Academic Journal of Science and Technology ISSN: 2771-3032 | Vol. 12, No. 2, 2024 242 Exploring the Influencing Factors of Net Ecosystem Productivity (NEP) Based on Random Forest and SHAP Zhen He 1, 2, Weibo Yuan 3, * 1 College of Computer, Qinghai Normal University, Xining 810016, China 2 The State Key Laboratory of Tibetan Intelligent Information Processing and Application, Xining 810008, China 3 Qinghai Provincial Key Laboratory of Medicinal Plant and Animal Resources of Qinghai-Tibet Plateau, School of Life Sciences, Qinghai Normal University, Xining, Qinghai, 810008, China * Corresponding author: Weibo Yuan Abstract: As global climate change intensifies, Net Ecosystem Productivity (NEP) serves as a crucial indicator for measuring the carbon absorption and release of ecosystems, playing a vital role in understanding the carbon cycle and shaping climate policy. The Northwestern region of China, characterized as a typical inland arid and semi-arid area, is particularly sensitive to global changes. Understanding the environmental drivers of NEP in this region is critical for both regional ecological protection and global environmental management. This study utilized environmental data from five provinces in Northwestern China, applying the Random Forest (RF) model and SHapley Additive exPlanation (SHAP) method to analyze the primary environmental factors influencing NEP. The research integrated climatic data (including temperature, precipitation, wind speed, and solar radiation), soil characteristics such as pH, organic carbon content, and soil texture, topographic attributes elevation and slope, and a Human Footprint. The RF model identified significant environmental factors impacting NEP, and SHAP values were used to explain the specific contributions of these factors. Furthermore, multiple linear regression analysis revealed interactions among environmental factors. The results indicate that solar radiation (Srad), precipitation, topsoil reference bulk density (T_REF_BULK), temperature, and the topsoil clay fraction (T_CLAY) significantly influence NEP. Notably, interactions between aspect and T_CLAY, aspect and T_REF_BULK, as well as the human footprint (HFP) and Srad, also show significant impacts on NEP. This study confirms that solar radiation, precipitation, soil characteristics, and human activities are the primary environmental drivers of NEP in the Northwest region of China, with solar radiation playing the most critical promotional role. The findings not only provide a scientific basis for the management of ecosystems in Northwestern China but also offer references for formulating global climate change mitigation strategies and estimating the global carbon budget. Future research should further explore the specific mechanisms and interactions of these drivers across different ecosystems to more comprehensively understand and predict the trends in NEP changes. Keywords: Net Ecosystem Productivity; Random Forest; SHAP Values; Environmental Drivers. 1. Introduction Northwestern China, encompassing Shaanxi Province, Ningxia Hui Autonomous Region, Gansu Province, Qinghai Province, and Xinjiang, is a region with diverse ecological types including cropland, grassland, desert steppe, desert, and plateau. As typical inland arid and semi-arid areas, Northwestern China is not only crucial for ecological construction in the country but also one of the area’s most sensitive to global changes [1]. This region serves as an important national security barrier, an ecological shield, and a strategic rear area for China. Covering about 30% of the country's area, it suffers from severe soil erosion and urgently needs ecological management [2]. The RF is an ensemble learning method that uses bootstrap resampling techniques to construct training sample sets for tree predictors by resampling the original training set with replacement. It then combines these tree predictors using an ensemble learning algorithm [3,4]. RF demonstrates excellent performance in dataset handling, with its key feature being the introduction of randomness, evident in the random selection of training samples for each tree and the random selection of attribute sets for node splitting in each tree. These randomness features help avoid the occurrence of overfitting [5], RF combines bagging and decision trees in a boosting method; while a single decision tree may have certain predictive accuracy, increasing the number of trees enhances overall predictive accuracy. This is because each tree in the random forest participates in making the final decision, which is the core idea of the algorithm [6]. Some scholars have used the RF regression algorithm to predict plant transpiration based on environmental parameters and leaf area index, finding that light intensity is the most critical factor influencing transpiration, followed by the relative humidity of the air [7]. Although many machine learning algorithms have built-in methods to assess the relative importance of predictive variables, such as feature importance plots, these methods generally evaluate global importance and do not provide insights into the predictive drivers for given observational points, presenting many limitations [8-10]. Interpretative methods in machine learning are developed to identify the predictive drivers of various predictive variables within a model. The SHAP algorithm posits that a model's prediction is the result of the cooperation of input features, and thus, the SHAP value of each feature for each sample indicates the contribution of each feature to the prediction [11,12]. SHAP provide a more intuitively appealing interpretation of variable importance [13]. SHAP has been widely applied in understanding machine learning models. JI Peng [14] used SHAP to explain four types of machine learning algorithms: RF, XGBoost, SVM, and ANN. DENG Menghua [12] used SHAP to elucidate the factors influencing 243 residents' willingness to pay for ecological compensation. In recent years, frequent human activities have led to substantial carbon dioxide emissions, accelerating global warming and ultimately impacting the global carbon cycle process [15]. The global carbon cycle plays a key role in regulating the Earth's climate system. Within this cycle, NEP serves as a critical measure of the net effect of carbon absorption and release by terrestrial ecosystems [16], and is an important tool for assessing ecosystem responses to climate change. NEP not only reflects the biogeochemical cycles within ecosystems but also forms the basis for understanding the global carbon balance. The definition of NEP is the difference between vegetation Net Primary Productivity (NPP) and carbon emissions from soil microbial respiration (RH) within an ecological area. Formula: HR-NPP=NEP NEP > 0 indicates that the carbon fixed by photosynthesis in the ecosystem exceeds the carbon released by total ecosystem respiration, meaning the ecosystem is a "carbon sink." This implies that the ecosystem net absorbs CO2 from the atmosphere, helping to reduce greenhouse gas levels and combat global warming. NEP < 0 indicates that the carbon released by total ecosystem respiration exceeds the carbon fixed by photosynthesis, meaning the ecosystem is a "carbon source." This implies that the ecosystem net releases CO2 into the atmosphere, increasing atmospheric greenhouse gas levels and exacerbating global warming. NEP = 0 indicates that the carbon fixed by photosynthesis is equal to the carbon released by total ecosystem respiration, achieving a balance in the ecosystem’s CO2 emissions and absorption. In this state, the ecosystem has a neutral impact on atmospheric CO2 concentrations. Scholars believe that climatic factors and anthropogenic influences are the two main drivers of changes in grassland primary productivity [17]. Previous studies have found a highly significant correlation between forest NEP and precipitation [18], and Xiaowei Yin et al.'s research highlights that understanding the impact of high-altitude flash droughts on vegetation net primary productivity (NPP) is crucial for ecosystem management [19]. The comprehensive effects of soil's physical, chemical, and biological characteristics influence plant growth, development, distribution, and productivity [20]. However, under rapidly changing climate conditions, the dynamics of NEP and its primary drivers are not fully understood, especially when considering the diversity of environmental factors. Therefore, this study delves into the specific contributions of various environmental variables to NEP using the Random Forest model and SHAP analysis method, with further analysis of the interactions between environmental factors through multiple linear regression. This aims to provide theoretical and empirical support for ecosystem management, climate change mitigation strategies, and global carbon budget estimation. 2. Materials and Methods 2.1. Data Collection In this study, NEP data were obtained from the National Earth System Science Data Center (https://www.geodata.cn/). The scope data for the five provinces in Northwestern China were sourced from the Resource and Environmental Science Data Platform (https://www.resdc.cn/). Climatic data including wind speed, solar radiation, average temperature, and precipitation were acquired from WorldClim (https://worldclim.org). Elevation data were imported into ArcGIS 10.8 to extract slope and aspect data. Soil attributes were derived from the World Soil Database (HWSD) (https://www.fao.org/soils-portal/). Human footprint data were taken from “A global record of annual terrestrial Human Footprint dataset from 2000 to 2018,” which consists of human activities such as built environments, population density, night lights, farmland, pasture, roads, railways, and navigable waterways. Refer to TABLE 1. 2.2. Data Processing Using R, random points were generated within the study area. Using ArcGIS, NEP values and various environmental factors were extracted for these locations. The extracted data were organized into an XLSX file, and outliers were removed using R. Descriptive statistical analysis and correlation analysis were performed to initially explore the relationships between factors and NEP, and to assess the data’s validity. Figure 1. Overview of NEP in the Study Area 2.3. Statistical Analysis Table 1. The meaning of environment variable data Name Mean Unit NEP Net Ecosystem Productivity gCꞏm-2ꞏa-1 aspect aspect degree Elev Elevation m slope slope degree T_CEC_SOIL Topsoil CEC (soil) Cmol/Kg T_CLAY Topsoil clay fraction %wt T_GRAVEL Topsoil gravel content %vol T_OC Topsoil organic carbon %weight T_PH_H2O Topsoil pH (H2O) -log(H+) T_REF_BULK Topsoil Reference Bulk Density Kg/dm3 T_SAND Topsoil sand fraction %wt T_SILT Topsoil silt fraction %wt HFP Human Footprint Level (1-50) Temperature average temperature ℃ Precipitation precipitation mm Wind wind speed m/s Srad solar radiation kJꞏm-2ꞏday-1 244 The “randomForest” package in R was used to train a Random Forest model with NEP as the target variable and all environmental factors as feature variables. The data were randomly split into a training set and a test set at a 70% ratio. A grid of parameters for the Random Forest was set up, and the trained model was used to predict the test set using the predict function. The coefficient of determination (R²) was calculated to evaluate the model's performance. To further understand the contributions of each predictor variable to the estimation of NEP, the iml package in R was utilized to compute SHAP values. SHAP values are a quantitative measure of the impact of features on the model output, providing a detailed and interpretable method for assessing feature importance. Variables with high importance in SHAP were selected for multiple linear regression analysis to further analyze the interactions between environmental factors. 3. Results 3.1. Descriptive Statistical Analysis Through the means, standard deviations, maximum values, and minimum values, we can understand the dispersion and range of the data. From TABLE 2 and FIGURE 2, it is evident that the data used in this study are reasonably distributed and cover a wide area, reflecting good capture of environmental diversity. Table 2. Basic Information of the Data Name average Standard Deviation minimum maximum NEP 347.74 189.39 69.04 745.14 aspect 154.39 104.22 12.65 358.62 Elev 2018.97 1093.69 377 4760 slope 3.48 3.17 0.06 13.64 T_CEC_SOIL 14.66 5.66 6 38 T_CLAY 19.4 7.38 5 47 T_GRAVEL 7.9 4.31 2 24 T_OC 1.06 0.64 0.36 3.02 T_PH_H2O 7.1 0.95 5.1 8.3 T_REF_BULK 1.43 0.09 1.22 1.7 T_SAND 42.85 14.66 10 89 T_SILT 37.75 9.97 6 54 HFP 14.45 9.29 0.44 39.32 Temperature 6.25 4.71 -6.88 14.72 Precipitation 38.11 17.96 3.33 92.17 Wind 2.39 0.49 1.52 3.7 Srad 15092.97 964.99 13104.25 17012.25 Figure 2. Boxplot of Basic Data Information Correlation analysis was conducted using SPSS software, and the results were visualized using R. The analysis showed that the correlation coefficient between Temperature and Elevation is -0.96, indicating a very strong negative correlation; as Elevation increases, temperature generally decreases. This finding is consistent with the fundamental principles of geography and climatic science, indirectly verifying the reliability and rationality of the data. 3.2. Random Forest Simulation Coupling the Random Forest model with the SHAP method, we analyzed the direction of influence and the relative importance of driving factors for NEP in the study area (as shown in FIGURE 4). The importance of NEP drivers in descending order is srad, Precipitation, T_REF_BULK, T_CLAY, Temperature, HFP, aspect, T_CEC_SOIL, T_GRABEL, Elev, wind, T_SAND, T_OC, T_SILT, T_PH_H2O, slope. Among surface climatic conditions, srad is the most important driver affecting NEP, followed by Precipitation, with slope being the least important. Regarding the direction of influence of the drivers, srad, T_REF_BULK, T_CLAY, and aspect. 3.3. Multiple Linear Regression Analysis From SHAP analysis, factors strongly driving NEP (chosen with SHAP values greater than 3) included seven variables: Srad, T_REF_BULK, T_CLAY, aspect, HFP, Temperature, and Precipitation. A multiple linear regression analysis was constructed using R to explore the interrelationships among these seven factors and their impact on NEP. Our results indicated that the interaction term between aspect and T_CLAY is statistically significant in the model, suggesting 245 that the combined effects of land slope and clay content significantly influence the ecosystem's NEP. The significance of the interaction term aspect:T_REF_BULK suggests that considering the slope or soil bulk density alone may not sufficiently reveal their impacts on NEP. Furthermore, the significant interaction between HFP and srad indicates that human activities might also interact with other environmental factors as a whole, collectively influencing NEP. Figure 3. Heatmap of Correlations Among Factors Table 3. SHAP Values of Each Factor Name SHAP_values Srad 5.762153583 T_REF_BULK 4.482324216 T_CLAY 4.455572117 aspect 3.342999929 T_CEC_SOIL 2.995941909 Elev 2.483418756 T_OC 1.224301556 T_PH_H2O 0.170039409 slope -0.060874055 T_SILT -0.231707001 T_SAND -1.82764097 Wind -2.01705201 T_GRAVEL -2.708125475 HFP -3.920637791 Temperature -4.373084215 Precipitation -4.555035252 Figure 4. SHAP Values of Each Factor 246 T_CEC_SOIL, Elev, T_OC, and T_PH_H2O positively influence NEP in the study area, whereas Precipitation, Temperature, HFP, T_GRAVEL, wind, T_SAND, T_SILT, and slope negatively impact NEP. Table 4. Results of Multiple Linear Regression Analysis for Each Factor Name Estimate t value Pr(>|t|) aspect:T_CLAY -0.709248833 -2.864361097 0.007990246 aspect:T_REF_BULK -54.06936287 -2.913398967 0.007096829 aspect:HFP -0.023210499 -0.865196939 0.394553322 aspect:Temperature -0.053385018 -0.915406973 0.368081757 aspect:Precipitation 0.042854289 1.638998156 0.11281572 aspect:Srad 0.000282929 0.642027445 0.526269366 T_CLAY:T_REF_BULK 201.2313493 1.743746325 0.092580889 T_CLAY:HFP 0.915171334 0.373240616 0.711883829 T_CLAY:Temperature 0.651757448 0.125393055 0.9011414 T_CLAY:Precipitation 0.948703075 0.477054314 0.637162238 T_CLAY:Srad -0.023621445 -0.579005987 0.567384684 T_REF_BULK:HFP -68.6928441 -0.503207728 0.618898863 T_REF_BULK:Temperature 73.1063777 0.208717775 0.83623519 T_REF_BULK:Precipitation 113.1619569 0.747191548 0.461405919 T_REF_BULK:Srad -0.244148379 -0.081775638 0.935428337 HFP:Temperature 1.474377564 1.510932876 0.142421856 HFP:Precipitation 0.361640752 0.882117978 0.385499159 HFP:Srad 0.012883504 2.43813225 0.021624542 Temperature:Precipitation -0.141047656 -0.173665057 0.863424194 Temperature:Srad -0.003088549 -0.214678229 0.831631409 Precipitation:Srad 0.001564595 0.989170883 0.331361857 Estimate: Indicates the size and direction (positive or negative) of the variable's impact. t value: A statistic used to test if the estimate is significantly different from zero. Pr(>|t|): Reflects the probability of statistical significance, where a value less than 0.05 typically means significant. 4. Discussion NEP is influenced by various factors including climatic elements, soil conditions, and land use [21,22]. This study examines the impacts of climate, soil, topography, and human activities on NEP. We found that temperature has a negative contribution to NEP, similar to the findings of HUA Langqin [23]. This may be due to Northwestern China's inland and arid location, where increased temperatures and insufficient precipitation lead to higher evapotranspiration, causing vegetation degradation and reduced plant net primary productivity [24,25], consistent with situations observed by Pang Zhaoyue et al. in central counties of Xinjiang [26]. Generally, studies suggest that more precipitation favors plant growth and enhances carbon fixation capabilities, but for arid and desert grassland areas in Northwestern China, some scholars have proposed different views that align with our findings, suggesting that only precipitation above 5mm is effective, and rainfall events under 7mm should be considered ineffective for the desert grasslands, as these minor precipitations likely evaporate after being intercepted by shrub layers or sand surfaces, and their retention in the soil surface layer has limited impact on vegetation growth. In plateau areas like the Tibetan Plateau, precipitation is often accompanied by temperature drops, and the water retained in the soil surface may exist in solid form, posing a freeze threat to plants. More importantly, small rainfall events may cause leaching or result in nutrient loss due to water restrictions, leading to inadequate nitrogen supply in nutrient-poor areas, which is detrimental to vegetation growth and productivity [27-29]. Scholars have also found that aspect influences the net primary productivity of Quercus mongolica Forests [30]. Solar radiation, being the sole energy source for photosynthesis, can enhance plant photosynthetic efficiency under constant variables; however, once the light intensity exceeds a threshold, the efficiency no longer increases. This phenomenon indicates that excessive light intensity may damage the protoplasts of plant cells, leading to chlorophyll degradation and excessive cell dehydration, which causes stomatal closure and ultimately reduces or stops photosynthesis. On the Tibetan Plateau, intense solar radiation significantly affects the content of chlorophyll, carotenoids, and flavonoids within plants; moreover, light quality is a key factor determining the development of plant roots, root growth, chlorophyll synthesis and accumulation, and protein synthesis [31]. Studies have shown that forest GPP is significantly positively correlated with photosynthetically active radiation, and grassland annual mean NPP is also affected by solar radiation [32,33], findings by SUN Qiang [34] that vegetation NPP in the Gannan Tibetan Autonomous Prefecture region correlates positively with solar radiation are consistent with our results. WANG Ziwen [35] using biomod2 simulations found that clay content and aspect significantly affect the growth and distribution of Datura stramonium. Additionally, researchers pointed out that effective soil water content and clay content are major factors affecting grassland NEE [36]. Soil, as a natural resource, is the growth mechanism for terrestrial vegetation. Soil particles, being specific geometric bodies, determine the fundamental physical properties of soil, such as bulk density, porosity, and water content, which are crucial indicators of soil structure, hydrological conditions, and soil quality assessment. These properties can directly or indirectly affect soil aeration, water retention, and nutrient preservation capacities, thereby significantly influencing the vegetation growing above [37]. In the inland areas of Northwestern China, the soil is relatively infertile and compact, excessive soil compaction might lead to poor aeration, and in some parts of Northwestern China, particularly in the Tibetan Plateau area, the high altitude and harsh environment cause plants to 247 be stunted, overly compact soil may hinder plant rooting and growth. Soil texture, which is the percentage composition of different soil particle sizes, also known as mechanical composition, affects soil water, air, heat, and nutrient conditions, thereby influencing the growth and development of plant roots and aerial parts [38]. Besides natural factors, as human activities intensify, anthropogenic factors are also gradually considered in environmental issues, with researchers using process-based terrestrial ecosystem models and remote sensing-based (CAS) methods to simulate the combined effects of climate change and human activities on grassland productivity in the Tibetan Plateau area from 1982 to 2011, finding that the factors driving regional productivity in the first 20 years shifted from climate change to human activities in the last 10 years, with over 20% of the area shifting from climate change dominance to human activity dominance, indicating that human activities are becoming a major factor affecting the terrestrial vegetation productivity on the Tibetan Plateau [39]. LU Xiaoquan[40] found that as urbanization intensity increases, ecosystem total primary productivity GPP shows a declining trend. This suggests that soil factors might influence NEP by affecting plant growth and development, while human factors also have a certain impact on NEP, and the impact brought by human factors may become more apparent with the development of the socio- economy in the future. This study also has certain limitations; although it considered human, topographic, soil, and climatic factors, some aspects like water body proportion and socio-economic factors were not covered. This paper primarily used Random Forest for analysis, and alternative approaches could be explored in the future to evaluate environmental effects on NEP. 5. Conclusion This study, through Random Forest and multiple linear regression, analyzed the impact of environmental and anthropogenic factors on NEP, with SHAP providing explanations for feature importance values in the Random Forest model. It was found that srad, T_REF_BULK, T_CLAY, Aspect, T_CEC_SOIL, Elev, T_OC, and T_PH_ H2O promote NEP in the study area, with the highest for srad at 5.762153583. Precipitation, Temperature, HFP, T_ GRAVEL, wind, T_SAND, T_SILT, and slope suppress NEP, with the highest for Precipitation at -4.555035252. Multiple linear regression analysis showed that the interaction between Aspect and T_CLAY and T_REF_BULK significantly affected the ecosystem's net primary productivity, while the significant interaction between HFP and srad indicates that human activities might also interact with other environmental factors as a whole, jointly affecting NEP. ORCID Weibo Yuan https://orcid.org/0000-0002-7233-9594. Competing Interests: All authors declare that they have no conflicts of interest. Author Agreement: All authors agree with the responses of Detailed Response to Reviewers and Revised Manuscrip. Acknowledgments Qinghai Provincial Science and Technology Department Applied Basic Research (2023-ZJ-749). References [1] Song Jiaying. Research on the spatial-temporal variation of vegetation NDVI and driving factors in Northwest China. 2021. Northwest Normal University. [2] Hong Wei., Study on the Ecological Footprint and Carrying Capacity of the Five Northwestern Provinces: Based on Time Series and Panel Data Spatial Econometric Analysis. 2021, Shaanxi Normal University. [3] Vens C, Costa F. Random forest based feature induction [C]// 2011 IEEE 11th international conference on data mining. IEEE, 2011: 744-753. [4] Tarroso P, Carvalho S B, Brito J C. Simapse–simulation maps for ecological niche modelling[J]. Methods in Ecology and Evolution, 2012, 3(5): 787-791. [5] Guisan A, Thuiller W, Zimmermann N E. Habitat suitability and distribution models: with applications in R[M]. Cambridge University Press, 2017. [6] Breiman L. Bagging predictors[J]. Machine learning, 1996, 24: 123-140. [7] Li L, Chen S, Yang C, et al. Prediction of plant transpiration from environmental parameters and relative leaf area index using the random forest regression algorithm[J]. Journal of Cleaner Production, 2020, 261: 121136. [8] Strobl C, Boulesteix A L, Kneib T, et al. Conditional variable importance for random forests[J]. BMC bioinformatics, 2008, 9: 1-11. [9] Altmann A, Toloşi L, Sander O, et al. Permutation importance: a corrected feature importance measure[J]. Bioinformatics, 2010, 26(10): 1340-1347. [10] Filippi P, Whelan B M, Vervoort R W, et al. Mid-season empirical cotton yield forecasts at fine resolutions using large yield mapping datasets and diverse spatial covariates[J]. Agricultural Systems, 2020, 184: 102894. [11] HUANG Si-hao, WANG Li-xia, LIU Yong-xia, et al. Banana yield prediction based on machine learning algorithm[J]. Journal of Tropical Biology. 2024, 1-11. [12] DENG Menghua, ZHANG Tianshu, CHEN Junfei., Study on influencing factors of residents' willingness to pay for eco- compensation based on XGBoost-SHAP in Taihu Basin[J]. Journal of Economics of Water Resources. 2024, 42(02):44-50. [13] Lundberg S M, Lee S I. A unified approach to interpreting model predictions[J]. Advances in neural information processing systems, 2017, 30. [14] JI Peng, YUAN Xing., Modeling the evapotranspiration and its long-term trend over Northwest China using different machine learning models[J]. Transactions of Atmospheric Sciences. 2023, 46(01):69-81. [15] Friedlingstein P, O'sullivan M, Jones M W, et al. Global carbon budget 2022[J]. Earth System Science Data, 2022, 14(11): 4811-4900. [16] Fernández-Martínez M, Sardans J, Chevallier F, et al. Global trends in carbon sinks and their relationships with CO2 and temperature[J]. Nature climate change, 2019, 9(1): 73-79. [17] JIA Ying, QI Zheng Peng,. The Response of Grassland Net Peimary Productivity to Climate Change[J]. Jiurnal of Animal Science and Veterinary Medicine. 2016, 35(06):87-89. 248 [18] Zheng J, Mao F, Du H, et al. Spatiotemporal simulation of net ecosystem productivity and its response to climate change in subtropical forests[J]. Forests, 2019, 10(8): 708. [19] Yin X, Wu Y, Zhao W, et al. Spatiotemporal responses of net primary productivity of alpine ecosystems to flash drought: The Qilian Mountains[J]. Journal of Hydrology, 2023, 624: 129865. [20] Fu G, Shen Z X. Response of alpine plants to nitrogen addition on the Tibetan Plateau: A meta-analysis[J]. Journal of plant growth regulation, 2016, 35: 974-979. [21] Han L, Wang Q F, Chen Z, et al. Spatial patterns and climate controls of seasonal variations in carbon fluxes in China's terrestrial ecosystems[J]. Global and planetary change, 2020, 189: 103175. [22] Chen Z, Yu G, Ge J, et al. Roles of climate, vegetation and soil in regulating the spatial variations in ecosystem carbon dioxide fluxes in the Northern Hemisphere[J]. PLoS One, 2015, 10(4): e0125265. [23] Shengheng W, Yuqin Z, Dongxin J, et al. Spatio-temporal changes and attribution analysis of net ecosystem productivity in forest ecosystem in Fujian province[J]. Ecology and Environment, 2023, 32(5): 845. [24] ZHANG Ru, SONG Xiaobin, ZHANG Jixun. Spatiotemporal responses of vegetation net primary productivity to drought evapotranspiration in Yellow River Basin. Water Resources and Hydropower Engineering. 2022, 53(09):57-69. [25] LI Huaihai, LI Chunbin, WU Jing, et al. Spatio-temporal dynamics and climate response of grassland net primary productivity in Shiyang River Basin from 2000 to 2020[J]. PRATACULTURAL SCIENCE. 2022, 39(10):2048-2061. [26] Pang Z, Wang J, Mamat A, et al. The Impact of External Factors on The Evolution Characteristics of Net Primary Productivity of Vegetation in the Kashi Region[J]. Polish Journal of Environmental Studies. [27] Xin-rong L I, Feng-yun M A, Li-qun L, et al. Soil water dynamics under sand-fixing vegetation in Shapotou area[J]. Journal of Desert Research, 2001, 21(3): 217. [28] Guo Q, Hu Z, Li S, et al. Contrasting responses of gross primary productivity to precipitation events in a water-limited and a temperature-limited grassland ecosystem[J]. Agricultural and Forest Meteorology, 2015, 214: 169-177. [29] Guo Q. Responses of grassland ecosystem productivity to altered precipitation regime: A review[J]. Ying Yong Sheng tai xue bao= The Journal of Applied Ecology, 2019, 30(7): 2201- 2210. [30] Hua M Y, Sun Z L, Yin Z B. Effects of slope direction on NPP and NEP of Quercus mongolica forests in southwest of Changbai Mountains[J]. Journal of Northeast Forestry University, 2023, 51(6): 13-19. [31] Wang Yan. Overview of Plant Diversity and Conservation Strategies in the Ecosystems of the Tibetan Plateau. Chinese Qinghai Journal of Animal and Veterinary Sciences, 2015, 45(03):47-49. [32] LI Min WANG Hongyu., Estimation and Analysis of Gross Primary Productivity of China's Forest Ecosystem[J]. Henan Science and Technology. 2022, 41(15):100-106. [33] Meiling Z, Quangong C, Wenlan J. Simulation and sensitivity analysis of net primary productivity (NPP) of different grassland types[J]. Arid Land Geography, 2021, 44(2): 369- 378. [34] SUN Qiang, ZHANG Lifeng, HE Yi, et al. Spatio-temporal changes and driving force analysis of vegetation net primary productivity in Gannan Tibetan Autonomous Prefecture[J]. PRATACULTURAL SCIENCE. 2023, 40(07):1729-1741. [35] WANG Ziwen, YIN Jin, WANG Xing, et al. Habitat suitability evaluation of invasive plant Datura stramonium in Liaoning Province: Based on Biomod2 combination model[J]. Chinese Journal of Applied Ecology. 2023, 34(05):1272-1280. [36] Zhou H, Shao J, Liu H, et al. Relative importance of climatic variables, soil properties and plant traits to spatial variability in net CO2 exchange across global forests and grasslands[J]. Agricultural and Forest Meteorology, 2021, 307: 108506. [37] Pérès G, Cluzeau D, Menasseri S, et al. Mechanisms linking plant community properties to soil aggregate stability in an experimental grassland plant diversity gradient[J]. Plant and soil, 2013, 373: 285-299. [38] ZHANG Ningning, BAI Xue,OUYANG Xiaoxue, et al. Research Progress of the Relationship between Soil Factors and Agricultural Product Quality[J]. Chinese Journal of Inorganic Analytical Chemistry. 2023,13(06):513-523. [39] Chen B, Zhang X, Tao J, et al. The impact of climate change and anthropogenic activities on alpine grassland over the Qinghai-Tibet Plateau[J]. Agricultural and Forest Meteorology, 2014, 189: 11-18. [40] LU Xiaoquan, LIU Ying, LI Bo, et al. Impact of Urbanization on the Total Primary Productivity of Terrestrial Ecosystems in Nanjing Area[J]. Climatic and Environmental Research. 2024, 29 (01):103-112.