In order to provide references for landslide hazard prevention and research on the sensitivity of landslide-causing factors, the Longyangxia-Jishixia section of the upper Yellow River basin was selected as the study area, and 16 factors such as elevation, slope, terrain roughness and lithology were taken as typical landslide hazard factors. The collinearity test was carried out by Spearman correlation coefficient method to select landslide hazard factors with strong correlation. GIS was used to reclassify landslide disaster-causing factors and analyze their weights with geographic detectors. The geographic detection model results were coupled with random forest model to obtain landslide prediction probabilities under different causative factors. ROC curve was used to verify the accuracy of prediction results. The results indicate that a）the explanatory power of the interaction between causative factors is greater than that of individual factors, with the synergistic effect of elevation and other topographic parameters being particularly significant. b）The importance of drainage density, topographic roughness, and profile curvature is nearly zero, suggesting that these features may not have a direct or significant correlation with landslide occurrence. c）There are notable differences in the contribution of causative factors to the prediction results, with the elevation-slope combination being the core driving unit for landslide development in the study area. d）The AUC value ofrandom forest model has achieved 0.93, indicating strong classification performance.

Due to its complex terrain and special climate, Qinghai Province is prone to sudden and destructive mountain floods. In order to provide a basis for the monitoring and prevention of mountain floods in the region, taking the Longyangxia-Jishixia section of the upper reaches of the Yellow River in Qinghai Province, where mountain flood disasters were relatively severe, as the study area, 10 influencing factors of mountain flood disasters were initially selected. Based on the data of 115 historical flood disaster points in the study area from 1958 to 2000, four factors with strong correlations were eliminated through Pearson correlation test. The remaining six influencing factors were classified. GIS spatial analysis technology was used to obtain the classified data of the six influencing factors. The entropy index method was adopted to calculate the weights of each factor and identify the main disaster-causing factors. The research results show that elevation, annual precipitation, terrain roughness, NDVI, distance from the river course, and aspect are the disaster-causing factors of mountain floods in the study area (with the weights of 0.571 6, 0.144 8, 0.107 9, 0.094 8, 0.071 9 and 0.009 0 respectively), among which, elevation, annual precipitation and terrain roughness are the main disaster-causing factors. According to the classification of the main disaster-causing factors, 91.30% of the historical mountain flood disasters in the study area occur in the areas with an elevation lower than 3 091 m, 99.14% occur in the areas with an annual precipitation greater than 317 mm, and 98.26% occur in the areas with a terrain roughness less than 1.10.

In order to screen the disaster-causing factors that induce debris flows, evaluate the performance of the three machine learning algorithms of RF, GBDT and XGBoost, in predicting the susceptibility of debris flows, and provide references for geological disaster prediction and disaster prevention and mitigation in the areas prone to debris flows, taking Hainan Tibetan Autonomous Prefecture, Qinghai Province as the study area and based on historical debris flow disaster data, based on the initially selected 17 influencing factors and the Pearson correlation coefficient of debris flow disasters, the disaster-causing factors that induce debris flow were screened. The remaining disaster-causing factors were classified and 8 combinations of disaster-causing factors were set. The three machine learning algorithm of RF, GBDT, and XGBoost were used to predict the susceptibility of debris flow. The prediction effect was evaluated by indicators such as accuracy rate, precision rate, recall rate, F1 score, and ROC-AUC. The results show that a) distance from the river channel, elevation, soil erodibility, topographic moisture index, annual rainfall, normalized vegetation index, coefficient of variation of elevation, lithology, topographic roughness, profile curvature, curvature, slope, coefficient of variation of elevation, and aspect are the disaster-causing factors that induce debris flows in the study area. Among them, elevation, distance from the river channel, soil erodibility, topographic moisture index, annual rainfall, and normalized vegetation index are the main disaster-causing factors. b) When the three algorithms of RF, GBDT and XGBoost are used to predict the likelihood of debris flow, the prediction effect based on the disaster-causing factor combination C7 (this combination does not consider slope direction) is the best, and the prediction effects based on the disaster-causing factor combinations C5 (this combination does not consider curvature, coefficient of variation of elevation and slope direction) and C8 (this combination considers all disaster-causing factors) are also good. c) When predicting the susceptibility of debris flows based on the disaster-causing factor combination C7, the ranking of the advantages and disadvantages of the three algorithms is XGBoost, GBDT and RF.

Concrete face rockfill dams (CFRDs) are widely utilized in water conservancy and hydropower engineering due to their excellent durability, strong impermeability, and cost-effectiveness. However, their seismic safety during service has raised significant concerns. In this study, a concrete face rockfill dam located on the upper reaches of the Yellow River was selected as the research object. Based on the structural characteristics of the dam and geological conditions, a three-dimensional numerical model was established to simulate the dam’s dynamic response under seismic action by integrating the Duncan-Chang model and the concrete damage plasticity constitutive model. Through this model, the dynamic responses of the dam under varying seismic intensities were analyzed, revealing the evolution of stress distribution and structural deformation characteristics. The results indicate that the critical zones prone to transverse damage under seismic action are located at two-fifths of the total height of the concrete face and the toe slab. Furthermore, based on the computational results, the correlation between failure variables and structural damage patterns under seismic action was investigated, and vulnerability curves under different seismic intensity levels were plotted. The findings demonstrate that with the increase of  seismic intensity, the vulnerability curves shows a rightward shift trend, and the failure probability of the dam body rises significantly.

Strong earthquakes will greatly reduce the stability of the reservoir slope and easily induce landslide geological disasters. The superposition of earthquake and landslide surge load greatly threatens the safety of the dam body. In order to explore the dynamic response and damage evolution of high arch dam under the superposition of earthquake-landslide surge, this paper took a high arch dam as the object, built  a fine finite element model, determined the calculation model of earthquake-landslide surge load, and analyzed the modal variation law, displacement response characteristics and damage evolution trend of arch dam under multiple working conditions. The results show that the hydrodynamic added mass significantly reduces the wet modal frequency of the arch dam by 18%-23%, and the high-order vibration modes change significantly. When the earthquake and surge are superimposed, the peak displacement Rd of the midpoint of the vault is positively correlated with the peak acceleration of the earthquake and the maximum surge height. When the peak acceleration of the earthquake increases from 0.2g to 0.6g, Rd increases by 89.7%. The maximum displacement time under different working conditions is affected by many factors. The damage degree of dam body varies greatly under different working conditions. The damage of upstream surface is sensitive to the change of surge height. When the surge height is different by 40 m, the weighted damage area ratio of upstream surface to RUWA changes by 27.1%. From condition one to condition three, the weighted damage area ratio of cantilever surface is increased from 9.99% to 25.76% compared with RFWA, and the dam body has penetrating cracks. The damage dissipation energy increases sharply with the increase of load strength.

For the landslide disaster in the section from Chada Village to Songba Village in the upper reaches of the Yellow River, a logistic regression model was established to conduct a systematic analysis of its susceptibility. Based on the analysis of the disaster-pregnant environment, eight influencing factors, including Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM), slope, aspect, precipitation, temperature, ground temperature and SBAS-InSAR time-series surface deformation data were selected to build an indicator system for landslide susceptibility analysis. Subsequently, the binary logistic regression model was utilized to delve into the quantitative relationships between these influencing factors and the occurrence of landslides. In order to further refine the research, spatial consistency processing was performed on local influencing factors, and 362 landslide grid points, covering a total area of approximately 2.855 km2 were established. The regression coefficients and regression equations for each influencing factor were then determined by using the binary logistic regression model. Finally, the occurrence probability of landslide hazards in the study area was derived by using the probabilistic implicit function. Based on the results obtained from the logistic regression model, the spatial distribution patterns of landslide susceptibility in the study area were discussed. It is found that high-risk areas are mainly concentrated in regions with steep terrain, poor vegetation cover and harsh meteorological conditions.

The geological structure of the upper Yellow River reservoir and dam cluster is complex, with frequent earthquakes, collapse, landslides, debris flows and torrential floods that threaten dam safety. In order to achieve full chain dynamic warning and emergency decision support for multiple disaster risks, we developed a multi-disaster big data early warning platform. The platform covered multiple functional modules such as dynamic monitoring of geological hazards, calculation of multi hazard warning models, warning release and display of real-life 3D models. By integrating GNSS station monitoring data, unmanned aerial vehicle remote sensing data, high-resolution and sentinel satellite remote sensing data, real-time monitoring and data visualization of disasters such as landslides, collapses and debris flows could be achieved. It could build a standardized and scalable multi-disaster risk early warning probability model library by breaking through the limitations of single disaster risk warning probability models. Using 3D visualization fusion of Cesium and UE5 engine, high-precision simulation of flood routing and landslide disaster processes was carried out, while provided spatial interactive tools such as inundation analysis and profile measurement. The platform can quickly respond to complex disaster risks and provide accurate and visual risk warning support for the reservoir dam group in the upper reaches of the Yellow River.

The reservoir-dam group on the main stream of the upper reaches of the Yellow River serves as a critical flood control barrier. While delivering comprehensive benefits in flood mitigation, water supply, power generation and ecological conservation, it faces complex and dynamic natural disaster risks. By analyzing the characteristics of natural disaster risks in this region, the paper discussed the main natural disaster risks faced by reservoir and dam groups and the challenges of risk management and response. The results indicate that natural disasters in the upper reaches of the Yellow River dam-reservoir system exhibit compound and systemic features, where extreme events such as mega-earthquakes, giant landslides, and super-standard floods may trigger disaster chains. Disaster prevention and mitigation are confronted with challenges such as insufficient identification of multi-source dynamic risks, the need to improve the joint regulation and emergency response mechanism, and the lagging adaptability to climate change.

The assessment of landslide susceptibility is an important basis for geological disaster prevention and control, but landslide sample data are often missing or incomplete, making it difficult for machine learning models to conduct accurate and reliable susceptibility modeling. Based on random forest (RF) and artificial neural network (ANN) models, this paper discussed the accuracy changes of landslide susceptibility assessment results and susceptibility zoning characteristics under different missing ratios (10%-50%) and regional missing conditions. After expanding the sample by using the synthetic minority oversampling technique (SMOTE), the prediction results were compared and analyzed to verify the effectiveness of sample expansion. The results show that with the increase of the proportion of missing samples, the model accuracy gradually decreases, but the decrease is limited. The predicted areas of the RF and ANN models in the high-risk areas above the higher level are reduced by up to 7.0% and 5.5% respectively. Under regional missing conditions, the accuracy of the evaluation results varies greatly, and the maximum predicted area of high-level prone areas is reduced by 11.1% and 11.2% respectively. After expanding the sample, the accuracy of the evaluation results decreases slightly with the increase of the supplement ratio. When 50% of the samples are supplemented, the predicted areas of the high-risk areas of the two models are reduced by 14.0% and 19.5% respectively. Generating landslide samples based on the SMOTE strategy can provide an effective method for landslide susceptibility evaluation modeling in areas with missing landslide data.

China is a country with severe earthquake disasters, with a wide range of affected areas, high frequency of occurrence, and high intensity of earthquake activity. Roads, as the "lifeline", play an important role in the transportation of materials and personnel. After an earthquake disaster occurs, quickly and accurately obtaining the location of road damage is of great significance for timely dredging of lifelines and carrying out post disaster rescue. In response to the issues of strong shadow interference, high fault missed detection rate, and severe fragmentation in remote sensing identification of road damage after earthquakes, this paper proposed a road damage layer extraction framework that integrated prior knowledge of OpenStreetMap (OSM). The effectiveness of the method was verified by using the 2023 Jishishan M6.2 earthquake as a typical case. By building a four layer technical system of "vector constraint-image segmentation-topology repair-damage detection", rapid localization of road damage in complex terrain areas had been achieved, providing assistance in improving rescue speed and reducing personnel and property losses.

The geomorphic and geological conditions in the upper reaches of the Yellow River are characterized by complexity and fragility. Disasters, such as debris flows induced by extreme rainfall events, are frequently triggered, leading to significant economic losses. The Erlian Village debris flow of Guide County, located on the upper reaches of the Yellow River, was selected as the primary research focus in this study. Based on the collection of relevant data, field investigations, and remote sensing interpretation, and employing physical and empirical statistical models, including the TRIGRS, Flow-R, and FLO-2D models, a connection between landslides and debris flows was established, with particular emphasis placed on the role of rainfall-induced landslides as the material source for debris flows. A coupled numerical model describing the dynamic evolution of rainfall-induced landslides and debris flows was developed, and the dynamic risk of Erlian Gully debris flow, under a 50-year rainfall scenario, was quantitatively assessed. The coupled model not only incorporated the influence of the background conditions of the disaster environment on the debris flow source but also accounted for the dynamic material source provided by rainfall-induced landslides.

The upstream mainstream dam group of the Yellow River is located in the transitional zone between the Qinghai Tibet Plateau and the Loess Plateau, requiring large-scale landslide hazard identification. Taking the Longyang Gorge-Lijia Gorge dam group section as the research area, SBAS-InSAR method was used to form small baseline interferograms, the amplitude deviation index and coherence coefficient method were used to select PS points(Permanent Scatterer), and the KS statistical test method was used to select DS points（Pistributed Scatterer）. The surface deformation rate and deformation amount in the study area were obtained  through time-series deformation calculation. Results show that the deformation rate in the study area ranges from -80 to 50 mm/a, with most areas having a deformation rate of -10 to 10 mm/a, indicating a relatively stable state. Combining slope and deformation rate analysis, identifies 13 landslide hazard points in the study area. The maximum deformation rate at the rear edge of the landslide below the top of the Guobu slope is 80 mm/a, which is only 1 km away from the right bank reservoir area of the Laxiwa hydropower station, and the deformation at the top of the slope is increasing year by year. The cumulative deformation value from 2021 to 2024 reaches 300 mm.

The operational status of the Sanmenxia Reservoir critically impacts water supply security and flood prevention for downstream cities along the Yellow River. However, conventional monitoring approaches relying on sensor networks and manual observations suffer from inadequate real-time performance and insufficient precision. To address these limitations, this study developed an integrated early-warning model incorporating multi-source data and dynamic response mechanisms. The methodology involved calculating warning thresholds for critical parameters (e.g., dam displacement and seepage pressure) using the Typical Small Probability Method, coupled with an LSTM-based prediction model for efficient warning through threshold comparison. Validation using the Sanmenxia Reservoir case demonstrates the model’s superior predictive accuracy compared to traditional methods.

This paper addressed the operational requirements for multi-hazard early warning systems in the Longyangxia to Liujiaxia section of the upper reaches of the Yellow River, conducting research on disaster data collection, management, and database system design. Firstly, various disaster data were collected and categorized into structured data, spatial data, and file-type data, with corresponding storage solutions: PostgreSQL for structured data, File-based Geodatabase for spatial data, and folder-based hierarchical storage for file-type data, while enabling data sharing through various methods. Subsequently, a data warehouse was built , organized into four thematic categories including public information on geological hazards and single-disaster prediction and forecasting. A dimensional model was established, and ETL operations were implemented using Kettle software to enable the analysis of multi-source data. Finally, a B/S architecture-based database management system of the upper reaches of the Yellow River was developed, comprising four layers of  data acquisition, application support, system application, and system users. The system incorporated functionalities such as basic data management, map visualization, and monitoring data analysis.

In response to the issue that natural disaster knowledge graphs for single disaster types had a narrow coverage of information, making it difficult to extract knowledge from massive and complex information, this study proposed a method for building a natural disaster knowledge graph for the Yellow River Basin oriented toward loss assessment. The natural disaster knowledge graph consisted of a data resources layer, a knowledge extraction layer (including a schema layer and a data layer), and an application service layer. The schema layer was built by using a top-down approach, centered on ontological models of natural disaster events, fundamental geographic information, and disaster loss assessment. Multi-source heterogeneous data were collected, including natural disaster event data, basic geographic information data, and disaster loss data, and the data layer was built by using a bottom-up approach, enabling knowledge extraction, knowledge fusion, and knowledge storage from these diverse data sources. The application situations show that the knowledge graph supports efficient spatial-temporal relationship queries and rapid identification of regional concurrent disasters. The analysis of 160 major disaster events in the Yellow River Basin from 1981 to 2018 reveals an increasing frequency of disasters over time, with floods being the most common disaster type. Additionally, 37 instances of concurrent disasters are identified. This map breaks through the limitations of the traditional single disaster knowledge system and improves the efficiency of disaster loss assessment.

Under the background of the rapid development of big data, Internet of Things and artificial intelligence technologies, the data of debris flow disasters is increasingly presenting the characteristics of being massive, multi-source and heterogeneous. This study primarily employed extraction model libraries such as jieba, NLPIR and LTP to parse and extract unstructured debris flow disaster data, subsequently aggregating the results into a database to achieve data fusion. By mapping words into a high-dimensional space through the Word2Vec model, textual vocabulary was converted into real-valued vector representations. t-SNE and Kernel PCA were then applied to reduce the dimensionality of these vectors, and the K-means algorithm was used for clustering and visualization. The results show that in data extraction evaluation, the average values of consistency, completeness and accuracy are all above 0.800, with mean square deviations below 0.050. Comparing the PCA and t-SNE dimensionality reduction methods using the Silhouette Score (SS) to evaluate clustering effectiveness, PCA achieves an SS value of 0.359, outperforming the t-SNE method, which has an SS value of 0.336. The findings demonstrate that the PCA method exhibits superior performance. Additionally, the Bert model-with its strong contextual understanding-is more suitable for debris flow disaster data extraction. By leveraging the CBOW architecture of the Word2Vec model to obtain word vectors, the results also reveal that PCA outperforms t-SNE in overall evaluation metrics. Targeting the challenges posed by multi-source debris flow disaster data and semantic consistency, this study conducts in-depth research on data extraction, dimensionality reduction, and clustering, ultimately providing a breakthrough technical method for debris flow disaster data knowledge classification and semantic consistency fusion, as well as an important technical solution for disaster data integration.

In order to solve the issue of inconsistent spatial-temporal correlation of multiple disaster causing factors, a method for building a knowledge graph of multiple disaster types in the upper reaches of the Yellow River based on multi-source information fusion was proposed. Through data cleaning, entity recognition and relationship extraction, structured and unstructured data such as remote sensing images, meteorological data and geological data are integrated to build the core concept of ontology definition and define six types of semantic associations such as causal relationship, spatial-temporal association and participation relationship. Knowledge storage and dynamic reasoning were carried out by using the Neo4j graph database, and a multi-source data mapping mechanism was designed to support multi-dimensional queries of entities such as disaster events, trigger factors and derivative disasters, as well as their relationships. Combining Bayesian network probabilistic reasoning and graph topology analysis algorithms, the propagation paths of disaster chains were quantitatively evaluated. Based on the landslide event in a certain area of the upper reaches of the Yellow River in 2023, the potential relationship chain between landslides and debris flows can be discovered through the inference function of the graph database. Compared to traditional databases, knowledge graphs have faster and more accurate information queries.

Currently, the frequency and intensity of extreme climate events have significantly increased, making the safety risks of reservoir overtopping and dam breaches more prominent. This study investigated the overtopping-induced breach process through a combination of generalized flume experiments and mathematical theoretical analysis. The results indicate that compared to coarser-grained dam materials, fine-grained dam bodies exhibit slower seepage rates and weaker permeability. As the upstream water level rises, the dam material gradually becomes saturated, with the saturation rate accelerating as the water level rises faster. The initiation of surface particle movement is identified as the key factor triggering erosion damage, while the upstream retreat of the scarp is critical to the occurrence of a breach. Based on the breach process, a geometric generalization model for the longitudinal and lateral development of the breach is proposed. Furthermore, by applying calculus principles, the flood discharge process during overtopping failure is analyzed, demonstrating that the flood release process can be regarded as the cumulative effect of various breach-influencing factors. Mathematical expressions for breach development and discharge variation at different stages are also derived. 

The evaluation of landslide susceptibility is of great significance for regional disaster prevention and mitigation. In view of the issues that the single classifier in the landslide susceptibility evaluation using machine learning algorithms had poor precision, and the selection of negative samples of landslides was relatively arbitrary, a landslide susceptibility evaluation model was proposed, which combined the selection method of negative samples of landslides based on the information quantity method with machine learning integration algorithms. Taking the section from Lijiaxia to Gongboxia in the upper reaches of the Yellow River as the study area, 13 evaluation factors such as elevation, slope gradient and precipitation were selected as the evaluation factors for landslide occurrence. Three selection methods for negative samples of landslides, namely buffer zone, low slope gradient and information quantity were adopted. By building the landslide susceptibility evaluation models of the classification and regression tree (CART) and three integrated learning algorithms (Bagging, Boosting, and random forest), the performance of the evaluation models under different integrated learning algorithms and different selection methods for negative samples of landslides was analyzed. The results show that the integrated learning algorithm can significantly improve the performance of the single base classifier, and the improvement effect of the Boosting algorithm is the most prominent. The selection method of negative samples based on the information quantity takes most of the evaluation factors into full consideration, and the reliability of the model is higher.

Severe rainfall events often trigger cascading hazards such as flash floods and debris flows, forming intricate disaster chains that pose significant threats to social economy and natural ecosystems. However, the existing studies are still insufficient in exploring the interaction relationships among multi-disaster nodes and their superimposed risks. In this study, we combined complex network theory with FLO-2D numerical simulation to analyze the rainstorm-flood-debris flow disaster chain. Taking Shishugou in Luanchuan County, Henan Province as a case study, the dynamic evolution characteristics of the disaster chain under the background of the extreme rainstorm disaster occurred in Zhengzhou on July 20, 2021 were analyzed. The results indicate that a) the numerical simulation results indicate that the maximum flow velocity of the debris flow can reach 6.56 m/s, with a maximum deposition depth of 7.30 m. The proportions of high, medium and low-risk areas are 3.61%, 17.65% and 78.74% respectively; b) damage or blockage of drainage systems, damage to electrical facilities, urban flooding, submerged vehicles, mudslides and backflow of river water represent the critical nodes in the disaster chain network; c) the key edges in the network include backflow of river water → urban flooding, urban flooding → submerged vehicles, urban flooding → damage to electrical facilities, damage to riverbanks → river channel blockage, submerged vehicles → casualties. These findings effectively identify the critical nodes and weak links within disaster chains from a dynamic evolution perspective. Furthermore, the study elucidates the interactions between multiple hazard nodes and their impact on disaster propagation, addressing key gaps in dynamic disaster chain evaluation and risk assessment. The results provide a scientific basis for disaster prevention, mitigation measures and emergency management planning.

An unsupervised anomaly detection model based on a deep autoencoder was proposed to address anomalous readings in concrete dam deformation monitoring, with the objective of enhancing detection accuracy and automation. The autoencoder was trained in an unsupervised manner on normal deformation data to learn low-dimensional feature representations and was subsequently employed to rebuild  incoming measurements. Measurements exhibiting significant deviations between observed and rebuild values were classified as anomalies. The result shows that the proposed model achieves over 97% accuracy in anomaly detection and is demonstrated to perform reliably under various testing conditions. Consequently, the deep autoencoder-based approach is capable of effectively identifying deformation anomalies in concrete dams, exhibiting robust and precise detection capabilities.

混凝土坝；变形监测；变形预测；LSTM；TCN

In order to enhance the accuracy of early warning in concrete dam safety monitoring, this study proposed an intelligent early warning analysis method based on multi-point monitoring data, aiming to overcome the susceptibility of traditional single-point methods to non-structural interference. Firstly, K-means clustering method was used to partition monitoring points with similar deformation patterns. Then, ConvLSTM model was employed to extract the spatial-temporal features of the deformation sequences from each cluster and make predictions. By analyzing the residual sequences and determining the early warning threshold based on the 3-Sigma principle, single-point early warning results were generated. Finally, the early warning results from all clusters were integrated to ensure that an early warning was triggered only when all monitoring points within a cluster exhibit anomalies simultaneously at the same time. Experimental results show that the proposed method reduces the false alarms and missed detections caused by external disturbances in single-point early warning methods by integrating multi-point information, thereby improving the reliability and stability of the early warning system.

Reservoir systems serve as critical ecological barriers and pivotal nodes in watershed water resources management. However, their stability is increasingly challenged by the growing frequency of extreme precipitation events induced by global warming. These events can trigger cascading disasters, posing significant threats to human life and socioeconomic development in affected regions. This study focused on the upper reaches of the Yellow River and utilized high-resolution data from the Coordinated Regional Climate Downscaling Experiment (CORDEX), specifically derived from the East Asia domain. Four bias correction methods of linear scaling, quantile mapping, Quantile-Q adjustment, and variance scaling-were applied to evaluate their effectiveness in correcting precipitation data. Based on comparative analysis, the Quantile-Q adjustment method demonstrated the best overall correction performance and was therefore selected to calibrate CORDEX future-period simulations. Under the RCP2.6 scenario, the spatial distribution of multi-year mean extreme precipitation indices during 2006-2099 remains consistent with the historical period, showing only slight increases. In contrast, under the high-emission RCP8.5 scenario, several indices rise significantly, with enhanced regions mainly located in central Qinghai, and south Inner Mongolia.