1. Introduction
Land-use activities, activities converting natural landscapes for human use or changing management practices on human-dominated lands, have transformed the planet’s land surface [
1,
2]. At the global scale, land is becoming a scarce resource, which leads to the fierce competition and conflict between various groups of people [
3]. The world has experienced a large-scale cropland expansion and intensification to meet food demands for the growing population in the last centuries [
4]. Also, the conversion of Earth’s land surface to urban uses drives the loss of farmland, affects local climate, fragments habitats, and threatens biodiversity [
5]. Meanwhile, afforestation/reforestation projects have been implemented to mitigate climate change and sustain land systems [
6]. For example, 16 sustainability programs have improved the sustainability of land systems in China [
7]. Land use, as a force of global importance, significantly influences the atmosphere, pedosphere, hydrosphere, and biosphere of the earth [
8], and its change results in climatic, biological, and socio-political forces [
9,
10]. With different demands for economic growth, population growth, and ecological restoration and protection, modeling and projecting land-use change processes is necessary to support the design and implementation of land-use planning and policy.
Previous studies have elaborately reviewed the multiple available models of simulating land-use change, which were developed and applied to provide a platform for both emulating mechanisms of land-change processes by the computer encodes and making projections of future land-cover and land-use patterns [
10,
11,
12,
13]. Models have ranged from those using pattern-based methods to structural or process-based methods [
14,
15]. Brown, Verburg, Pontius and Lange [
14] identified five key types of modeling approaches according to the model input, output, and land use conversion rule design. The models quantify the complex relationships between physical factors, human activities, and land-use dynamics, and simulate the land system process under different strategies [
14,
15]. Under the combined influences of these factors with different spatial patterns, the land use change would constitute an obviously spatial-explicit discrimination. In this way, the land use simulation models, setting the uniform rules of land use change in the model design, tend to lose the detailed spatial characteristics of land use conversion in the previous studies. Therefore, it is necessary to incorporate relevant regional difference information for simulating the land use dynamics.
To better project the future land-use change, precisely quantifying the complicated impact of the driving factors is necessary. However, land-use changes differ with driving factors by region [
12,
16]. Global scale assessments may therefore conflict with the findings of micro- or meso-scale data sets [
9]. In this case, the demands for land resources would be different and conflicted at different stakeholders and scales [
17]. To take the scale dependency into account, a separate simulation method was used for different world regions in the global land-use projection [
16,
18]. Within regional research of land-use change, few studies have focused on this issue, which leads to scale biases and simulation uncertainties. In particular, mountain ecosystems, characterized by their topographic and climatic variety, presented shifts in land-use change across their elevation gradient [
19]. Meteorological variables, soil properties, and vegetative functions change with the elevation in the mountainous area [
20,
21], and further the human activities and demands on the land use are influenced the elevation [
19,
22]. Previous study only used the elevation as one of driving forces, but limited studies measure the difference of land use dynamics within different elevation gradients. To quantify the regional differences of land-use changes [
23], it call for an insight into spatially varying process of land-use change in the model design. Using the elevation gradients as zoning constraints, demands for the land and land use functional orientations are spatial-differentially depicted, and further regional relationships between driving factors and land use are quantified by the stratified strategy in this study. This can help better understand the land system dynamics and improve the simulation accuracy.
Stratifications provide a useful approach to simplifying heterogeneity, which divides environmental gradients into convenient units and then to uses these as sub-regions with relatively consistent characteristics [
24]. Traditional stratification is subjectively based on expert knowledge [
25] or a statistical clustering of environmental variables [
26,
27]. Knowledge of stratification approaches for the projection of land-use change is limited and unclear. Object-based segmentation methods provide an innovative way for image analysis to develop observation units with a set of similar pixels [
28,
29], which would be potentially useful in a stratified strategy for simulating land system dynamics. Thus, the objective of this study was to (1) develop an enhanced land-use change simulation model with an elevation-based stratification strategy; and (2) apply the new model to simulate land-use change and investigate the model performance. The model contributes to the existing body of literature on getting a better approximation for spatial land use change pattern and an improvement of model simulation performance. This would enhance the understanding of land use dynamics process and provide a more reliable output to support for the decision making.
4. Discussion
The novelty of our established SLUCS model is that it uses the strategy of elevation-based stratification to get a better approximation for the land use dynamics process. Elevation gradients reflects the differences in meteorological variables, soil properties, and vegetative functions [
20,
21], and further influences the functional orientation of land use, as well as the activities and human demands on the land use [
19,
22]. Along the elevation-based stratifications, the model used the quantitative method to better quantify stratified land-use demands and the stratified relationship of driving factors to land use to improve model performance. Not unlike the method of subjective division used in previous studies [
16,
18], the elevation-based stratification module of the SLUCS model developed a quantitative method of dividing the study area into multiple stratifications. Based on the proposed ASD
intra, LV
inter, and OSS in this study, the intra-segment homogeneity and the inter-segment heterogeneity of the land-use characteristics within different elevation gradients were comprehensively taken into consideration to perform the segmentation and generate stratifications at the optimal segmentation size.
The elevation-based strategy produced stratifications along the elevation gradients relating to different spatial patterns of physical and social-economic factors, which can better present the spatial land use characteristics and simulate the land use change. Stratifications can ensure that region-level results from macro-scale models are spatially allocated [
18]. Comparing to the administrative unit with the fixed boundaries, the SLUCS model flexibly divided the entire study area to multiple stratifications based on the elevation and land use characteristics. The municipal boundaries consisting of sixteen units (
Figure S2) was used to calculate the inter-segment heterogeneity of land use in 2000 (LV
inter). LV
inter of municipal boundaries is 0.048, which is nearly less than one third of that for elevation-based stratifications of our model (0.074). It indicated a larger spatial variation of land use characteristics for the elevation-based stratifications than that of municipal boundaries. Within different stratifications from the elevation-based stratification module, the land-use characteristics and historical changes were different from each other, indicating the spatial difference of land utilization orientation (
Figure 7). For example, the two stratifications with the high elevations of 1732 m and 1714 m located in the northwest of the study area would have a higher increased rate of forests of 3.02% and 2.75%, which are much higher than the average value of 0.60% in the entire study area. In contrast, the stratification with the lowest average elevation of 144 m located in the southwest had the highest built-up land area proportion of 6.37% in 2000 and presented the highest increase area of 486 km
2 from 2000 to 2015. Thus, the non-spatial land-use demand module would better present the spatial differences in land-use actives within the elevation gradients, forecast stratified land-use demand, and set zoning constraints for simulating land use [
62]. Better quantifying the relationship of the driving factors with land use would significantly help improve the simulation accuracy of the land-use change [
46,
63]. Using the dummy variables to indicate different stratifications in the logistic regression model, the stratified analysis could reveal the regional differences in the impacts of the driving factors on land use [
12]. The stratified suitability estimation can provide new spatial information regarding land-use change to improve model performance.
The application of the SLUCS model in the study area validated its effectiveness to project land-use change (
Figure 8 and
Table 6). Compared to the traditional method, without the stratifications, the stratification strategy of the SLUCS model exhibited superior spatial consistency with the reference land use (from the visualized interpretation) and higher accuracy assessment (from the statistical metrics). Previous studies validated the idea that the traditional method could account for land-use persistence simulated the area without land use change in the simulation period well [
18,
35,
58]. As only 2.4% of the reported locations changed from 2000 to 2015, both presented a high kappa coefficient. However, the high K
Simulation indicated the much better performance in simulated areas experiencing land-use conversion when the stratified method was used than the traditional method [
58]. In particular, the stratified method resulted in a better spatially visualized fit (
Figure 8) with the reference and a higher Modified Lee and Sallee metric (
Table 6) for simulating the built-up land expansion. The urban growth was constrained by stricter planning and management efforts in China within different administrative units [
64]. Thus, there was a clear spatial gradient in the urbanization ratio and the urbanization gaps among different regions persisted [
65]. The traditional method was not able to fully simulate the regional differences in the urbanization process and projected a spatial dispersion distribution of built-up land. In contrast, the stratified method had the advantage of a stratified consideration of the regional urbanization process and built-up land expansion. The improvement of SLUCS model in projecting land use change would provide a more reliable output to support for the decision making.
To a certain degree, our model had limitations and did not perform well in local areas of land-use conversion. The model showed that the simulation accuracies of the built-up land were still lower than other land-use types and presented a spatial disparity within the referenced land use in tiny conversion areas. Comparing to ROC values of around 0.85 in previous studies of predicting urban growth [
66,
67], the ROC of 0.78 is relative lower in the logistic model for built-up land. This indicated that the model cannot fully explain the spatial variability of built-up land and would further influence the local performance of the model simulation. Land use change is determined by the interaction of driving factors, where different spatial resolutions of data would affect the simulated results and may cause uncertainties. The aggregated data at coarse scales may obscure the local variability, but can show patterns invisible at fine scales, and vice versa [
68]. Different biophysical and socio-economic processes influence the land use at their own dominant scale [
69], but there is not an absolutely optimal scale for the system [
70]. The data source determined the basic analysis pixel size [
68]. As the spatial resolution in most of driving factors is 1000 × 1000 m, other factors were resampled at the same pixel size. Incorporation of the fine-scale socio-economic process is difficult, due to the high data requirements at this scale. Quantification of the land-use planning and policy, significantly influencing the built-up land expansion, was difficult even with detailed spatial information [
71]. Thus, this study could only characterize the different macro-scale land-use demands driven by the land-use policy and planning within different stratifications. Also, there were difficulties in determining a straightforward quantitative approach for human demands and activities on land [
72,
73]. The use of several proxies for them in this study (
Table 1), which have proven to be acceptable in recent studies, would influence the local performance of the model. As only seven independent variables at the significant level of 0.05 were used to build the logistic model, the spatial-explicitly detailed socio-economic factors would help improve the model performance in the future study.
In addition, this study hypothesized that the optimal segmentation was identified as the one with the highest OSS, and seventeen dummy variables were added into the regression model. With more independent variables, there may be statistical uncertainty for the estimation with additional parameters [
74]. The application of SLUCS model in the other study areas could be tested to produce a reasonable number of stratifications and validate the performance of the stratified strategy, where the number is not too large or small. Also, this study only used the elevation data to perform the segmentation to examine the effect of the elevation gradient on the land use. Other features could be selected and tested to generate the stratification with multiple environmental variables [
26,
27]. Considering the data availability, only temporal scales of partial driving factors with obviously changes from 2000 to 2015 were updated to simulate the land use in 2030 (
Figure 9). Other factors with gradual and minor changes can be updated in the future study to better project the land use.
The SLUCS model can project the spatial pattern of future land use change under different land use scenarios. With multiple demands in terms of economic development, food security, and ecological protection, frequent conflict and competition occurs between multiple land-use types [
75]. Under the three scenarios, the outputs of the SLUCS model demonstrated the potential land-use change and competition with different land-use priorities and strategies. Consistent for most regions in China [
76], the model indicated an increasing threat to food security in the future with a continuing decline in farmlands in all the scenarios; this crisis will raise the priority on socio-economic growth and ecological protection in the planning and protect scenarios. The planning scenario presented a more scattered pattern of built-up land, with the largest increased area of built-up land under the three scenarios. This scenario implied that the limited flat land resources in the mountainous areas would restrict the expansion of built-up land near cities and part of the future built-up land will have to be dispersed. Over-urbanization in the mountains may easily damage the ecosystem [
77] and calls for the scientific planning of urbanization in the study area [
78]. The protect scenario showed that a significant increase in forests could be realized at the huge expense of low increase rate of built-up land and high decrease rate of farmlands and grasslands. Investments into ecological protection and willingness to make socio-economic sacrifices should be taken into consideration during decision making [
79,
80].
Limited land resources suitable for living and production in the karst mountain region would accelerate the different demands for land-use activities [
31]. The tradeoffs among the different land-use types were suggested to support land-use planning and management. Our model visualized the land system dynamic process and projected the land-use change trajectory at the spatial dimension. With different target land-use scenarios, areas of land-use conversion can be identified using our model, which can then be used to evaluate their potential impacts, such as food production, soil conservation, water conservation, and biodiversity protection [
2,
9,
81,
82]. The conversion areas with high values of land-use should be the focus of land-use planning. For example, strict farmland protection policies should be implemented in the high production of farmlands occupied by built-up land, as per the model [
83]. Comparing the three scenarios, areas of conversion from other land-use types to forests could be the priority of afforestation, because of the high suitability for forest planting. To coordinate multiple conflicting land-use types, information regarding different land-use change trajectories could provide potential options for sustainable land use.
5. Conclusions
Better understanding and projecting of land system dynamics will support the design and implementation of land-use planning and policy. To quantify the regional differences in land-use changes within the elevation gradient, this study built a novel SLUCS model using the elevation-based stratification strategy. Along the elevation gradients, the stratifications can make zoning constraints for the simulation with different land utilization orientations and quantify the stratified relationship of the driving factors with land uses, which help better approximate the land use dynamics process. The model included four modules: Elevation-based stratification, non-spatial land-use demand, stratified suitability estimation, and spatial allocation of land use. The first module developed a quantitative method of generating stratifications at the optimal segmentation size using ASDintra, LVinter, and OSS, which considered the intra-segment homogeneity and inter-segment heterogeneity of land-use characteristics. The second module presented the regional demands for land use and made zoning constraints for simulating land use. The third module utilized the stratified logistic regression model, with dummy variables to indicate different stratifications in order to reveal the regional differences in the relationship between driving factors and land use. Taking the Guizhou and Guangxi Karst Mountainous Region as the case study area, the model was executed to validate its performance. The effectiveness of the SLUCS model for projecting land-use change was validated. Compared to the traditional method without stratifications, the stratification strategy demonstrated an improved model performance than the traditional method, including a better spatial consistency with the reference and a higher accuracy assessment. Particularly, a much better model performance of simulating land use conversion areas was seen in the SLUCS model, where the KSimulation was higher (0.52) than that of the traditional model (0.19). Further, a better built-up land expansion simulation accuracy was seen, based on the Modified Lee and Sallee metric. The historic-condition, planning and protect scenarios from 2015 to 2030 were designed with different land-use priorities and management strategies. Under the three scenarios, the outputs of the SLUCS model projected the potential land-use changes and visualized the competition between different land uses. Consistent with the historical change trend from 2000 to 2015, the three scenarios presented the same change trends, but with different magnitudes. The HISTROY scenario forecasted the historical trend of land use according to the change in the past 15 years. The planning scenario presented the highest increase in built-up land among three scenarios from conversions of a large area of paddy field, dry land, and grassland. The protect scenario presented the highest increase in forests at the expense of low increase in built-up land and high decrease in farmland and grassland. Various management and tradeoff strategies for the multiple land-use types were suggested, based on the different scenarios. The results validated the effectiveness of the SLUCS model and the significance of supporting sustainable land use. The limitations and possible improvements of the SLUCS model mentioned in this paper can be tested in future research, including the data acquisition of spatial-explicitly detailed socio-economic factors for better explaining land use patterns and improvement of the stratification producing process with the optimal segmentation and multiple environmental variables.