the area below this level is subject to seasonal submersion. Therefore, the LUT of this area was determined to be water area that would not change. The restricted area was then changed to be the reservoir region outside the areas below the water level of 175 m.

4) Determination of conversion elasticity coefficient and conversion rate matrix. The coefficient of elasticity of land-use conversion represented the probability that a given LUT will be converted to another LUT. It can be defined by the model parameter ELAS and its values fall within the range of 0 to 1. The higher the conversion elasticity coefficient of a LUT, the more stable the LUT is and the lower the probability that it will be turned into another LUT. Additionally, as the CLUE-S model is sensitive to changes in this coefficient, the conversion elasticity coefficients of the LUTs in the TGRR were determined after a series of efforts, including an overview of LUC in this region, creating a matrix of conversion rates for the period from 2000 to 2010, and adjustments to the model testing. The results are listed in Table 1. The conversion rate matrix shows the pattern of conversion between different LUTs. A conversion rate of 0 means that the LUT cannot be converted, while a conversion rate of 1 indicates that the LUT is changeable.

Figure 2 shows the pattern of land-use conversion in the TGRR. For a sparsely populated region, the prediction of LUC normally does not take into account the conversion of construction land unless the future demand for construction land decreases. It is difficult to convert the dryland in the TGRR to paddy field, primarily because the site conditions of the dryland are unsuitable for paddy field farming and the cost of paddy field farming is much higher than dryland farming. This, together with the growing wages paid for nonfarm jobs, drives rural workers to shift away from agriculture to non-agricultural sectors on a large scale. In response, part of existing paddy field is currently cultivated in ways that require less labor input, primarily dryland farming. Moreover, the sloping farmland in the TGRR does not allow for mechanized farming and delivers just paltry returns. Due to these shortages and the shrinking labor force, arable land is left uncultivated and converted to woodland or grassland across the TGRR. The outflow of rural workers and the aging of those left behind further intensify this trend. Water area change is time dependent. During the period before 2010, the land along the banks of the Yangtze River and its main tributaries had been submerged by the periodic water storage in the Three Gorges Project. Therefore, the TGRR underwent greater expansion of water area than its shrinkage. Since 2010 when the project was nearly completed, the zone located between the designed maximum and minimum water levels had experienced repeated land conversion due to its fluctuating water levels, which makes it difficult to simulate and forecast local LUC. Therefore, the zone below the designed maximum water level, at 175 m, was treated as a changeless zone dominated by water area.

1.3.3 Validation of the simulation results

The results of LUC simulation were evaluated using the Kappa coefficient. The Kappa coefficient is frequently used to assess the accuracy of image classification. If two images differ significantly, the Kappa coefficient should be small (Liu et al., 2009). It can be expressed as:

K = (P₀-P_c) / (1-P_c) (3)

P₀ = s/n (4)

P_c = (a₁·b₁ + a₀·b₀)/n² (5)

where K is the Kappa coefficient; P₀ is the observed agreement between the raster graphics resulting from the actual land-use map and the simulation results; P_c is the expected agreement; N is the total number of pixels in each raster graphic; a₁ and a₀ represent the numbers of pixels with values of 1 and 0, respectively, in the actual raster graphic; b₁ and b₀ denote the numbers of pixels with values of 1 and 0, respectively, in the simulated raster graphic; and s is the number of pixels in the actual raster graphic that have the same values as the corresponding ones in the simulated raster graphic. The values of Kappa fall within the interval [0, 1]. The agreement can be divided into five levels based on the values of Kappa. The five levels and the corresponding value ranges of Kappa are as follows: 0.0-0.20: slight agreement; 0.21-0.40: fair agreement; 0.41-0.60: moderate agreement; 0.61-0.80: substantial agreement; 0.81-1: almost perfect agreement (Peng, 2013).

1.3.4 Design of land-use scenarios

Future scenarios were designed to analyze the possible patterns and processes of LUC in the TGRR. The TGRR is characterized by big cities and villages, massive mountains and a large reservoir. Therefore, it has experienced rapid and profound LUC driven by complex and varying factors during the process from damming the Yangtze River, to storing reservoir water, and then to the post-resettlement development. The LUC in each of the three stages has and will have much bearing on the regional and even national interests. In this context, governments in this region should not only ensure the food security, but also improve the well-being of the resettled people without sacrificing local ecological health. The land-use pattern and process in the TGRR are inevitably subject to the combined effects of the region’s economic development, fragility of ecosystem, and post-resettlement development and construction. After careful consideration of the status quo of land use as well as future socio-economic development strategies, four land-use scenarios were designed for forecasting the region’s future land demand; these scenarios emphasized natural growth, food security, migration-related construction, and ecological conservation, respectively. Land-use simulation was then performed in these scenarios for the period from 2020 to 2030, which was determined primarily based on the Plan prepared in 2010. Details of the four scenarios are given in Table 2.

1.3.5 Extraction of the patterns of LUC in different scenarios

Spatial overlays of the actual raster graphic for 2010 and the simulated ones for 2020 and 2030 from different scenarios were performed through an algebraic operation on the digital raster graphics, in order to extract polygons that can reflect the dynamics of land use over the three years. The operation can be written as:

code = (code₂₀₁₀+1) × 100 + (code₂₀₂₀+1) × 10 + code₂₀₃₀+1 (6)

where code is a code indicating the LUC occurring in a polygon between 2010 and 2030, and code₂₀₁₀, code₂₀₂₀, and code₂₀₃₀ are the original codes of the six LUTs (0 through to 5) in 2010, 2020, and 2030, respectively. Therefore, in the code yielded by the algebraic operation, the number in the hundreds place denoted the LUT in 2010, the number in the tens place denoted the LUT in 2020, and the number in the ones place represented the LUT in 2030; the original codes of paddy field, woodland, grassland, construction land, and water area were then replaced by numbers 1 through to 6. The LUC code represents the LUC occurring during the periods from 2010 to 2020 and from 2020 to 2030. For example, the code “112” indicates that the LUT is paddy field in both 2010 and 2020 and then it changes to dryland from 2020 to 2030.

Given its spatial and temporal variability, the LUC occurring between 2010 and 2030 in the TGRR was divided into three types: early change, late change and continuous change. A polygon showing early change experienced LUC only in the period between 2010 and 2020. For example, the LUC represented by the code “122” is classified as early change. Late changes only took place between 2020 and 2030. Codes like “112” reflect this LUC type. Continuous changes mean that LUC occurred in both periods. Codes like “121” and “124” represent this LUC type. The total areas subject to the three types of LUC (measured in km²) were calculated separately. The future spatiotemporal evolutions of the TGRR’s land use in different scenarios were then predicted from the results.

The results of the ROC test (Table 3) show that the goodness of fit of the model for all LUTs were higher than 0.8. This demonstrates that the selected driving factors had good explanatory power and can be used to simulate and forecast the probabilities of LUTs in the TGRR in the future.

*Note: The type of factors represented by the encoded sc1gr0-sc1gr13 are sc1gr0 (elevation/m), sc1gr1 (slope degree/°), sc1gr2 (slope aspect), sc1gr3 (average annual rainfall/mm), sc1gr4 (average annual temperature/℃), sc1gr5 (the nearest distance to town/m), sc1gr6 (the nearest distance to rural settlements/m), sc1gr7 (the nearest distance to water area/m), sc1gr8 (the nearest distance to the main road/m), sc1gr9 (population density/(×10⁴ person/km²)), sc1gr10 (urbanization rate/%), sc1gr11 (per capita GDP/(yuan/person)), sc1gr12 (per unit area grain production/(t/km²)), and sc1gr13 (chemical fertilizer input (pure calculation)/×10⁴t), respectively. Lyfd0-Lyfd5 are the neighborhood abundance of the paddy field, dry land, woodland, grassland, construction land and water area, respectively. Beta coefficient is the coefficient of relationship diagnosed by logistic regression equation. Exp (B) is Beta coefficient taking as a natural power index coefficient to e as the base. Its value is equal to the ratio of the event, i.e., winning rate. ROC value is the goodness of fit.

The study found that the probability of paddy field had close correlations with topographic, climatic, and socio-economic factors. Elevation and slope gradient were negatively correlated with the probability of paddy field. Slope gradient exerted a dominant influence on the distribution of paddy field. With each 1° increase in slope gradient, the probability of paddy field decreased by 4.42%, suggesting that a steeper slope tended to have a smaller paddy field area. The probability of annual precipitation was approximately 1, suggesting that it contributed little to the distribution of paddy field. Population density had a strong negative correlation with the probability of paddy field, and the corresponding regression coefficient was -7.95. Each 10,000 people per km² increase in population density was associated with a 99.96% decrease in the probability of paddy field, indicating that population had a great effect on the distribution of paddy field. Normally, paddy field is primarily distributed in densely populated areas. However, as the arable land per worker substantially increased due to the mass migration of rural workers seeking better-paid non-farm jobs, the relatively elderly left-behind workers in rural areas had to choose less labor-intensive LUTs. They usually transformed paddy field to dryland, which was one of the most common types of land conversion in this region. As a result, the dryland area increased while the paddy field area decreased. By contrast, the distribution of dryland was correlated only with elevation and distance factors, and their correlations were not strong. The probability of elevation was 0.9995. With each 1 m increase in elevation, the probability of dryland dropped by only 0.05%. The distance factors affecting the distribution of dryland included the distances to the nearest rural residential area, water area and key roads. Their probabilities were all 1, suggesting no significant influence on the distribution of dryland. In terms of neighborhood with other LUTs, the distribution of dryland was positively correlated with the distribution of grassland.

Major factors influencing the distribution of woodland included elevation, slope gradient and aspect, annual average temperature, distance to the nearest rural residential area, population density and food production per unit land area. Population distribution had the strongest influence; the probability of population density was as high as 11.9375. Annual average temperature was the second largest factor affecting the distribution of woodland, with every 1℃ rise in annual average temperature being associated with a 16.01% increase in the probability of woodland. Each 1° increase in slope gradient led to a 2.19% increase in the probability of woodland. The probabilities of elevation, slope aspect, distance to the nearest rural residential area, and food production per unit land area were around 1, indicating that the four factors contributed little to the distribution of woodland. Some socio-economic factors, such as population density, urbanization rate, and fertilizer input, were found to be capable of explaining, and significantly correlated with, the distribution of grassland, while elevation and per capita GDP had little influence. The probability of population density approached 0, suggesting that an area with a higher population density tended to have a lower probability of grassland. Urbanization rate affected the distribution of grassland in a way similar to how it affected the distribution of paddy field. Its probability was 1.0132, lower than that of urbanization rate. An additional 10,000 tons of fertilizer input tended to increase the probability of grassland by 27.3%.

Compared to other LUTs, the distribution of construction land and water area could be explained by fewer driving factors and their correlations were not significant. Elevation affected the distribution of construction land and water area in similar ways, as the two LUTs were located in similar topographic settings. In addition, the distribution of construction land was also affected by distance to the nearest rural residential area and urbanization rate. Distance to the nearest water area was another factor influencing the distribution of water area; the longer the distance, the lower the probability of water area.

2.2.1 Characteristics of LUC

Table 4 shows that a total of 1,202.33 km² of land in the TGRR, equivalent to 2.06% of the region’s area, was subject to LUC during the period from 2000 to 2010. This demonstrates that only localized LUC occurred in this region, which is consistent with the findings by Shao et al. (2013). The areas of paddy field and dryland lost were significantly larger than the areas of newly created paddy field and dryland. During the 10-year period, the TGRR experienced a huge loss of farmland, primarily due to the conversion of dryland (577.9 km²). Meanwhile, dryland was the major source of land for expansion of other LUTs. The area of newly created woodland (472.62 km²) exceeded the area of woodland lost (144.47 km²), as a result of vigorous growth of woodland. The newly created woodland mainly resulted from the conversion of farmland, especially dryland (242.78 km²). The newly formed water area (200.26 km²) was much larger than the water area lost (1.91 km²). The construction of the Three Gorges Project led to extensive submersion of the land along the banks of the Yangtze River. In this process, conversion of all other LUTs into water area occurred. The expansion of construction land (314.03 km²), which was attributed mostly to the conversion of farmland, far outpaced its shrinkage (14.15 km²). The lost farmland was primarily converted into woodland during the nationwide implementation of the “Grain for Green” program or occupied by construction land during the post-resettlement development and supporting infrastructure construction. The conversion of farmland to woodland mostly occurred on steeply sloping dryland, while the farmland occupied by construction land was mostly gentle sloping dryland and paddy field. In addition to sloping dryland, new woodland was also established on grassland by afforestation of barren hills. The enlargement of water area was primarily a result of reservoir construction. Paddy field, dryland and grassland can be easily converted and were thus at a disadvantage in the competition with other LUTs in the TGRR. Woodland possessed advantage in the competition for space due to the protection offered by relevant policies. Despite not being competitive enough, water area and construction land were the major products of the conversion of other LUTs and they have been at an advantage over other LUTs in the competition during the ten years. This was attributable to the development and construction in this region.

Table 4 shows that a total of 1,202.33 km² of land in the TGRR, equivalent to 2.06% of the region’s area, was subject to LUC during the period from 2000 to 2010. This demonstrates that only localized LUC occurred in this region, which is consistent with the findings by Shao et al. (2013). The areas of paddy field and dryland lost were significantly larger than the areas of newly created paddy field and dryland. During the 10-year period, the TGRR experienced a huge loss of farmland, primarily due to the conversion of dryland (577.9 km²). Meanwhile, dryland was the major source of land for expansion of other LUTs. The area of newly created woodland (472.62 km²) exceeded the area of woodland lost (144.47 km²), as a result of vigorous growth of woodland. The newly created woodland mainly resulted from the conversion of farmland, especially dryland (242.78 km²). The newly formed water area (200.26 km²) was much larger than the water area lost (1.91 km²). The construction of the Three Gorges Project led to extensive submersion of the land along the banks of the Yangtze River. In this process, conversion of all other LUTs into water area occurred. The expansion of construction land (314.03 km²), which was attributed mostly to the conversion of farmland, far outpaced its shrinkage (14.15 km²). The lost farmland was primarily converted into woodland during the nationwide implementation of the “Grain for Green” program or occupied by construction land during the post-resettlement development and supporting infrastructure construction. The conversion of farmland to woodland mostly occurred on steeply sloping dryland, while the farmland occupied by construction land was mostly gentle sloping dryland and paddy field. In addition to sloping dryland, new woodland was also established on grassland by afforestation of barren hills. The enlargement of water area was primarily a result of reservoir construction. Paddy field, dryland and grassland can be easily converted and were thus at a disadvantage in the competition with other LUTs in the TGRR. Woodland possessed advantage in the competition for space due to the protection offered by relevant policies. Despite not being competitive enough, water area and construction land were the major products of the conversion of other LUTs and they have been at an advantage over other LUTs in the competition during the ten years. This was attributable to the development and construction in this region.

Table 4 shows that a total of 1,202.33 km² of land in the TGRR, equivalent to 2.06% of the region’s area, was subject to LUC during the period from 2000 to 2010. This demonstrates that only localized LUC occurred in this region, which is consistent with the findings by Shao et al. (2013). The areas of paddy field and dryland lost were significantly larger than the areas of newly created paddy field and dryland. During the 10-year period, the TGRR experienced a huge loss of farmland, primarily due to the conversion of dryland (577.9 km²). Meanwhile, dryland was the major source of land for expansion of other LUTs. The area of newly created woodland (472.62 km²) exceeded the area of woodland lost (144.47 km²), as a result of vigorous growth of woodland. The newly created woodland mainly resulted from the conversion of farmland, especially dryland (242.78 km²). The newly formed water area (200.26 km²) was much larger than the water area lost (1.91 km²). The construction of the Three Gorges Project led to extensive submersion of the land along the banks of the Yangtze River. In this process, conversion of all other LUTs into water area occurred. The expansion of construction land (314.03 km²), which was attributed mostly to the conversion of farmland, far outpaced its shrinkage (14.15 km²). The lost farmland was primarily converted into woodland during the nationwide implementation of the “Grain for Green” program or occupied by construction land during the post-resettlement development and supporting infrastructure construction. The conversion of farmland to woodland mostly occurred on steeply sloping dryland, while the farmland occupied by construction land was mostly gentle sloping dryland and paddy field. In addition to sloping dryland, new woodland was also established on grassland by afforestation of barren hills. The enlargement of water area was primarily a result of reservoir construction. Paddy field, dryland and grassland can be easily converted and were thus at a disadvantage in the competition with other LUTs in the TGRR. Woodland possessed advantage in the competition for space due to the protection offered by relevant policies. Despite not being competitive enough, water area and construction land were the major products of the conversion of other LUTs and they have been at an advantage over other LUTs in the competition during the ten years. This was attributable to the development and construction in this region.

2.2.2 Validation of the simulation results

A comparison of the actual land-use map and the simulation results for 2010 (Figure 3) shows that the spatial locations of the six LUTs in the former were generally consistent with those in the latter, with no significant positional error. Only localized differences were observed between them.

Table 5 reveals some differences in the areas of different LUTs between the actual land-use map and the simulated results. The simulated area of paddy field increased 681.7 km² from its actual area, while the simulated area of dryland decreased 683.75 km² from its actual area. As the increase in the paddy field area roughly offset the decrease in the dryland area, there was almost no net change in the farmland area. Compared to their actual areas, the simulated woodland and grassland expanded by 11 km² and 1 km², respectively, while the simulated construction land and water area shrunk by 9.5 km² and 3 km², respectively. The differences in these four LUTs’ areas were much smaller than those in paddy-field and dryland areas. This finding suggests that the conversion of paddy field to dryland and restoration of paddy field from dryland frequently occurred in the largely mountainous TGRR, due to the difficulty in the mechanization of farming and the increased opportunity cost of farming. This greatly complicated the simulation.

The six LUTs were ranked in descending order of overlap ratio: woodland, paddy field, dryland, grassland, construction land, and water area. The Kappa coefficients of woodland, dryland, paddy field, grassland and construction land were higher than 0.81, indicating that the simulated spatial patterns of the five LUTs approximately coincided with their actual spatial patterns on a regional scale. The Kappa coefficient of water area was between 0.61 and 0.80. This relatively lower agreement was primarily due to the significant change in water area during the period from 2000 to 2010, i.e. the stage of storing reservoir water of the project.

2.3.1 Scenario 1: natural growth

In the scenario of natural growth, the main LUTs in the TGRR were woodland, grassland, dryland, and paddy field over the two decades between 2010 and 2030, and the spatial patterns of land use in 2020 and 2030 were largely consistent (Figures 4a and 4b).

In this scenario, large-scale conversions of paddy field and grassland were the main types of land conversion in both the early period (from 2010 to 2020) and the late period (from 2020 to 2030). These conversions produced new dryland (from paddy field), construction land, woodland, and water area. As shown in Figures 4c and 4d, the conversion of paddy field into dryland was the most prominent form of conversion; a total of 1411.25 km² of paddy field was replaced by dryland over the two periods. It took place in all districts and counties of Chongqing and was especially active in low mountainous and hilly areas. Extension of construction land was the second most prominent form of conversion; it primarily involved occupying paddy field (443.75 km²) and grassland (55 km²). The paddy field and grassland in gentle sloping areas were the first to be occupied during the construction of the TGRR, especially the sprawl or discontinuous expansion of urban areas. Abandoned dryland or paddy field was the major source of land for establishment of new woodland. The TGRR saw woodland area increases of 208.25 km² in the early period and 185.5 km² in the late period. The new woodland was largely distributed in the belt along the Yangtze River in Jiangjin District (at the tail of the reservoir region), the parallel ridges in the main urban districts of Chongqing, the hilly area in southern Chongqing, Wanzhou District and Kaixian County at the heart of the TGRR. The new water areas spread mainly over paddy field (379.75 km²) and grassland (139 km²) and sporadically over small areas of dryland and woodland.

2.3.2 Scenario 2: food security

The land-use simulation in the scenario of food security focused on the conversions of, and conversions to, paddy field and dryland and it was intended to raise the minimum requirements for food security in the future when food self-sufficiency occurs in the TGRR. Contiguous high-quality farmland was protected from conversion for other purposes that could pose a threat to food security. Figures 5a and 5b demonstrate that the localized expansion and shrinkage of paddy field and dryland did not affect the overall spatial pattern of farmland. The spatial pattern of contiguous high-quality farmland was consistent to its actual spatial pattern in 2010 and the simulated pattern in scenario 1. Similarly, no noticeable change was detected in the spatial patterns of woodland, grassland, construction land, and water area.

In the scenario of food security, the TGRR mainly underwent conversions from and to paddy field and dryland in both the early period (2010-2020) and the late period (2020-2030). The main forms of the conversions included the conversion of paddy field to dryland, expansion of construction land, leaving farmland uncultivated, and enlargement of water area. As can be seen in Figures 5c and 5d, the conversion of paddy field to dryland occurred only in the early period and produced 666 km² of new dryland, which was less than half that in scenario 1. This suggests that this conversion was not as widespread as in scenario 1. The expansion of construction land was also not as significant as in scenario 1. The new construction land came primarily from the conversions of paddy field and dryland in both periods, which totaled 462.75 km². The conversion of paddy field to dryland and expansion of construction land in scenario 2 were relatively moderate in comparison with scenario 1, because part of the paddy field and dryland were turned into grassland. This is the most profound contrast to scenario 1 where the grassland shrunk substantially. Additionally, some farmland was left uncultivated in the scenario of food security. The areas of new grassland created from uncultivated farmland in the early and the late periods were 163 km² and 173.5 km², respectively. The new grassland was distributed in Kaixian County, Wanzhou District, Shizhu County, and southern Jiangjin District of Chongqing as well as Badong, Yiling and other remote mountainous areas in Hubei.

2.3.3 Scenario 3: migration-related construction

In the scenario of migration-related construction, the LUC in the TGRR was primarily characterized by continuous expansion of construction land and the accompanying remarkable shrinkage of paddy field, dryland and grassland. Figures 6a and 6b reveal that the expansion of construction land was attributed largely to the expansion of urban areas, especially the counties and market towns along the Yangtze River in Chongqing. These included Changshou and Fuling districts which are close to the main urban districts of Chongqing, counties along the Yangtze River in the central part of the TGRR, and their surrounding villages and towns. The rapid expansion of these urban areas was largely propelled by the steady new urbanization throughout the country, which necessarily occupied surrounding farmland and grassland and thus put them at a disadvantage in the competition for space. The spatial pattern of woodland was relatively stable over the two periods, owing to its remoteness from urban areas and relevant regulatory protection.

In this scenario, the major forms of land conversion in the TGRR over the two decades were the expansion of construction land, conversion of paddy field and dryland, establishment of new woodland and expansion of water area. The expansion of construction land was much more active in scenario 3 (Figures 6c and 6d) than in scenarios 1 and 2. The new construction land (2146.25 km²) was derived mainly from paddy field, grassland and dryland, leading to shrinkage of the three LUTs. The land involved was concentrated in the cities and towns around the segment of the Yangtze River within the key urban districts and Wanzhou District of Chongqing, and scattered in urban areas along the primary and secondary tributaries of the Yangtze River. The conversion of paddy field to dryland (633.25 km²) took place only between 2010 and 2020, like in scenario 2. The new dryland derived from paddy field was found in all districts (counties) of Chongqing. Meanwhile, the growth of woodland was also noticeable. Over the two periods, 1166.75 km² of new woodland was created from paddy field, dryland and grassland.

2.3.4 Scenario 4: ecological conservation

The land-use simulation in the scenario of ecological conservation focused on LUTs of high ecological value, such as woodland, grassland and water area. As illustrated in Figures 7a and 7b, the woodland, grassland and water area kept spreading from their actual scopes in 2010 outward to the surrounding paddy field and dryland. The wide distribution of paddy field and dryland made them vulnerable to the expansion of woodland, grassland and water area. In this scenario, therefore, the competition for space among the LUTs was reflected in the outward spreading land for ecological conservation (LEC) and the retreating cropland. Like in the previous three scenarios, the TGRR in this scenario also underwent the following five forms of land conversion: turning farmland back into forests, leaving farmland uncultivated, conversion of paddy field to dryland, expansion of construction land, and enlargement of water area. However, the land conversions usually occurred continuously in the two decades, which is the most striking difference from the other three scenarios.

As can be seen in Figures 7c and 7d, these five forms of conversion were more noticeable in the early period (2010-2020) than in the late period (2020-2030). The region’s LUC was dominated by extensive conversions of paddy field and dryland in both periods. Farmland suitable for forestation was turned back into forests, and steep or scattered farmland was usually left uncultivated. During the period between 2010 and 2020, a total of 2251 km² of woodland was created from farmland, and the new woodland was concentrated in the mountainous areas in eastern Jiangjin, the southern part of Chongqing’s main urban district, Wanzhou District, and Kaixian County. In the period between 2020 and 2030, the region primarily underwent the conversion of paddy field into woodland (87.25 km²). Uncultivated paddy field and dryland were primarily turned into grassland. This conversion created in roughly equal areas of new grassland in the two periods: 261 km² in the early period and 262 km² in the late period. Compared to the other three scenarios, the conversion of paddy field to dryland in this scenario (395.25 km²) was less significant and more concentrated. Construction land also expanded over farmland; the area of farmland occupied by construction land was 702.25 km², smaller than that area observed in scenario 3. Figure 7e demonstrates that the dominant continuous LUC was the change of farmland in 2010 to woodland in 2020 and then to other LUTs by 2030. The continuous changes occurred in limited areas compared to early changes and late changes and these areas were scattered in the TGRR. There were two continuous changes involving over 100 km² of land: the change of dryland to woodland then to water area, and the change of dryland to woodland and then dryland. The competition between the LEC and the land protected for food security was more intense in this scenario than in other three scenarios.

2.3.5 Comparative analysis

A comparison of Figures 4-7 reveals the five dominant forms of land conversion the TGRR experienced in the four scenarios: conversion of paddy field to dryland, expansion of construction land, turning farmland back into forests, enlargement of water area, and leaving farmland uncultivated. The land areas involved in these land conversions between 2010 and 2030 were calculated from the data in Figures 4-7, for a comparison of the outcomes obtained from the four scenarios (Figure 8). As shown in Figure 8, the conversion of paddy field to dryland was most intense in the scenario of natural growth; the area of land involved in this conversion was 1411.25 km², much greater than those in other scenarios (< 1,000 km²). The expansion of construction land was most dramatic in the scenario of migration-related construction; the construction land area increased by 1939.5 km² to meet the land demand for construction. In the scenario of ecological conservation, the TGRR underwent large-scale conversion of farmland to woodland; the area of woodland created from farmland reached 2338.25 km², larger than the areas of land involved in other conversions.

This is because no additional limitation was imposed on this conversion (the conversion elasticity coefficient of woodland was set high, at 0.8). The expansion of water area and uncultivated land was also noticeable in the scenario of ecological conservation, primarily due to the limitations on the development of water area and grassland.

Driving factors behind LUC have been discussed in many studies. Among various models used in relevant research, the CLUE-S model is able to capture the combined effects of natural factors (elevation, slope gradient, precipitation, temperature, etc.) and socio-economic factors (population, income, urbanization, food, etc.) on multiple spatio-temporal scales and reveal the relationships between the spatial pattern of land use and the driving factors selected. The suitability of different LUTs in each spatial unit can be inferred from the results of simulation. This study found that elevation, slope gradient and precipitation had different degrees of negative correlation with the distribution of paddy field and dryland in the TGRR. This finding is similar to the result by Peng (2013) about the quantitative relationship between the TGRR’s land-use pattern and the factors influencing it. Slope gradient was found to be the major factor controlling the distribution of woodland in this region, which supports Zhang and Lou’s (2011) and Li et al.’s (2013) findings that “low lying and gentle sloping areas tend to have low densities of vegetation”. Compared to other LUTs, construction land and water area made up relatively small proportions of total land area. For this reason, there were relatively fewer factors influencing their overall distribution on a broad spatial scale. In this study, therefore, elevation and their distances to the nearest city and town, rural residential area, and water area were selected as the only factors influencing the distribution of the two LUTs. However, construction land and water area are extremely susceptible to impacts of human activities. Research has shown that the wave of development and construction triggered by the establishment of the TGRR, including the construction of the Three Gorges Project itself, has greatly promoted the expansion of construction land and water area in this region (Wang et al., 2013; Xiu et al., 2013).

The driving factor analysis described in this paper has two weaknesses. First, the CLUE-S model is unable to give full consideration to dynamic factors that are likely to vary sharply within a short period, such as policy- and project-related factors, due to its low sensitivity to such factors. It is therefore incapable of simulating abrupt LUCs, including those caused by the establishment of urban development zones and industrial parks. For example, the simulated evolution of the construction land in the main urban districts of Chongqing between 2000 and 2010 did not indicate the accurate direction of the city’s urban expansion during this period. Second, the simulation in this study failed to visually demonstrate the influences of policy- and project-related factors on land use, due to the difficulty in spatialization of these dynamic factors. For example, before the reservoir water level reached the designed maximum level (175 m), the LUC in the region took place within a very short span of time and thus did not allow for direct spatialization. Therefore, future studies should choose longer time spans and more diverse driving factors in order to achieve more accurate simulation.

The biggest advantage of the CLUE-S model in the LUC simulation is that it can allocate non-spatial land demand to particular spatial locations and define the level of difficulty in converting different LUTs. The four land-use scenarios described in this paper took into account both the general patterns of LUC, and the characteristics typical of the TGRR. For example, the conversion of paddy field into dryland, leaving farmland uncultivated and other forms of land conversion are common trends of LUC in China, and strict protection of woodland, grassland and other types of LEC, and active expansion of construction land and water area is obvious trend that is typical of the TGRR.

The conversion of paddy field to dryland occurred in the four scenarios. The land area involved in this conversion was substantially larger in the scenario of natural growth than in other three scenarios, which is consistent with the trend discovered by Shao et al. (2013) in a remote sensing study of the LUC in the TGRR during the hydraulic engineering construction of the Three Gorges Project. This phenomenon occurred because the land demand in this scenario was calculated using the rate of change observed over the period between 2000 and 2010, during which the conversion of paddy field to dryland was widespread while the conversions of paddy field to woodland and grassland are involved in smaller areas of land. Moreover, existing literature suggests that the conversion of paddy field and dryland into grassland had become a common trend in the TGRR. This trend has been accelerated by the rapid urbanization and industrialization after 2010 and the low comparative efficiency of farming (Shao et al., 2015; 2016). Therefore, this form of land conversion was included in the scenario design.

Strict protection of woodland, grassland and other types of LEC has been implemented in the TGRR since 2000 as a part of ecological improvement efforts. Efforts included turning farmland back into forests, forest engineering, etc. More importantly, the concept of ecological civilization defined at the 18th National Congress of the Communist Party of China and the strategic decision to protect the ecological environment of the TGRR made by General Secretary Xi Jinping at the beginning of 2016 offered policy support for the retention or extension of LEC. However, the expansion of LEC is bound to cause shrinkage of farmland, especially steep sloping farmland. The massive movement of rural young and prime working-age labor force to urban areas and the difficulty in mechanization of farming on steep sloping farmland were another two important drivers behind the conversion of farmland into woodland and grassland. Therefore, the expanding woodland and grassland occupied large tracts of farmland in the scenario of ecological conservation.

Expansion of construction land was most vigorous in the scenario of migration-related construction. It was propelled by a range of driving forces, including the resettlement of migrants to new settlements nearby, removal and reconstruction of cities and towns, supporting infrastructure construction, industrialization, new urbanization, construction of new socialist countryside, and the policy to relocate poor villages deep in high mountains. During the migration, large numbers of rural people moved to cities or leapfrogged from mountainous areas with rugged terrain to regions with gentler terrain and convenient transportation, which necessarily boosted the expansion of construction land. As the expansion of construction land needs to occupy large expanses of high-quality farmland or gentle sloping farmland on low mountains and hills, the farmland area was inevitably reduced. However, the future expansion of construction land in the TGRR will be subject to restrictions imposed by the policy guidance on “building an ecological protective screen in the upper reaches of the Yangtze River” from the central government.

In terms of model processing, the CLUE-S model used in this study has two major shortcomings as follows:

(1) To execute the CLUE-S model, the raster data must be converted to the ASCII format. The raster size and the degree of information integrity during the conversion inevitably affect the accuracy of computation. In this study, the selected driving factors and the primary raster data on land use were included in the Binary logistic model and the ROC curve was used to examine the regression results for different raster sizes. Based on this, the raster size was determined to be 500 m by 500 m. In addition, this study did not analyze the influence of raster size and conversion accuracy on the results of land-use simulation in different scenarios. This needs to be addressed with CLUE-S models capable of multi-scale analysis in future research.

(2) As the land demand calculation is independent of the CLUE-S model’s input, a plugin model is needed to perform the calculation. In this study, the land demand was calculated through interpolation, the most frequently used method in previous studies. The calculation did not consider how the relationship between land use and economic development affected land demand, primarily because the variation in land demand on a broad (regional) scale depends on multiple factors, including dynamic socio-economic factors that can not be related to individual raster cells. Therefore, this study used district (county)-level administrative divisions of the TGRR as the statistical units and each administrative division was treated as a spatial unit with homogeneous attributes. In this way, potential errors were avoided and the model was adapted to the land-use simulation on a broad scale. As the combination of spatial pattern simulation and land demand prediction can improve the accuracy of land-use simulation, the combination of land demand calculation model and the CLUE-S model is expected to be a focus of future research.

Driving factor analysis is not only an important part of research on LUC, but also provides the key basis for formulating proper regulatory policies on land use. Analysis of driving factors behind LUC greatly depends on the scale for analysis, and the scale for analysis is strongly correlated with the study area. If a study is conducted on a regional or national scale, the driving factor analysis is normally regional or national in scale (Dong et al., 2009; Li et al., 2010). If a study looks at a landscape or community, the driving factor analysis should be performed on the corresponding scale. The analysis in the former case is a macro-scale analysis, while that in the latter case is a micro-scale analysis (Shao et al., 2008; Shao et al., 2016). Studies conducted on different scales require different data bases and differ in the depth of information about the driving mechanism behind LUC they can reveal. This study was conducted on a regional scale to analyze future land use in the whole TGRR (or the catchment basin of the reservoir). Therefore, the analysis of factors influencing land use and the selection of parameters for the land-use simulation in different scenarios were also done on a regional scale. Besides, to facilitate collection and preparation of time series data for a long period with multiple time points, the study used districts (counties) as the spatial units to identify the socio-economic factors that affected land use.

LUC usually occurs locally and the driving factors behind it necessarily have close relationship with local basic environments and their surroundings, such as special site conditions, socio-economic background, etc. Though regional economic environment and policies can lead to local LUC, the surroundings of an area undergoing LUC remain the source of most essential driving forces. In this study, therefore, the driving mechanism behind LUC in the TGRR was analyzed on a regional (county) scale and the region’s future land use was then simulated and forecast in several scenarios based on the analysis results. For this reason, the analysis was not precise enough to illustrate the driving factors in depth. This could affect the accuracy of analysis of the LUC process and path under the effects of the selected driving factors, and thereby the results of land-use forecasting.

The scale selection for driving factor analysis exposed another weakness of this study. As the socio-economic data collected on district (county) scale was allocated to particular spatial locations and then converted to rasters of the same size, different forms of land conversion in a district (county) could exhibit the same value for a certain socio-economic factor. This can reduce the accuracy of the driving factor analysis and thereby the accuracy of the land-use simulation in future scenarios.