The Yanshangou landslide, located in the Baihetan Reservoir area, poses severe potential threats to the normal operation of the reservoir due to its distinct deformation characteristics and high sensitivity to reservoir water level fluctuations. This study systematically investigates the geological background, deformation characteristics, stability evolution, and landslide-induced surge hazards of the Yanshangou landslide in the Baihetan Reservoir area. This work only considers the influence of reservoir water level fluctuations, which is the dominant factor controlling the current progressive deformation of the landslide. Field surveys and GNSS/deep displacement monitoring results revealed that the Yanshangou landslide exhibits obvious staged deformation characteristics, and the landslide deformation rate was closely coupled with the dynamic changes in reservoir water level. A slope stability evaluation method integrating the Morgenstern-Price limit equilibrium method and Richard's equation was established, and the results indicated that the Yanshangou landslide has low saturated permeability. Therefore, its factor of safety (FOS) presents a clear four-stage variation trend in response to reservoir water level fluctuations. A Smoothed Particle Hydrodynamics (SPH)-based numerical model was further developed to simulate the landslide-induced surges under two typical reservoir water level scenarios (815 m and 765 m). The simulation results demonstrated that a high reservoir water level led to more intense surges with greater height and higher velocity, while a low reservoir water level resulted in surges with a wider propagation range along the reservoir bank. The research findings of this study provide a comprehensive theoretical basis and detailed data support for the prevention and mitigation of geological hazards in the Baihetan Reservoir area, and also offer a reference for the hazard management of similar reservoir landslides worldwide.

Rock mass quality classification (RMQC) plays a crucial role in rock mass stability analysis and in the design and construction planning of rock engineering projects. However, current RMQC methods rely on expert experience, which makes it difficult for RMQC to be intelligent, scientific, and interpretable, and is not conducive to understanding rock mass characteristics in engineering applications. Therefore, this study proposes an interpretable rock mass quality intelligent classification model (IRICM) by coupling random forest (RF) and genetic algorithm (GA) to refine decision rules, aiming to enhance the intelligence, scientificity, and interpretability of RMQC. Based on 318 tunnel section data, the RMQC dataset was constructed using rock mass rating (RMR) parameters obtained from field investigations and laboratory experiments. By coupling RF and GA, the rules from all decision trees were selected, combined, and optimized to refine decision rules, achieving a classification accuracy of 87.50 % with only five rules per class. Interpretability analysis of the refined decision rules revealed that rock quality designation (RQD), intact rock strength (IRS), joint spacing (JS), and groundwater (GW) were the most frequently used features, confirming their importance in RMQC. Further analysis using post-hoc interpretability techniques also indicated that RQD, IRS, JS, and GW contributed most significantly to RMQC, especially in distinguishing poor rock mass quality (classes IV and V). The model was applied to the RMQC of tunnels and rock slopes, and the results demonstrated consistency with classification outcomes from the Q, RMR, and geological strength index (GSI) systems, validating its reliability and stability.

Accurate and rapid recognition of weathering degree (WD) and groundwater condition (GC) is essential for evaluating rock mass quality and conducting stability analyses in underground engineering. Conventional WD and GC recognition methods often rely on subjective evaluation by field experts, supplemented by field sampling and laboratory testing. These methods are frequently complex and timeconsuming, making it challenging to meet the rapidly evolving demands of underground engineering. Therefore, this study proposes a rock non-geometric parameter classification network (RNPC-net) to rapidly achieve the recognition and mapping of WD and GC of tunnel faces. The hybrid feature extraction module (HFEM) in RNPC-net can fully extract, fuse, and utilize multi-scale features of images, enhancing the network's classification performance. Moreover, the designed adaptive weighting auxiliary classifier (AC) helps the network learn features more efficiently. Experimental results show that RNPC-net achieved classification accuracies of 0.8756 and 0.8710 for WD and GC, respectively, representing an improvement of approximately 2%-10% compared to other methods. Both quantitative and qualitative experiments confirm the effectiveness and superiority of RNPC-net. Furthermore, for WD and GC mapping, RNPC-net outperformed other methods by achieving the highest mean intersection over union (mIOU) across most tunnel faces. The mapping results closely align with measurements provided by field experts. The application of WD and GC mapping results to the rock mass rating (RMR) system achieved a transition from conventional qualitative to quantitative evaluation. This advancement enables more accurate and reliable rock mass quality evaluations, particularly under critical conditions of RMR.< br /> (c) 2026 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/ 4.0/).

Rock mass quality evaluation provides the fundamental basis for stability analysis, support design, and engineering decision-making. However, traditional empirical methods mainly rely on localized field investigations and subjective interpretation, resulting in limited objectivity and spatial representativeness. To overcome these limitations, this study proposes an automated and refined rock mass rating (AR-RMR) system integrating validated artificial intelligence (AI) modules with multi-source heterogeneous data. The AR-RMR framework provides three major capabilities: (1) AI-assisted extraction of key RMR parameters through automated recognition of planar and linear discontinuities; (2) refined quantification of key parameters, including rock quality designation, spacing, persistence, aperture, roughness, infilling, weathering degree, and groundwater condition; and (3) integration of automated parameter extraction, refined parameter quantification, and RMR scoring into a unified rock mass quality evaluation workflow, thereby improving the objectivity and spatial representativeness of conventional RMR assessment. Validation on two representative tunnel sections demonstrates good agreement between AR-RMR and conventional field evaluation. Further application to 47 tunnel sections yields a coefficient of determination (R2) of 0.729 and a root mean square error (RMSE) of 3.656 compared with field RMR, while independent validation using geological radar prediction further confirms the reliability of the proposed framework. The proposed AR-RMR provides an effective pathway toward objective, refined, and spatially representative rock mass quality evaluation, offering reliable technical support for stability assessment, support design, and engineering decision-making in rock engineering.

Understanding the anisotropic characteristics and failure mechanisms of jointed rock masses is essential for reliable stability evaluation in rock engineering. However, accurately quantifying the influence of in-situ stress on the mechanical properties and failure modes of such systems remains a challenging and unresolved issue. To address this gap, we developed a discrete fracture network (DFN) model based on outcrop data from a dam site in southeastern Tibet. The representative elementary volume (REV) was determined using discrete-element analysis with mechanical upscaling, and a series of numerical triaxial compression tests were conducted on REV-scale models employing the synthetic rock mass (SRM) approach. Based on the anisotropy index, an advanced method is proposed for the quantitative evaluation of rock mass anisotropy and the analysis of the failure mechanisms in 3D jointed rock masses. Our results reveal that low confining pressures primarily induce joint slippage, whereas higher confining pressures reduce the anisotropy index of the jointed rock mass. Additionally, variations in crack orientation under different stress conditions highlight the pivotal role of confinement in governing fracture development. These findings offer new insights for enhanced stability evaluation in rock engineering.

Accurately estimating the mean trace length (mtl) of discontinuities is crucial for three-dimensional (3D) network simulations of random discontinuities. Currently, the mtl is mainly estimated by laying out scan lines or sampling windows, based on the intersection of discontinuities with these scan lines or sampling windows. However, these methods project discontinuity traces interpreted by 3D point clouds or models onto a two-dimensional (2D) sampling window, which leads to the loss of the trace length. This study aims to address this limitation and proposes a 3D mean trace length estimation model (3D-MTLEM) by extending the 2D rectangular sampling window to a 3D cuboid sampling block, enabling the full utilization of the 3D discontinuity information. Firstly, the proposed 3D-MTLEM is applied to three simulated trace datasets. The results indicate that compared to traditional 2D window estimation methods, the 3D-MTLEM provides more accurate and stable estimates of mtl. Furthermore, the 3D-MTLEM conducts an analysis of the influence of four key factors: trace length (tl), trace direction (td), trace length distribution function (tldf), and cuboid sampling block size (csbs). Subsequently, the 3D-MTLEM is applied to two case studies of an artificial quarry slope, further confirming its applicability in practical engineering.

Interferometric Synthetic Aperture Radar (InSAR) technology provides high-resolution time-series observation capabilities for monitoring surface slow-moving landslide deformation. Accurately identifying subtle yet critical change points (CPs)—acceleration, destabilization, or stagnation—remains a key challenge in improving the timeliness and reliability of geological hazard early warning systems. To validate the performance limits of CP detection methods under complex environmental conditions, this study integrates InSAR data with advanced Change Point Detection (CPD) techniques. Three slow-moving landslides (Xiongba, Sela, and Gongba) along the Jinsha River Basin were selected, with their subtle surface deformation histories from 2014 to 2022 reconstructed using SBAS-InSAR analysis. The first InSAR-based landslide critical point validation set was established, comprising 752 independently annotated deformation datasets from four geologists. Bayesian optimization was employed to quantitatively delineate the applicability boundaries of the three algorithms—Breaks For Additive Season and Trend (BFAST), Bayesian Estimator of Abrupt change, Seasonal and Trend change (BEAST) and Parallel Autoencoder (PAE). The results reveal that BEAST achieves superior precision (0.93) in noise-free, seasonally dominated environments, whereas PAE demonstrates robust recall (0.62) under real-world. BFAST prioritizes sensitivity to trend breaks but exhibits higher false positive rates. By employing a multi-model consensus approach, we precisely identified deformation processes associated with seismic swarms and natural dam-break events, and confirmed that the dam-break-induced displacement threshold closely aligns with previous findings. Notably, this study provides the first evidence that local earthquake swarms (Mw 4.6–5.3) can trigger abrupt landslide displacement. Overall, integrating CPD with InSAR time-series effectively detects subtle nonlinear changes missed by threshold methods. Our hybrid validation (simulated data, expert annotations, geological records) shows method effectiveness depends on geodynamic context, requiring scenario-specific selection. Future integration of CPD with high-precision, real-time InSAR will greatly enhance landslide early warning.

Accurately predicting landslides is critical for effective warning and management, but remains challenging due to unpredictable triggering events and the spatial heterogeneity of soil and slope structures. Existing prediction methods often rely on point-sampled data, neglecting the heterogeneity in landslide evolution. To address this, we propose integrating Spatiotemporal Graph Convolutional Networks (STGCN) with Synthetic Aperture Radar Interferometry (InSAR) to capture the spatiotemporal characteristics of landslide events. The STGCN processes spatial features through its Graph Neural Network (GNN) layer and analyzes temporal dynamics using the Gated Recurrent Unit (GRU) layer. This allows for a more precise extraction of displacement features associated with landslides. An application of this approach in the Sela Mountain region of the Jinsha River on the Tibetan Plateau (China) demonstrated that the STGCN model significantly improves prediction accuracy compared to traditional deep learning models, with Mean Squared Error (MSE) and Mean Absolute Error (MAE) reduced to 25.51 and 2.34, respectively. This represents a 35 % and 50 % improvement over the best-performing traditional model in similar tests. Notably, this method, notably the first to incorporate the direction of material migration in landslide predictions, effectively addresses the challenge of spatial heterogeneity and expands the predictive framework from merely temporal to both spatial and temporal dimensions. Our findings highlight that this integrated approach provides a powerful tool for more accurate and comprehensive landslide prediction.

Estimating fracture size is a fundamental aspect of rock engineering. However, determining the most probable diameter (MPD) from a fracture’s surface trace remains challenging in the scientific community. The prevailing methodologies typically infer statistical distributions of fracture sizes rather than specific values. This research presents a novel approach to inferring the MPD and the true spatial distribution pattern of each fracture. The challenge lies in linking the inference process with the trace length of each fracture and the statistical characteristics of the entire outcrop. Additionally, it is necessary to address the non-unique inverse problem. The methodology comprises several key steps. Firstly, the issue of censoring bias is addressed by considering the lengths of the traces contained. Secondly, the orientation bias is corrected using the vector method, and the true mean trace length and standard deviation are estimated and derived. Thirdly, assuming a lognormal distribution for fracture sizes, the mean and standard deviation of diameters are derived through a high-order moment relationship between trace lengths and diameters, validated by Crofton’s theorem. Finally, the MPDs of all trace samples are determined by relating MPDs to trace lengths and the standard deviation of diameters using stereology techniques. Furthermore, the true fracture spatial patterns are inverted based on spatial geometric relationships. The proposed methodology is validated through rigorous Monte Carlo simulation and applied in a practical engineering case study, demonstrating its potential for use in rock engineering applications.

Accurate and rapid measurement of fractures is crucial for rock mass quality evaluation and stability analysis. Due to rock mass fractures often being considered linear for reasons of simplification and exhibiting characteristics of randomness and complexity in their distribution, traditional manual measurement methods are timeconsuming and labor-intensive. Therefore, this study proposes a novel fracture point detection network (FPDnet) for automatically detect fracture points from rock mass three-dimensional (3D) point clouds. By an integrated post-processing technology (IPPT), the detected fracture points are further connected into complete fractures, enabling fast and automated fracture extraction. FPD-net improves Pointnet++, a deep learning framework for hierarchical feature extraction from 3D point clouds, by incorporating a weighted feature sampling (WFS) method and a dual attention module (DAM), enabling it to detect point clouds of rock mass fractures, i.e., fracture points. The ablation analysis results show that WFS and DAM in FPD-net have a positive impact on the model performance, with the precision improved by 0.018 and 0.024, respectively. Compared to Pointnet++, tensor voting, and curvature-based methods, FPD-net detected fracture points with higher precision on each side wall of the tunnel, with fewer misclassified and noisy points. Furthermore, IPPT was used to connect the fracture points detected by FPD-net. The results indicate that the extracted fractures closely resemble manually marked fractures, across all side walls of the tunnel, demonstrating consistently high extraction accuracy. Additionally, a sensitivity analysis of the seven parameters in the IPPT was conducted, quantitatively determining the impact of each parameter on fracture extraction accuracy. This method offers a new approach for automatically extracting fractures directly from 3D point clouds, presenting distinct advantages over traditional methods and promoting the development of non-contact measurement methods in rock mass fracture extraction.

Rock fractures are one of the main factors leading to rock failure. Accurately extracting fracture characteristics is crucial for understanding the rock failure mechanism. Inspired by the latest developments in computer vision, we introduce a state-of-the-art deep learning model YOLACT++ for the automated interpretation of rock fractures. YOLACT++ inherits the basic architecture of YOLACT (You Only Look At CoefficienTs) and optimizes the backbone network, which improves segmentation accuracy while ensuring real-time performance. Based on Unmanned Aerial Vehicle multi-angled proximity photography, the dataset is collected from various rocky slopes for model training and validation. We propose performance evaluation metrics for the model, including intersection over union, precision, and recall, as well as quantitative parameters for describing fractures, including orientation, trace length, roughness, aperture, spacing, and fracture intensity. The segmentation results of YOLACT++ are compared with two other classic instance segmentation models, the Mask Region-based Convolutional Neural Network (Mask R-CNN) and the You Only Look Once (YOLO) V8. The results show that YOLACT++ has a stronger generalization ability, with more accurate segmentation results at image boundaries. With the ResNet-101 backbone network, YOLACT++ achieves 93.8 %, 87.1 % and 92.2 % for precision, intersection over union and recall, respectively. This represents improvements of 5.4 %, 3.6 %, and 8.3 % compared to Mask R-CNN, and 3.3 %, 7.8 %, and 4.2 % compared to YOLO V8. Overall, the deep learning-based YOLACT++ model proposed in this study provides an efficient and reliable approach for the automated interpretation of rock fractures. It can also be applied to crack recognition in other materials.

The discontinuity system exerts significant control over slope deformation and failure. Nevertheless, the automatic identification of these discontinuities remains a challenging task, particularly concerning linear discontinuities. Current methodologies are insufficient in detecting linear discontinuities, let alone conducting an analysis of their internal parameters. Consequently, this leads to imprecise guidance for rock mass engineering endeavors. This paper proposes a method for identifying linear discontinuities and researching their thermal characteristics, leveraging thermal infrared technology in conjunction with unmanned aerial vehicles. A comprehensive 24-h thermal infrared survey was conducted on the rock mass situated at the dam site of the Xulong Power Station in Yunnan Province, China, and the thermal radiation law of rock mass at the dam site was preliminarily summarized. The feasibility of identifying the linear discontinuities through the thermal radiation characteristics has been demonstrated. Furthermore, the correlation between discontinuity parameters and their thermal characteristics is established. The results indicated that the thermal characteristics of linear discontinuities (faults and joints) are closely related to their physical and geometric parameters. These findings enable the study of rock mass weathering and inference of the internal characteristics of discontinuities, thereby facilitating the analysis of slope deformation and failure.

Active tectonic movements and geological disasters frequently occur in the upper reaches of the Jinsha River,increasing the likelihood of landslides obstructing the river.Taking the ancient Rongcharong landslide dam failure events in the Suwalong reach as an example,this pa-per first analyzes the accuracy and applicability of the commonly used methods in calculating the peak flow of the dam failure,such as the empirical formula,the numerical method based on the physical mechanism,and the computational fluid dynamics(CFD)method.Then,the peak flood flow of the Rongcharong-dammed lake when it overflows the dam is determined to be 28 393-64 272 m3/s.At the same time,the failure process of landslide dam due to flood erosion was eluci-dated using the CFD method,which can be divided into three stages:gradual erosion in the initial stage,rapid development in the middle stage,and gradual expansion in the final stage.Finally,the factors that affect the peak flow of floods are analyzed,and suggestions for emergency treatment of landslide dams are put forward.The findings of this research can serve as a valuable reference for disaster prevention and mitigation strategies to adapt to the increasing frequency of landslide-in-duced river blockages.

The Eastern Himalayan Syntaxis is a high denudation zones, with mass flow events occurring frequently. The Jiaobunong paleolandslide (JPL) river blocking event is a typical prehistoric event that can be identified by special topography and upstream lacustrine sediments. This study applies the topographic reconstruction technique to revive the paleolandscape prior to the landslide occurrence, and further adopts discrete element software to invert the kinematic process of the JPL. The numerical simulation results reveal the whole landslide process to last for 130 s, with a peak average velocity of 51 m/s after sliding for 45 s. Following this, a large deposit with an extremely thick front of approximately 260 m was formed. The landslide material not only blocks the Lulang River, but also climbs onto the opposite platform, forming a large dammed lake upstream with a maximum area of 7.12 km2. On-site investigation showed that the initial morphology of the JPL dam was altered by river incision, so parameter sensitivity analysis was used to obtain the optimal combination of morphological parameters. The JPL is particularly prominent in regional landforms due to the rapidly shrinking topography, which controls the formation of a slope-break knickpoint in the river profile. The formation of the knickpoint can be attributed to the topography control and landslide morphology of the region. This study can provide guidance for landslide prediction, disaster emergency response, and ecological restoration in similar regions.

Rock mass quality evaluation is a critical preliminary step in rock engineering, yet conventional methods often overlook the analytical connectivity of non-persistent fracture networks, reducing their precision and scope. To bridge this gap, this study proposes a novel connectivity-based geological strength index (CGSI) for rock mass quality assessment and establishes explicit and quantitative conversion relationships between different classification systems, thereby improving reliability across diverse engineering contexts. Using the representative elementary volume (REV) as the fundamental scale, an extensive quality evaluation of homogeneous domain #7 on a high and steep slope was conducted. Samples were optimized based on the REV, and four established assessment systems were applied to formulate conversion expressions. By integrating their core principles, we developed CGSI and established twelve quantitative conversion relationships spanning different rock strength ranges. CGSI's novelty lies in its capacity to quantify rock mass quality at the REV scale by incorporating surface conditions, structural distribution, fracture connectivity and fracture orientation effects to capture the threedimensional fracture network. Field investigations and comparative analyses validated CGSI, identifying grade III as the optimal quality for domain #7. Results demonstrate CGSI's superior conversion accuracy for fractured rock masses and its strong applicability in engineering practice. Extending the method to other homogeneous domains yielded overall quality ratings of II similar to IV. Information entropy analysis revealed that the number of moderately dipping fracture sets, REV size and mean fracture size are the dominant controlling factors. Notably, domains #1, #4, #6, #8, #12 and #14 were classified as grade IV, indicating elevated instability risks and the need for targeted reinforcement.

Fracture failure is a critical factor triggering rockslides in the deeply incised valleys of the Hengduan Mountains. Clarifying the failure mechanism of valley fractures associated with river downcutting is fundamental for risk assessment in these areas. This paper proposes a systematic framework to explore the failure mechanisms of valley fractures during river downcutting processes. Initially, a proportional elevation stretching algorithm is designed to reconstruct the paleo-geomorphology of valley at different geological periods based on current valley geomorphic features. Subsequently, a valley fracture network is constructed using unmanned aerial vehicle (UAV) photogrammetry and Monte Carlo simulations. Then, a multiscale DFN–DEM (discrete fracture network–discrete element method) approach is used to develop an equivalent geomechanical model of the valley. Finally, a continuous downcutting method is designed to simulate the evolution of valley fractures under non-disturbed and steady river downcutting process. The results indicate that valley fractures predominantly occur in shallow subsurface regions, where rock masses experience significant deformation due to lateral unloading and vertical rebound uplift. Besides, key findings highlight that the orientation of shallow tensile fractures is correlated with the topography, and this correlation becomes increasingly prominent as river downcutting deepens, fundamentally due to the shift in the role of gravity (from fracture-derived restraining force to fracture-derived driving force). The framework effectively overcomes the drawbacks of excavation disturbances in numerical simulations of river downcutting, accurately replicating the evolution of large-scale fractures that govern valley slope stability. This provides valuable guidance for construction projects in the deeply incised canyon regions of the Hengduan Mountains.

Accurate and objective regional landslide risk assessment is crucial for the precise prevention of regional disasters. This study proposes an integrated landslide risk assessment via a landslide susceptibility model based on intelligent optimization algorithms. By simulating the process of rime frost formation, it effectively selects features and assigns weights, overcoming the overfitting issue faced by XGBoost in handling high-dimensional features. By integrating the concepts of landslide susceptibility, dynamic landslide factors, and social vulnerability, an integrated landslide risk index was developed. Further investigation was conducted on how landslide susceptibility results influence risk, identifying regions with varying levels of landslide risk due to spatial heterogeneity in geological background, natural environment, and socio-economic conditions. This study’s results demonstrate that the RIME-XGBoost landslide susceptibility model exhibits superior stability and accuracy, achieving an AUC score of 0.947, which represents an improvement of 0.064 compared to the unoptimized XGBoost model, while the accuracy shows a maximum increase of 0.15 relative to other models. Additionally, an analysis using cloud theory indicates that the model’s expectation and hyper-entropy are minimized. High-risk-level areas, constituting only 1.26% of the total area, are predominantly located in densely populated, economically developed urban regions, where roads and rivers are the key influencing factors. In contrast, low-risk areas, which cover approximately 72% of the total area, are more broadly distributed. The landslide susceptibility predictions notably influence high-risk regions with concentrated populations.

The spatial distribution of discontinuities and the size of rock blocks are the key indicators for rock mass quality evaluation and rockfall risk assessment. Traditional manual measurement is often dangerous or unreachable at some high-steep rock slopes. In contrast, unmanned aerial vehicle (UAV) photogrammetry is not limited by terrain conditions, and can efficiently collect high-precision three-dimensional (3D) point clouds of rock masses through all-round and multiangle photography for rock mass characterization. In this paper, a new method based on a 3D point cloud is proposed for discontinuity identification and refined rock block modeling. The method is based on four steps: (1) Establish a point cloud spatial topology, and calculate the point cloud normal vector and average point spacing based on several machine learning algorithms; (2) Extract discontinuities using the density-based spatial clustering of applications with noise (DBSCAN) algorithm and fit the discontinuity plane by combining principal component analysis (PCA) with the natural breaks (NB) method; (3) Propose a method of inserting points in the line segment to generate an embedded discontinuity point cloud; and (4) Adopt a Poisson reconstruction method for refined rock block modeling. The proposed method was applied to an outcrop of an ultrahigh steep rock slope and compared with the results of previous studies and manual surveys. The results show that the method can eliminate the influence of discontinuity undulations on the orientation measurement and describe the local concave-convex characteristics on the modeling of rock blocks. The calculation results are accurate and reliable, which can meet the practical requirements of engineering.