Forecasting Urban Land Use Change Based on Cellular Automata and the PLUS Model

: Nowadays, cities meet numerous sustainable development challenges in facing growing urban populations and expanding urban areas. The monitoring and simulation of land use and land-cover change have become essential tools for understanding and managing urbanization. This paper interprets and predicts the expansion of seven different land use types in the study area, using the PLUS model, which combines the Land use Expansion Analysis Strategy (LEAS) and the CA model, based on the multi-class random patch seed (CARS) model. By choosing a variety of driving factors, the PLUS model simulates urban expansion in the metropolitan area of Hangzhou. The accuracy of the simulation, manifested as the kappa coefﬁcient of urban land, increased to more than 84%, and the kappa coefﬁcient of other land use types was more than 90%. To a certain extent, the PLUS model used in this study solves the CA model’s deﬁciencies in conversion rule mining strategy and landscape dynamic change simulation strategy. The results show that various types of land use changes obtained using this method have a high degree of accuracy and can be used to simulate urban expansion, especially over short periods.


Introduction
Since 2014, more than half of the world's population now live in urban areas, and this proportion is expected to increase to 66 percent by 2050 (UN, 2018) [1]. Cities face numerous challenges in meeting the needs of their growing urban populations, and managing urban areas has become one of the most critical development challenges worldwide [2]. Monitoring and simulating land use and land cover (LULC) are important for understanding the trend of urbanization within an area [3], as they can indicate anthropogenic impacts, identify land use problems such as degradation and deforestation, and act as basic supporting information for land use plans aiming towards the sustainable development of a city [4].
Presently, three key areas advocate new research internationally on LULC: (1) The temporal and spatial changes in land use. The application of remote sensing and GIS technology allows scholars to catch up with land use changes throughout the world [5]. Evidence all over the world has illustrated the uneven distribution of land use changes in different areas, from European countries [6] to western Kenya [7], and eastern China [8]. (2) The environmental impact of land use. Scholars began to link land use changes with air pollution [9][10][11], temperature change [12,13], water-logging disasters [14], and so on.  The latitude and longitude span of the study area is 28°15′~30°22′ N, 117°0 E~121°16′. Located in the subtropical monsoon area, it contains a subtropical monso climate, with four distinct seasons and abundant rainfall [40]. At present, the total area the study area is 76,521 square kilometers, including Jiande City, Chun'an County, Tong County, Haining City, Tongxiang City, Anji County, Deqing County, She County, Zh City, and other county-level cities.
This paper selected the MCD12Q1 data of the study area from 2001 to 2019 as t basic urban land use data. It used these data to perform artificial-intelligence-based urb land use interpretation. For this purpose, the characteristics of the development of oth first-tier cities in China and the acquisition of remote-sensing data, were comparative analyzed. At the same time, the LULC data of the study area in 2005, 2010, 2015, and 20 were selected as the basic urban land use data. The urban expansion simulation was bas on cellular automata.

Data and Criteria
The data collected and used in this study included data graphs of driving factors su as GDP, subway, digital elevation model (EDM), elevation, highway, river, slope, pop lation, soil, school, and hospital, from 2005, 2010, 2015, and 2018. Moreover, four LUL maps and LULC maps with a resolution of 500 m × 500 m from 2001 to 2019 were use The driving-factor data maps were from the Institute of Remote Sensing and Digital Ear The latitude and longitude span of the study area is 28 • 15 ~30 • 22 N, 117 • 02 E~121 • 16 . Located in the subtropical monsoon area, it contains a subtropical monsoon climate, with four distinct seasons and abundant rainfall [40]. At present, the total area of the study area is 76,521 square kilometers, including Jiande City, Chun'an County, Tonglu County, Haining City, Tongxiang City, Anji County, Deqing County, She County, Zhuji City, and other county-level cities.
This paper selected the MCD12Q1 data of the study area from 2001 to 2019 as the basic urban land use data. It used these data to perform artificial-intelligence-based urban land use interpretation. For this purpose, the characteristics of the development of other first-tier cities in China and the acquisition of remote-sensing data, were comparatively analyzed. At the same time, the LULC data of the study area in 2005, 2010, 2015, and 2018 were selected as the basic urban land use data. The urban expansion simulation was based on cellular automata.

Data and Criteria
The data collected and used in this study included data graphs of driving factors such as GDP, subway, digital elevation model (EDM), elevation, highway, river, slope, population, soil, school, and hospital, from 2005, 2010, 2015, and 2018. Moreover, four LULC maps and LULC maps with a resolution of 500 m × 500 m from 2001 to 2019 were used. The driving-factor data maps were from the Institute of Remote Sensing and Digital Earth, the Chinese Academy of Sciences, and the LULC maps were from USGS. Specifics are shown in Table 1.

LULC Data Processing
The obtained LULC map was spliced in ENVI, then the spliced data were rendered using ArcMap10.7 (the main application used in ArcGIS). Finally, the data type was converted into the corresponding type that the model was able to handle through the PLUS model to complete the preprocessing of the data.

Introduction of the Patch-Generating Land Use Simulation (PLUS) Model
Step 1: Land use Expansion Analysis Strategy (LEAS) [40] (see Figure 2). This strategy extracts the expansion of various types of land between the two phases of land use change. Moreover, sampling from the added part uses a random forest algorithm to mine the factors of various types of land use expansion and driving forces. As a result, the development probabilities of various types of land and the contribution of driving factors to the expansion of various types of land during this period are obtained. This strategy integrates advantages of the existing TAS and PAS, avoids the analysis of conversion types that increase exponentially with the number of categories, retains the model's ability to analyze the mechanism of land use change within a certain period, and has better interpretability.
Step 2: The CA model is based on multi-class random patch seed (CARS) [40] (see Figure 3). Combining the random seed generation and threshold-decreasing mechanism, the PLUS model can dynamically simulate the automatic generation of plaques in time and space, under the constraints of development probability. analyze the mechanism of land use change within a certain period, and has better in pretability.

Figure 2. Land use Expansion Analysis Strategy (LEAS).
Step 2: The CA model is based on multi-class random patch seed (CARS) [40] (  Figure 3). Combining the random seed generation and threshold-decreasing mechani the PLUS model can dynamically simulate the automatic generation of plaques in ti and space, under the constraints of development probability.
The basic formula for the total probability of each land use type is as follows [40] , where , =1 represents the growth probability of type k land use at i; represents impact of future land use demand of type k, which depends on the gap between the c rent amount of land at t and the land use target demand k; and , represents the dom effect of unit i, which is the following neighbor of the coverage ratio of the land use co ponent of k within the domain [40].
Among them, ( −1 = ) represents the last iteration. In the window of n× n, represents different land use types having different neighborhood effects. The defa value of is 1, but it can be freely defined. The adaptation method is as follows [41] = where the sum is the difference between the number of land use types k, in iterati −1 and −2 and the future demand. Finally, according to the overall probability of Land 2022, 11, x FOR PEER REVIEW 6 of 16 land use types, a wheel is constructed to select the land use status in the next iteration [42]. At the same time, in order to simulate the patch evolution of multiple land use types, a multi-type random patch-seeding mechanism based on threshold drop is adopted, which is realized through a calculation process. , Among them, r is a random value from 0 to 1, and is the threshold for generating a new land use patch for k-type land use, which the model user determines in order to control the organic growth and spontaneous growth of multiple land use types. If a new land use type wins a round of competition, a lowered threshold τ is used to evaluate the selected candidate land use types, as follows: Among them, the step length is the step length of the PLUS model, to approximate the land use demand; δ is the attenuation factor of the decreasing threshold τ, ranging from 0 to 1, which the expert sets; r1 is the random value of the normal distribution, the mean value is 1, the range is 0~2; l is the number of attenuation steps.
, is the transition matrix, which defines whether the land use type K-type is allowed to be converted to C-type. Through this descending threshold, new land use patches will spontaneously grow and develop freely under the constraints of growth probability [43].

Geomorphology and Geology
The terrain of Zhejiang is high in the southwest and slopes from the southwest to the northeast. The northeast is a low and flat alluvial plain, hills and coastal plains dominate the east, the middle is dominated by hills and basins, and the southwest is dominated by The basic formula for the total probability of each land use type is as follows [40]: where P d=1 i,k represents the growth probability of type k land use at i; D t k represents the impact of future land use demand of type k, which depends on the gap between the current Land 2022, 11, 652 6 of 16 amount of land at t and the land use target demand k; and Ω t i,k represents the domain effect of unit i, which is the following neighbor of the coverage ratio of the land use component of k within the domain [40].
Among them, con C t−1 i = k represents the last iteration. In the window of n × n, w k represents different land use types having different neighborhood effects. The default value of w k is 1, but it can be freely defined. The adaptation method is as follows [41]: where the sum is the difference between the number of land use types k, in iterations G t−1 k and G t−2 k and the future demand. Finally, according to the overall probability of all land use types, a wheel is constructed to select the land use status in the next iteration [42].
At the same time, in order to simulate the patch evolution of multiple land use types, a multi-type random patch-seeding mechanism based on threshold drop is adopted, which is realized through a calculation process.
Among them, r is a random value from 0 to 1, and µ k is the threshold for generating a new land use patch for k-type land use, which the model user determines in order to control the organic growth and spontaneous growth of multiple land use types. If a new land use type wins a round of competition, a lowered threshold τ is used to evaluate the selected candidate land use types, as follows: Among them, the step length is the step length of the PLUS model, to approximate the land use demand; δ is the attenuation factor of the decreasing threshold τ, ranging from 0 to 1, which the expert sets; r1 is the random value of the normal distribution, the mean value is 1, the range is 0~2; l is the number of attenuation steps. TM k,c is the transition matrix, which defines whether the land use type K-type is allowed to be converted to C-type. Through this descending threshold, new land use patches will spontaneously grow and develop freely under the constraints of growth probability [43].

Geomorphology and Geology
The terrain of Zhejiang is high in the southwest and slopes from the southwest to the northeast. The northeast is a low and flat alluvial plain, hills and coastal plains dominate the east, the middle is dominated by hills and basins, and the southwest is dominated by mountains and hills. It can be roughly divided into six terrain areas: the North Zhejiang Plain, the hills and basins in the southwest of Zhejiang, the coastal plains, and the coastal islands in the southeast of Zhejiang. Alluvial plains with dense water networks in northern Zhejiang, coastal plains and hills in eastern Zhejiang, basins in central Zhejiang, hills and mountains in southwestern Zhejiang, and island landforms in Zhoushan all include various types of landforms. Mountains with sizes over a kilometer mostly surround the southwest. Huangmaojian, located in Longquan, has an elevation of 1929 m, and it is the highest peak in Zhejiang Province. Again, plains, hills, and basins dominate the terrain.
By combining the results of land use classification in the study area, the main built-up areas of the study area are concentrated in the eastern region dominated by plains. In contrast, the cultivated land is concentrated in the southern and northern regions and dominated by low mountains and hills, and woodlands are widely distributed throughout the entire research area. For this reason, the expansion of land use in the study area is mainly concentrated within the eastern, southern, and northern plains and the middle-low mountain areas.

Climatic Settings
Due to the subtropical monsoon climate, the climate characteristics of the study area are high in temperature: the annual average temperature is 15~23 • C. Taking July as the highest temperature, there is much precipitation, with the annual average precipitation being 1200~2000 mm. The rainfall distribution during the year is relatively even, with there being no drought and less rain areas. There is much sunshine, with the annual average sunshine period being more than 1800 h. Sunshine is most abundant from June to December. The wind speed is low: the average wind speed in urban areas is not high, with an average speed of only 1.9 m/s. Except for the wind direction, which directly affects the layout of the industrial land and residential land, other climatic conditions indirectly influence land use distribution in the study area.

Economic Development
Urban expansion and economic development present a cyclical relationship, and most of the expansion of construction land is influenced by economic development. In the Hangzhou metropolitan area, four rounds of urban expansion have been observed. From 1984 to 1985, the first round of rapid economic development since the economic reform caused the first instance of out-of-control land use in most parts of Hangzhou and the surrounding cities. In the subsequent years from 1988 to 1989, from 1993 to 1994, and after 2000, there were three rounds of urban land expansion under the influence of excessive economic growth.

Industrialization Policies
In China, with mixed-use urban planning, the issue of land transfer permits, and government investment in infrastructures and public amenities occurring, governments in Chinese cities have proven their robust control in defining the development of a city [44]. In Hangzhou and the surrounding cities, the local governments have introduced measures for the macro-control of urban expansion, from limiting the supply of construction land, to land planning and design. Although most measures have failed to address the increasing need for construction land during rapid economic development, these policies and measures have defined the direction of legal urban expansion, and they are decisive for certain issues, such as "return farmland to forest".

Land Use Interpretation
Based on GIS and LULC raster data, processing and analysis were performed to describe land-cover types. Generally speaking, the data set contained 17 main land-cover types. According to the International Geosphere-Biosphere Program (IGBP), this included 11 natural vegetation and 3 land development land types, mosaic land types, and 3 nonvegetable land types. According to the analysis of the study area, it was found that the land-cover types in the study area were mainly limited to 12 to 14 types, as shown in  Based on GIS and LULC raster data, processing and analysis were performed to describe land-cover types. Generally speaking, the data set contained 17 main land-cover types. According to the International Geosphere-Biosphere Program (IGBP), this included 11 natural vegetation and 3 land development land types, mosaic land types, and 3 nonvegetable land types. According to the analysis of the study area, it was found that the land-cover types in the study area were mainly limited to 12 to 14 types, as shown in Figure 4. The specific area changes of each type are shown in Figure 5. The results showed a tremendous increase in urban built-up land from 2001 to 2019, which corresponded to a decrease in the area of farmland and other land types. At the same time, the forest land also increased significantly.

Urban Expansion Simulation Based on PLUS Model
By inputting the LULC map in 2005 and 2010, 2010 and 2015, and 2015 and 2018, and by extracting the expansion of various types of land between the two phases of land use changes in different years in the study area, this study extracted the land use expansion for Zhejiang for different periods, as shown in Figure 6.

Urban Expansion Simulation Based on PLUS Model
By  We sampled the additional part obtained from Step 1 and used the random forest algorithm. The sampling methods were uniform sampling and random sampling, and the sampling rate was set to 0.01 to excavate various land uses and driving forces one by one. Eight driving factors were selected for this excavation, including GDP, subway, EDM, elevation, highway, river, slope, population, soil, school, and hospital.
The contribution of all these driving factors to the expansion of various land use types is outputted, as shown in Figures 7-10. The higher the brightness, the greater the contribution rate. As shown, the driving factors had the highest contribution rates toward grassland, forest, farmland, and urban land, and the contribution rates to forest and urban land kept rising. We sampled the additional part obtained from Step 1 and used the random forest algorithm. The sampling methods were uniform sampling and random sampling, and the sampling rate was set to 0.01 to excavate various land uses and driving forces one by one. Eight driving factors were selected for this excavation, including GDP, subway, EDM, elevation, highway, river, slope, population, soil, school, and hospital.
The contribution of all these driving factors to the expansion of various land use types is outputted, as shown in Figures 7-10. The higher the brightness, the greater the contribution rate. As shown, the driving factors had the highest contribution rates toward grassland, forest, farmland, and urban land, and the contribution rates to forest and urban land kept rising.

Verification of PLUS Model
Combining random seed generation and the threshold-decreasing mechanism, the PLUS model in this study dynamically simulated the automatic generation of plaques in time and space, under the constraints of development probability. The simulation was implemented for 2005-2010, 2010-2015, and 2015-2018 in turn, and the results are shown in Figure 10.
According to the above results, the simulation accuracy was verified, and the results are shown in Table 2. According to the advanced comparison between the PLUS model simulation results and the real land use type data, the results are shown in Table 3.

Verification of PLUS Model
Combining random seed generation and the threshold-decreasing mechanism, the PLUS model in this study dynamically simulated the automatic generation of plaques in time and space, under the constraints of development probability. The simulation was implemented for 2005-2010, 2010-2015, and 2015-2018 in turn, and the results are shown in Figure 10.
According to the above results, the simulation accuracy was verified, and the results are shown in Table 2. According to the advanced comparison between the PLUS model simulation results and the real land use type data, the results are shown in Table 3. According to the above data, the results of the three simulations reached over 80%, the accuracy of centralized urban land use reached over 84%, and the kappa coefficient reached over 91%. Therefore, the accuracy of the model was high.

Application of PLUS Model for Prediction
According to the data, the study area has, in recent years, reduced mainly in farmland and increased in forest area, as a result of the protection and development of the ecological environment. At the same time, the area of urban land in the study area has been increasing year by year, and the level of urbanization has increased significantly. Therefore, some environmental issues are echoed in the strengthening of ecological civilization within the study area. Based on the above results from the simulation of land use types in 2024, 2027, and 2030, these results are shown in Figure 11, and the specific values are shown in Table 4. According to the simulation of the future years, 2024, 2027, and 2030, it can be seen that the area of urban land in the study area will continue to increase significantly, which may be related to the rapid economic development of the study area in the future. The increase in land use in these cities was mainly achieved by reducing the grassland area, while the changes in area for other types of land use were relatively small.

Conclusions and Discussion
Scholars have long pursued models for better accuracy in predicting land use change. For example, Salem et al. used the logistic regression model to analyze Delhi over the past 30 years, based on various driving factors, but the maximum regression coefficient only reached 0.8646 [45]. Gantumur et al. used CA and an artificial neural network (ANN) to study the urban growth rate of Ulan Bator, but the kappa coefficient only reached barely higher than 70% [46]. The kappa coefficient was raised to 80% and 78% when Pijanowski et al. established a UEM model with GIS, ANN, and RS. ROC curve and kappa statistics were used to conduct a calibration for the study of land use changes in the Tehran metropolitan area [47]. Since then, it seems extremely difficult to increase the accuracy.
This paper interprets and predicts the expansion of seven land use types in the study area from 2001 to 2019, based on the PLUS model, which combines LEAS and CARS. By choosing a variety of driving factors, the PLUS model simulates urban expansion in the metropolitan area of Hangzhou. The results show that the accuracy of the simulation, manifested as the kappa coefficient of urban land, has increased to more than 84%, and the kappa coefficient of other land use types is more than 90%. Compared with other models, such as the cellular automata and Markov chain, this model has better accuracy and can better simulate the future land use changes of the city. Cellular automata and Markov chain are used to study the driving mechanisms of land use change. The results show that economy and policy lead to urban change, compared with population and natural factors. This study further simulates and predicts land use over the next eight years and finds that except for the fast growth (more than 10.6%) of urban land, grassland and evergreen forest would also grow, at the cost of reducing other types of land use.
To a certain extent, the PLUS model used in this paper solves the CA model's deficiencies in the conversion rule-mining strategy and landscape dynamic change simulation strategy. A new analysis strategy is applied to better uncover the inducements of various land use changes by containing a new multi-type seed growth mechanism that can better simulate multi-type land use patch-level changes. At the same time, coupled with multiobjective optimization algorithms, the simulation results can better support planning policies to achieve sustainable development.

Conclusions and Discussion
Scholars have long pursued models for better accuracy in predicting land use change. For example, Salem et al. used the logistic regression model to analyze Delhi over the past 30 years, based on various driving factors, but the maximum regression coefficient only reached 0.8646 [45]. Gantumur et al. used CA and an artificial neural network (ANN) to study the urban growth rate of Ulan Bator, but the kappa coefficient only reached barely higher than 70% [46]. The kappa coefficient was raised to 80% and 78% when Pijanowski et al. established a UEM model with GIS, ANN, and RS. ROC curve and kappa statistics were used to conduct a calibration for the study of land use changes in the Tehran metropolitan area [47]. Since then, it seems extremely difficult to increase the accuracy. This paper interprets and predicts the expansion of seven land use types in the study area from 2001 to 2019, based on the PLUS model, which combines LEAS and CARS. By choosing a variety of driving factors, the PLUS model simulates urban expansion in the metropolitan area of Hangzhou. The results show that the accuracy of the simulation, manifested as the kappa coefficient of urban land, has increased to more than 84%, and the kappa coefficient of other land use types is more than 90%. Compared with other models, such as the cellular automata and Markov chain, this model has better accuracy and can better simulate the future land use changes of the city. Cellular automata and Markov chain are used to study the driving mechanisms of land use change. The results show that economy and policy lead to urban change, compared with population and natural factors. This study further simulates and predicts land use over the next eight years and finds that except for the fast growth (more than 10.6%) of urban land, grassland and evergreen forest would also grow, at the cost of reducing other types of land use.
To a certain extent, the PLUS model used in this paper solves the CA model's deficiencies in the conversion rule-mining strategy and landscape dynamic change simulation strategy. A new analysis strategy is applied to better uncover the inducements of various land use changes by containing a new multi-type seed growth mechanism that can better simulate multi-type land use patch-level changes. At the same time, coupled with multi-objective optimization algorithms, the simulation results can better support planning policies to achieve sustainable development.
However, this study lacks consideration towards many restrictive factors when considering driving factors, such as topography, which would affect the final accuracy to a certain extent. In the future, there should be more consideration toward restrictive factors. Furthermore, during the research process, it was found that the PLUS model could not effectively simulate long-term land use change. Therefore, a new method for the prediction of long-term land use is still needed, and neural networks or deep-learning methods might be helpful in this regard.