3.1. Spatiotemporal Changes in Ecological Sustainability during 2003–2022
Based on the 20-year average ecological sustainability values calculated at the raster scale, the land in the NFYM was classified into five levels of ecological sustainability—great, good, moderate, low, and poor—using the natural breaks method (
Table 6).
The spatial distribution of ecological sustainability in the NFYM is illustrated in
Figure 4a, revealing a general decreasing trend from east to west. The southern and eastern counties exhibit great ecological sustainability (i.e., WC, ZL, DL, and the eastern part of AB). The transition area from the east to the central region (i.e., GY, the southern part of DMJ, the southern part of SZW, and the western part of AB) shows moderate to good sustainability. The central and northern regions (i.e., ER, SR, SL, the northern part of SZW, and the northern part of DMJ) and the western part of UM exhibit low ecological sustainability. The western part of UB shows poor ecological sustainability. Overall, areas with great, good, moderate, low, and poor sustainability accounted for 13.37%, 23.96%, 45.65%, 12.10%, and 4.92%, respectively, of the NFYM in 2022.
Ecological sustainability in the NFYM experienced an initial decline followed by a subsequent increase from 2003 to 2022. Changes in average ecological sustainability were calculated for each county (
Figure 4b). Specifically, all counties exhibited a fluctuating downward trend in ecological sustainability from 2003 to 2014, with decreases ranging from 11.49% to 29.68%. Fourteen of the eighteen counties experienced decreases of more than 15%. However, ecological sustainability improved from 2014 to 2022. Every county experienced an increase in ecological sustainability, ranging from 11.68% to 39.65%. Fifteen of the eighteen counties experienced increases of more than 15%. Overall, by 2022, ecological sustainability in the NFYM had returned to levels close to those of 2003.
Rapid population growth and relatively low productivity levels have led to predatory exploitation and inefficient use of water resources since 2003. The collection of fuelwood, overgrazing, and excessive cultivation have severely damaged existing forest and grassland ecosystems, resulting in a significant decline in ecological sustainability between 2003 and 2014. Since around 2010, the government has placed greater emphasis on ecological environment construction and protection, advancing projects such as the Three-North Shelterbelt Program and the Beijing-Tianjin Sandstorm Source Control Project. The NFYM has implemented ecological projects such as artificial reforestation and grassland restoration, and the restoration and management of degraded grasslands. Additionally, the government formulated water-saving plans, including regulating water demand based on availability, effectively preventing the overuse of water resources, and improving the water environment. Furthermore, the Sandstorm Source Control Project and stricter factory emission controls have alleviated air pollution issues. Consequently, ecological sustainability recovered between 2014 and 2022. However, average ecological sustainability values in some counties dominated by desert grasslands remained low in 2022 (e.g., ER 0.41, SR 0.39, DMJ 0.40, UB 0.30, UM 0.39). Restoring fragile ecosystems requires continuous effort, and any relaxation can result in setbacks. Therefore, projecting future ecological sustainability changes in the NFYM based on SSPs is important to help address future risks and enhance ecosystem resilience in the NFYM.
3.2. Selection of Optimal Future Climate Mode and Projecting Human Activities
Ecological sustainability is influenced by a multitude of factors pertaining to climate conditions and human activities. Variations in temperature and precipitation directly influence the hydrothermal conditions of ecosystems. Population size is a significant factor impacting ecosystem stress. Agriculture and livestock husbandry are the primary industries upon which people in the arid farming–pastoral zone rely. Fluctuations in population and livestock numbers result in changes in the demand for natural resources, with livestock numbers being a crucial factor impacting the sustainability of grassland ecosystems. Changes in cropland area directly influence land use structure and productivity, thereby impacting ecological sustainability. Accordingly, this study identifies temperature, precipitation, population, cropland area, and livestock numbers as the primary driving factors of ecological sustainability in the arid farming–pastoral zone. The data processing methods for each driving factor are provided in
Appendix A.1.
Initially, the simulation accuracy of the CMIP6 mode was compared to observed data. Observed data of temperature and precipitation from seven national meteorological stations in the NFYM, spanning January 2015 to July 2023, were used as the observed values. The data projected by the various CMIP6 modes were extracted based on the latitude and longitude of the meteorological stations and used as the projected values. On this basis, the MAE, RMSE, and R2 for the temperature and precipitation data over 104 months were calculate.
All seven modes present relatively good performance in simulating temperature in the NFYM (
Figure 5a–c). The CAMS-CSM1-0 (MAE = 1.94 °C, RMSE = 2.49, R
2 = 0.96), GFDL-ESM4 (MAE = 2.01 °C, RMSE = 2.59, R
2 = 0.96), and NorESM2-MM (MAE = 2.05 °C, RMSE = 2.71, R
2 = 0.96) modes demonstrate excellent temperature simulation performance. However, the simulation performance of the modes for precipitation in this region varies greatly, ranked from high to low as follows: GFDL-ESM4, NorESM2-MM, CESM2-WACCM, BCC-CSM2-MR, CAMS-CSM1-0, CIESM, and CAS-ESM2-0 (
Figure 5d–f). Compared to other modes, GFDL-ESM4 has the lowest MAE (19.97 mm) and RMSE (34.48), and a significantly higher R
2 (0.36). Meanwhile, CIESM (MAE = 24.11 mm, RMSE = 39.50, R
2 = 0.15) and CAS-ESM2-0 (MAE = 23.89 mm, RMSE = 40.29, R
2 = 0.12) demonstrate poorer simulation performance.
The temperature simulation performance of each mode is superior to their precipitation simulation performance. The reason for this is the significant intra-annual and inter-annual variability observed in the NFYM continental monsoon climate zone, which makes precipitation simulation more challenging than temperature simulation. Additionally, the precipitation process is more complex than temperature changes, involving atmospheric water vapor condensation, cloud formation, and microphysical processes of precipitation. The GFDL-ESM4 mode couples comprehensive physical, chemical, and biogeochemical processes, allowing for a more thorough simulation of climate changes in the earth system. Therefore, it can provide more accurate simulations of future precipitation in the NFYM. The GFDL-ESM4 mode improved temperature simulation metrics (MAE, RMSE, R2) by an average of 15.60%, 13.58%, and 1.47% compared to the other six climate modes. For precipitation simulation, the improvements were 10.86%, 8.85%, and 59.55%. Based on the comparison of simulation performance, the GFDL-ESM4 mode is the only one that shows excellent performance in both temperature (MAE = 2.01 °C, RMSE = 2.59, R2 = 0.96) and precipitation (MAE = 19.97 mm, RMSE = 34.48, R2 = 0.36) simulations among all modes. Therefore, the GFDL-ESM4 mode is selected as the climate simulation mode.
Based on the GFDL-ESM4 mode, the temperature is projected to increase consistently during 2022–2099 (
Figure 6a). The SSP5-8.5 scenario shows the strongest increase, with a rate of 76.00%, followed by the SSP2-4.5 (29.24%) and SSP1-2.6 (5.49%) scenarios. The historical multi-year average temperature is 4.86 °C. From 2023 to 2060, the multi-year average temperatures under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios increase to 5.51 °C, 6.07 °C, and 6.53 °C, respectively. From 2061 to 2099, temperatures continue to rise, with the multi-year average values under the three scenarios further increasing to 5.82 °C, 6.98 °C, and 8.41 °C, respectively.
The GFDL-ESM4 mode projects that precipitation will also increase under all three future scenarios, with small discrepancies of less than 10 mm in annual averages (
Figure 6b). The historical multi-year average precipitation is 244.83 mm. Under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios, the future multi-year average precipitation is projected to be 330.35 mm, 320.77 mm, and 320.31 mm, respectively, corresponding to increases of 34.93%, 31.02%, and 30.83% relative to historical data.
Future cropland areas show a decreasing trend, with the rate of decline increasing from the SSP1-2.6 to the SSP5-8.5 scenarios (
Figure 6c). The cropland area is 1.41 × 10
4 km
2 in 2022, decreasing to 1.21 × 10
4, 1.11 × 10
4, and 0.97 × 10
4 km
2, respectively, under the three scenarios by 2099. The number of livestock also shows a decreasing trend, with the rate of decline increasing from the SSP1-2.6 to the SSP5-8.5 scenarios (
Figure 6d). In 2022, the number of livestock is 973.39 × 10
4 heads. By 2099, the number of livestock under the three scenarios is projected to be 651.43 × 10
4, 670.73 × 10
4, and 748.58 × 10
4 heads, respectively.
The population shows an increasing trend from 2003 to 2007, reaching 2.41 × 10
6 people in 2007, and a decreasing trend from 2008 to 2022, with the population decreasing to 2.22 × 10
6 people by 2022 (
Figure 6e). Projections indicate that the population will continue to decline in the future. From 2023 to 2050, the rate of decline is similar across all three scenarios. From 2050 to 2099, the SSP2-4.5 scenario shows a slower rate of decline, with the population decreasing to 1.20 × 10
6 people by 2099. Both the SSP1-2.6 and SSP5-8.5 scenarios show a faster rate of decline, with the population decreasing to 0.90 × 10
6 and 0.97 × 10
6 people, respectively, by 2099.
3.3. Selection of the Optimal Machine Learning Model
The dataset of annual ecological sustainability and its driving factors for each county from 2003 to 2022 was divided into training and validation sets with a ratio of 70% and 30%, respectively. Simulations of ecological sustainability were conducted separately for each machine learning model and statistical model PLS (
Figure 7). The accuracy of the training set projections, ranked from highest to lowest, is as follows: CNN, RF, BPNN, LSTM, RBF, and PLS. The projection performance of the various models on the validation set is relatively close, with R
2 values ranging from 0.46 to 0.66. Among these, the RF and RBF models exhibit superior projection performance. The accuracy of the validation set projections, ranked from highest to lowest, is as follows: RBF, RF, CNN, LSTM, PLS, and BPNN.
The CNN (MAE = 0.025, RMSE = 0.031, R
2 = 0.86) and the RF (MAE = 0.026, RMSE = 0.032, R
2 = 0.86) demonstrated superior performance in projecting the training set data. The RBF (MAE = 0.040, RMSE = 0.050, R
2 = 0.66) and the RF (MAE = 0.037, RMSE = 0.045, R
2 = 0.65) demonstrated the most effective performance in projecting the validation set data (
Figure 8). The characteristics of a given dataset can influence the extent to which different models can leverage their respective strengths. Among the six models, only the RF consistently demonstrated superior performance in projecting both the training and validation set data. The RF model improved the MAE, RMSE, and R
2 on the training set by an average of 30.74%, 30.40%, and 25.43%, respectively, compared to the other five models. On the validation set, these metrics improved by 13.11%, 13.41%, and 13.81%, respectively. RF reduces model variance and improves generalization ability by constructing multiple decision trees and performing voting or averaging. RF uses randomly selected subsets of features when building each decision tree. This approach increases model diversity and improves performance. Compared to CNN, RF can automatically perform feature selection, has a clear structure, and trains quickly. Additionally, RF is more stable and has better generalization ability compared to PLS. Therefore, RF demonstrates superior performance in simulating ecological sustainability. In general, RF stands out with relatively strong performance and is used to project the future ecological sustainability of the NFYM.
3.4. Projected Changes in Ecological Sustainability during 2023–2099
Changes in ecological sustainability for each county in the NFYM from 2023 to 2099 were simulated using the driving factors and the RF model. The annual average value of ecological sustainability in the NFYM was calculated (
Figure 9a). Future ecological sustainability shows an upward trend under the SSP1-2.6 scenario, a slight decline under the SSP2-4.5 scenario, and a significant decline under the SSP5-8.5 scenario. Specifically, under the SSP1-2.6 scenario, ecological sustainability maintains a growth trend similar to that of 2015–2022. The average value of ecological sustainability in the NFYM in 2022 was 0.42, which is projected to increase to 0.57 by 2099, representing a growth of 34.92%. The growth trend is projected to slow down relatively after the mid-21st century, with an average change rate of 2.23 × 10
−3/year from 2023 to 2050 and a slightly slower growth rate of 2.02 × 10
−3/year from 2051 to 2099. Under the SSP2-4.5 scenario, future ecological sustainability shows a declining trend but remains generally stable. There is a slight decline from 2023 to 2070 at an average change rate of −0.93 × 10
−3/year. After 2070, despite large interannual fluctuations, the overall trend tends to stabilize. Under the SSP5-8.5 scenario, ecological sustainability shows a declining trend, with a relatively large decline from 2023 to 2060 at an average change rate of −5.68 × 10
−3/year. From 2061 to 2099, the trend tends to stabilize but continues to decline at an average change rate of −0.50 × 10
−3/year, reaching 0.27 by 2099, a decrease of 35.80% compared to the current situation.
The spatial distribution maps of ecological sustainability in the NFYM for the years 2030, 2050, 2070, and 2099 are shown in
Figure 9b. Overall, ecological sustainability in each county of the NFYM continues to recover under the SSP1-2.6 scenario. Under the SSP2-4.5 scenario, ecological sustainability slightly deteriorates from 2030 to 2070 and then slightly improves by 2099. Ecological sustainability in each county continues to deteriorate under the SSP5-8.5 scenario.
In 2030, as projected by the SSP1-2.6 scenario, the central region shows moderate ecological sustainability, with average values of 0.41 for SL and 0.43 for SR. In contrast, the western region has low ecological sustainability, with values of 0.35 for UM and 0.33 for UB, while other counties exhibit good ecological sustainability. Ecological sustainability in the central and western regions improves year by year. By 2099, the western region (0.47 for UM and 0.44 for UB) reaches a good level of ecological sustainability, while the average ecological sustainability value of the other 16 counties ranges from 0.41 to 0.56, all at a high level.
Under the SSP2-4.5 scenario, areas with moderate or lower ecological sustainability show a gradual expansion trend from 2030 to 2070. Nine counties in the eastern and southern regions have good ecological sustainability in 2030. However, only two counties maintain this level by 2070, while the ecological sustainability of the western region also decreases. Ecological sustainability improves from 2070 to 2099, with the number of counties with good ecological sustainability increasing to five, while most other counties have moderate ecological sustainability.
Ecological sustainability continues to deteriorate under the SSP5-8.5 scenario. In 2030, ecological sustainability shows a trend of gradually decreasing from northeast to southwest. Two counties in the western region have poor ecological sustainability, the central region has low ecological sustainability, and the eastern and southern regions have moderate ecological sustainability. From 2030 to 2099, ecological sustainability in each county continues to decline. By 2099, all 18 counties have poor ecological sustainability, with the lowest being UB (0.22) and the highest being DL (0.29).
The differences among the three future scenarios illustrate the impact of human activities on the future ecological sustainability of the NFYM. Reduced reliance on fossil fuels and sustainable climate policies decrease greenhouse gas emissions and mitigate the negative impact of human activities on the environment under the SSP1-2.6 scenario. Humans adopt a sustainable development policy path, including effective emission reduction actions and ecological protection measures, such as artificial reforestation and grassland restoration. Additionally, climate change leads to a slight increase in temperature and increased precipitation [
46]. The warmer and wetter climate benefits vegetation growth. The reduction in human impact and favorable natural climate conditions lead to ecological recovery and increased sustainability. In the SSP2-4.5 scenario, there is moderate reliance on fossil fuels. This results in a slightly higher warming trend. Moderate sustainability policies cause a smaller negative impact on the environment from human use of natural resources [
47]. Therefore, ecological sustainability slightly decreases between 2023 and 2070. Furthermore, under the SSP2-4.5 scenario, increased precipitation and reduced human activity pressure between 2070 and 2100 lead to slight ecological recovery [
48]. Weaker climate actions and higher greenhouse gas emissions exacerbate environmental degradation under the SSP5-8.5 scenario. The intensity of human agricultural, pastoral, and industrial activities is greater, negatively impacting vegetation, atmosphere, soil, and water cycles, and leading to resource depletion [
49]. The focus on short-term economic interests and the lack of effective ecological protection measures results in a vicious cycle. This cycle leads to overexploitation and worsening resource depletion.
Further analysis was conducted on the magnitude of changes in ecological sustainability for each county from 2022 to 2099 (
Figure 10). The ecological sustainability of all counties increases under the SSP1-2.6 scenario. The three counties with the smallest increases are DL (rising from 0.58 to 0.64, an increase of 10.77%), UM (rising from 0.39 to 0.47, an increase of 19.55%), and WC (rising from 0.47 to 0.56, an increase of 19.68%). Under SSP2-4.5 scenario, except for UB and XH, the ecological sustainability of the remaining counties declines slightly. The three counties with the largest decreases are DL (falling from 0.58 to 0.50, a decrease of 14.35%), HD (falling from 0.48 to 0.43, a decrease of 10.45%), and WC (falling from 0.47 to 0.43, a decrease of 8.23%). The ecological sustainability of all counties decreases under the SSP5-8.5 scenario. Five counties experience a decrease of more than 40%, namely DL (49.56%), HD (46.62%), ZL (42.97%), WC (41.93%), and ZXB (40.61%).
The results indicate that counties with higher risks of future ecological sustainability decline are all located in the southern part of the NFYM. The southern counties of the NFYM face higher risks of ecological degradation under future SSPs. The southern region currently has higher ecological sustainability, providing more room for decline, which could lead to more severe ecological degradation consequences. Additionally, the southern region has higher economic development levels and population density than the northern region, resulting in greater ecological pressure. The southern ecosystem is mainly composed of forests and farmland, making it more sensitive to human activities. Excessive agricultural development is more likely to lead to land degradation. The DL, HD, and WC counties show the greatest risks of ecological sustainability decline. It is recommended to prioritize strengthening ecological environment protection policies in areas with high ecological degradation risks.
In conclusion, the SSP2-4.5 and SSP5-8.5 scenarios project a heightened risk of future ecological sustainability decline within the NFYM region. However, the anticipated increase in ecological sustainability under the SSP1-2.6 scenario suggests that proactive human attitudes towards the environment can have a significant impact. The NFYM faces substantial risks of future ecological sustainability decline. Decision-makers can address these challenges by implementing enhanced climate adaptation measures, promoting sustainable land management practices, and optimizing the utilization of sustainable water resources.
This study projects the spatiotemporal changes in ecological sustainability within the NFYM from 2023 to 2099. The advantages of the research framework are as follows: (1) the selection of the most accurate CMIP6 mode for simulating future climate, (2) five distinct machine learning models were employed to project ecological sustainability, with the RF model identified as the optimal model based on validation set results, enhancing projection accuracy and reliability, and (3) the spatiotemporal changes in ecological sustainability under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios were analyzed, with corresponding policy recommendations proposed. Furthermore, the calculation methods for NDVI, TWS, and PM2.5 concentration data utilized in this study are well established. Additionally, data on population, cropland area, and livestock numbers can be readily obtained from yearbooks. This makes the proposed research framework more feasible for practical application in regional ecological sustainability assessments and projections compared to other complex models. The GFDL-ESM4 climate mode selected from the CMIP6, along with the RF model chosen from the machine learning models, have demonstrated high accuracy in their application to the NFYM region. This can serve as a reference when selecting models for other arid farming–pastoral zones.
Nevertheless, this study is subject to several limitations and uncertainties. This study assesses ecological sustainability based on remote sensing technology. In the future, we will explore the development of field measurement methods. Field measurements will be used to evaluate and validate the accuracy of remote sensing assessments. Additionally, future ecological sustainability projections depend on the accuracy of SSPs, which inherently possess uncertainties. These limitations necessitate further validation and adjustment in subsequent research.
In future studies, the ecological changes identified in this study can be used to propose water resource allocation plans or sustainable land use strategies. This approach will furnish decision-makers with additional references and options to enhance ecological sustainability.