Spatiotemporal evolution and driving factors of water conservation capacity in Lanzhou City Authors: Huimin Hou, Feng Guo, Pengquan Wang, Di Lu, Changjie Chen, Junxing Bai, Haohao Li, Zhiqiang Bao, Mingyang Qin, Yufei Liu, Xinjian Fan Published: 2026 ABSTRACT AbstractWater conservation services serve as a pivotal ecosystem service for water regulation and storage, which is of great significance for alleviating water scarcity and addressing climate change. Taking Lanzhou City as the research area, this study adopted long-term sequential land use, meteorological and socio-economic datasets from 2000 to 2023. By integrating the PLUS model, InVEST model and the interpretable machine learning framework of XGBoost-SHAP, this paper systematically explored the spatio-temporal evolution, driving mechanisms and future scenario responses of water conservation capacity. The results showed that: (1) Water conservation capacity in Lanzhou experienced prominent interannual fluctuations from 2000 to 2023 without any statistically significant long-term trend. Spatially, it generally presented a distribution pattern with high values concentrated in the northwestern and southern mountainous areas and low values distributed in central river valleys and southeastern regions. (2) Soil saturated water content, precipitation and NDVI were the core positive driving factors, while human activity intensity was the dominant negative driving factor. Complex nonlinear synergistic and antagonistic interactions existed among multiple influencing factors, and their influencing directions varied dynamically with factor intensity. (3) By 2050, the spatial pattern of water conservation capacity under different climatic and socio-economic scenarios will remain basically consistent with that in 2020, while its total volume is highly sensitive to scenario selection. Under the SSP1-2.6 pathway, the ecological protection scenario exhibits the highest water conservation capacity (2872.82 × 104 m3). This study provides reliable references for water resource management, ecological conservation and territorial spatial planning in Lanzhou City.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56962-z.Keywords: Water conservation capacity, Land use change, CMIP6, InVEST model, XGBoost-SHAP modelSubject terms: Climate sciences, Ecology, Ecology, Environmental sciences, Hydrology, Water resources Introduction As a core component of ecosystem service functions, water conservation services reflect the ability of ecosystems to regulate and store water through processes such as vegetation interception, soil water storage, and surface infiltration1. Driven by climate change and intensified human activities2,3, water scarcity and hydrological process imbalance have become prominent bottlenecks restricting regional sustainable development. Especially in arid and semi-arid regions with sensitive ecological environments and prominent contradictions between water supply and demand, the maintenance and improvement of water conservation capacity are even more urgent4,5. Lanzhou is located in the semi-arid region of the upper Yellow River basin. Driven by the Western Development Strategy and the Belt and Road Initiative, the city has undergone rapid urbanization and dramatic shifts in land-use patterns. Characterized by mountain‑valley topography, fragmented landscapes, and intensive human activities, its water conservation capacity is highly sensitive to land-use change and climate change, posing severe challenges to the maintenance of hydrological ecosystem services6. Therefore, exploring the impacts of climate and land use changes on water conservation capacity, scientifically identifying the long-term evolution laws of water conservation capacity in Lanzhou City and the interaction of natural and anthropogenic driving factors, has become an urgent demand for the sustainable utilization of water resources and ecological protection and restoration in this region. How to quantify the impacts of land use and climate change on the variation of water conservation capacity is one of the research hotspots in water resources management. Many scholars have adopted the PLUS model and InVEST model to evaluate the impacts of different land use scenarios and climate change on the spatiotemporal pattern changes of water conservation capacity7,8. With the rapid development of remote sensing technology, geographic information systems, and eco-hydrological models, significant progress has been made in the quantitative assessment and mechanism analysis of water conservation capacity9–11. Among them, the water yield module of the InVEST model has become an important tool for assessing regional water yield, and can effectively simulate water yield services under different land-use scenarios12,13. It should be noted that the direct output of this module is water yield, rather than water conservation capacity in the strict sense. In this study, water conservation capacity was further derived by correcting water yield using the topographic index, soil saturated hydraulic conductivity, and flow velocity coefficient. At the same time, the impact of land use change on water conservation services has attracted increasing attention, and the use of the PLUS model for predicting land use change has been widely applied in eco-environmental research at different scales14,15. Machine learning methods such as XGBoost16 have shown great potential in the field of eco-hydrology. This model can effectively capture the complex nonlinear relationships and interaction effects between variables, while the SHAP interpretation framework17 , based on game theory, provides a unified and reliable attribution explanation for model prediction results, effectively solving the “black box” problem of machine learning models. G.R. et al.18 analyzed the influencing factors of ecosystem services in the Chishui River Basin through the XGBoost-SHAP model, revealing the contribution degree and direction of action of each characteristic variable. V.K. et al.19 used five machine learning models to evaluate the impacts of climate and land use change on water resources in the Kharun River Basin in central-eastern India. Performance evaluation showed that XGBoost was superior to other models and could be used for optimization and scenario prediction. Based on the simulation of hydrological physical processes by hydrological models, coupled with the XGBoost machine learning model to improve prediction accuracy, and using the SHAP method for feature attribution explanation, the accurate prediction and mechanism analysis of changes in water conservation capacity can be achieved20. Combining the PLUS model, InVEST model, and XGBoost-SHAP framework can realize the leap from “accurate assessment” to “mechanism analysis”, providing a new methodological support for deeply understanding the driving mechanism of water conservation capacity. At present, although some relevant studies on water conservation capacity have been carried out in Lanzhou City21,22, they have mainly focused on historical periods. Some studies have assessed the impacts of future land-use development scenarios, yet few have simultaneously considered the coupled driving mechanisms of future climate change and land-use change. In addition, the current status and spatiotemporal characteristics of water conservation capacity remain unclear, the factors governing regional water conservation capacity dynamics are insufficiently identified, and nonlinear explanations for the interactive effects of natural and anthropogenic drivers are still lacking. Therefore, this study takes Lanzhou City as the research object. Based on long-term multi-source data (meteorology, land use, etc.) from 2000 to 2023, we constructed a coupled PLUS–InVEST–XGBoost-SHAP model chain to conduct a continuous analysis of future scenario simulation, ecosystem service response assessment, and driving mechanism identification of water conservation capacity in Lanzhou. This study aims to achieve the following objectives: To quantitatively evaluate the water conservation capacity amount in Lanzhou and analyze its spatio-temporal evolution characteristics; To quantitatively evaluate the water conservation capacity amount in Lanzhou and analyze its spatio-temporal evolution characteristics; To explore the impacts of natural factors and human activity intensity on water conservation capacity based on the XGBoost—SHAP model, and clarify the key driving factors dominating its dynamic changes; To explore the impacts of natural factors and human activity intensity on water conservation capacity based on the XGBoost—SHAP model, and clarify the key driving factors dominating its dynamic changes; To couple the PLUS and InVEST models to predict the regional water conservation capacity in 2050 under the dual background of climate change and land use change. To couple the PLUS and InVEST models to predict the regional water conservation capacity in 2050 under the dual background of climate change and land use change. This study provides scientific support for the improvement of regional water conservation capacity, ecological environment protection and restoration practices. Meanwhile, it offers practical references for water resource governance and territorial spatial planning in analogous arid and semi-arid regions. Study area overview and methods Study area overview Lanzhou City is located in the inland northwest of China, at the junction and transition zone of three major physical geographical regions: the Loess Plateau, the Qinghai-Tibet Plateau, and the Inner Mongolia Plateau. Its total area is about 13,100 km2, and the Yellow River flows through the city, forming a typical river valley basin landform. The city has a temperate continental semi-arid climate, with vegetation mainly consisting of grasslands and xerophytic shrubs, and the ecosystem is fragile. The average annual precipitation is 250–350 mm, and the average annual potential evaporation is more than 1500 mm. Affected by the combined impacts of natural conditions and rapid urbanization, Lanzhou faces a prominent contradiction between water resource supply and demand. It has become a typical representative area in arid and water-deficient regions for investigating water conservation capacity and its responses to climate change and human activities (Fig. 1). Overview of the study area. Data sources The required data and their sources are shown in Table 1. In order to ensure data consistency and comparability, all data were resampled to a unified spatial resolution of 1 km, and the grid data are uniformly projected to the coordinate system Albers_Conic_Equal_Area. Table 1Data sources.Data categoryData nameData sourceDescriptionNatural Resource DataLand use dataChina Land Cover DatasetTiff format (30 m)ElevationGeospatial Data CloudTiff format (30 m)NDVIGeospatial Data CloudTiff format (30 m)Precipitation, temperature, potential evapotranspirationResource and Environmental Science Data CenterAnnual valueSaturated vapor pressureCalculated from temperatureCalculated by Magnus formulaSolar radiationhttps://cds.climate.copernicus.eu/Soil typeWorld Soil DatabaseTiff format (1 km)Soil saturated water contentCalculated by empirical formulaSocio Economic DataNight-time lightEarth Observation GroupDMSP-OLSPopulation densityhttps://landscan.ornl.gov/Tiff format (1 km)Future population dataXinyu, W.23Tiff format (1 km)Future GDP dataDaisuke, M.24Tiff format (1 km)Distance to expresswayOpen Street MapTiff format (30 m)Distance to roadTiff format (30 m)CMIP6 Climate DataPrecipitation, temperature, potential evapotranspirationNational Tibetan Plateau Science Data CenterNC format (1km) Data sources. Research methods This study develops a comprehensive framework to investigate the spatiotemporal evolution of water conservation capacity and its driving mechanisms in Lanzhou City. The framework comprises three integrated modules, as illustrated in the Fig. 2 Flow Chart of Research Methods: trend and spatial autocorrelation analysis, driving factor analysis, and future multi-scenario prediction. Flow chart of research methods. Note: HAI is human activity intensity, NDVI is normalized difference vegetation index, pre is precipitation, DEM is elevation, Slope is slope, E is saturated vapor pressure, sr is solar radiation, and theta_s is soil saturated water content. PLUS model The PLUS model is a comprehensive prediction tool integrating land expansion analysis and patch dynamic simulation25. This study uses the model to carry out multi-scenario simulation research on land use change. Based on historical land use data, the model adopts a land expansion analysis strategy and combines the random forest algorithm to analyze the impact mechanism of multi-dimensional driving factors such as physical geography, socioeconomic factors, and spatial constraints on the expansion of various land types. An adaptive cellular automaton framework is introduced to comprehensively consider inertia coefficients, neighborhood weights, and conversion rules to simulate the patch-level evolution process of land use types in spatial competition26. The Kappa coefficient is used to verify the accuracy of simulation results in historical periods. Selection of Driving Factors. In this paper, five driving factors are selected from natural factors: elevation, slope, precipitation, soil type, and temperature; two driving factors from socioeconomic factors: GDP and population density; and seven driving factors from accessibility factors: distance to city center, distance to railway, distance to tertiary roads, distance to secondary roads, distance to primary roads, distance to county government residences, and distance to rivers. A total of fourteen driving factors were adopted to calculate the land use expansion probability of Lanzhou City27,28. Based on the land use data of Lanzhou City in 2010 and 2015, the spatial distribution map of land use types in Lanzhou City in 2020 is simulated. The Kappa statistical analysis is used to test the accuracy of the simulated and actual results of land use in Lanzhou City in 2020. The obtained Kappa coefficient is 0.803 (> 0.8), FoM is 0.360, and the overall accuracy is 0.931, indicating that the model’s prediction of land use change has high credibility. Scenario Combination. To systematically reveal the combined impact of climate and socioeconomic paths on land use patterns, the study designs an orthogonal experimental framework of climate scenarios and land use strategies. CMIP6 provides climate projections based on the combined scenario framework of Shared Socioeconomic Pathways (SSPs) and Representative Concentration Pathways (RCPs). These scenarios reflect different trajectories of socioeconomic development and corresponding changes in radiative forcing29. This study selects three scenarios: SSP1-2.6, SSP2-4.5, and SSP5-8.5. SSP1-2.6 represents sustainable socioeconomic development with low greenhouse gas emissions, SSP2-4.5 represents socioeconomic development scenario with medium greenhouse gas emissions30, and SSP5-8.5 represents rapid growth scenario with large-scale use of fossil fuels and high greenhouse gas emissions31. In terms of land use scenarios, three development modes are formulated in accordance with regional development orientations, namely the natural development scenario, ecological protection scenario and economic development scenario. The natural development scenario follows the historical land use evolution trends of Lanzhou without additional policy interventions and external constraints. The ecological protection scenario aims to mitigate ecological damage and maintain ecosystem stability. Under this scenario, forestland, grassland and water bodies are restricted from conversion into other land types; the conversion probability of construction land to other land categories rises by 20%, while the probability of other land types being converted into construction land decreases by 30%; additionally, the conversion probability of cultivated land and grassland into forestland increases by 20%. The economic development scenario prioritizes maximum economic benefits and improves land economic value, in which construction land is prohibited from being converted to other land types with its outward conversion probability reduced by 30%, and the conversion probability of other land types into construction land rises by 20%. Full orthogonal combinations of the three climate scenarios and three land use scenarios were established, forming a total of nine future scenarios. These scenarios are applied to simulate the spatial pattern of land use in Lanzhou in 2050, so as to provide basic datasets for subsequent water conservation capacity assessment. Calculation of water conservation capacity This study identified and evaluated water conservation capacity in 2000, 2005, 2010, 2015, 2020, and 2023. The Annual Water Yield Model of InVEST model is used to calculate the water yield of the study area. Based on the Budyko hydrothermal coupling balance hypothesis and water balance principle, this module can effectively simulate the water yield service function of ecosystems under different land uses through grid-scale calculations32. The specific formula is as follows: where represents the annual water yield of each grid cell; is the actual annual average evapotranspiration in grid cell x; is the annual average precipitation of grid cell x. where is the ratio of potential evapotranspiration to precipitation, is a non-physical parameter related to climate and soil properties. where Z represents the Zhang coefficient, an empirical constant, combined with the precipitation characteristics of the study area and relevant literature, the Zhang coefficient in this region was determined to be 2.233,34; denotes plant available water content. where is the vegetation evapotranspiration coefficient of grid cell x; refers to the potential evapotranspiration of grid cell x. The water conservation capacity is obtained by correcting the water yield (Y) using topographic index (TI), soil saturated hydraulic conductivity (), and flow velocity coefficient (velocity)35,36. where is the regional water catchment (mm); is the depth of soil layer (mm); is the percentage of slope (%)37. The flow velocity coefficient is obtained by referring to existing literature21,38 and combining with the actual situation of the study area, and the specific assignment values are shown in Table 2. Soil saturated hydraulic conductivity(, cm/d) is calculated by the Cosby soil transfer formula39. Table 2Flow velocity coefficient of the InVEST model.Land use typeFarmlandForestGrasslandWaterConstruction landUnused landVelocity800180500201220121500 Flow velocity coefficient of the InVEST model. Intensity of human activities The intensity of human activities reflects the degree of development and influence of human production and living activities on the Earth’s land surface, exerting profound and multi-scale impacts on regional ecosystems. The study selects 4 indicators: population density, land use, night lights, and roads to calculate the human activity intensity model of Lanzhou City40. The construction method of human activity intensity refers to the human footprint assessment framework proposed by Venter et al.41. By integrating multi-source human activity pressure indicators, it realizes the spatial and quantitative evaluation of human activity intensity. where HAI represents the human activity intensity index of each grid cell, with a value range of 0–40. , , and denote the contribution scores of four indicators to human activity intensity, namely land use type, population density, nighttime light and road density, respectively. Each indicator is assigned an independent score ranging from 0 to 10. The assigned results are positively correlated with the degree of human activity impact, and higher values indicate greater intensity of human footprint impact. The comprehensive HAI value is calculated by direct summation. Before summation, all indicators are standardized to a unified dimension of 0–10 through a nonlinear scoring method. Referring to the research of Zheng Chengwei et al.42, the intensity of human activities reflected by land use types is assigned: 10 for construction land, 7 for cultivated land, 5 for forestland, 3 for water areas, 2 for grassland, and 1 for unused land. For the DMSP-OLS night-time light data after preprocessing, the grids with DN value equal to 0 are assigned 0, and then the grids with DN value greater than 0 are assigned 1–10 using the quantile classification method43. Roads were classified into four categories according to traffic flow and influence scope: expressways, national highways, provincial highways, and other roads (including county roads, township roads, and unclassified local roads). Buffer analysis was conducted for each road category by setting six distance intervals44: 0–50 m, 50–100 m, 100–300 m, 300–500 m, 500–1000 m, and 1000–5000 m. Human activity intensity was assigned corresponding values based on the distance to roads on both sides, with specific assignment rules listed in Table 3. The assignment followed the principle that higher road grades and shorter distances correspond to higher scores, thereby reflecting stronger human footprint impacts. Referring to Venter et al.'s method41 for assigning population density, areas with a population exceeding 1000 people/km2 are assigned 10, and areas with a small population are assigned using a logarithmic equation: where represents the score assigned to population density, and PD represents the population density value of the grid pixel. Table 3Assignment of traffic accessibility.Category0–50m50–100m100–300m300–500m500–1000m1000–5000mExpressway1097531National Highway1085310Provincial Highway865320Road864210 Assignment of traffic accessibility. The intensity of human activities is divided as follows using the natural breakpoint method: low intensity (0 < HAI ≤ 6), relatively low intensity (6 < HAI ≤ 10), medium intensity (10 < HAI ≤ 14), relatively high intensity (14 < HAI ≤ 20), and high intensity (20 < HAI ≤ 40). Spatial correlation This study employed GeoDa software (Version 1.20) to analyze the local spatial autocorrelation of water conservation capacity in Lanzhou City45. Water conservation capacity is calculated at a 1 km × 1 km grid scale, and a row-standardized first-order Queen contiguity spatial weight matrix was constructed. The significance level is set at α = 0.05, and significance tests are conducted using 999 random permutations as default settings. The null distribution of Moran’s I statistic is established via random permutation. On this basis, the local Moran’s I index of each grid unit is calculated, and the LISA cluster map is further generated. In accordance with local spatial association patterns, grid units passing the significance test (p < 0.05) are classified into four spatial aggregation types: high-high, low-low, high-low and low–high clusters. Units failing to pass the test are categorized as non-significant. The LISA cluster map can visually identify the local spatial pattern and spatial heterogeneity of water conservation capacity across Lanzhou City. XGBoost model XGBoost is an advanced integrated learning algorithm. By serially constructing multiple decision trees and integrating their prediction results, it can excellently capture the nonlinear relationships and complex interaction effects between independent and dependent variables46. To ensure the robustness and reliability of the model, the study divides the data set into training set and test set, optimizes the key hyperparameters of the model using cross-validation and grid search technology, and finally evaluates the model performance with indicators such as coefficient of determination (R2) and root mean square error (RMSE). SHAP value algorithm In the model interpretation stage, to solve the “black box” problem commonly existing in machine learning models, this study introduces the SHAP interpretability framework47. It assigns a specific value to each feature of each sample, which represents the contribution of the feature to the final prediction result of the model. Its additive property enables any prediction output of the model to be decomposed into the sum of SHAP values of all features, thereby realizing the interpretable deciphering of the prediction result48. Based on the calculated SHAP values, a global importance analysis is performed. By calculating the average absolute value of SHAP values of each feature across all samples and sorting them accordingly, the dominant factors that have the most significant impact on the water conservation capacity of Lanzhou City are objectively and quantitatively identified. And by drawing the SHAP dependence diagram of each key factor, the relationship between the value of the factor and its SHAP value is intuitively displayed. Results and analysis Spatiotemporal distribution characteristics of water conservation capacity in Lanzhou City The total water conservation capacity of Lanzhou City in 2000, 2005, 2010, 2015, 2020, and 2023 was 1209.94 × 104 m3, 1377.28 × 104 m3, 1289.93 × 104 m3, 1196.28 × 104 m3, 1362.98 × 104 m3, 1009.79 × 104 m3. From the perspective of spatial distribution (Fig. 3), the water conservation capacity of Lanzhou City shows obvious regional differences and local agglomeration characteristics, generally presenting a spatial differentiation pattern of "higher in the northwestern and southern mountainous areas, and lower in the central river valley and southeastern regions". Distribution of average unit water conservation capacity in Lanzhou City. The high-value areas in each year are mainly distributed in the northwest of Lanzhou City. This area is dominated by the mountainous landform of the eastern extension of the Qilian Mountains. The low-value areas are concentrated in the southeastern river valley plain and urban built-up areas. Overall, the spatial pattern of water conservation capacity in Lanzhou City presents a relatively stable regional structure. The spatial distribution of high, medium, and low value areas is highly related to physical geographical conditions and human activity intensity, reflecting the comprehensive impact of terrain, vegetation coverage, and urban development. Spatiotemporal change analysis of water conservation capacity From 2000 to 2023, the annual average water conservation capacity in Lanzhou fluctuates greatly, and the interannual difference is significant (Fig. 4). From the perspective of interannual variation, the fluctuation amplitude of water conservation capacity is extremely significant, and the numerical range is about 78~159 mm. This severe interannual oscillation clearly indicates the instability of regional hydrological processes. Further analysis of its phased characteristics shows that from 2000 to 2010, there was frequent alternation of positive and negative values, with the most severe fluctuations, which may be related to the frequent occurrence of extreme climate events (such as droughts and rainy years) or rapid changes in land use during this period. After 2010, the fluctuation range narrowed, but there were still obvious interannual differences, and the changes tended to be relatively flat until 2020–2023. Trend of average unit water conservation capacity in Lanzhou City (2000–2023). Based on 24 periods of water conservation capacity data from 2000 to 2023, this study uses the combination of Sen’s slope estimation and Mann–Kendall trend test to conduct a pixel-by-pixel analysis of the spatiotemporal change pattern of water conservation capacity in Lanzhou City49. Based on distinct slope variations and significance levels, the trends were classified into 9 categories (Table 4). Table 4Classification of trend changes.Sen’s slope tsetM–K trend testSignificance of change trendsβ > 0Z ≥ 2.58Extremely significant increase2.58 > Z ≥ 1.96Significant increase1.96 > Z ≥ 1.645Slight increaseZ < 1.645Increased non-significantlyβ = 0ZNo changeβ < 0Z ≥ 2.58Extremely significant decrease2.58 > Z ≥ 1.96Significant decrease1.96 > Z ≥ 1.645Slight decreaseZ < 1.645Decreased non-significant Classification of trend changes. The research results (Fig. 5) show that from 2000 to 2023, the water conservation capacity of Lanzhou City showed an insignificant increasing trend overall. Specifically, the southern mountainous areas and parts of the eastern region are the core areas where water conservation capacity has improved, mainly showing significant and extremely significant increasing trends. The central river valley area of the study area and parts of the northern region, as the most active areas for population agglomeration, urban expansion, and economic activities, have experienced large-scale extremely significant and significant decreasing trends. The substantial increase in impervious surfaces caused by urbanization has seriously weakened the surface infiltration capacity; at the same time, the high-intensity water extraction for industry, agriculture, and domestic use has also brought great pressure on regional water resources, which together have led to the decline of water conservation functions in this area, directly threatening urban water supply security. Areas with insignificant/slight changes are widely distributed in farmland, grassland, and transition zones, reflecting the complexity and uncertainty of the response of ecosystems in this area to climate change and human activities. Spatial change characteristics of water conservation capacity. Spatial analysis of water conservation capacity The local spatial autocorrelation of water conservation capacity in Lanzhou City is studied using GeoDa software, as shown in Fig. 6. The global Moran’s I indices of water conservation capacity in 2000, 2005, 2010, 2015, 2020, and 2023 were 0.667, 0.660, 0.660, 0.671, 0.641, and 0.646 respectively, indicating that the spatial pattern of water conservation capacity in Lanzhou City has a strong spatial correlation. LISA cluster maps of water conservation capacity in Lanzhou City. The water conservation capacity of Lanzhou City shows significant spatial agglomeration characteristics. The high-high agglomeration areas are mainly distributed in mountainous forest areas such as the Liancheng National Nature Reserve in the northwest of Lanzhou City. This area has high vegetation coverage and superior topographic water storage conditions, and the surrounding areas also have high conservation capacity, forming a significant positive spatial correlation. The low-low agglomeration areas are distributed in the urbanized areas and surrounding dry farmland areas in the central and eastern parts of Lanzhou City. The hardening of underlying surfaces caused by urbanization and the low vegetation coverage of dry farmland are the core drivers of the low-value agglomeration of water conservation capacity. The low–high agglomeration areas are mostly located in the edge transition zones of high-high agglomeration areas. These areas show low water conservation values due to local human activity interference (such as reclamation and urbanization erosion), but are surrounded by mountains with high conservation capacity; the high-low agglomeration areas are mainly distributed in the edges of low-low agglomeration areas close to mountains. Their own limited vegetation or water areas form high water conservation values, but are surrounded by low-value areas. Spatial distribution characteristics of human activity intensity in Lanzhou City From 2000 to 2023, the intensity of human activities in Lanzhou City showed an overall increasing tendency, accompanied by the spatial expansion and agglomeration of high-intensity areas. From the perspective of the change in the proportion of different intensity levels (Table 5), the hierarchical structure of human activity intensity has undergone obvious changes. Compared with 2000, by 2023 the proportions of extremely low-intensity areas and low-intensity areas have decreased from 67.1 and 21.2% to 55.9 and 16.5% respectively, while the proportions of medium, high, and extremely high-intensity areas have increased by 4.0%, 6.3%, and 5.5% respectively, reflecting the overall trend of expanding scope and increasing intensity of human activity interference. Table 5Proportion of different levels of human activity intensity.Extremely lowLowMediumHighExtremely high200067.05%21.09%6.97%3.57%1.31%200576.64%15.42%4.70%2.08%1.17%201075.53%15.04%5.35%2.68%1.40%201571.39%15.22%6.68%4.44%2.26%202060.69%16.52%9.60%8.30%4.89%202355.87%16.47%10.94%9.88%6.84% Proportion of different levels of human activity intensity. From the perspective of spatial distribution, the intensity of human activities in Lanzhou City shows a significant characteristic of "high density in the central urban area and low density in the peripheral areas"(Fig. 7). High-intensity and extremely high-intensity areas are obviously agglomerated in the central urban area and along the Yellow River. Among them, the extremely high-intensity areas are concentrated in the core urban areas such as Chengguan District and Qilihe District. With the continuous advancement of urbanization and infrastructure construction, the scope of such areas continues to expand; the high-intensity areas are mainly distributed around the extremely high-intensity areas and gradually spread to peripheral areas such as Anning District and Xigu District; the scope of medium-intensity areas has expanded significantly, mostly distributed in the periphery of the urban area and along major transportation lines, showing an outward expansion from the urban core; the low-intensity areas are mainly scattered in the periphery of medium and high-intensity areas, with less influenced by the central urban area; the extremely low-intensity areas are mainly distributed in the mountainous and hilly areas in the northern and southern parts of the city, where the ecological environment is less disturbed, but their overall proportion shows a decreasing trend during the study period. In general, the spatial distribution of human activity intensity in Lanzhou City not only reflects the radiation effect and driving role of the central urban area, but also reveals the characteristics of human activities spreading along transportation routes and expanding to the periphery. It can be note that areas with high human activity intensity tend to exhibit a certain degree of spatial overlap with regions of relatively low water conservation capacity. This relationship will be quantitatively analyzed using the XGBoost-SHAP model. Spatial distribution of human activity intensity in Lanzhou City from 2000 to 2023. Analysis of driving factors of water conservation capacity Analysis of the importance of influencing factors Eight indicators are selected from both social and natural factors to analyze the driving factors of water conservation capacity in Lanzhou City: human activity intensity (HAI), precipitation (pre), elevation (DEM), slope (Slope), saturated vapor pressure (E), solar radiation (sr), normalized difference vegetation index (NDVI), and soil saturated water content (theta_s). Systematic training and predictive performance evaluation of the XGBoost model are conducted at a 1 km × 1 km grid scale using time-series datasets spanning 2000 to 2023. The entire dataset is divided into 80% training set and 20% test set, and five-fold cross-validation is adopted on the training set for parameter optimization. The coefficient of determination (R2) and root mean square error (RMSE) are employed to quantify model performance. The results show that the XGBoost model achieved satisfactory fitting performance and stable predictive capability across all years. As shown in Table 6, the R2 values exceeded 0.87 in all periods, indicating that the model possessed strong explanatory power for data variation. The RMSE values ranged from 2.153 to 3.545 mm, demonstrating high prediction accuracy. According to the average outcomes of five-fold cross-validation, no obvious overfitting was observed, and the model exhibited favorable generalization ability, which is well applicable to identify the driving factors of water conservation capacity in Lanzhou City. Table 6Evaluation metrics of XGBoost model (average values of fivefold cross validation).Year200020052010201520202023R20.9130.8790.8640.9110.8920.925RMSE (mm)2.7233.4793.5452.7603.3552.153 Evaluation metrics of XGBoost model (average values of fivefold cross validation). Driving factors affecting water conservation capacity in Lanzhou were quantitatively analyzed via the XGBoost-SHAP model, with detailed results presented in Table 7 and Fig. 8. From the temporal perspective, soil saturated water content (theta_s) achieves the highest annual mean SHAP value, confirming its dominant influence on the spatiotemporal variations of water conservation capacity. Precipitation (pre) ranks second consistently across all years and acts as a stable core driving factor. The explanatory contribution of the normalized difference vegetation index (NDVI) shows an overall fluctuating upward trend. The impact of human activity intensity (HAI) displays distinct phased characteristics. Its influence stays relatively weak in 2005 and 2010, while the mean SHAP values rise markedly in 2000, 2015, 2020 and 2023. This indicates that the interfering effect of human activities on water conservation capacity rises intermittently alongside ongoing urbanization. Slope presents a relatively strong effect in 2010 and maintains a moderate and stable secondary driving role in other periods. In comparison, elevation (DEM), solar radiation (sr) and saturation vapor pressure (E) keep low mean SHAP values throughout the study period, suggesting their limited contributions to the dynamics of water conservation capacity. Table 7Contribution rate of SHAP value of driving factors of water conservation capacity in Lanzhou (2000–2023). Note: HAI is human activity intensity, NDVI is normalized difference vegetation index, pre is precipitation, DEM is elevation, Slope is slope, E is saturated vapor pressure, sr is solar radiation, and theta_s is soil saturated water content.Feature200020052010201520202023theta_s39.34%41.52%42.31%40.85%40.01%40.22%pre15.82%17.78%16.81%16.13%14.08%15.61%HAI14.46%7.11%7.32%14.44%12.52%11.95%NDVI13.99%14.05%9.06%10.26%12.83%11.48%Slope7.52%10.02%10.59%6.91%7.78%8.32%DEM3.67%4.20%5.83%5.26%6.00%5.40%sr3.00%2.92%4.23%2.31%3.99%3.65%E2.19%2.41%3.85%3.83%2.79%3.36% Contribution rate of SHAP value of driving factors of water conservation capacity in Lanzhou (2000–2023). Note: HAI is human activity intensity, NDVI is normalized difference vegetation index, pre is precipitation, DEM is elevation, Slope is slope, E is saturated vapor pressure, sr is solar radiation, and theta_s is soil saturated water content. SHAP Importance Evaluation of driving factors from 2000 to 2023. Note: HAI is human activity intensity, NDVI is normalized difference vegetation index, pre is precipitation, DEM is elevation, Slope is slope, E is saturated vapor pressure, sr is solar radiation, and theta_s is soil saturated water content. Interaction analysis of influencing factors To deeply explore the combined impact mechanism of various driving factors on water conservation capacity, the study ranked all factor pairs in each year in descending order of interaction intensity based on the average absolute values of SHAP interaction values. The pair with the highest intensity was defined as the most significant annual interaction pair, and further analyzes the intensity and direction of their interaction effects. A SHAP value of 0 indicates no interaction effect; greater than or less than 0 indicates positive or negative interaction effects respectively. The results of the most significant interaction effects from 2000 to 2023 are shown in Fig. 9. The dominant interaction combinations in each year are significantly different, and the interaction effects show significant threshold dependence and direction reversibility. Interaction analysis of the most significant influencing factors from 2000 to 2023. Note: HAI is human activity intensity, NDVI is normalized difference vegetation index, pre is precipitation, and theta_s is soil saturated water content. The most prominent interactive combination in 2000, 2005, 2020 and 2023 consistently lies between NDVI and soil saturated water content. When NDVI falls below 0.6, the SHAP interaction value approaches zero and the interactive effect remains weak. Once NDVI exceeds 0.6, the interaction diverges distinctly: high soil saturated water content generates positive interaction effects, while low soil saturated water content imposes inhibitory impacts. In 2010, precipitation and soil saturated water content dominate the interaction relationship. When annual precipitation is lower than 400 mm, the increase in soil saturated water content exerts positive interactive effects on water conservation capacity. When precipitation exceeds 400 mm, their joint effect shifts from positive to negative with rising precipitation. In 2015, human activity intensity (HAI) and NDVI form the primary interactive pair. Slight positive interaction exists under low-intensity human disturbance, whereas the interaction turns negative once HAI surpasses 25. Notably, HAI is a comprehensive index integrated from multiple indicators including land use type, population density, nighttime light intensity and road distribution, and its values present discrete distributions. Accordingly, the vertical point bands and local clustering patterns observed in the figures partially reflect the discrete attribute of HAI rather than purely ecological processes. Dependence relationship between influencing factors and water conservation To identify the nonlinear response characteristics of water conservation capacity to various driving factors, this study focused on the 2023 dataset and performed quadratic polynomial fitting for each factor based on SHAP values. The point where the first derivative of the fitted curve equals zero was defined as the candidate threshold50. A candidate threshold was marked as an effective threshold (blue dashed lines in Fig. 10) only if it fell within the actual observed range of the corresponding factor and the quadratic term coefficient was non-zero. No dashed lines were plotted for variables without obvious threshold effects. This unified criterion was applied to all driving factors to guarantee the comparability of analytical results. Dependence relationship between driving factors and water conservation capacity in 2023. Note: HAI is human activity intensity, NDVI is normalized difference vegetation index, pre is precipitation, DEM is elevation, Slope is slope, E is saturated vapor pressure, sr is solar radiation, and theta_s is soil saturated water content. The red solid line represents the trend line of second-order polynomial fitting. The blue vertical dashed line indicates the identified threshold. As shown in Fig. 10, each driving factor exhibits distinct nonlinear responses to water conservation capacity. In terms of topographic and soil factors, slope presents a unimodal pattern with an inflection point at 56.82°. The SHAP value rises gently with increasing slope below this threshold, yet declines gradually beyond it. Elevation (DEM) presents a sustained positive SHAP contribution. Soil saturated water content (theta_s) changes at the threshold of 0.54 m3/ m3, the driving effect strengthens with the growth of soil saturated water content below this value. After exceeding the threshold, the driving effect tends to flatten out. For climatic factors, the positive contribution of precipitation (pre) increases as precipitation rises below 432.82 mm, and slightly declines once this value is surpassed. No obvious threshold is detected for solar radiation (sr) within the research range, which shows a weak negative linear correlation and exerts gradually diminishing inhibitory impacts with increasing radiation. Taking 10.39 mm as the dividing point, the SHAP value of saturation vapor pressure (E) decreases with its rise below this value and tends to be stable afterwards. With regard to human activity and ecological factors, human activity intensity (HAI) generally imposes negative effects on water conservation capacity, with a critical point at 7.68. Below this value, its SHAP value rises slightly with minor fluctuations, while above it, the SHAP value decreases sharply and the negative influence becomes stronger. The influencing trend of NDVI alters at 0.30, and its positive promotion effect on water conservation capacity is markedly enhanced beyond this threshold. Based on the above analysis using data of 2023, soil saturated water content, precipitation, NDVI and elevation within high-value ranges act as positive driving factors. By contrast, human activity intensity, saturation vapor pressure and solar radiation in high-value ranges serve as negative driving factors, while slope displays interval-dependent influencing characteristics. Multi-scenario simulation of land use and water conservation capacity for 2050 Analysis of Multi-scenario Land Use changes in 2050 Based on the orthogonal experimental framework, the spatial distribution of land use in Lanzhou City under 9 scenarios in 2050 is simulated (Fig. 11, Table 8). The simulation results reveal that under the same development scenario, differences in land use patterns among alternative SSP climate pathways are relatively minor. In contrast, under the same climate pathway, land use patterns exhibit much more pronounced discrepancies across different development scenarios. Spatial distribution of land use under different climate scenarios and development scenarios in 2050. Notes: NDS is Natural development scenario; EPS is Ecological protection scenario; EDS is Economic development scenario. Table 8Land use area and proportion in Lanzhou City in 2020 and under Different Scenarios in 2050. Notes: NDS is Natural development scenario; EPS is Ecological protection scenario; EDS is Economic development scenario.FarmlandForestGrasslandWaterConstruction landUnused land2020 year Area(km2)1837.82553.3410,479.7431.81167.84130.27 Proportion (%)13.92%4.19%79.39%0.24%1.27%0.99%NDS2050ssp126 Area (km2)1757.21697.2910,224.4231.81237.40252.69 Proportion (%)13.31%5.28%77.45%0.24%1.80%1.91%2050ssp245 Area (km2)1760.14697.2910,221.4931.81237.40252.69 Proportion (%)13.33%5.28%77.43%0.24%1.80%1.91%2050ssp585 Area (km2)1760.14697.2910,221.4931.81237.40252.69 Proportion (%)13.33%5.28%77.43%0.24%1.80%1.91%EPS2050ssp126 Area (km2)1820.94583.9610,542.8420.87129.95102.27 Proportion (%)13.79%4.42%79.87%0.16%0.98%0.77%2050ssp245 Area (km2)1820.94583.9610,510.6122.18145.49117.64 Proportion (%)13.79%4.42%79.62%0.17%1.10%0.89%2050ssp585 Area (km2)1820.94583.9610,510.1822.31145.50117.93 Proportion (%)13.79%4.42%79.62%0.17%1.10%0.89%ENS2050ssp126 Area (km2)1886.61556.9810,416.3131.81180.66128.45 Proportion (%)14.29%4.22%78.91%0.24%1.37%0.97%2050ssp245 Area (km2)1894.87548.2410,416.3131.81180.66128.92 Proportion (%)14.35%4.15%78.91%0.24%1.37%0.98%2050ssp585 Area (km2)1892.41554.8710,416.3131.81180.66124.76 Proportion (%)14.34%4.20%78.91%0.24%1.37%0.95% Land use area and proportion in Lanzhou City in 2020 and under Different Scenarios in 2050. Notes: NDS is Natural development scenario; EPS is Ecological protection scenario; EDS is Economic development scenario. 2050 ssp126 2050 ssp245 2050 ssp585 2050 ssp126 2050 ssp245 2050 ssp585 2050 ssp126 2050 ssp245 2050 ssp585 Under the natural development scenario, the areas of cultivated land and grassland decrease under all SSP paths, while construction land and forestland expand simultaneously, with area proportions of 1.80% and 5.28% respectively. This indicates that without specific policy intervention, may place pressure on agricultural production space and ecological stability. The ecological protection scenario prioritizes ecological conservation and ecosystem stability. Under the SSP585 scenario, the grassland area accounts for about 79.62%. Under the SSP126 and SSP245 scenarios, the grassland area growth is lower than that under the SSP585 scenario, but still shows an increasing trend; under the three climate scenarios, the construction land shows a decreasing trend. The economic development scenario is characterized by the simultaneous expansion of construction land and cultivated land. Under this scenario, construction land expands mainly by occupying cultivated land, while the area of cultivated land is maintained and even increased through the conversion of grassland and unused land. The cultivated land area is the only one that achieves growth among all scenarios, with the largest increase under the SSP245 scenario, increasing to 14.35%; the expansion of construction land is relatively smaller than that under the natural development scenario, reflecting the emphasis on intensive land use and agricultural security under the economic growth target. Multi-scenario analysis of water conservation capacity changes in 2050 By combining land use data simulated by the PLUS model with the InVEST model, this study evaluates the spatiotemporal variations in average grid-scale water conservation capacity of Lanzhou in 2050 under nine orthogonal scenarios (Fig. 12, Table 9). The results indicate that the spatial distribution of water conservation capacity in Lanzhou in 2050 remains similar to that in 2020, generally presenting a pattern of ‘higher in the northwestern and southern mountainous areas, and lower in the central river valley and southeastern regions.’ Although the total volume varies across scenarios, the overall spatial pattern stays relatively stable. Spatial distribution of average unit water conservation capacity in Lanzhou City under different climate scenarios and land use patterns in 2050. Notes: NDS is Natural development scenario; EPS is Ecological protection scenario; EDS is Economic development scenario. Table 9Total water conservation amount in 2020 and under different scenarios in 2050, with change rates of 2050 relative to 2020.Total amount/104 m3Change rate from 20202020 Year1362.98NDS ssp1262878.56111.20% ssp2451561.4014.56% ssp5851650.4921.09%EPS ssp1262874.82110.92% ssp2451566.7314.95% ssp5851656.7321.55%EDS ssp1262851.27109.19% ssp2451551.0913.80% ssp5851640.7420.38% Total water conservation amount in 2020 and under different scenarios in 2050, with change rates of 2050 relative to 2020. Under the natural development scenario, the total water conservation capacity shows an increasing trend under all three climate pathways compared with 2020. The growth rate of water conservation capacity under the SSP585 scenario is lower than that under the SSP126 scenario; the increase under the SSP245 scenario is relatively small, increasing by about 14.56% compared with the water conservation capacity in 2020. Under the ecological protection scenario, the water conservation capacity under SSP126 (2872.82 × 104 m3) is the highest among the three climate pathways, which is also the highest value across all nine scenarios, the high-value areas are concentrated in ecological reserves and key water sources. Under the economic development scenario, the total amount of water conservation under ssp245 scenario is the lowest among the three climate paths, about 1551.09 × 104 m3, and the water conservation capacity under the SSP126 scenario exceeds that under the SSP585 scenario. Discussion Spatio-temporal pattern of water conservation capacity Water conservation quantity in Lanzhou City presents distinct spatial heterogeneity and temporal fluctuations. Spatially, high-value areas are concentrated in the mountainous and forested zones in the northwest, while low-value areas are steadily distributed in the central river valley basin as well as built-up areas including Chengguan District and Qilihe District. This spatial pattern indicates that natural endowments and human activities jointly shape the spatial differentiation of water conservation capacity51, and further reveals the substantial spatial overlap between low-value zones and areas with intense human activities. Temporally, the total water conservation volume in Lanzhou City exhibited obvious interannual fluctuations during 2000–2023, with only a slight and statistically insignificant increasing trend. This finding is consistent with the trend analysis results of water yield in Lanzhou conducted by Zhang Tianhou et al.34, and also highlights the necessity of long-term dynamic monitoring in rapidly urbanizing regions. Notably, the overall water conservation level was relatively high in 2005 and 2020, which was probably positively affected by precipitation conditions and ecological restoration initiatives. By contrast, obvious declines occurred in 2010 and 2023; in particular, fluctuating water conservation functions emerged in some transitional zones, which may be attributed to climatic drought and land use changes. Spatial autocorrelation analysis further verified that water conservation quantity exhibits significant high-high and low-low agglomeration characteristics, reflecting the dual spatial structure of "ecological barrier-urban development" typical of arid valley cities. Accordingly, regional water conservation protection should be based on natural geographical conditions, prioritize the protection of core mountainous ecological areas, and strictly restrict the encroachment of urban expansion on ecological space. This study can provide a scientific reference for zoned management and control of ecological space in arid regions. Dominant effects of natural factors and negative correlations of human activities From an ecohydrological perspective, water conservation capacity is a comprehensive manifestation of water cycles within the Soil–Plant-Atmosphere Continuum (SPAC), and its evolution is governed by the coupling of multiple internal system factors and synergistic effects of external human disturbances52. In this study, soil saturated water content, precipitation and NDVI were identified as core positive driving factors, while human activity intensity served as the dominant negative driving factor. These findings are highly consistent with the hydrological mechanisms of SPAC systems in arid and semi-arid regions53,54. Precipitation stands out as the most critical climatic driver, indicating that its interannual variability is the pivotal climatic factor dominating the temporal dynamics of water conservation capacity55,56 . Under the dominance of natural factors, the negative disturbance effects of human activities become increasingly prominent. From 2000 to 2023, human activity intensity in Lanzhou kept rising, and high-intensity areas such as Chengguan District and Qilihe District showed obvious spatial agglomeration characteristics. Such spatial agglomeration pattern is highly coupled with low-low clustering areas of water conservation capacity, forming a distinct spatial correlation characterized by high-intensity disturbance and low hydrological function. This spatial linkage suggests that intensive human activities are a core contributor to the decline of local water conservation capacity57,58.The XGBoost-SHAP model further confirmed a negative correlation between human activity intensity and water conservation capacity. Its multi-year average SHAP contribution rate reaches 11.30%, which makes it a crucial negative driving factor. This result quantitatively clarifies the inhibitory effect of urban expansion on ecosystem water conservation function59,60. SHAP interaction analysis further reveals the complex nonlinear interaction mechanisms among multiple influencing factors. The interactive effects between NDVI and soil saturated water content show evident spatial differentiation, which essentially reflects the synergistic and competitive relationships between vegetation and soil moisture, and directly embodies the water redistribution process within the Soil–Plant-Atmosphere Continuum (SPAC). When NDVI exceeds 0.6 with high vegetation coverage, the hydrological functions of vegetation are significantly enhanced and differentiated. In areas with high soil saturated water content, sufficient water resources sustain vegetation growth, and vegetation improves water conservation capacity via soil consolidation, water retention and infiltration promotion, thereby generating a synergistic effect. By contrast, in regions with low soil saturated water content, water scarcity occurs, and transpiration water consumption of densely covered vegetation aggravates soil water deficit, leading to competitive negative interactions. There exists a distinct threshold for the interaction between precipitation and soil saturated water content, which is consistent with the regional division west of the 400 mm isohyet specified in China’s 14th Five-Year Plan Outline. When annual precipitation is below 400 mm, effective rainfall infiltration replenishes soil moisture, and the two factors jointly promote water conservation capacity. Once precipitation exceeds 400 mm, the soil gradually reaches saturation, accompanied by increased invalid surface runoff and evaporative loss, thus turning into an inhibitory effect. A prominent threshold effect is also observed in the interaction between human activity intensity and NDVI. Weak synergism occurs under low-intensity human disturbance, while intensive human activities weaken the water conservation function of vegetation and transform the interaction into a negative effect. These findings demonstrate that variations in water conservation capacity are not driven linearly by individual factors, but result from synergistic or antagonistic interactions among multiple factors, whose action direction and magnitude change dynamically with variations in factor levels61,62. To further verify the reliability of long-term interactive relationships, this study selected 2023, the year with the lowest water conservation capacity during the research period, to conduct single-factor dependence analysis. This year exhibited the most prominent eco-hydrological stress characteristics, which effectively eliminates the masking effect of ecological buffering in high-value years on threshold identification. Its factor distribution and interactive characteristics are highly consistent with long-term trends, ensuring the credibility of threshold determination. The single-factor dependence results of 2023 further confirm the core conclusions derived from long-term analysis. Driving factors show distinctly different response patterns towards water conservation capacity. Clear critical thresholds exist for human activity intensity, precipitation, NDVI and soil saturated water content. The influencing direction and intensity change in segmented forms alongside factor variations, which are fully consistent with the threshold characteristics obtained from long-term interaction analysis. This further verifies the inherent law of threshold sensitivity and multi-factor coupling governing water conservation capacity in arid regions. Future management of water conservation capacity should shift from single-factor regulation to comprehensive regulation under threshold constraints. It is essential to prioritize improving soil water storage capacity in core ecological areas, increasing vegetation coverage in ecological restoration zones, and strictly limiting human activity intensity in urban areas, so as to realize differentiated and targeted ecological governance. Mechanisms and implications of responses to future scenarios By 2050, the spatial pattern of water conservation capacity under diverse climatic and socio-economic development scenarios will remain broadly consistent with that in 2020. Nevertheless, the total water conservation volume is highly sensitive to development pathways and will increase to varying degrees compared with the level in 2020, demonstrating the synergistic regulation effects of climate conditions and development strategies. Under the same climatic pathway, the total water conservation capacity reaches the highest level under the ecological protection scenario, followed by the natural development scenario, and is the lowest under the economic development scenario22. This outcome reveals the potential impacts of diverse land use policies on water conservation capacity, indicating that the ecological protection scenario is more applicable to the future development of Lanzhou City. For the same development pathway, water conservation capacity under the SSP126 scenario is superior to that under the SSP245 and SSP585 scenarios63, suggesting that low-carbon development serves as a long-term strategic option to guarantee regional water resource security. Despite the uncertainties induced by climate change, proactive land use management and ecological protection policies, such as restricting the expansion of construction land and promoting the restoration of forest and grassland, are the most effective interventions to strengthen regional climate resilience and improve hydrological regulation services. The land use allocation scheme corresponding to the ecological protection scenario under the low-emission SSP126 pathway can be taken as a priority reference for multi-scenario comparison in the territorial spatial planning of Lanzhou City. It facilitates the coordinated delineation of ecological protection redlines, regulation of urban development boundaries and protection of permanent basic farmland. Limitations and prospects The integrated framework consisting of the InVEST model, XGBoost-SHAP and PLUS model adopted in this study exhibits favorable applicability in quantifying water conservation capacity, clarifying driving mechanisms and predicting future land use changes, and provides scientific references for water resource management in arid and semi-arid regions. Nevertheless, several limitations still exist in this research. Firstly, the water yield module of the InVEST model essentially simulates water yield rather than water conservation capacity in a strict sense. Although corrections were made using topographic index, soil hydraulic conductivity and other parameters, the model fails to fully consider the influences of groundwater, surface runoff and real-time watershed monitoring data. In addition, the flow velocity coefficient was determined based on empirical values, which inevitably introduces certain simulation errors. Further integration of groundwater modules and watershed monitoring data is recommended in future studies to improve the simulation accuracy and physical consistency of water conservation capacity, so as to more accurately characterize regional hydrological processes. Secondly, the XGBoost-SHAP model achieves satisfactory fitting performance for interannual variations in water conservation capacity with the coefficient of determination R2 exceeding 0.87 in test datasets, and can effectively capture nonlinear relationships among variables. However, this approach cannot explicitly address spatial autocorrelation and spatial heterogeneity64, and is incapable of revealing the spatial differentiation of driving patterns across different sub-regions. Geographically weighted machine learning65 and the GeoShapley algorithm66 can be introduced for in-depth exploration in follow-up research. Thirdly, although the selected driving factors are representative, some potentially crucial factors such as soil texture, groundwater level and agricultural irrigation intensity are not incorporated. Moreover, the coupling between CMIP6 datasets and the PLUS model adopted for future climate scenario prediction remains preliminary, without fully accounting for the dynamic feedback of human water consumption behaviors under diverse scenarios. Accordingly, future studies are supposed to expand the driving factor system and incorporate more geographical and socioeconomic variables to systematically reveal the multi-dimensional driving mechanisms behind the evolution of water conservation capacity. Combined with multi-source high-resolution remote sensing data and diverse CMIP6 climate models, the simulation accuracy and depth of mechanism analysis can be further enhanced. In practical policy-making and spatial planning, the application of model results should be closely integrated with regional planning practices, ecological restoration projects and local water resource management experience. Conclusions This study systematically evaluated the spatiotemporal evolution (2000–2023) and future changes (2050) of water conservation capacity in Lanzhou City using an integrated PLUS‑InVEST‑XGBoost‑SHAP framework. The main conclusions are as follows: From 2000 to 2023, the variations in water conservation capacity across Lanzhou were dominated by interannual fluctuations without a stable long-term trend. Its spatial pattern remained stable, presenting a distribution pattern of high values in the northwestern and southern mountainous areas and low values in the central river valleys and southeastern regions. Significant high-high and low-low spatial agglomeration characteristics were observed. Spatial differentiation was primarily governed by natural endowments, while urbanization intensified the agglomeration of low-value areas. From 2000 to 2023, the variations in water conservation capacity across Lanzhou were dominated by interannual fluctuations without a stable long-term trend. Its spatial pattern remained stable, presenting a distribution pattern of high values in the northwestern and southern mountainous areas and low values in the central river valleys and southeastern regions. Significant high-high and low-low spatial agglomeration characteristics were observed. Spatial differentiation was primarily governed by natural endowments, while urbanization intensified the agglomeration of low-value areas. Soil saturated water content, precipitation and NDVI were identified as the core positive driving factors of water conservation capacity, with their multi-year average SHAP contribution rates reaching 40.71%, 16.04% and 11.94% respectively, among which precipitation acted as the most vital climatic driving factor. Human activity intensity functioned as the predominant negative driving factor. Complex nonlinear synergistic and antagonistic interactions existed among multiple factors, and their influencing directions shifted dynamically with changes in factor intensity. Soil saturated water content, precipitation and NDVI were identified as the core positive driving factors of water conservation capacity, with their multi-year average SHAP contribution rates reaching 40.71%, 16.04% and 11.94% respectively, among which precipitation acted as the most vital climatic driving factor. Human activity intensity functioned as the predominant negative driving factor. Complex nonlinear synergistic and antagonistic interactions existed among multiple factors, and their influencing directions shifted dynamically with changes in factor intensity. By 2050, the spatial pattern of water conservation capacity under various scenarios will be generally consistent with that in 2020, whereas its total amount is highly sensitive to scenario settings. Among all scenarios, the ecological protection scenario under SSP1-2.6 shows the highest projected water conservation volume (2872.82 × 104 m3), which is regarded as the optimal development pathway for regional water resource security and ecological conservation. By 2050, the spatial pattern of water conservation capacity under various scenarios will be generally consistent with that in 2020, whereas its total amount is highly sensitive to scenario settings. Among all scenarios, the ecological protection scenario under SSP1-2.6 shows the highest projected water conservation volume (2872.82 × 104 m3), which is regarded as the optimal development pathway for regional water resource security and ecological conservation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Author contributions Conceptualization, H.H. and F.G.; methodology, H.H. and F.G.; software, P.W.; validation, D.L., J.B. and C.C.; formal analysis, C.C. and H.L.; investigation, data curation, F.G.; writing—original draft preparation, H.L and Z.B.; writing—review and editing, Y.L and M.Q.; visualization, P.W.; supervision, H.H.; project administration, H.H.; funding acquisition, H.H, P.W. and X.F. All authors have read and agreed to the published version of the manuscript. Funding This study was jointly supported by the Gansu Provincial Water Conservancy Scientific Experimental Research and technology promotion project (23GSLK039; 25GSLK045; 25GSLK070); Major special plan for science and technology in Gansu Province(22ZD6WA043); Lanzhou Science and Technology Plan Project (Science and Technology Support Special) (2024–3-84). Data availability The data supporting this study are available from the corresponding author upon reasonable request. Competing interests The authors declare no conflicts of interest. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement The data supporting this study are available from the corresponding author upon reasonable request.