Development of a parallel computing-based Futureland model for multiple land-use simulation: a case study in Shanghai

Abstract

Predicting future urban land-use scenarios and assessing their encroachment on ecological land is important for sustainability policy formulation. This study develops a new Futureland model using cellular automata with user-defined transformation rules, graphical neighborhood configuration and matrixed transformation cost to simulate multiple land-use changes. Compared with existing models, Futureland applies different factors to different land-use types in a single simulation experiment. Futureland was developed using Geospatial Data Abstraction Library and parallel computing, which significantly improve the implementation efficiency. The case study of Shanghai 2010–2020 illustrates an overall accuracy of 86.6% and a Kappa simulation of 0.79. The land-use scenarios for 2020–2035 were projected under greenspace planning constraints using Futureland. The results indicate that Shanghai’s urban sprawl will gradually slow down by 2030, and the increased urban areas will be mainly in the urban fringes and suburban regions. Futureland can help decision-makers to manage future land-use and optimize urban planning policies.

Keywords:

• Land-use modeling
• cellular automata
• generalized logistic regression
• ecological land
• Futureland

1. Introduction

Widespread land-use and land cover change and the attendant side effects pose challenges to sustainable urban development globally (Nagendra et al. 2018). It is estimated that by 2030, urban land will reach about 1.2 million km², almost three times the size of the world’s urban land area in 2000 (Seto et al. 2012). Urbanization is a major feature of land-use change, especially in rapidly developing areas (Li et al. 2022), which has resulted in a continuous decline in farmland and greenspace, thus underlying degradation of urban ecosystems and environments (Güneralp and Seto 2013; Nor et al. 2017). Among many land-use types, urban greenspace is a typical ecological land type and an essential part of the ecosystems, which is important for maintaining sustainable urban development and ecological health (Jennings et al. 2016). In the context of the implementation of the urban master planning, the impact on ecological space is an important concern for future urban development (Wu et al. 2019). Therefore, spatially projecting urban land-use scenarios and their impacts on ecological space is crucial for spatial planning and harmonious urban development.

Spatial simulation approaches of urban sprawl and land-use change are typically constructed through cellular automata (CA) models, agent-based models and hydrology-driven models in geographical information systems. Among these, agent-based models have been used to simulate land use changes by considering authorities, developers and buyers as different agents and taking into account their different decision processes (Valbuena et al. 2010). In specific studies, hydrology-driven models have also been used to study the complex interactions between land-use and hydrological processes (e.g. variable infiltration capacity, and soil and water assessment tools). For example, such models have been employed to simulate agricultural land-use change in different scenarios (Srivastava et al. 2020). In contrast to these models, CA is a self-organizing spatial modeling framework that has been widely adopted to simulate land-use change, especially in rapidly urbanizing areas (Feng and Tong 2018; Liang et al. 2018a).

Commonly applied CA models and software include CA-Markov, CLUE-S, SLEUTH, FLUS, and UrbanCA (Halmy et al. 2015; Feng and Tong 2020; Wang et al. 2022), to name a few, the last of which provides many new methods for the retrieval of transformation rules and has been applied worldwide. In CA models, statistical regression, neural networks, deep learning, machine learning and heuristic optimization have been proposed to retrieve their transformation rules. In fact, Logistic regression (LR) is the classical method for retrieving transformation rules (Li et al. 2014; Shahbazian et al. 2019), and many consider LR an inaccurate method (Mustafa et al. 2018), while others consider LR a stable method and are constantly working to improve it (Feng 2017; Mirbagheri and Alimohammadi 2017). Even though LR has its limitations in some applications, with improvements it can be more useful and has great application value.

At the driving force aspect, it is widely recognized that land-use change is influenced by a variety of factors, including human disturbance, socioeconomic elements, and physical elements (Chen et al. 2020b). Also, land-use has a significant impact on urban sprawl due to changes in the physical characteristics of the ground surface, soil and vegetation (Srivastava et al. 2020; Wei et al. 2021). For the spatial simulation of multiple land-use, most CA models use the same influencing factors for different land-use types (Firozjaei et al. 2019). In terms of land-use dynamics, the influencing factors of one type are different from those of another one; therefore, when constructing the CA transformation rules, the statistical significance of each factor for different types should be tested to determine which factors influence each type. This requires modifying and refining the standard LR or proposing new approaches that can accept different combinations of factors for each land-use type. Thus, it is essential to remove unnecessary factors to reduce the multicollinearity for improving the simulation accuracy (Feng and Tong 2017).

At the computational efficiency aspect, CA models require iterative computation of each pixel, especially for large-scale, high-precision simulations, and it requires computational acceleration to achieve better performance. The computational load of CA models is largely influenced by the massive spatial datasets and the CA's complexities. However, it is very difficult to achieve rapid implementation when using sequential computation to simulate land-use change (Zhuang et al. 2022). In recent years, modelers have proposed various methods to improve the operational efficiency of CA tools. For example, graphical processing units (GPUs) can improve the efficiency of CA models in a lower computational cost and higher computational power (Li et al. 2012). CA models with integrated multiple GPUs and central processing units (CPUs) can also simulate urban sprawl with massive spatial data, providing more memory and threads to process the data compared to a single GPU (Guan et al. 2016). Meta-modeling approaches have also been proposed to reduce model iterations and achieve stable sensitivity estimates (Şalap-Ayça et al. 2018). Parallel computing is another preferred method that distributes tasks to multiple subtasks under a multi-threaded CPU to achieve simultaneous computation (Newland et al. 2018). This does not rely on the powerful hardware of the computer and has a great potential to scale up in building more efficient CA models.

At the application aspect, CA models are constantly improved and applied to more accurately simulate the spatiotemporal characteristics of complex urban systems and evaluate their impacts on the environment. Using various CA models, land-use change and its impact on ecosystems and the environment can be simulated spatially (Jennings et al. 2012). For example, by integrating CA and regional planning, a future land-use simulation (FLUS) model has been proposed to delineate urban sprawl boundaries and their impacts on the environment (Liang et al. 2018b). In addition, many others have examined the impacts of land-use change on prime farmland, urban greenspace, landscape patterns, and the loss of ecosystem service values (Al-Sharif and Pradhan 2015; Chen et al. 2020a; Gomes et al. 2021). Urban planning policies and greening plans have been developed at different levels (e.g. national and city) to improve habitat quality and significantly increase green coverage (Kanniah 2017). It is known that there are deviations between urban greening planning and practical implementation (Tong and Feng 2019), hence it is necessary to analyze the ecological impacts of future land-use change to help decision-makers undertake a risk assessment and ecological planning. However, the modeling of ecological impacts of future land-use scenarios from existing studies is still limited in considering issued planning, especially in megacities where urban quality is highly appreciated.

The objective of this study is to develop a new CA model of land-use change to address the following questions: (1) Whether different combinations of influencing factors can be used for different land-use types when constructing CA models to simulate land-use change more accurately?, (2) How can parallel computing be incorporated into CA models to improve their computational efficiency?, and (3) How to accurately assess the spatial impact of future urban sprawl on ecological land considering the practical planning? To answer these questions, we developed a new Futureland model based on CA, the Geospatial Data Abstraction Library (GDAL), and parallel computing. Futureland extends the conventional CA methodology by allowing developers to produce their own transformation rules that can be used for simulations and projections of land-use worldwide. To test the model and software, we chose Shanghai as a case study area to project its land-use scenarios. Based on the scenarios, we evaluated the impact of Shanghai’s urban sprawl on ecological space, which would contribute to the optimization of urban greenspace planning and sustainable urban development.

2. Methods: the Futureland model

2.1. The Futureland simulation model

2.1.1. Total transformation rules

The developed Futureland model consists of four core elements including the probability-of-occurrence, neighborhood effects, transformation costs, and constraints. The total transformation rules can be given by: (1) $$P_{\mathit{all}}=P_{\mathit{con}}\times P_{n(h,u)}\times (1-{\mathit{Cost}}_{c\to p})\times \left(1-\mathit{Res}\right)\times P_{\mathit{edge}}$$(1) where $P_{\mathit{all}}$ is the total probability for a cell to change from one state to another; $P_{\mathit{con}}$ is the probability-of-occurrence defined by influencing factors; $P_{n(h,u)}$ is the effect of neighboring cells at iteration time t; ${\mathit{Cost}}_{c\to p}$ is the transformation cost to from type-c to type-p; $\mathit{Res}$ is the local constraints; $P_{\mathit{edge}}$ is the edge enhancement factor, an optional element.

2.1.2. Probability-of-occurrence estimation

The estimation of the probability-of-occurrence is of great importance because it reflects the dynamics of land-use. Among many methods used to derive the transformation rules, LR has been applied for three decades and is very stable in many cases worldwide (Siddiqui et al. 2018). Although machine learning methods have been increasingly combined with CA models in recent years, variants of LR-based CA models continue to emerge (Lin et al. 2011). In this study, we proposed the generalized LR (GLR) to estimate the probability-of-occurrence, where different land types are influenced by the same or different factors for the land-use simulation of the same area. The GLR method can be given by: (2) $$P_{\mathit{con},r}={\left(\frac{e^{{\beta }_0+\sum _{i=0}^n{{\lambda }_i\beta }_i{X}_i}}{1+e^{{\beta }_0+\sum _{i=0}^n{{\lambda }_i\beta }_i{X}_i}}\right)}_r$$(2) where P_con,r is the probability-of-occurrence of land-use type r; n is the number of factors; X_i is the i^th factor; β₀ is a constant; β_i is the regression coefficient of factor X_i. The coefficient reflects the influencing degree of a factor on land-use change. To represent the GLR’s generalization, we introduced a parameter λ_i in the regression, where λ_i=0 means that factor X_i is included; otherwise, the factor is excluded.

GLR statistics in Futureland include factor weight estimates, standard deviations, z-values, and p-values. Depending on p-values, our Futureland model can select different influencing factors for different land-use types, thus improving the predictive ability of the CA model.

2.1.3. Neighborhood effects

Neighborhood influences the state transformation direction of the central cell, but many CA studies do not sufficiently consider the weights of different types of neighbors (Hamad et al. 2018). The weights represent the ability of each land-use type in the neighborhood to influence the central cell. The weight ranges from 0 to 1, where 0 indicates no influence and 1 indicates the strongest influence. In this study, the neighborhood weight matrix can be given by: (3) $$\mathit{W}=\left[\begin{array}{cccc}{w}_{i\to i}& w_{j\to i}& \cdots & w_{m\to i}\\ w_{i\to j}& w_{j\to j}& \cdots & w_{m\to j}\\ ⋮& ⋮& \ddots & ⋮\\ w_{i\to m}& w_{j\to m}& \cdots & w_{m\to m}\end{array}\right]$$(3) where $w_{j\to i}$ is the neighborhood weight of land-use type j on central cell of land-use type i. Thus, for land-use type u, the final neighborhood effects of the central cell h can be given by: (4) $$P_{n(h,u)}={\left(\frac{\sum _{k=1}^{N\times N-1}\mathit{con}\left({\mathit{w}}_{k\to u}\right)}{N\times N-1}\right)}_h$$(4) where $N\times N$ is the neighborhood size; $\mathit{con}(w_{k\to u})$ is the total neighborhood effects on the central cell. Figure 1 shows how to define the neighborhood weights and calculate the consequent neighborhood effects.

Figure 1. The neighborhood shape definition tool, and a case of defining the neighborhood weights and effects. For example, for land-use transition to type B, the weights for types A, B, C and D are w₁, w₂, w₃, and w₄, respectively. In a 3 × 3 square neighborhood, the total effects for type B are (3w₁+3w₂+w₃+2w₄)/8.

2.1.4. The transformation cost

The transformation of one land-use type to another is associated with socioeconomic costs (Verburg and Overmars 2009). In CA-based land-use change modeling, the transformation cost matrix is defined based on the context, land attributes, and socioeconomic conditions of the study area. The transformation cost (${\mathit{Cost}}_{c\to p}$) in Equ.(1) can be given by: #(5) $${\mathit{Cost}}_{c\to p}=\{\begin{array}{c}0, s.t. \mathit{type}-c=\mathit{type}-p\\ CC, s.t. \mathit{type}-c\ne \mathit{type}-p \end{array}$$#(5) where CC is the cost that ranges from 0 to 1, with a lower cost indicating a higher likelihood of transformation. Figure 2 shows the definition of the transformation cost matrix, where Cost_(A→B) donates the transformation cost from the current (t) land-use type A to the next (t + 1) possible land-use type B.

Figure 2. The definition of the transformation cost between any two land-use types.

2.1.5. The edge enhancement

In Futureland, we identified the edge position to enhance the edge shape of each land patch. Figure 3 shows that a pixel can be changeable if its surrounding pixels are of a different type ($P_{\mathit{edge}}$=1); otherwise, a pixel can be non-changeable if its surrounding pixels are of the same type ($P_{\mathit{edge}}$=0). This approach keeps land-use patterns well-marginalized, with better differentiation between different land-use types. The edge enhancement is typically applicable to areas where urban planning regulations are strictly enforced (e.g. Shanghai).

Figure 3. The processing of edge enhancement in the Futureland model.

2.1.6. The roulette selection

The selection of the target land-use type to be transferred is the final step in CA modeling. In the literature, many have chosen the type with the highest transformation probability as the target, but this denies the transformation possibility of cells with a lower probability. In our Futureland model, we applied the roulette-wheel approach commonly used in genetic algorithms to model the complex relationship of probabilistic choices (Lipowski and Lipowska 2012). The application of this method in geographic modeling has shown that it is effective and can yield better simulation results (Liu et al. 2017). In roulette-wheel selection, the land-use type with the highest transformation probability is most likely to be assigned as the target, but types with low probability also have a chance. For example, if the total probability of a pixel transforming from farmland to urban is 0.9, and the total probability to the water body is 0.2; in the conventional CA models, because the probability to urban is greater, the result is that the pixel transforms to urban; in our model, however, the possibility to urban is 90% and the possibility to water body is 20%. As the roulette-wheel selection incorporates stochastic parameters, it can well reflect the uncertainty of land-use dynamics and real-world land-use alternatives, and can better integrate the competition among different land-use types.

2.2. Global and local constraints

2.2.1. The modified Markov chain

The total amount of land demand for each type is a global constraint for future scenarios. The number of iterations is assigned according to the final land state to match the land demand. The standard Markov chain is a stochastic process in which time and state are discrete, and it evolves the future state of random variables through a transformation probability matrix (Mansour et al. 2020). In this study, a modified Markov chain that considers the probability-of-occurrence was used to project the total future land demand. The urban sprawl process conforms to an S-shaped curve, i.e. the urban sprawl rate slows down in the later urbanization stages (Mesta et al. 2022), we thus added a factor to the standard Markov chain to reduce the rate. The land demand based on the modified Markov chain can be given by: (5) $$L_{\mathit{DEMAND}}^{t+1}=L_{\mathit{DEMAND}}^t\times \left[\begin{array}{ccc}{P}_{11}& \cdots & P_{1n}\\ ⋮& \ddots & ⋮\\ P_{n1}& \cdots & P_{nn}\end{array}\right]\times P_{\mathit{occ}}$$(5) where $L_{\mathit{DEMAND}}^t$ and $L_{\mathit{DEMAND}}^{t+1}$ are the land quantities at time t and time t + 1, respectively; $P_{ij}$ is the land transformation probability matrix at time t; $P_{ij}$ satisfies the requirement of 0 ≤ $P_{ij}$ ≤ 1; the sum of elements in each line is equal to 1; $P_{\mathit{occ}}$ is an adjustment factor based on the probability-of-occurrence vector.

2.2.2. Restricted areas

The restricted areas are local constraints that do not change during the land-use simulation. The constraint term Res has a value of 0 or 1, where 0 means available and 1 means unavailable. The restricted areas typically include water bodies, nature reserves, and reserved areas in urban planning. By using the restriction layer, the simulation results can be made more consistent with the actual situation.

2.3. Simulation assessment

Futureland includes an accuracy assessment tool that uses an error matrix to measure how well the simulated land-use pattern matches the real land-use pattern. The error matrix shows the number of correctly and incorrectly simulated pixels associated with each land-use type. A few accuracy metrics were applied to evaluate the model and results, which typically include overall accuracy, producer’s accuracy, user’s accuracy, and Kappa simulation. We also applied fuzzy Kappa and aggregated cells to measure similarities and differences in land-use simulations (Table 1).

Table 1. The definition of assessment metrics of CA models and results.

Type	Metric	Definition
Simulation accuracy	Producer’s accuracy	Percentage of correctly simulated pixels in the ground reference results corresponding to a particular land-use type.
	User’s accuracy	The number of correctly simulated pixels associated with a particular land-use type, as a percentage of the predicted results for that type.
	Overall accuracy	The sum of correctly simulated pixels divided by the total number of pixels.
	Kappa simulation	The coefficient of agreement between the simulated land-use transitions and the actual land-use transitions (Van Vliet et al. 2011).
	Fuzzy Kappa	A variant of the Kappa coefficient that considers different degrees of similarity between pairs of cells on two maps (Hagen 2003).
	Aggregated cells	Similarities of simulated and actual land-use patterns at different scales (Pontius et al. 2004).

3. Materials and model framework

3.1. Study area and datasets

Shanghai located in eastern China is the most developed city of the country with an area of 6,340 km². As of 2021, the total population has reached 24.8 million, of which 89.3% were urban residents and only 10.7% were rural residents. On the one hand, urban sprawl affects urban population flows, e.g. the construction of rail transit in a region can attract people to live there; on the other hand, population flows also promote urban sprawl, e.g. population growth also drives the construction of infrastructure. Over the past decade, Shanghai’s rapid urban sprawl (from 87,327 ha in 2010 to 141,207 ha in 2020) has led to dramatic land-use change and encroachment on various ecological lands. Given the scarcity of ecological space, Shanghai has proposed to build a sound system of three major ecological lands including parks, forests, wetlands and two major networks of corridors and greenways, aiming to create the four types of ecological space (Figure 4a) raised in the Shanghai Urban Master Plan (2017–2035).

Figure 4. The study area of Shanghai: (a) Ecological spatial plan (2017–2035); (b) Land-use pattern in 2010.

Figure 4 shows the ecological spatial plan and land-use pattern in 2010, where Eco-space Type-I and Type-II are of significant ecological functions (e.g. wetland nature reserves and the core areas of forest parks), and development is prohibited to maintain ecological functions. The Eco-space Type-III refers to the ecological space outside the urban boundary, including basic farmland, forest land and ecological corridors, to name a few. Eco-space Type-IV indicates ecological space within the urban boundary, such as green belts of the Outer Ring Highway and suburban parks.

In this study, we used Landsat-5 TM and Landsat-8 OLI level-2 images, and performed radiometric and geometrical corrections. These images then have been applied to produce land-use maps for 2010, 2015, and 2020, using the supervised neural network approach in ENVI 5.3. Five land-use categories were considered according to the classification criteria developed by the Chinese Academy of Sciences (CAS; Table 2). We applied a systematic sampling method to select 52,503 pixels, of which 70% were used to calibrate the classifier and 30% for validation, with an overall accuracy higher than 95.7%; then, we selected 17,418 pixels to calibrate the Futureland model. The land-use patterns show that urban land accounted for 24.9% in 2010, while water area, greenspace, farmland and special land respectively accounted for 13.9%, 10.1%, 50.6% and 0.5%.

Table 2. Land-use classification criteria and definitions in the Shanghai study area.

New category	Initial category	Definition
Urban land	Built-up land	Land for urban and rural settlement, and man-made buildings and projects
Water area	Water body	Land covered by natural water bodies such as rivers and lakes or land with water storage and irrigation functions
Greenspace	Woodland & grassland	Land covered with vegetation, including natural forests, grasslands, wetlands and artificially planted orchards
Farmland	Cropland	Farmland for crops, including paddy land and dry land
Special land	Other built-up land	Land such as airports and harbors

Table 3 summarizes the remote sensing images and geographic information data used in this study. In addition to Landsat images, we collected Digital Elevation Model (DEM; from ASTER GDEM 2) data from the Geospatial Data Clouds. The socioeconomic data include the GDP density provided by CAS and the population density provided by Worldpop. The POI datasets including road and river networks were collected from OpenStreetMap. We performed a uniform coordinate transformation of all datasets using ArcGIS 10.2, then extracted 11 factors from the above dataset by following earlier works (Cloern et al. 2016; Chen et al. 2017; Gao et al. 2020). All distance factors were extracted from the facility data, such as various POIs, roads and rivers. All factors were converted to grid data with a resolution of 30 meters by resampling.

Table 3. Remote sensing images and geographical information data used in this study.

Type	Dataset	Factor	Scale	Date	Data source
Land	Landsat-5 TM	N/A	30m	2010-12-27	earthexplorer.usgs.gov
	Landsat-8 OLI	N/A	30m	2015-03-12
	Landsat-8 OLI	N/A	30m	2020-02-22
Terrain	ASTER GDEM 2	DEM	30m	2015	www.gscloud.cn
Socioeconomic data	Population density	POP	1000m	2015	www.worldpop.org
	Gross domestic product	GDP	1000m	2015	www.resdc.cn
Facilities	City center POI	D-city	30m	2020	www.openstreetmap.org
	District POI	D-district
	Town center POI	D-town
Roads and rivers	Primary road	D-road
	Highway	D-highway
	Subway	D-subway
	Railway	D-railway
	River	D-river

3.2. Modeling procedure based on Futureland

We developed a new software package (i.e. Futureland) for modeling and projecting complex land-use change based on user-defined transformation rules and parallel computing. Figure 5 shows the overall process of Futureland, which has several functional advantages including: (1) adopting parallel computing based on data blocks, which greatly accelerates the CA modeling process and the speed is related to the CPU threads of the modeler’s computer; (2) accepting user-defined CA transformation rules developed using the R language; (3) providing format conversion, image classification and data normalization tools to prepare the input layers; (4) independently specifying the central cell, the neighborhood shape, and the weight of each neighborhood cell to facilitate the definition of special neighborhoods; (5) allowing users to define transformation costs for different types of land-use transformations based on socio-economic conditions and land-use policies; (6) allowing for edge enhancement of land-use patches to reinforce the boundaries of different land-use types. The software is freely available (www.researchgate.net/publication/356978377) for international users to build their models.

Figure 5. Futureland for complex land-use change simulation and future scenario projections.

Figure 6. The probability-of-occurrence maps for the five land-use types in Shanghai.

Figure 7. The simulated land-use patterns for 2015 and 2020 in comparison to the actual patterns.

Figure 8. Future land-use scenarios of Shanghai in 2025 to 2035 projected using Futureland.

4. Results

4.1. Probability-of-occurrence

We identified the best combination of factors using GLR and calculated the CA coefficients to generate the probability-of-occurrence maps. Table 4 shows the final factors and their coefficients for each land-use type, where most factors are related to urban land, indicating the complexity of urban sprawl. Of all eleven factors, POP and GDP are the most influential for urban sprawl; whereas, the distance from urban centers (D-city) could resist urban sprawl. In contrast, all factors are not statistically significant for special lands because airports and seaports experienced only minor changes. DEM is the most influential factor contributing to the increase of watershed area and greenspace, and the road network (D-road) is the most influential factor hindering the increase of greenspace. The river network (D-river) is typically adjacent to farmland, as evidenced by the high absolute coefficients.

Table 4. The coefficients of the factors tested for statistical significance.

Factor	Urban land	Water area	Greenspace	Farmland	Special land
D-city	2.08	−2.18	/	3.04	/
D-district	−0.87	2.57	/	/	/
D-town	−1.15	/	/	/	/
D-road	/	/	3.87	/	/
D-highway	−0.73	/	−1.21	1.95	/
D-subway	−1.51	/	/	/	/
D-railway	2.22	/	−0.66	/	/
D-river	−2.26	/	/	−4.19	/
DEM	1.96	−7.20	−4.95	6.06	/
POP	−3.91	/	−12.00	/	/
GDP	−4.21	−117.75	−39.76	/	/

Inputting the land-use maps (both 2010 and 2015) and the most relevant factors into Futureland, we generated the probability-of-occurrence of the five land-use types (Figure 6). The probability maps reflect the possibility of each cell transforming into the relevant land-use type. The probability of urban land (urban sprawl) is relatively concentrated, as most of the high values are outside the city center and around the Outer Ring Highway. The low average and maximum probabilities for the waters are probably because they have not changed much in the last decade. The high probability of greenspace is mainly concentrated in the northeast coastal and southwest areas. Although the average probability of farmland is also low, the high probability areas are mainly distributed outside the Suburban Ring Expressway. The probability of special land is zero throughout the study area because of their weak variation.

4.2. Simulation results and assessment

Prior to the model implementation, we defined a neighborhood weight matrix (Table 5) and a transformation cost matrix (Table 6) considering the variation in the proportion of each land-use type, then assigned the weights and costs based on their interactions. For example, in the greenspace row of Table 5, it is possible to convert from water area and farmland to greenspace, thus they are assigned slightly higher neighborhood weights (0.2), while the others are 0.1. For the conversion costs, we assumed that the greenspace type (the greenspace row in Table 6) could be easily converted into urban land (0.2), but difficult to convert into water area (0.9), farmland (0.9), and special land (1.0).

Table 5. The definition of the neighborhood weight matrix.

The central cell type	The neighboring cell type
	Urban land	Water area	Greenspace	Farmland	Special land
Urban land	1.0	0.1	0.2	0.5	0.1
Water area	0.1	1.0	0.1	0.1	0.1
Greenspace	0.1	0.2	1.0	0.2	0.1
Farmland	0.1	0.1	0.1	1.0	0.1
Special land	0.1	0.1	0.1	0.1	1.0

Table 6. The definition of the transformation cost matrix.

Transformation from	Transformation into
	Urban land	Water area	Greenspace	Farmland	Special land
Urban land	0.0	1.0	1.0	1.0	1.0
Water area	1.0	0.0	0.2	0.9	1.0
Greenspace	0.2	0.9	0.0	0.9	1.0
Farmland	0.1	0.8	0.2	0.0	1.0
Special land	1.0	1.0	1.0	1.0	0.0

In this study, we used the 2010 land-use pattern (Figure 7a) as the start map and the 2015 pattern (Figure 7c) as the end map to calibrate the Futureland model. Visual inspection reveals that the simulated and actual land-use patterns are overall similar, with rapid urban sprawl from the city center to the suburbs over the past decade. We selected an area north of the city center and along the Huangpu River and zoomed on them to show more details. Figure 7a1–e1 display that farmland was obviously occupied by urban land while urban greenspace was gradually decreasing. This indicates that Futureland has effectively captured the trend of urban sprawl, but there are still some mis-simulations, such as the greenspace in the upper right corner (c.f. Figure 7c1).

Table 7 shows the quantitative evaluation for the calibration (2010–2015) and validation (2015–2020) phases based on Futureland and the MAP COMPARISON KIT software (Visser and De Nijs 2006). For 2015, the overall accuracy was 86.6%, the Kappas were close to 0.8, and the aggregation cells were about 0.9. These indicate high agreement between the actual land-use maps and the simulated results. In comparison, the validation (2020) decreases slightly by about 2.5% in overall accuracy, about 1–2% in Kappas, and about 1% in aggregation cells. This is a ubiquitous phenomenon in CA-based land-use and urban sprawl simulations, and our simulations show less reduction in accuracy from calibration to validation. According to the producer’s and user’s accuracies, the prediction reliability of greenspace is lower than other types. Combined with a visual inspection, we found that greenspace is relatively dispersed in urban areas, making it difficult for Futureland to capture all changes accurately.

Table 7. The quantitative assessment of the simulated results for both calibration and validation.

Peroid	Type	Producer’s accuracy (%)	User’s accuracy (%)	Overall accuracy (%)	Kappa simulation	Fuzzy Kappa	Aggregated cells
2010–2015 (Calibration)	Urban land	85.0	82.6	0.866	0.796	0.801	0.894
	Water area	89.1	95.9
	Greenspace	77.1	77.0
	Farmland	89.1	89.0
	Special land	100.0	100.0
2015–2020 (Validation)	Urban land	85.8	84.3	0.845	0.780	0.783	0.886
	Water area	92.1	94.1
	Greenspace	68.2	72.2
	Farmland	84.1	84.0
	Special land	100.0	100.0

4.3. Future scenario prediction

Land demand in 2025, 2030, and 2035 was projected using the modified Markov chain (Table 8). Considering that the urban sprawl does not increase linearly and will decrease over time, the land demand was calibrated to a downward trend. As compared to 2020, the urban land in Shanghai is expected to increase by 3.6%, 7.0%, and 10.2% in 2025, 2030, and 2035, respectively. Table 8 shows that among the ecological lands, farmland will see the largest reduction, greenspace may likely decrease slowly, and water area will remain stable largely. The predicted land demand was input into Futureland as a global constraint for future scenario projections.

Table 8. Total land demand in 2025, 2030 and 2035 based on modified Markov chain for Shanghai.

Year/Period	Land-use type (%)
	Urban land	Water area	Greenspace	Farmland	Special land
2020	40.3	12.4	8.3	38.5	0.5
2025	43.9	12.2	7.3	36.1	0.5
2030	47.3	12.0	6.4	33.8	0.5
2035	50.5	11.6	5.8	31.6	0.5
2020–2025 Change	3.6	−0.2	−1.0	−2.4	0.0
2020–2030 Change	7.0	−0.4	−1.9	−4.7	0.0
2020–2035 Change	10.2	−0.8	−2.5	−6.9	0.0

Figure 8 shows the future land-use scenarios projected using the validated Futureland model and the spatial dynamics of future land-use change, showing primary land-use change occurring in the southern part. The projection shows that urban sprawl in Shanghai will likely occur around the urban fringes and mainly occupies farmland; however, this does not make much change in the overall land-use patterns over the next 15 years. Meanwhile, urban sprawl outside Shanghai’s Outer Ring Highway and around the Suburban Ring Highway will be more significant than in other areas of Shanghai.

Two typical areas were chosen to show the details of the land-use scenario over the next 15 years (Figure 8a1–c2), where the first is around Wujing Town in the southern part (enlarged area-1) and the second is in the southeastern-most part of the Dishui Lake area (enlarged area-2). In Wujing Town, the northwestern area along the Huangpu River is close to the edges of the existing built-up area, where the land type changes mainly from farmland to urban land. In Dishui Lake, more greenspace is converted to urban type, lying around the Dishui Lake, while farmland in the north and west will be gradually occupied and some greenspace near the sea will be converted to urban type.

4.4. Urban sprawl on ecological space

Based on the future scenarios, we analyzed urban sprawl on farmland and greenspace in the next 15 years, using 1 km concentric buffer belts starting from the Outer Ring Highway located at Buffer-14 (Figure 9). For 2020–2035, urban sprawl at the expense of farmland (change farmland to urban) is mainly found in the belt area with a width of 12 km (Region-I; Figure 9a–b), where the city center and the far suburbs do not change much from farmland to urban, then decreases further out to the administrative boundary of Shanghai. The changes from greenspace to urban have two peak belts that both have a width of 5 km (Regions II and III; Figure 9c,d). Of these, the region-III may be due to urban sprawl at the expense of coastal wetlands, forests, and riverbank greenspace. The amount of greenspace being occupied within the Outer Ring Highway is small compared to that outside the highway.

Figure 9. Buffer-based analysis of urban sprawl on farmland and greenspace in 2025, 2030, and 2035.

Figure 10 shows the 2035 urban land and greenspace distribution related to the four eco-space types defined in urban planning. The results indicate that the urban sprawl 2020–2035 mainly encroaches on Type-III and Type-IV eco-spaces; whereas, greenspace and water areas also gain from the Type-I and Type-II eco-spaces. Quantitative analysis shows that Type-I eco-space in urban planning is mainly occupied by greenspace and water area, Type-II is mainly by water area, Type-III is mainly by farmland and urban land, and Type-IV is mainly by urban land. These indicate that the Type-IV eco-space near the city center is at high risk of being occupied by urban land. Most of the Type-I and Type-II eco-spaces are preserved as greenspace and water area, and farmland remains the main land resource for new urban development in Shanghai. On Chongming Island, the Type-II eco-space that may likely be developed as urban land is located in the north-central part (upper circle in Figure 10a), which is in fact under the jurisdiction of Jiangsu rather than Shanghai.

Figure 10. Urban land and greenspace scenarios for 2035. (a) Urban land scenario associated with the four eco-space types in urban planning; (b) greenspace scenario associated with the four eco-space types; (c) the proportion of land-use types associated with the four eco-space types.

5. Discussion

In this study, we developed a new Futureland model by providing relevant software and applied it in Shanghai. The model was used to simulate multiple land-use changes and analyze urban sprawl on greenspace and farmland from 2010–2035. One of the Futureland’s innovations is the CA modeling approach we developed, namely, the GLR method by defining different factors for different land-use types, and the modified Markov chain based on the probability-of-occurrence. Research has shown that commonly applied neural networks and deep learning have limitations in explaining land-use mechanisms due to their black-box property (Karimi et al. 2019; Saputra and Lee 2019; Xing et al. 2020). Compared to these methods, our GLR approach yields parameters that can well explain the dynamic processes of land-use and exclude irrelevant candidate factors. In fact, high precision of simulation is not the only purpose of CA modeling; it is also important to capture land-use dynamics and predict different future scenarios (Li et al. 2017). Also, the modified Markov chain can achieve a more accurate land demand forecast, overcoming the shortcomings of standard Markov chains that exaggerate land demand.

In addition, our software employs parallel computing to increase the computational speed and processing capability in simulating large-scale land-use change. CA models typically require long computational times to predict land-use change over long time series and/or large areas. Some early studies proposed that the time-consuming computation of CA models could be handled by parallelization (Schweitzer et al. 2011) or by executing from the command line to increase the runtime speed (Mas et al. 2014). Our model uses parallel computing based on data slicing techniques, resulting in an exponential speedup over conventional computing, and the speed is proportional to the CPU threads of the user’s computer. For example, in the Shanghai case study, Futureland was executed in less than 20 s, using a computer with Intel i7 12 threads CPU and 16GB memory.

Land-use scenario in 2030 for Shanghai indicates that future areas of change are concentrated in the urban fringes and suburban areas, where urban areas will continue to expand, particularly in the southwest suburbs. This is attributed to that Shanghai’s central area has been largely occupied by buildings and impervious surfaces, while in the west and south there exists a large amount of undeveloped farmland. As a result, urban land expanded to Shanghai’s suburbs (especially to the south) to satisfy the land demand for population growth and industrialization (Zhou et al. 2020). We found that farmland encroached upon by the city is concentrated in the suburbs between the Outer Ring Highway and the Suburban Ring Highway, while the green belt in the central city is also at risk of loss. If local governments aim to achieve the greenspace identified in the urban plan in 2035, they should strengthen the regulation of new land development and ensure permanent farmland (Liu et al. 2021), thereby alleviating the conflict between farmland/greenspace and construction land. Also, urban greenspace and coastal wetlands are very important ecological spaces, and their protection is an important policy for Shanghai to achieve ecological balance and high-quality sustainable development.

Our study also has limitations, including the resolution of Landsat images and the fact that Landsat images are affected by topographic illumination, shadow effects, and sun angle (Kumari et al. 2020; Martín-Ortega et al. 2020). These may affect the research results but not the validity of our Futureland model. Since the resolution of Landsat images is only 30 meters, there are also some limitations for fine mapping and classification; therefore, the granularity of the data can also be improved if it can be fused with high-resolution optical images such as Sentinel and SPOT. In future work, we shall test the applicability of Futureland under higher resolution data by jointly using different images to produce land-use maps and future scenarios with high spatiotemporal resolution. Also, we shall define better model parameters by adding spatiotemporal constraints of planning policies and considering ancillary datasets.

6. Conclusions

Effective simulation of land-use change and future scenarios to assess urban sprawl on ecological lands is important for urban planning and management. We developed a new Futureland model that can effectively project future land-use change and provide timely warning of the risk of ecological space loss. Our contributions are three folds: (1) we combined CA with GLR for the first time for constructing transformation rules, which can better describe the mechanisms of land-use change and urban sprawl by modeling different land-use types using different factors in a single simulation experiment; (2) we proposed a modified Markov chain that considers the probability-of-occurrence to more accurately project the total future land demand; (3) Futureland was developed using GDAL and parallel computing, which significantly improve the implementation efficiency. The software also integrates unique features such as graphical neighborhood configuration, matrixed conversion costs and edge enhancement. Although it does not provide enough options in CA conversion rules, it is an open simulation software tool that allows modelers to write different scripts to develop their new models using the R language, with scripting rules provided in the manual.

The 2010–2020 Shanghai case study verified the applicability of the Futureland and displayed its ability to simulate complex land-use change, as confirmed by the high overall accuracy (86.6%) and Kappa simulation (79.6%). We predicted land-use scenarios 2020–2035 under the urban ecological planning, showing that the ecological space in Shanghai in the future is at risk of being occupied by urban sprawl, especially the urban greenspace and suburban farmland around the Outer Ring Highway. The results provide an important reference for policy-making on urban development in Shanghai and can be applied to long-term green space planning and conservation. Because Futureland uses parallel computing, it can be applied to simulate land-use change and its impacts on a global scale.

Data availability statement

The Futureland software we developed is available at www.researchgate.net/publication/356978377 and the data we used is available at figshare.com/s/af1c7c14170fc44c39eb.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This study was supported by the National Natural Science Foundation of China (42071371) and the National Key R&D Program of China (2018YFB0505400).