Nonlinear and Threshold Effects of Urban Green Space Landscape Patterns on Carbon Sequestration Capacity: Evidence from Lanzhou and Baotou

Abstract

Urban green spaces (UGS) are critical regulators of carbon sequestration in industrial cities; however, the configuration mechanisms underlying their carbon dynamics remain insufficiently understood. This study investigates how landscape configuration influences carbon sequestration capacity in Lanzhou and Baotou using multi-temporal datasets from 2000, 2011, and 2022. Net primary productivity (NPP) derived from the CASA model was employed to represent carbon sequestration capacity. An integrated XGBoost-SHAP framework was applied to identify dominant configuration metrics, nonlinear responses, and structural thresholds. The XGBoost model showed stable predictive performance across the three periods, with test-set R² values ranging from 0.470 to 0.510 in Lanzhou and from 0.325 to 0.379 in Baotou. The results reveal systematic and persistent differences in configuration-driven controls between the two cities. In Lanzhou, aggregation-related metrics, particularly COHESION, consistently exert the strongest influence across all three periods, indicating that spatial cohesion and connectivity function as primary stabilizing mechanisms in a mountainous, valley-constrained urban system. Carbon sequestration performance increases once sufficient structural integration is achieved, with aggregation thresholds remaining relatively stable, for example AI values of approximately 0.31–0.34 across 2000–2022, reflecting the importance of maintaining ecological continuity under semi-arid climatic stress. In contrast, Baotou is more strongly regulated by fragmentation-related metrics, especially edge density (ED) and division index (DIVISION), suggesting that its relatively open terrain and industrial spatial structure render carbon sequestration more sensitive to patch separation and edge proliferation. Here, fragmentation acts as a dominant structural constraint, limiting vegetation productivity once spatial disintegration intensifies; for example, ED thresholds shifted from approximately −0.23 in 2000 to −0.56 in 2022. Landscape–carbon relationships exhibit pronounced nonlinear and threshold-dependent behavior in both cities. Rather than responding gradually to structural modification, NPP shifts across identifiable transition points that remain broadly stable over time; for instance, Lanzhou’s AI threshold remains within 0.31–0.34, whereas Baotou’s ED threshold changes from −0.23 to −0.56 across 2000–2022, indicating that these thresholds represent intrinsic structural characteristics of the respective urban ecological systems. However, the magnitude and configuration logic of these thresholds differ between Lanzhou and Baotou, confirming the existence of city-specific nonlinear regimes. These findings demonstrate that urban carbon sequestration operates through context-dependent configuration pathways shaped by terrain, climatic constraints, and long-term spatial organization. The study advances understanding of how structural heterogeneity governs carbon dynamics in arid and semi-arid industrial cities and provides a quantitative basis for configuration-sensitive land planning.

1. Introduction

Empirical evidence suggests that carbon emissions in industrial cities are substantially higher than the global average [1]. Accordingly, mitigating carbon emissions from industrial cities is a key priority for achieving China’s carbon peaking and carbon neutrality goals. Existing studies primarily address this challenge through two pathways: reducing carbon emissions [2,3] and enhancing carbon sequestration capacity [4,5]. However, structural constraints associated with heavy industry often limit the effectiveness of emission reduction strategies. In contrast, enhancing carbon sequestration has been shown to offer greater practical efficiency and ecological benefits [6], with vegetation-based sequestration attracting particular attention due to its high efficiency, broad spatial coverage, and minimal ecological disturbance [7]. In the context of China’s carbon peaking and carbon neutrality goals, improving vegetation-based carbon sequestration has increasingly been regarded as an important complement to direct emission reduction, particularly in industrial cities where structural constraints may limit the effectiveness of conventional mitigation measures. This is also consistent with the broader policy emphasis on improving carbon reduction efficiency and ecological benefits within urban systems [8].

Within urban ecosystems, carbon sequestration is achieved through both natural and artificial processes, with vegetation, soil, and water bodies serving as the primary carbon sinks [9]. Among these components, Urban green spaces (UGS) play a dominant role, while simultaneously supporting ecosystem regulation services that enhance overall carbon sink functions [10]. As the principal spatial carrier of urban vegetation, UGS constitute a critical foundation for improving urban carbon sequestration capacity [11].

Net primary productivity (NPP) is widely used as a proxy for aboveground vegetation carbon sequestration capacity in terrestrial ecosystem studies [12,13,14]. NPP has been extensively applied in urban and regional-scale assessments. For instance, Wang et al. [15] evaluated vegetation carbon sequestration in Yunnan Province using NPP and examined trade-offs between carbon storage and other ecosystem services. In terms of methodological approaches, Jiang et al. [16] employed the dynamic ecosystem model LPJ-GUESS to demonstrate the positive effects of irrigation on soil carbon accumulation. Numerous studies have further explored variations in NPP associated with land-use change, climate variability, vegetation types, and phenological cycles across cropland, forest, grassland, and construction [17,18].

Despite these advances, most existing studies rely on relatively simple linear or monotonic relationships to interpret interactions between landscape patterns and ecosystem functions [19]. Such approaches often fail to capture nonlinear responses, threshold effects, and saturation phenomena, thereby limiting their applicability for precise spatial planning. Consequently, uniform or “one-size-fits-all” UGS planning strategies may result in inefficient resource allocation and uneven ecological benefits, underscoring the necessity for more refined and differentiated analytical frameworks [20].

However, existing studies rarely move beyond pattern description to explicitly identify nonlinear responses, threshold behaviors, and interaction mechanisms linking urban green space landscape patterns to carbon sequestration capacity. This limitation is particularly evident in industrial cities of arid and semi-arid regions, where ecological constraints and development pressures interact to shape distinctive landscape–carbon relationships [21]. In arid and semi-arid industrial cities, the effects of landscape configuration on carbon sequestration are unlikely to be purely geometric. Patch aggregation, connectivity, and fragmentation may alter local hydrothermal conditions by affecting moisture retention, evapotranspiration loss, edge exposure, and thermal buffering, which in turn influence vegetation productivity under water-limited environments [22]. Such processes are particularly relevant in dry urban settings, where heat stress and water limitation may amplify the ecological consequences of fragmentation and reduce the capacity of green patches to maintain stable carbon assimilation [23]. Moreover, comparative evidence across cities with similar developmental backgrounds remains scarce, hindering a deeper understanding of why carbon sequestration capacity responds differently under varying landscape configurations [24]. This gap is particularly important in the Yellow River Basin, where ecological protection and high-quality development are being advanced simultaneously. Understanding how landscape structure mediates ecological functions at the urban scale can provide useful complementary evidence for basin-scale ecological management [25]. In particular, the lack of interpretable analytical frameworks limits the translation of remote sensing-based assessments into actionable sustainability strategies [26,27]. To address these gaps, this study integrates process-based NPP estimation with explainable machine learning to systematically compare Lanzhou and Baotou—two representative industrial cities in the upper Yellow River region—thereby enabling transparent identification of dominant drivers, nonlinear effects, and critical thresholds that underpin inter-city differences in urban green space carbon sequestration. By clarifying the mechanisms behind these differences, the study provides a basis for differentiated planning strategies applicable to industrial cities in arid and semi-arid environments. To investigate relationships between spatial patterns and ecosystem functions, commonly used methods include Pearson correlation analysis [28], Spearman’s rank correlation [29], and geographically weighted regression (GWR). While Pearson and Spearman coefficients quantify the strength of associations between variables [30], they provide limited insight into the mechanisms through which explanatory variables influence response variables. GWR offers improved spatial interpretability [31,32] but is constrained by computational complexity and its reliance on spatial weighting matrices, making it less suitable for large-scale datasets. Given the pronounced heterogeneity and fragmentation of UGS [33], robust analysis requires extensive sampling to accurately quantify landscape-carbon relationships [34,35].

Recent advances in machine learning have effectively addressed limitations related to data scale and complex nonlinear relationships, enabling more accurate modeling of multivariate interactions [28,36,37]. However, conventional machine learning models are often criticized for their “black-box” nature, which obscures the interpretability of prediction outcomes [38]. To overcome this limitation, Lundberg et al. [39] proposed the SHapley Additive exPlanations (SHAP) framework, which quantifies the contribution of individual features to model predictions and provides both global and local interpretability. SHAP offers a powerful tool for revealing nonlinear relationships, interaction effects, and context-dependent mechanisms underlying urban green space carbon sequestration.

Existing research has largely focused on individual land-use types, overall vegetation carbon sequestration capacity, or regions characterized by either rapid economic growth or urgent ecological protection needs. However, relatively few studies have systematically examined the relationship between UGS landscape pattern heterogeneity and a single ecosystem service, particularly carbon sequestration, in industrial-dominated, high-emission cities [5,40]. Lanzhou and Baotou share similar functional orientations but differ in natural setting, urban development pattern, industrial background, and green space configuration. This contrast provides a useful comparative context for examining how landscape configuration regulates carbon sequestration under different urban conditions [41,42]. Enhancing urban carbon sequestration through optimized green space configuration represents a cost-effective and land-efficient pathway for mitigating climate impacts in industrial cities [43]. Identifying threshold values and nonlinear responses is therefore critical for avoiding inefficient or counterproductive landscape interventions, particularly under conditions of limited land resources and competing urban demands [34]. Against this background, this study conducts a comparative analysis of the central urban areas of Lanzhou and Baotou, two representative heavy industrial cities in the upper reaches of the Yellow River. Although both cities have distinct industrial development histories, the present analysis focuses on landscape configuration itself, while industrial background is treated as a contextual condition rather than an explicitly modeled explanatory factor. UGS, including cropland, forest, and grassland, are identified as key components of the urban carbon sink system. NPP, estimated using the Carnegie–Ames–Stanford Approach (CASA) model, is adopted as a proxy for aboveground vegetation carbon sequestration capacity. Landscape pattern metrics at both class and landscape levels are quantified using Fragstats 4.2, and a large set of randomly distributed sampling points is generated to construct spatially explicit datasets. An integrated XGBoost–SHAP framework is then employed to examine nonlinear landscape–carbon relationships, identify dominant configuration drivers, and detect critical thresholds. Specifically, this study aims to address three core questions:

(1)

How do the dominant landscape configuration metrics contributing to carbon sequestration differ between Lanzhou and Baotou?

(2)

Do landscape–carbon relationships exhibit city-specific nonlinear and threshold effects?

(3)

What configuration mechanisms underlie the divergent carbon sequestration responses between the two cities?

The overall research framework is illustrated in Figure 1.

2. Materials and Methods

2.1. Study Area

Lanzhou and Baotou, both located in the upper reaches of the Yellow River, are representative industrial cities in northwestern China. They provide a useful comparative setting for examining urban green space carbon sequestration in arid and semi-arid environments, where rapid urban expansion and ecological constraints interact strongly. Lanzhou is situated in a mountainous river valley setting, whereas Baotou is distributed across flatter plains and adjacent uplands. Their contrasting urban morphologies, together with their long-term industrial development backgrounds, make them suitable for investigating how landscape configuration influences carbon sequestration in urban green spaces (UGS). Lanzhou, the capital of Gansu Province, is situated in a narrow river valley setting characterized by a pronounced “mountains–river–urban corridor” topography. The city extends primarily along the Yellow River, with higher elevations to the south and lower terrain to the north. This constrained geomorphological context has contributed to concentrated urban development and relatively fragmented green spaces within the valley corridor. Under semi-arid climatic conditions, such a topographic setting also increases the ecological importance of patch continuity and local hydrothermal buffering. These characteristics make Lanzhou a representative case for examining how aggregation and connectivity influence carbon sequestration in valley-type industrial cities [44].

Baotou, located in western Inner Mongolia, represents a contrasting urban form characterized by broader spatial expansion and a long-term resource-based industrial background. The city spans plains and mountainous areas, with extensive built-up and industrial land that has been associated with elevated environmental pressure. Compared with Lanzhou, Baotou exhibits a broader spatial extent and more heterogeneous green space distribution, providing a useful counterpart for examining how fragmentation, edge effects, and landscape division regulate UGS carbon sequestration under different urban structural conditions [45].

Historically, both Lanzhou and Baotou were designated as key industrial centers during China’s First Five-Year Plan (1953–1957) and were major recipients of projects under the Soviet-assisted “156 Projects” initiative. Lanzhou developed as a national petrochemical base [44], while Baotou became a major center for the machinery and metallurgical industries [46]. This shared industrial legacy, together with their strategic positions in the upper Yellow River Basin, supports their value as representative comparative cases for this study. In the present study, this industrial legacy is used as a contextual background for inter-city comparison rather than as a directly quantified explanatory factor.

In addition, the two cities also differ in dominant vegetation composition and soil background, which may affect the ecological performance and carbon sequestration potential of urban green spaces. The main climatic and ecological characteristics of the two study areas are summarized in

Table 1. Climatic and ecological characteristics of the study areas in Lanzhou and Baotou.

Variable	Lanzhou	Baotou
climate type	Temperate continental semi-arid climate	Temperate continental arid to semi-arid climate
Geographic setting	Upper reaches of the Yellow River; narrow river valley urban corridor bordered by mountains	Upper reaches of the Yellow River; broader urban area distributed across plains and adjacent uplands
Dominant vegetation	Coniferous forest; broad-leaved forest; shrubland; grassland; meadow; desert	Coniferous forest; broad-leaved forest; shrubland; meadow; desert
Soil types	Loess; Gray-calcareous soil; Lithosol; Meadow soil	Chestnut soil; Gray-brown soil; Meadow soil

.

Given their contrasting urban morphologies, shared industrial backgrounds, and critical roles in the Yellow River Basin’s ecological governance framework, Lanzhou and Baotou offer a compelling comparative setting for this study. Following the Territorial Spatial Master Plan of Lanzhou (2021–2035) and the Territorial Spatial Master Plan of Baotou (2021–2035), the central urban areas of Lanzhou (including Chengguan, Xigu, Anning, Qilihe districts, and Jiuhe and Zhonghe towns) and Baotou (including Kundulun, Qingshan, Jiuyuan, and Donghe districts) were selected as the study areas (Figure 2).

2.2. Data Sources and Processing

Land use and vegetation type data for the central urban areas of Lanzhou and Baotou in 2000, 2011 and 2022 were obtained from the GLC_FCS30 dataset (https://data.casearth.cn/, accessed on 19 November 2024), which provides global fine land cover classification at a spatial resolution of 30 m × 30 m [47]. Based on a first-level vegetation classification scheme, land cover was categorized into cropland, forest, grassland, water, construction, and unused.

Given the relatively strong carbon sequestration capacity of cropland, forest, and grassland ecosystems, these land cover types were reclassified as UGS, while construction, water, and unused were treated as non-green spaces. This classification scheme follows established approaches in previous studies [48,49,50] and provides a consistent basis for subsequent landscape pattern analysis. Reclassification was performed using the Reclassify tool in ArcMap 10.8.

Meteorological variables, including temperature, precipitation, and solar radiation, were derived from the TerraClimate dataset (https://developers.google.cn/earth-engine/datasets, accessed on 25 November 2024). The Normalized Difference Vegetation Index (NDVI) was obtained from the China Meteorological Science Data Center (https://data.cma.cn/, accessed on 1 December 2024). Monthly NDVI composites were calculated using Landsat series products through the layered stacking function in ENVI 5.6, with a spatial resolution of 30 m × 30 m. DEM data were sourced from the European Space Agency (https://panda.copernicus.eu/panda, accessed on 1 December 2024) at the same spatial resolution. All spatial datasets were projected to the WGS_1984_World_Mercator coordinate system to ensure spatial consistency and facilitate subsequent analysis.

2.3. Estimation of Net Primary Productivity Using the CASA Model

NPP of UGS in Lanzhou and Baotou was estimated using the Carnegie–Ames–Stanford Approach (CASA) model proposed by Potter et al. [51]. The CASA model quantifies vegetation productivity based on the effective absorption of photosynthetically active radiation and light use efficiency, making it well-suited for large-scale and spatially explicit NPP estimation. In this study, carbon sequestration capacity refers specifically to aboveground vegetation productivity as represented by NPP, rather than total ecosystem carbon storage. Moreover, in the CASA framework, NPP is regulated not only by absorbed photosynthetically active radiation but also by light use efficiency, which is constrained by temperature and water stress. In arid and semi-arid urban environments, landscape configuration may indirectly influence these stress terms by modifying local hydrothermal conditions, including moisture retention, evapotranspiration loss, and edge-related thermal exposure.

In this study, vegetation type, solar radiation, NDVI, air temperature, and precipitation were integrated to estimate annual NPP for urban green spaces, which was used as a proxy for carbon sequestration capacity. The model formulation follows established practices and can be expressed as

$$NPP=APAR\left(x,y\right)\times \epsilon \left(x,y\right).$$

(1)

$NPP$ represents the net primary productivity value (g/m/a) of vegetation in Lanzhou and Baotou within a certain period. $APAR\left(x,y\right)$ represents the photosynthetically active radiation (MJ·m⁻²·month⁻¹) of the vegetation for pixel $x$ in month $y$; $\epsilon \left(x,y\right)$ represents the vegetation light energy conversion rate (gC·MJ⁻¹) of pixel $x$ in month $y$.

$$APAR\left(x,y\right)=SOL\left(x,y\right)\times FPAR\left(x,y\right)\times k$$

(2)

$FPAR\left(x,y\right)$ represents the effective absorption rate of solar radiation by vegetation; $SOL\left(x,y\right)$ represents the solar radiation (MJ/m²) per month, and $k$ is 0.5.

$$FPAR\left(x,y\right)=a\times {FPAR\left(x,y\right)}_{NDVI}+\left(1-a\right)\times {FPAR\left(x,y\right)}_{SR}$$

(3)

The adjustment parameter $a$ is set to 0.5. The specific calculation process of ${FPAR\left(x,y\right)}_{NDVI}$ and ${FPAR\left(x,y\right)}_{SR}$ uses the $NPP$ remote sensing estimation model [52] and, respectively, represent the effective absorption rates based on $NDVI$ and $SR$ solar radiation:

$$\epsilon \left(x,y\right)=T_{\epsilon 1}\left(x,y\right)\times T_{\epsilon 2}\left(x,y\right)\times W_{\epsilon }\left(x,y\right)\times {\epsilon }_{max},$$

(4)

where ${\epsilon }_{\mathrm{m}\mathrm{a}\mathrm{x}}$ is the maximum light use efficiency, $T_{\epsilon 1}$ is the low-temperature stress scalar, $T_{\epsilon 2}$ is the high-temperature (or optimum-temperature deviation) stress scalar, and $W_{\epsilon }$ is the water stress scalar, with a value range of 0.5 to 1. $T_{\epsilon 1}$, $T_{\epsilon 2}$, and $W_{\epsilon }$ are dimensionless coefficients, and their multiplicative form follows the standard CASA framework in which temperature and water constraints jointly regulate actual light use efficiency.

2.4. Quantification of Urban Green Space Landscape Pattern Metrics

Landscape pattern metrics provide an objective representation of the spatial organization of natural elements and human activities. Following established studies [53,54], six commonly used landscape metrics were selected at the class and landscape levels to characterize UGS patterns: area-weighted mean fractal dimension (FRAC_AM), percentage of landscape (PLAND), edge density (ED), division index (DIVISION), cohesion index (COHESION), and aggregation index (AI).

FRAC_AM and PLAND describe patch shape complexity and proportional area, respectively, whereas ED and DIVISION reflect landscape fragmentation and edge effects. COHESION and AI characterize the degree of spatial aggregation and physical connectivity among green space patches. Together, these metrics capture landscape pattern characteristics from three complementary dimensions: patch area and shape, fragmentation and edge complexity, and aggregation and connectivity. The definitions and calculation formulas of these indices are consistent with those reported in previous studies [55]. Landscape metrics were computed using Fragstats 4.2.

2.5. XGBoost-SHAP

2.5.1. XGBoost-Based Regression Modeling

XGBoost is a scalable boosting framework built upon decision trees, designed to enhance model accuracy through sequential optimization, as introduced by Chen et al. [56]. By iteratively optimizing weak learners and incorporating regularization terms, XGBoost effectively improves predictive accuracy while reducing the risk of overfitting, making it suitable for modeling complex nonlinear relationships.

To establish the relationship between UGS landscape patterns and carbon sequestration capacity, 20,000 random sampling points were generated using the random point tool in ArcMap 10.8. Sampling points located outside UGS or exhibiting excessive spatial clustering were removed to ensure randomness and spatial representativeness. Specifically, spatial clustering was evaluated using a nearest-neighbor distance criterion, and points with distances smaller than one raster cell were excluded to ensure spatial independence. Furthermore, due to the significant spatial heterogeneity and scale differences in UGS landscapes. Therefore, we chose to analyze various green space scales, observed the changes in the UGS landscape patterns, and combined them with existing research [57], finding that when the UGS landscape scale was between 40 and 60 m, the landscape pattern indicators tended to stabilize. Therefore, we have decided to select 50 m as the scale of the green space for our study. The size of the sliding window was determined to be 400 m based on the actual reference to existing studies in the research area and the relevant conditions of the UGS landscape pattern in the research area, as an appropriate scale [23,58]. Separate datasets were constructed for Lanzhou and Baotou. Landscape pattern metrics were used as explanatory variables, while NPP served as the response variable. All six landscape pattern metrics satisfied the variance inflation factor (VIF) criterion, with values below the accepted threshold (VIF < 10), suggesting that multicollinearity was not a concern in the model.

The full dataset was partitioned into independent subsets for model calibration and evaluation, with 70% of the samples used for training and the remaining 30% reserved for testing. Model optimization in XGBoost is achieved by minimizing a composite objective that integrates prediction error with a regularization component, formulated as follows:

$$y_i^{\prime }=\sum _{t=1}^n{f}_t\left(x_i\right)=y_i^{\left(\mathrm{n}-1\right)}+f_n\left(x_i\right).$$

(5)

The algorithm includes $n$ tree models, and $y_i^{\prime }$ represents the algorithmically predicted values of the $n$ trees corresponding to $x_i$, $y_i^{(\mathrm{n}-1)}$ represents the prediction result of the first $(\mathrm{n}-1)$ trees, and $f_n\left(x_i\right)$ represents the model of the $n$ tree. The objective function that constitutes XGBoost consists of a loss function and a regularization function, and the formula is as follows:

$$Obj=\sum _{t=1}^{k}l\left(y_i,y_i^{\prime }\right)+\sum _{t=1}^n\mathsf{\Omega }\left(f_t\right),$$

(6)

$$\mathsf{\Omega }\left(f_t\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda \sum _{j=1}^T{w}_j^2.$$

(7)

$Obj$ represents the objective function; $\sum _{t=1}^{k}l\left(y_i,y_i^{\prime }\right)$ is the loss function; $\sum _{t=1}^n\mathsf{\Omega }\left(f_t\right)$ represents the regularization function; $k$ represents the sample size.; $T$ represents the number of leaves, which determines the complexity of the tree. $w_j^2$ represents the weight of each leaf node; $\lambda$ is a constant term, representing the regularization coefficient. $\gamma$ is the splitting threshold, which is used to control the complexity of the model as follows:

$${\mathrm{y}}_i^{\left(t\right)}=\sum _{n=1}^t{f}_n\left(x_i\right)=y_i^{\left(t-1\right)}+f_t\left(x_i\right).$$

(8)

$y_i^{\left(t\right)}$ represents the prediction result after iteration; $y_i^{\left(t-1\right)}$ represents the prediction result; $f_t\left(x_i\right)$ represents the tree model. At this point, the objective function is

$$Obj=\sum _{i=1}^{k}l\left(y_i,y^{\left(t-1\right)}+f_t\left(x_i\right)\right)+\sum _{t=1}^n\mathsf{\Omega }\left(f_t\right).$$

(9)

Meanwhile, perform a second-order Taylor expansion on the objective function ${Obj}_1$ at this time and remove the constant term to obtain the following results:

$${Obj}_1=\sum _{i=1}^k\left[g_i{f}_t\left(x_i\right)+{\displaystyle \frac{1}{2}}{h}_i{f}_t^2\left(x_i\right)\right]+\mathsf{\Omega }\left(f_t\right).$$

(10)

$g_i$ represents the first derivative of the $t-1$ round loss function ${(y}_i,y^{\left(t-1\right)})$; $h_i$ represents the second derivative of the $t-1$ round loss function ${(y}_i,y^{\left(t-1\right)})$. And continuously optimize the parameters through the objective function ${Obj}_1$, and finally output the result when the experimental requirements are met.

To improve model performance, Bayesian optimization combined with ten-fold cross-validation was employed for hyperparameter tuning. Spatial cross-validation is further applied to evaluate the generalization ability and robustness of the model. The optimal hyperparameters were determined as: max_depth = 3, learning_rate = 0.1, n_estimators = 50, random_state = 666, and colsample_bytree = 0.7.

To ensure that the identified dominant drivers and threshold patterns were not sensitive to a specific parameter configuration, additional robustness tests were conducted. Key hyperparameters (e.g., max_depth and learning_rate) were perturbed within reasonable ranges around the optimal values, and alternative random seeds were tested. The ranking of dominant landscape metrics and the identified nonlinear and threshold-dependent patterns remained largely consistent across these alternative settings, indicating that the core findings are stable and not driven by a particular hyperparameter combination.

Model performance was evaluated using the mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²). Lower MAE and RMSE values indicate higher predictive accuracy, while R² values closer to 1 suggest stronger agreement between predicted and observed values.

2.5.2. SHAP-Based Model Interpretation

To enhance the interpretability of the XGBoost model, the SHAP method was employed. SHAP is grounded in cooperative game theory and quantifies the contribution of each feature to individual model predictions [39]. By decomposing prediction outcomes into additive feature contributions, SHAP enables a transparent assessment of both the magnitude and direction of variable effects.

In this study, SHAP was used to interpret the XGBoost regression results and to reveal the heterogeneous and nonlinear influences of UGS landscape patterns on carbon sequestration capacity. Importantly, SHAP facilitates the identification of thresholds and turning points associated with landscape pattern changes, thereby providing a quantitative basis for understanding how variations in UGS configuration influence NPP.

3. Results

3.1. Urban Green Space Pattern Dynamics in Lanzhou and Baotou (2000–2022)

UGS in the central districts of Lanzhou and Baotou are primarily composed of cropland, grassland, and forest, yet their spatial configurations exhibit both inter-city differences and temporal evolution from 2000 to 2022 (Figure 3).

In Lanzhou, UGS display a pronounced river valley-constrained pattern throughout the study period. In 2000, green spaces were largely distributed along the peripheral mountainous zones, while the central valley corridor was dominated by built-up land. Between 2000 and 2011, moderate expansion of construction land occurred along the valley axis, leading to increased fragmentation of adjacent green patches. By 2022, green spaces in mountainous areas remained relatively stable, whereas the central urban corridor exhibited intensified spatial heterogeneity due to continued urban infill and redevelopment. This pattern reflects Lanzhou’s topographically constrained urban expansion and the combined effects of urban growth control and ecological restoration policies implemented after 2010.

In contrast, Baotou presents a more horizontally expanded and terrain-influenced configuration. In 2000, green spaces were mainly concentrated in the northern and eastern elevated zones, with relatively continuous patches. During 2000–2011, rapid urban development in central districts increased landscape fragmentation, particularly around industrial and residential clusters. From 2011 to 2022, partial stabilization and greening efforts contributed to localized improvements in patch continuity; however, green space distribution remained more dispersed compared to Lanzhou. The broader spatial extent of urban expansion and industrial land use in Baotou contributed to a more fragmented landscape structure.

Overall, while both cities are dominated by grassland and cropland, their spatial evolution trajectories differ. Lanzhou maintains a valley-oriented, topographically confined configuration with relatively concentrated green spaces in peripheral zones, whereas Baotou exhibits a more dispersed and expansion-driven pattern. These contrasting spatial dynamics provide the structural basis for the subsequent analysis of differences in NPP responses and landscape-driven carbon sequestration mechanisms.

3.2. Spatiotemporal Dynamics of Urban Green Space NPP in Lanzhou and Baotou (2000–2022)

Based on the spatial configuration of UGS, NPP was estimated for 2000, 2011, and 2022 using the CASA model (Figure 4), revealing distinct temporal trajectories and inter-city contrasts in carbon sequestration capacity.

In Lanzhou, NPP consistently exhibits a pronounced elevation-driven gradient across all three periods. High productivity is concentrated in the southern mountainous zones, while lower values persist within the central valley corridor along the Yellow River. This pattern suggests that topographic position affects not only land development intensity but also local hydrothermal conditions. In the mountainous belts, relatively continuous green patches may improve moisture retention and reduce thermal exposure, whereas the densely built valley corridor is more vulnerable to heat accumulation and water stress. From 2000 to 2011, NPP increased primarily in peripheral upland areas, corresponding with ecological restoration initiatives and reduced land disturbance. By 2022, productivity gains further expanded in the southern and southeastern mountainous belts. However, the valley floor—characterized by dense construction, limited vegetation continuity, and intensified anthropogenic pressure—remained relatively low in NPP. The persistence of this spatial contrast likely reflects the combined effects of topographic confinement, semi-arid climatic stress, and concentrated urban development within the valley system. Under semi-arid valley conditions, such fragmented and exposed green spaces are less able to buffer evapotranspiration loss and thermal stress, which likely constrains vegetation productivity. This provides an ecological explanation for the persistent low-NPP pattern in the central corridor.

Baotou displays a comparatively broader and more horizontally distributed NPP pattern. In 2000, moderate-to-high productivity zones were primarily located in northern plains and western mountainous areas, while lower NPP values were concentrated in the central built-up districts. In Baotou, this contrast likely reflects not only differences in land-use intensity but also the greater exposure of central green spaces to thermal stress and moisture limitation under flatter and more open terrain conditions. Such settings may amplify the ecological consequences of fragmented green patterns for vegetation productivity. During 2000–2011, peripheral green spaces showed gradual productivity enhancement, likely associated with vegetation recovery and urban greening policies, whereas central zones experienced spatial fluctuation likely associated with land redevelopment and changes in urban land-use intensity. By 2022, higher NPP areas expanded across northern and western regions, yet relatively low productivity persisted in densely built and industrially intensive urban cores. Compared with Lanzhou, Baotou’s NPP distribution appears more sensitive to land development patterns and the broader urban spatial context than to topographic confinement alone. This suggests that, in an arid and industrially exposed landscape, carbon sequestration is more directly affected by the spatial reorganization of urban land. Once green spaces become more isolated or edge-dominated, local hydrothermal stress may intensify more rapidly than in the topographically buffered landscape of Lanzhou.

Overall, although both cities demonstrate an upward trend in NPP from 2000 to 2022, their spatial structures differ markedly. Lanzhou exhibits a terrain-constrained productivity regime dominated by mountainous ecological cores, whereas Baotou presents a more development-sensitive pattern associated with lateral urban expansion and land-use restructuring. These spatiotemporal differences establish the structural context for the subsequent analysis of how landscape configuration regulates carbon sequestration capacity under distinct urban systems.

3.3. Spatiotemporal Heterogeneity of Urban Green Space Landscape Metrics (2000–2022)

Landscape pattern metrics of UGS in Lanzhou and Baotou exhibit persistent spatial heterogeneity across 2000, 2011, and 2022, while also revealing distinct structural contrasts between the two cities (Figure 5 and Figure 6).

In Lanzhou, aggregation-related metrics, including AI and cohesion, consistently display higher values in the southern and peripheral mountainous areas and markedly lower values within the central valley corridor. This pattern remains stable across all three periods and reflects the city’s valley-constrained morphology, where built-up land is concentrated along the Yellow River axis under a long-term industrial development context. The semi-arid climate further reinforces this spatial segregation, as vegetation continuity is more easily maintained in less-developed upland areas than in the densely built valley floor. Fragmentation-related metrics (ED and DIVISION) show elevated values within the central corridor and transitional slopes, with intensified fragmentation observed during 2000–2011, coinciding with accelerated urban expansion and infrastructure development. By 2022, partial structural stabilization is evident, though fragmentation remains prominent in low-elevation urban zones. FRAC_AM indicates higher patch shape complexity in mountainous regions, while PLAND remains substantially greater in peripheral green belts than in the compact urban core.

Baotou presents a comparatively broader and laterally distributed configuration pattern. Aggregation metrics are higher in northern and western green belts, whereas the central plains—characterized by dense built-up land and intensive transportation infrastructure within a long-term industrial development background—exhibit persistently lower aggregation and connectivity. Fragmentation metrics (ED and DIVISION) highlight stronger structural disintegration within built-up plains, particularly during 2000–2011, a period associated with rapid outward urban growth and land-use reorganization. By 2022, moderate structural adjustment is observable, although spatial separation remains evident in intensively developed districts. Unlike Lanzhou, Baotou’s configuration patterns appear more closely aligned with development intensity and land-use restructuring than with strong topographic confinement.

Overall, both cities maintain a stable contrast between aggregated peripheral green spaces and fragmented urban cores over time. However, Lanzhou exhibits a terrain-dominated structural polarization, whereas Baotou demonstrates a more development-sensitive and laterally dispersed pattern. In this section, the industrial background is considered as part of the broader urban development context, whereas the analysis itself focuses on the spatial configuration of UGS. These spatiotemporal differences in landscape configuration provide the structural foundation for understanding the distinct nonlinear and threshold-dependent carbon sequestration responses identified in the subsequent analysis.

3.4. XGBoost-SHAP Analysis of Landscape Pattern Effects on Urban Green Space NPP

3.4.1. Comparison of Model Performance

To ensure the robustness of the regression analysis, three machine learning algorithms-XGBoost, Random Forest (RF), and Gradient Boosting Decision Trees (GBDT) were evaluated across the three study years. Furthermore, we also selected the Ordinary Least Squares (OLS) model of the traditional regression model for result comparison. Among the tested models, XGBoost consistently demonstrated superior predictive performance and more stable generalization ability (

Table 2. Performance comparison of machine learning models for predicting NPP in Lanzhou and Baotou.

Year	Type	Method	Lanzhou		Baotou
			Training	Testing	Training	Testing
2000	XGBoost	R2	0.651	0.470	0.358	0.325
		MAE	42.587	48.257	72.512	75.651
		RMSE	75.549	94.561	91.264	97.263
	Random Forest (RF)	R2	0.488	0.462	0.327	0.306
		MAE	69.089	70.442	60.761	83.521
		RMSE	92.242	94.045	85.403	99.101
	Gradient Boosting Decision Trees (GBDT)	R2	0.548	0.403	0.302	0.297
		MAE	43.358	53.391	68.901	73.981
		RMSE	93.189	95.902	88.182	86.367
	Ordinary Least Squares (OLS)	R2	0.195		0.090
		MAE	182.904		289.190
		RMSE	209.198		387.275
2011	XGBoost	R2	0.520	0.495	0.427	0.358
		MAE	41.657	50.261	69.581	60.274
		RMSE	90.268	92.657	63.517	65.213
	Random Forest (RF)	R2	0.438	0.423	0.429	0.325
		MAE	64.562	66.874	54.532	52.169
		RMSE	88.917	83.820	80.718	88.172
	Gradient Boosting Decision Trees (GBDT)	R2	0.455	0.432	0.356	0.338
		MAE	58.419	57.372	60.901	58.772
		RMSE	79.609	85.385	83.714	98.209
	Ordinary Least Squares (OLS)	R2	0.182		0.076
		MAE	338.180		376.150
		RMSE	359.378		389.290
2022	XGBoost	R2	0.643	0.510	0.457	0.379
		MAE	55.35	46.93	65.168	82.772
		RMSE	76.977	90.441	79.088	99.642
	Random Forest (RF)	R2	0.488	0.462	0.477	0.326
		MAE	69.089	70.442	60.761	83.521
		RMSE	92.242	94.045	85.403	99.101
	Gradient Boosting Decision Trees (GBDT)	R2	0.555	0.469	0.474	0.306
		MAE	64.051	69.749	60.535	76.903
		RMSE	85.945	93.475	85.423	91.516
	Ordinary Least Squares (OLS)	R2	0.223		0.118
		MAE	129.109		159.170
		RMSE	148.392		183.253

). This also indicates that the UGS landscape pattern may have complex nonlinear effects on the spatial response of NPP that traditional regression models may not be able to accommodate [59].

Specifically, in 2000, XGBoost achieved test-set R² values of 0.470 in Lanzhou and 0.325 in Baotou, with corresponding RMSE values of 94.561 and 97.263. In 2011, the test-set R² values reached 0.495 and 0.358, respectively, accompanied by RMSE values of 92.657 and 65.213. By 2022, XGBoost further improved predictive accuracy, yielding test-set R² values of 0.510 in Lanzhou and 0.379 in Baotou, with RMSE values of 90.441 and 81.642. It should be noted that the R² values from the regression results generally range between 0.325 and 0.510, indicating a relatively weak overall correlation. This is because the dataset utilized in this study consists of statistical data derived from spatial sampling; consequently, the regression results do not readily demonstrate a strong correlational relationship.

Across all years, XGBoost maintained relatively low MAE values and stable explanatory power, confirming its suitability for modeling the nonlinear relationships between UGS landscape patterns and carbon sequestration capacity. Based on this evaluation, XGBoost was selected for subsequent SHAP-based interpretation analyses.

3.4.2. Temporal Evolution of Landscape Metric Contributions to NPP

The SHAP summary plots (Figure 7) indicate that the relative importance of landscape configuration metrics in regulating NPP varies not only between Lanzhou and Baotou but also across the three study years, reflecting dynamic structural controls on urban green space carbon sequestration. These temporal shifts in feature importance also imply that landscape metrics do not affect NPP through purely geometric properties. In this analysis, differences between the two cities are interpreted primarily through landscape configuration and hydrothermal constraints, while industrial background is treated as a contextual condition rather than a directly quantified explanatory factor. In arid and semi-arid urban systems, aggregation, fragmentation, and edge exposure can alter local hydrothermal conditions by influencing moisture retention, evapotranspiration loss, and thermal buffering. Therefore, changes in metric rankings may reflect changing ecological pathways through which landscape configuration regulates vegetation productivity.

In Lanzhou, aggregation-related metrics dominate in the early stage. In 2000, AI and COHESION ranked as the two most influential predictors, suggesting that NPP is primarily regulated by the spatial continuity and connectivity of green patches. Under Lanzhou’s mountainous valley setting, such aggregation-related control likely reflects the role of spatial cohesion in stabilizing local hydrothermal conditions. More continuous green patches can better retain soil moisture, reduce edge-related heat exposure, and form relatively buffered microclimatic units, which is particularly important under semi-arid water limitation. In this sense, the dominance of AI and COHESION suggests that NPP in Lanzhou depends not only on the amount of green space but also on whether patch structure is sufficiently integrated to mitigate moisture loss and thermal stress. This pattern is consistent with Lanzhou’s terrain-constrained urban morphology, where mountainous belts surrounding the valley corridor maintain relatively intact ecological cores that strongly determine carbon productivity. By 2011, ED rises to the second position in importance, indicating that increasing edge complexity associated with urban infill and transitional development begins to exert a stronger influence on NPP. In 2022, although COHESION regains prominence and DIVISION increases in relative contribution, aggregation-related controls remain fundamental. These shifts imply that carbon sequestration in Lanzhou is jointly shaped by stable peripheral aggregation and evolving fragmentation processes within the central valley floor under continued urban development. The mountainous topographic framework buffers large-scale structural disruption, allowing aggregation metrics to maintain consistent explanatory power over time.

In contrast, Baotou displays a fragmentation-dominated contribution structure throughout most of the study period. In 2000, ED and PLAND appear among the leading predictors, reflecting the sensitivity of NPP to edge density and proportional green space coverage under dispersed urban expansion. By 2011, PLAND became the most influential factor, followed by ED and DIVISION, indicating that both green space proportion and patch separation play critical roles during a phase of rapid urban expansion and land-use intensification. In 2022, DIVISION rose to the top rank, with AI and PLAND following, highlighting the increasing importance of patch isolation and structural discontinuity. Compared with Lanzhou, this pattern indicates that Baotou’s carbon sequestration is more strongly constrained by fragmentation-mediated ecological stress. In a flatter and more open urban landscape, green patches are less protected by topographic buffering and are more directly exposed to heat, wind, and moisture loss. Under such conditions, urban expansion can more readily increase patch separation and edge interfaces, causing NPP to become more sensitive to structural disintegration than to aggregation enhancement alone. Unlike Lanzhou, Baotou’s relatively flat terrain permits lateral urban expansion, leading to more widespread patch subdivision and stronger edge effects. Under such conditions, carbon sequestration capacity becomes highly responsive to fragmentation dynamics rather than aggregation stability.

These differences suggest that the same landscape metric may carry different ecological meanings under distinct urban structures and hydrothermal constraints. In the present study, such differences are interpreted mainly in relation to terrain, hydrothermal regulation, and urban spatial structure, while industrial background provides contextual support. Taken together, the temporal shifts in metric rankings reveal two contrasting structural pathways. Lanzhou’s NPP remains closely tied to aggregation integrity constrained by valley-mountain interactions, whereas Baotou’s carbon dynamics are more strongly governed by fragmentation intensity under lateral urban expansion and land-use restructuring. These evolving contribution patterns provide critical evidence that the mechanisms linking landscape configuration and carbon sequestration are city-specific and co-develop with urban spatial restructuring processes.

3.4.3. Nonlinear and Threshold-Dependent Responses of NPP to Landscape Metrics

The SHAP dependence results for 2000, 2011, and 2022 reveal that the relationships between landscape configuration and NPP in Lanzhou and Baotou are consistently nonlinear and characterized by distinct threshold transitions rather than gradual linear gradients (Figure 8 and Figure 9). Across all three periods, each landscape metric exhibits identifiable turning points at which the direction or intensity of its influence on carbon sequestration changes abruptly. These thresholds shift over time and differ markedly between the two cities, indicating that vegetation productivity responds to landscape restructuring through context-dependent ecological regimes. In this section, urban development context is considered primarily through its effects on landscape structure, whereas industrial disturbance is treated as a background condition rather than a directly quantified explanatory variable. More specifically, these thresholds can be interpreted as points at which landscape structure begins to alter local hydrothermal regulation in a nonlinear manner. In arid and semi-arid cities, aggregation, fragmentation, and edge exposure may influence vegetation productivity by modifying moisture retention, evapotranspiration loss, and thermal buffering. Therefore, threshold shifts across years and between cities likely reflect changes in the ecological capacity of green space networks to mitigate water limitation and heat stress under different urban development conditions.

(1)

Lanzhou: Increasing Structural Sensitivity under Topographic Constraint

In Lanzhou, aggregation-related metrics play a dominant regulatory role, and their thresholds exhibit relative stability over time. For example, the threshold of AI remains around 0.31–0.34 from 2000 to 2022, with a consistently positive post-threshold effect. Below this level, increases in aggregation yield limited gains in NPP, suggesting that dispersed patches in mountainous terrain are insufficient to alter microclimatic stress. Once AI surpasses the threshold, however, NPP rises sharply, indicating that a minimum degree of spatial clustering is required to overcome moisture limitation and enhance ecological cohesion in a semi-arid valley setting. This response suggests that patch aggregation does not simply increase spatial continuity, but also strengthens local ecological buffering. Once clustering exceeds the identified threshold, green patches are more capable of conserving soil moisture, reducing edge-related thermal exposure, and stabilizing the microclimatic conditions that support vegetation growth. Under Lanzhou’s valley-constrained morphology, such buffering is especially important because topographic enclosure tends to intensify heat accumulation in fragmented lowland spaces.

COHESION shows a similar nonlinear response, with thresholds shifting slightly from 0.17 (2000) to approximately 0.13 (2022). The declining threshold suggests that over time, Lanzhou’s green system became more sensitive to connectivity improvements, likely reflecting policy-driven greening and slope ecological restoration that increased patch linkage in peripheral mountainous zones. As connectivity improves, evapotranspiration stress is mitigated and soil moisture retention stabilizes, strengthening carbon assimilation.

By contrast, fragmentation-related metrics such as ED and DIVISION exhibit negative threshold responses. ED thresholds remain consistently around −0.7, beyond which NPP declines markedly. This pattern reflects the ecological cost of excessive edge exposure in a valley city characterized by strong heat accumulation and limited ventilation. Increased edge density intensifies solar radiation exposure and desiccation in fragmented patches. Similarly, DIVISION shows turning points around 1.4 in 2000 and 2022, indicating that once spatial separation exceeds a critical level, vegetation systems lose functional integration. Together, these negative thresholds indicate that fragmented green patches in Lanzhou become functionally ineffective once edge exposure and spatial separation exceed certain levels. In a semi-arid valley system, such fragmentation accelerates surface drying and weakens the ability of UGS to moderate local thermal stress, thereby reducing the light use efficiency captured in the CASA-based NPP estimates.

PLAND displays a dynamic shift: the threshold moves from −1.96 (2000) to approximately 0.58 (2011–2022), suggesting that the ecological meaning of area proportion changes over time. In the early stage, increases in green proportion primarily compensated for ecological deficits in built-up cores. In later stages, however, simple expansion of green area becomes insufficient unless accompanied by improved configuration, indicating a transition from quantity-driven to structure-driven carbon enhancement.

Overall, Lanzhou’s nonlinear responses are strongly shaped by its river valley morphology and surrounding mountainous terrain. In a semi-arid environment with complex topography, aggregation and connectivity determine whether green patches can form effective microclimatic buffers. Threshold stability over time indicates structural control dominated by terrain and moisture constraint, while the industrial background is better understood as part of the broader urban development context.

Baotou: Fragmentation-Regulated Regime under Arid Urban Context. Baotou exhibits a contrasting response pattern in both threshold magnitude and temporal dynamics. In this more open and arid setting, fragmentation-related metrics exert stronger control over NPP variability.

ED thresholds shift from −0.23 (2000) to −0.56 (2022), indicating increasing sensitivity to edge effects over time. In early stages, moderate edge density slightly enhances productivity, likely due to increased light penetration and airflow in relatively flat terrain. However, under continued urban expansion and land-use pressure, excessive edge formation amplifies ecological stress, leading to steeper post-threshold declines in NPP. This pattern suggests that edge-related thresholds in Baotou are closely tied to the amplification of hydrothermal stress under open and arid conditions. Compared with Lanzhou, the flatter terrain offers less topographic buffering, so fragmented patches are more directly exposed to solar radiation, wind-driven moisture loss, and surrounding built-up land pressure. As a result, the ecological cost of edge proliferation becomes more pronounced once the structural threshold is crossed.

DIVISION demonstrates pronounced regime shifts, with thresholds around −0.11 (2000), 1.54 (2011), and 1.44 (2022). The upward shift suggests that Baotou’s vegetation system became increasingly vulnerable to spatial separation during the period of rapid urban expansion and land-use reorganization from 2000 to 2011. In an arid climate with limited precipitation and strong wind erosion, fragmentation disrupts resource redistribution and reduces core habitat stability, leading to sharper productivity loss once the separation threshold is exceeded. This effect may also be reinforced by the spatial organization of built-up land and infrastructure, which tends to subdivide green spaces into more isolated units. Although industrial disturbance was not explicitly quantified in this study, the threshold behavior suggests that spatial separation in Baotou is more likely to translate into ecological stress than in the more topographically buffered Lanzhou system. Accordingly, the industrial background of Baotou is interpreted here as part of the broader urban development context that may reinforce fragmentation, rather than as a directly modeled driver of threshold behavior.

AI and COHESION display lower threshold magnitudes compared with Lanzhou (e.g., AI around 0.05 in 2000 and 0.30 in 2022), implying that aggregation effects are weaker and more context-dependent. Unlike Lanzhou, where clustering directly enhances moisture retention in mountainous microenvironments, Baotou’s flatter terrain allows broader ecological exchange, making aggregation less structurally decisive.

PLAND shows consistent positive thresholds around 0.3–0.5, but the slope beyond the threshold flattens in later years, suggesting diminishing marginal returns of area expansion. In an arid urban landscape with long-term industrial development pressure, simply increasing the green proportion cannot fully offset climatic and anthropogenic stress without structural optimization

3.4.4. Divergent Ecological Regimes Between Two Arid Industrial Cities

Although both Lanzhou and Baotou are located in dry climatic zones and share industrial development backgrounds, their threshold behaviors reflect distinct ecological control mechanisms.

Lanzhou’s responses are aggregation-dominated and topography-mediated. As shown in (Figure 8), aggregation-related thresholds remain comparatively stable across the three periods, with AI consistently concentrated around 0.31–0.34 and COHESION shifting only slightly from 0.17 in 2000 to approximately 0.13 in 2022. By contrast, fragmentation-related thresholds such as ED remain negative (around −0.7), indicating that excessive edge exposure persistently suppresses NPP once critical structural limits are crossed. These relatively stable threshold magnitudes suggest that terrain-induced hydrothermal buffering remains the primary regulator in Lanzhou. The mountainous valley structure amplifies the ecological importance of connectivity and clustering, because more cohesive patches are better able to conserve soil moisture, reduce edge-related heat exposure, and stabilize microclimatic conditions. Baotou’s responses are more fragmentation-sensitive. As shown in Figure 9), ED thresholds shift from approximately −0.23 in 2000 to −0.56 in 2022, while DIVISION exhibits larger temporal variation, with threshold values around −0.11 in 2000, 1.54 in 2011, and 1.44 in 2022. These larger shifts indicate that Baotou’s vegetation system is more sensitive to changes in patch separation and edge proliferation than Lanzhou’s. In a flatter and more wind-exposed landscape, fragmented green patches are less protected by topographic buffering and more directly exposed to solar radiation, wind-driven moisture loss, and other forms of urban land-use pressure. This helps explain why fragmentation thresholds in Baotou correspond more readily to ecological decline. Importantly, the city-specific threshold patterns indicate that the same landscape metric may correspond to different ecological consequences depending on the capacity of local green space systems to buffer water limitation and heat stress. Thus, NPP responses in both cities are governed not by universal landscape rules but by threshold-mediated ecological regimes shaped by climatic aridity and terrain configuration. Carbon sequestration enhancement in arid and semi-arid industrial cities, therefore, requires city-specific structural optimization strategies rather than uniform greening targets.

4. Discussion

4.1. Implications of Landscape Configuration for Urban Green Space Carbon Sequestration

The multi-temporal analysis (2000-2011-2022) demonstrates that landscape configuration regulates urban green space carbon sequestration through nonlinear and threshold-dependent mechanisms, and that these mechanisms differ systematically between Lanzhou and Baotou despite their shared industrial background and location in arid to semi-arid regions. The findings indicate that carbon sequestration responses are not determined solely by green space quantity, but by configuration-specific structural regimes shaped primarily by terrain, climatic stress, and urban spatial organization under a long-term industrial development background. More specifically, the regulatory role of landscape configuration in these dry urban environments is not purely structural [60]. Aggregation, connectivity, edge exposure, and fragmentation can modify local hydrothermal conditions by altering moisture retention, evapotranspiration loss, and thermal buffering, which in turn affect vegetation productivity under the water- and heat-limited conditions represented in the CASA-based NPP estimates [61]. In this sense, the identified thresholds can be interpreted as boundaries at which the UGS structure begins to either sustain or undermine local hydrothermal regulation. In the present discussion, industrial development is considered as part of the broader urban context that may reinforce landscape restructuring, rather than as a directly quantified driver of threshold behavior.

4.1.1. Divergent Structural Controls Under Arid and Semi-Arid Conditions

In Lanzhou, aggregation- and connectivity-related metrics consistently exhibit stable and positive threshold effects across all three periods. The relatively stable AI and COHESION thresholds suggest that once a minimum level of spatial cohesion is achieved, NPP increases rapidly and remains structurally supported. This threshold behavior suggests that aggregation in Lanzhou functions as a hydrothermal stabilization mechanism rather than merely a spatial clustering effect. This interpretation is consistent with international studies showing that connected and clustered green spaces can improve ecological stability and support vegetation productivity [62]. However, the present results extend this understanding by showing that, under semi-arid valley conditions, the benefits of aggregation are strongly threshold-dependent and only become pronounced once a minimum level of spatial cohesion is achieved. Once cohesion exceeds a minimum level, green patches are more capable of forming effective ecological cores that retain soil moisture, reduce edge-related thermal exposure, and buffer evapotranspiration loss. Under valley-constrained semi-arid conditions, such structural integration is particularly important because fragmented lowland patches are more vulnerable to heat accumulation and desiccation. This pattern reflects the ecological constraints imposed by Lanzhou’s mountainous valley morphology and semi-arid climate. In such environments, vegetation productivity depends heavily on microclimatic buffering capacity. When green patches are sufficiently aggregated, internal humidity retention improves, heat stress is moderated, and evapotranspiration losses are reduced.

Over time, slight shifts in threshold magnitudes indicate increasing structural sensitivity. As urban expansion progressed between 2000 and 2022, simple increases in green area became less effective unless accompanied by improved spatial integration. At the same time, the growing importance of fragmentation-related thresholds implies that structural disruption increasingly offsets the benefits of area expansion. As edge density and spatial separation increase within the valley floor, green patches lose core-area function and become less able to regulate local heat and moisture conditions [63]. This helps explain why additional green area alone produces limited gains unless accompanied by stronger spatial consolidation. This transition suggests that in topographically constrained semi-arid cities, configuration quality gradually becomes more decisive than area quantity for sustaining ecosystem function.

By contrast, Baotou exhibits fragmentation-sensitive regulation. Across the three periods, ED and DIVISION demonstrate stronger and more temporally variable threshold behaviors, while aggregation-related metrics exert comparatively weaker control. This indicates that Baotou’s green system is less buffered against structural disturbance than Lanzhou’s. This finding is in line with international evidence that fragmentation and edge proliferation tend to reduce ecological performance [18]. At the same time, our results suggest that in arid and more openly exposed industrial cities, these negative effects may be intensified by hydrothermal stress, leading to stronger threshold responses than those commonly reported in more humid urban settings. In a flatter and more open landscape, patch subdivision more directly translates into edge exposure, wind-driven moisture loss, and thermal stress. Although industrial disturbance was not explicitly quantified, Baotou’s long-term industrial development background may provide a contextual explanation for this sensitivity by reinforcing patch isolation and reducing the continuity of vegetated surfaces through continued urban land transformation. As a result, fragmentation thresholds in Baotou more readily correspond to ecological decline. In an arid and relatively flat urban landscape with a long-term industrial development background, vegetation systems are more exposed to wind erosion, thermal stress, and anthropogenic disturbance. Moderate edge interfaces may temporarily enhance environmental exchange; however, once fragmentation surpasses specific limits, NPP declines sharply. The upward shift and amplification of fragmentation thresholds between 2000 and 2022 indicate increasing vulnerability of Baotou’s green space network under continued urban restructuring and land-use pressure.

4.1.2. Context-Dependent Ecological Pathways

The comparative results reveal two distinct configuration-productivity pathways within arid industrial cities. This overall pattern is broadly consistent with international research showing that greater landscape connectivity and lower fragmentation generally support ecosystem functioning, while excessive edge proliferation and patch isolation tend to weaken ecological performance and carbon-related services [3,64]. The present results extend this literature by showing that, in arid and semi-arid industrial cities, these relationships are strongly threshold-dependent and that the same landscape metric may correspond to different ecological consequences under different terrain and hydrothermal settings. Importantly, the same landscape metric may have different ecological meanings under different urban-environmental settings. In Lanzhou, aggregation metrics primarily represent the capacity of UGS to form buffered microclimatic units, whereas in Baotou, fragmentation metrics more directly reflect exposure to hydrothermal stress. This distinction helps explain why threshold magnitudes and response directions differ between the two cities even when the same metrics are examined. In the present analysis, industrial development is interpreted as a contextual factor that may reinforce fragmentation and ecological stress, rather than as a directly quantified explanatory variable. In this sense, Lanzhou and Baotou do not contradict international findings on connectivity and fragmentation; rather, they show that these widely recognized relationships operate through different threshold pathways under dry-climate urban conditions [22,58]. Lanzhou highlights the threshold-dependent ecological benefits of aggregation and connectivity, whereas Baotou shows that fragmentation effects can become more severe in flatter and more exposed environments. Lanzhou follows an aggregation-dominated regime, where carbon sequestration improves once spatial cohesion exceeds critical thresholds. Terrain complexity amplifies the ecological benefits of clustering, making connectivity enhancement a primary strategy. Baotou follows a fragmentation-constrained regime, where excessive spatial separation and edge proliferation quickly undermine vegetation productivity. Under strong climatic stress and a long-term industrial development context, maintaining structural integrity becomes more important than increasing aggregation alone. Importantly, the temporal consistency of nonlinear responses across 2000–2022 indicates that these threshold mechanisms are not short-term phenomena but structural characteristics of the local ecological systems. The thresholds identified in this study, therefore, represent adaptive limits of urban vegetation under arid and semi-arid stress conditions.

From a broader urban carbon balance perspective, it should be acknowledged that UGS carbon sequestration alone cannot offset the full magnitude of anthropogenic CO₂ emissions generated by transportation, heating, industry, and other urban activities. This is especially true for large industrial cities, where total emissions are structurally high and often far exceed the biophysical uptake capacity of urban vegetation [65]. However, this limitation does not reduce the relevance of UGS as a mitigation pathway. Rather, it indicates that UGS should be understood as a complementary component of urban low-carbon transition, alongside direct emission reduction and industrial decarbonization [66]. In this context, the significance of landscape configuration lies not in suggesting that urban greening can neutralize all urban emissions, but in demonstrating that the carbon sequestration efficiency of existing green space systems can vary substantially depending on their spatial structure [54]. For land-constrained industrial cities such as Lanzhou and Baotou, optimizing aggregation, connectivity, and fragmentation control may therefore improve the marginal carbon benefit of UGS while simultaneously enhancing ecological stability, thermal regulation, and environmental quality.

4.2. City-Specific Planning Implications for Lanzhou and Baotou

The multi-period threshold analysis indicates that landscape configuration should be managed within defined structural boundaries to maintain stable carbon sequestration performance. These boundaries can be translated into operational planning indicators rather than abstract ecological principles.

4.2.1. Lanzhou: Maintaining Minimum Aggregation and Limiting Fragmentation

In Lanzhou, aggregation functions as an activation mechanism. The results indicate a minimum aggregation threshold (AI = 0.34 in 2022) below which NPP remains weakly responsive. This suggests that green patches must achieve a baseline level of spatial cohesion before measurable carbon sequestration gains occur. Planning practice should therefore prioritize maintaining AI above this minimum threshold, particularly in valley-floor redevelopment areas where patch fragmentation is common. At the same time, fragmentation-related metrics show clear upper control limits. Edge density stabilizes only beyond approximately −0.74, while DIVISION intensifies its negative impact once exceeding roughly 1.4. These values imply that: new infrastructure should avoid subdividing continuous green cores; road and development layouts should minimize additional edge interfaces within existing UGS; spatial separation of patches should be kept below the identified division threshold. PLAND contributes positively up to approximately 0.58, but further increases yield, diminishing returns if the configuration remains poor. Therefore, area expansion should be accompanied by aggregation enhancement rather than implemented independently. In summary, for Lanzhou, practical implementation requires: maintaining a minimum aggregation level, preventing fragmentation indices from exceeding suppressive thresholds, and integrating area increase with spatial consolidation.

4.2.2. Baotou: Controlling Edge Density and Fragmentation Upper Limits

In Baotou, the dominant constraint is fragmentation rather than aggregation. Edge density shows a clear upper transition point (ED = −0.56 in 2022), beyond which carbon sequestration gains weaken and eventually decline. This indicates that edge density should be treated as a controlled upper limit in planning decisions. Similarly, DIVISION exhibits a critical boundary near zero, after which NPP decreases sharply. Even moderate increases in patch separation impose ecological penalties under arid and industrial conditions. Planning measures should therefore: restrict further subdivision of existing green parcels, prioritize connectivity among isolated patches, and avoid creating narrow, elongated green strips that inflate edge length without increasing core area. Aggregation (AI = 0.05) becomes ecologically meaningful only after surpassing its minimum cohesion threshold, suggesting that spatial consolidation is beneficial but secondary to controlling fragmentation intensity. For Baotou, practical planning emphasis should therefore focus on: setting an upper control limit for edge density, keeping spatial separation below the identified fragmentation boundary, and strengthening network continuity before pursuing additional green area expansion.

4.2.3. Threshold-Guided Implementation Under Industrial Constraints

Under conditions of limited land supply and frequent redevelopment, threshold values should be directly incorporated into spatial planning review and design evaluation procedures rather than treated as abstract ecological findings. Specifically, planning proposals involving green space adjustment, road construction, or land-use conversion should be assessed against configuration-based boundary conditions derived from the identified thresholds.

For Lanzhou, project evaluation should verify whether spatial reorganization increases AI beyond its minimum activation level and ensures that DIVISION and ED remain below their suppressive boundaries. Interventions that subdivide existing continuous patches or reduce aggregation below the identified threshold should be reconsidered or redesigned. In practice, this means prioritizing patch consolidation, corridor extension, and avoidance of internal segmentation within valley-floor green systems.

For Baotou, planning control should emphasize limiting ED and DIVISION within acceptable ranges. New developments should be examined for their potential to increase edge interfaces or isolate existing patches beyond identified transition levels. Projects that create elongated, narrow, or discontinuous green forms should be adjusted to favor compact, connected configurations. Aggregation improvement may be encouraged, but only insofar as it contributes to reducing fragmentation intensity.

In both cities, configuration thresholds can function as quantitative screening criteria during redevelopment approval, ecological redline coordination, and blue-green infrastructure design. By embedding these structural boundaries into planning procedures, landscape restructuring can be guided to remain within ecologically stable regimes, reducing the likelihood of abrupt declines in carbon sequestration capacity while accommodating necessary urban growth.

4.3. Limitations and Future Research Directions

Despite incorporating multi-temporal data (2000, 2011, and 2022), several limitations remain. First, carbon sequestration capacity was represented by NPP derived from the CASA model, which primarily reflects aboveground productivity. Soil carbon pools and belowground processes were not explicitly considered. Although NPP is widely accepted as a proxy for vegetation carbon uptake, it does not fully capture total ecosystem carbon dynamics. Second, the temporal design includes three discrete time points rather than continuous monitoring. While this improves robustness compared with single-year analysis, it does not account for intra-annual phenological variation or long-term climate-driven fluctuations. The identified nonlinear relationships and thresholds therefore represent structurally persistent patterns rather than short-term responses. Third, although Lanzhou and Baotou were selected as representative industrial cities, industrial land distribution and industrial activity intensity were not explicitly incorporated into the model. As a result, industrial development was treated in this study as a contextual condition rather than a directly quantified explanatory variable. This means that the observed inter-city differences were interpreted primarily through landscape configuration, terrain, and hydrothermal constraints, while the role of industrial disturbance was considered indirect and inferential rather than directly tested. Future research should integrate industrial land-use indicators, industrial activity intensity, or proxy measures of anthropogenic disturbance to better clarify how industrial development reshapes green space structure and thereby influences carbon sequestration capacity. Fourth, landscape configuration was assessed using static spatial metrics at each time node. The study does not explicitly model dynamic transition processes such as patch evolution trajectories or temporal connectivity shifts. This also limits the ability to examine how industrial redevelopment or land-use conversion may dynamically alter green space structure over time. Future research should integrate longer and higher-frequency time series, incorporate soil carbon dynamics, explicitly quantify industrial land-use and disturbance intensity, and extend comparative analysis to additional arid and semi-arid industrial cities. Such efforts would improve understanding of threshold stability and the long-term resilience of urban green space carbon systems under climatic and developmental pressures.

5. Conclusions

This study investigated how UGS landscape configuration regulates carbon sequestration capacity in two industrial cities in the upper Yellow River region—Lanzhou and Baotou—using multi-temporal data (2000, 2011, and 2022) and an integrated XGBoost-SHAP framework. The results provide clear answers to the three research questions.

First, the dominant landscape configuration metrics contributing to NPP differ systematically between the two cities. In Lanzhou, aggregation-related metrics, particularly COHESION, consistently exert the strongest influence across all three periods, indicating that spatial cohesion and connectivity are primary determinants of carbon sequestration performance. In contrast, Baotou is more strongly regulated by fragmentation-related metrics, especially ED and DIVISION, reflecting greater structural sensitivity to patch separation and edge dynamics.

Second, landscape-carbon relationships exhibit pronounced nonlinear and threshold-dependent behavior in both cities. Rather than responding gradually to structural change, NPP shifts across identifiable transition points. These thresholds remain broadly consistent across 2000–2022, suggesting that they represent structural characteristics of the urban ecological systems rather than short-term fluctuations. However, the magnitude and direction of threshold responses differ between Lanzhou and Baotou, confirming the presence of city-specific nonlinear regimes.

Third, the divergent carbon sequestration responses are underpinned by distinct configuration mechanisms. Lanzhou follows an aggregation-dominated pathway shaped by mountainous valley morphology and semi-arid climatic constraints, where sufficient spatial cohesion activates microclimatic buffering and stabilizes productivity. Baotou, characterized by flatter terrain and stronger industrial background disturbance, exhibits a fragmentation-constrained regime in which excessive edge proliferation and spatial separation rapidly suppress vegetation productivity. These findings demonstrate that landscape configuration influences urban carbon regulation through context-dependent structural pathways rather than universal rules.

Overall, this study shows that carbon sequestration capacity in arid and semi-arid industrial cities is governed primarily by landscape configuration quality rather than green space quantity alone. By identifying dominant drivers, nonlinear responses, and structural thresholds, the proposed analytical framework advances understanding of how spatial organization shapes urban ecological function and provides a quantitative basis for configuration-sensitive planning strategies.