Glacial lakes, as the primary carriers of glacier meltwater, can postpone the loss of local glacier freshwater resources to some degree. However, they also offer a breeding ground for Glacial Lake Outburst Floods (GLOFs) and other mountain natural disasters (e.g., landslides, mudslides, etc.). In the mountain glacier zones, glacial lakes play a crucial role in the chain of glacier-related disaster risk. The sudden release of a massive volume of water occurs when a glacial lake dam breaches, is overtopped, or is influenced by other events such as earthquakes and avalanches of ice or rock, which poses a major danger to the downstream infrastructure, possessions, and lives of residents living in high-altitude mountains. Glacial lake evolution and glacial changes are closely related to each other. As glaciers shrink and recede, glacial lakes develop and expand. Effective prevention and management of glacial lake disaster risk requires knowledge of glacial lake changes, in addition to retrospective and investigative studies on past glacial lake outburst flood events. However, due to the distribution of glacial lakes in high-altitude mountain regions, its susceptibility to global warming, and the difficulty in accessing these areas, remote sensing monitoring has emerged as the most practical technical method and provides opportunities for analyzing global climate change and assessing natural disasters. Recent research has indicated an increase in the frequency and impact of GLOFs incidents, emphasizing the growing significance of studying these disasters. Based on this, in this study, we first identified key research areas in recent years through the metrological analysis of the literature on the remote sensing monitoring of glacial lakes and GLOFs. Second, focusing on three main directions of the research on glacial lakes and GLOFs (109 important research literatures), namely remote sensing monitoring of glacial lakes and GLOFs, response analysis of glacial lake evolution in the context of climate change, and glacial lake risk assessment with case studies of GLOFs, ten essential topics of recent research advances at home and abroad as well as the shortcomings of current studies are systematically summarized and analyzed. Finally, the direction of future research is prospected, including extraction of glacial lake morphology using artificial intelligence and GLOFs events inventory, glacier-glacial lake (especially for proglacial, supraglacial lake) system evolution and its relationship to climate change, glacial lake monitoring, and early warning and disaster prevention. Our review offers references for the management and adaptive planning of glacial lake and mountain glacier related catastrophes.

Remote sensing images contain rich semantic information and play an important role in landslide disaster monitoring. Traditional landslide recognition is mainly based on remote sensing visual interpretation and human-computer interaction recognition, which is time and labor consuming, with strong subjectivity and low extraction accuracy. Semantic segmentation, as an important task in deep learning, has played an important role in automatic recognition tasks using remote sensing images due to its end-to-end, pixel-level classification capability and has great potential in automatic recognition of landslides. The existing semantic segmentation models for landslides using remote sensing images usually lack the feature information of multi-scale ground objects, and the boundary will be blurred with the increase of network depth. In this paper, Attention combined with Encoder-Decoder Network (AED-Net) is proposed for landslide recognition. A shallow feature extraction network is used to alleviate the boundary ambiguity caused by deep neural network. Multi-scale feature extraction capability of convolution pool pyramid structure in void space is utilized. Combined with the feature restoring ability of the encoder-decoder structure, the boundary information is restored, and the channel attention mechanism is used to enhance the key feature learning ability of the model. The focal-loss function is used to alleviate the imbalance of positive and negative samples. In our study, firstly, the GID-5 data set is used to conduct comparative tests on the expansion rate setting of void convolutions and the selection of channel attention mechanism in the model to get the optimal solution. Then, the feature weight is transferred to the semantic segmentation task for landslide disaster by using transfer learning method, and the hyperparameter discussion and ablation experiment are carried out. The resulting model achieves the optimal segmentation performance on the landslide disaster data set of Bijie City, with a Pixel Accuracy (PA) of 95.58%, the Mean Pixel Accuracy (MPA) of 89.24%, and the Mean Intersection over Union (MIoU) of 82.68%. Compared with classical semantic segmentation networks such as PSP-Net, Attention U-Net, DeeplabV3+ with ECA attention mechanism, and semantic segmentation models such as PA-Fov and LandsNet for classfifying landslide disasters, the pixel accuracy of our model increases by 0.73%~1.97%. The average pixel accuracy of all categories increases by 1.0%~2.84%, and the average interaction ratio increases by 2.25%~5.11%. Moreover, the edge information of landslide image is smoother and the multi-scale landslide segmentation accuracy is better than other deep learning models, which demonstrates the effectiveness of the proposed model in semantic segmentation of landslides from remote sensing images.

Landslide identification is a fundamental work in geological survey. Characterizing the distribution of earthquake-induced landslides can help define where areas are prone to landslides and where areas are safe, which is important for subsequent post-earthquake reconstruction. At present, the co-seismic landslide is usually identified by remote sensing interpretation and field survey based on color changes and textual characteristics such as crown cracks, accumulation, and flank. However, the identification and segmentation of large-area earthquake-induced landslides often takes a lot of time and resources, making it difficult to provide timely relevant data to rescuers in disasters. Recently, the combination of Google Earth Engine (GEE), Colab, and Cloud storage provides new technique support for post-earthquake landslide interpretation. Remote sensing images can be quickly extracted in GEE, learned in the Colab platform, and stored in Cloud storage. Based on these technologies, a U-net model built on Tensorflow was integrated in the GEE to identify co-seismic landslides of Wenchuan. We tested three combinations of different remote sensing data to identify landslides, namely Landsat 5 (LT05), LT05+slope+DEM, and LT05+slope+DEM+NDVI. The results showed that all these three combinations had an accuracy of 0.71~0.87, precision of 0.66~0.75, recall of 0.29~0.57, and mIoU of 0.50~0.64. Due to the complexity of landslides in remote sensing images, the U-net model still resulted in misclassification and cannot replace human interpretation completely. However, this model can help us in interpreting earthquake-induced landslides to some extent. From the detailed images, there is a high possibility for areas that have been marked as landslides, but there may also be a large number of landslides for areas that are not marked as landslides. The accuracy and precision increased about 1% in all three data combinations with the increasing number of parameters, but the number of false positives also increased with the increasing number of parameters. The NDVI, a derivative image from LT05, enhanced learning efficiency and accuracy, which indicated that the prior knowledge was very important in a neural network model. By comparison, the model can segment large-scale landslides well in the case of using the combination of LT05+NDVI+DEM+slope. However, due to the limited image resolution, the segmentation accuracy of small landslides was still not good. We need higher spatial resolution images to improve segmentation accuracy. This study demonstrates the computation efficiency of the GEE+Colab+Cloud storage platform, which provides a technique support for researchers to conduct object recognition quickly using remote sensing data based on neural networks.

China is one of the countries most severely affected by geological disasters. Researching high-precision and highly reliable methods for monitoring and predicting landslide deformation holds practical significance for disaster prevention and mitigation efforts. Using the massive Outang landslide in the Three Gorges Reservoir Area as a case study, this paper addresses the issue of the atmospheric interference in extracting landslide deformation using time-series InSAR technology. To correct for atmospheric effects, the GACOS model is introduced and validated against GNSS observation data. To address the often-overlooked temporal-spatial analysis before landslide deformation prediction, the Moran index and Hurst index are calculated to analyze the spatiotemporal characteristics of landslide deformation. Recognizing that landslide deformation is influenced not only by historical deformation but also by various external factors, this paper proposes coupling landslide influencing factors with deformation data for prediction. A Long Short-Term Memory (LSTM) model, optimized by Variational Mode Decomposition (VMD) and the Sparrow Search Algorithm (SSA), is employed for the prediction. By decomposing landslide displacement data into trend, periodic, and random components using VMD, the LSTM network structure is constructed. SSA is used to determine the optimal number of hidden units, maximum training periods, and the initial learning rate of the LSTM model. Additionally, methods such as data normalization, regularization, and model evaluation are employed to enhance the performance and stability of the LSTM model. Finally, the model is trained using the influencing factors and decomposed displacement data to predict landslide deformation. The results indicate that: (1) From January 2021 to June 2023, the maximum and minimum deformation rates of the Outang landslide were -72.75 mm/a and 50.74 mm/a, respectively; (2) The deformation in the study area exhibits positive spatial autocorrelation, with the landslide in the settlement area showing a persistent trend; (3) The prediction error of the LSTM model optimized by VMD and SSA is only 0.37 mm, representing an 11.004% accuracy improvement compared to the standard LSTM model. Based on time-series InSAR technology and spatiotemporal analysis results, this paper constructs a high-precision prediction model for landslide deformation, incorporating multiple influencing factors. This model provides a valuable reference for the prevention and control of landslide disasters.

Land subsidence is an important factor that influences the sustainable development of a region. Due to the complexity of land subsidence, the uncertainty and risk caused by land subsidence disasters are increasing. Therefore, new methods need to be developed to quantify the nonlinear land subsidence processes, identify emerging risk, and improve urban resilience. In this paper, the necessity of introducing peridynamic to land subsidence modeling is discussed by analyzing the progress and shortcomings of current land subsidence modeling. For natural discontinuous structures such as fractures and faults, current deterministic models based on differential equations are insufficient to describe land subsidence. Therefore, the peridynamic theory which is suitable for discontinuous and nonlinear characteristics is introduced. The peridynamic theory (PD) describes the mechanical behavior of matter by solving integral equations and has advantages in analyzing discontinuous and multi-scale problems. The applicability of peridynamic in land subsidence is analyzed from the aspects of material properties and modeling methods, respectively. By establishing a peridynamic model of land subsidence, discontinuous disasters such as ground crack and ground collapse can be included, so as to realize the multi-field and multi-scale recognition of land subsidence under a unified framework. In the light of the “Higher-bigger-deeper” urban construction, combined with the CAS-ESM, the simulation of future evolution of ground subsidence and ground fractures can be carried out. However, there are still problems to be solved in the interdisciplinary research, such as the reasonable generalization of material properties, material structure, and the balance between operation accuracy and operation cost. Then, based on theoretical principles, the modeling method, solving process, and optimization method of peridynamic land subsidence model are given. Besides the establishment, solution and optimization of the model, a variety of spatial monitoring methods and data are also needed, e.g., subsidence data monitored by InSAR technology, the underground structure and density information obtained by Seismic Frequency Resonance Technology (SFRT), bedrock and stratified scale data, groundwater level data, building information data, and road network data. In this paper, a peridynamic land subsidence model with a range of 4km*6km and a depth of 0.2 km is established in Liyuan-Taihu -Zhangjiawan area in the eastern Beijing, and the evolution process of land subsidence is simulated by using the monthly average rate of groundwater level decline from 2007 to 2010 as the boundary condition. The mean absolute error between the simulated and the measured values is 18mm, which verifies the effectiveness of this interdisciplinary research. The peridynamic theory has superiority in the field of materials and the study of fatigue, damage, fracture, and so on. Our study provides new ideas and new methods for regional land subsidence modeling. Furthermore, with the support of big data, cloud computing platforms, and Geo-AI, new opportunities are emerging for preventing, controlling, slowing down, and avoiding land subsidence hazards.

Landslides frequently occur in the mountainous areas of western China. Accurate mapping of landslide susceptibility is essential for geohazard management. Integrated models combining statistical methods and machine learning models have been widely applied to landslide susceptibility mapping. However, further optimization of their results is still worth investigation. This study proposes a comprehensive assessment method that couples statistical methods, machine learning models, and clustering algorithms. The effectiveness of the proposed method on improving the accuracy of landslide susceptibility mapping in Ningnan County is investigated. Firstly, the landslide influencing factors are selected from five aspects: geological environment, topography and geomorphology, meteorology and hydrology, vegetation and soil, and human engineering activities in the study area. Indicators are initially selected based on correlation analysis using the Pearson correlation coefficient method, and highly correlated factors are eliminated to establish the landslide susceptibility mapping index system. Next, the Information Value (IV), Certainty Factor (CF), and Frequency Ratio (FR) methods are combined with Random Forest (RF) model respectively to obtain three integrated models (IV-RF, CF-RF, and FR-RF). Then, the ISO clustering algorithm, Natural Breaks clustering, and Kmeans clustering algorithms are introduced to classify the results of the three integrated models, obtaining nine coupled assessment models (IV-RF-ISO, CF-RF-ISO, FR-RF-ISO, IV-RF-NBC, CF-RF-NBC, FR-RF-NBC, IV-RF- Kmeans, CF-RF- Kmeans, and FR-RF- Kmeans). Lastly, Area Under the Curve value (AUC), accuracy, F1 score, and Seed Cell Area Indexes (SCAI) are used to evaluate the accuracy of the models. The results demonstrate that all the integrated models outperform single models. The accuracy and F1 score of all integrated models both exceed 0.85, and their AUC values exceed 0.9. The integrated models effectively address the misclassification of non-landslide samples, which is especially prominent in single IV and CF models. Among the integrated models, the FR-RF model performs the best. The accuracy (0.911), F1 score (0.912), and AUC value (0.965) of FR-RF model improves by 0.095, 0.096, and 0.074, respectively, compared to the FR model. Compared with the natural break and Kmeans clustering methods, the coupled FR-RF-ISO model exhibits the optimal classification results, and the difference in SCAI values between its high and low susceptibility zones is more significant. The extremely high landslide susceptibility zones are primarily concentrated in the southern, eastern, and central parts of Ningnan County. The study demonstrates the high accuracy of the integrated assessment method that couples statistical methods, machine learning, and clustering algorithms, and provides insights for improving the accuracy of landslide susceptibility mapping.

Hazard assessment of sudden geological disasters is of great significance for disaster prevention and risk management. Due to different factors affecting the occurrence of disasters in different regions, it is difficult to select appropriate factors comprehensively and objectively in an actual evaluation process. Machine learning has unique advantages in dealing with high-dimensional nonlinear problems of disaster systems, but its evaluation performance is limited because the model is difficult to tune. This paper attempted to propose a two-way optimization method for landslide hazard assessment. Based on a factor sensitivity index built for quantitative sensitivity analysis, combining importance analysis, correlation analysis, and collinearity analysis, and following the principle of “guarantee sensitivity, retain importance, eliminate correlation, and avoid collinearity", a four-dimensional (4D) feature screening method was constructed to evaluate the comprehensive optimization of factors. In order to overcome the problem that the model is difficult to tune, the Differential Evolution (DE) algorithm was further introduced. Two machine learning models with strong generalization ability, i.e., Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP), were optimized. Finally, we took the landslide in Fujian Province as an example to verify the proposed evaluation method. We found that the 4D feature screening method can more objectively and comprehensively select suitable hazard assessment factors, thereby reducing the data dimension and reducing information redundancy to improve the performance of the assessment model. Ten suitability assessment factors were finally used for landslide hazard assessment in Fujian Province including aspect, variance coefficient in elevation, land use type, average annual rainfall, surface cutting depth, distance to river, distance to road, engineering geological rock group, topographic wetness index, and stream power index. The DE algorithm can obtain better hyperparameters from global search and has a significant optimization effect on SVM and MLP, which is beneficial to improve the evaluation accuracy of the landslide hazard of the model. Compared with the unoptimized models, the AUC values of DE-SVM and DE-MLP increased by 4.43% and 4.37%, respectively. The results of landslide hazard assessment based on two-way optimization show that rainfall and land use types have an important impact on the occurrence of landslides in Fujian Province. Terrain curvature elements, terrain variability elements, and fault structures have little impact on landslide occurrence. The extremely high-hazard areas generally have high annual rainfall and complex and changeful terrain. The extremely low-hazard areas are mainly located along the southeast coast and on both sides of the Minjiang River Basin. This research provides some ideas for objective selection of influencing factors in landslide hazard assessment and machine learning model tuning.

The single machine learning-based landslide susceptibility prediction model has different focuses of features and a weak classification ability, and also the accuracy of traditional random sampling of non-landslide is low. To solve these problems, this study optimized Non-Landslide Samples (NLS) based on the information value model and utilized Stacking heterogeneous ensemble models to evaluate the landslide susceptibility of Fengjie County in the Three Gorges Reservoir. Firstly, 16 evaluation indexes were extracted based on multiple sources of topographic, geologic, and remote sensing data, including elevation, slope, aspect, profile curvature, plan curvature, lithology, distance to fault, topographic wetness index, stream power index, distance to river, normalized difference vegetation index, distance to road, and land use, and the correlation analysis was carried out to exclude high correlation indicators and construct the landslide susceptibility evaluation criteria system. Then, the NLS index was proposed based on the information value model to divide the non-landslide samples into two categories: information values less than or equal to 0, and greater than 0. Finally, the logistic regression model was used to compare the non-landslide samples under different NLS conditions, and the NLS index was used to obtain optimized non-landslide samples, which forms the training set with the same number of landslide samples. Finally, Random Forest (RF), Light Gradient Boosting Machine (LGBM), Gradient Boosting Decision Tree (GBDT), and homogeneous (Boosting-RF, Boosting-LGBM, Boosting-GBDT) and heterogeneous (Stacking) ensemble methods based on these three models were compared for susceptibility evaluation. The results show that non-landslide sampling using NLS can produce non-landslide samples of high quality and generalization ability, which in turn improves the learning ability of the model and the accuracy of susceptibility evaluation. The Stacking heterogeneous ensemble model has the best accuracy of 0.941, which is better than the Boosting homogeneous ensemble models (an accuracy of 0.902, 0.897, 0.870, respectively) and other single models (an accuracy of 0.882, 0.864, 0.855, respectively). These results indicate that the Stacking heterogeneous ensemble algorithm is capable of extracting landslide and non-landslide features from various spatial angles, realizing the complementary advantages and disadvantages of the models, significantly improving the performance of machine learning, and obtaining better predictions, and thus is a reliable landslide susceptibility evaluation model. This study contributes to a better understanding of the landslide activity, improves the reliability of regional landslide hazard risk assessment, and provides support for carrying out reasonable land use planning, disaster prevention, and mitigation strategies.

Landslide disasters pose a grave threat to national infrastructure projects, making landslide disaster risk assessment in engineering areas crucial for ensuring the safety of railway operations. Typically, such assessments involve employing frequency analysis, probability analysis, and deterministic analysis methods. Among these methods, the establishment of a physical deterministic model based on the mechanisms of rainfall infiltration and water accumulation yields more objective evaluation results and exhibits excellent applicability. However, physically deterministic models often necessitate the inclusion of numerous geotechnical parameters in their calculations, leading to certain limitations. Factors like temporal and spatial variability, as well as the uncertainty in geotechnical parameters, influence these models. To enhance the prediction accuracy of landslide risk assessments, this study focuses on the Gaojiawan landslide as the research area. By utilizing the particle filter algorithm, the study assimilates safety factor (Fs) data from the TRIGRS model, incorporating SBAS-InSAR observation data. Additionally, it updates the internal friction angle parameter of the model. The results reveal a gradual decrease in the safety factor of the Gaojiawan landslide following assimilation. Moreover, the safety factor at the front edge of the slope is significantly lower than that at the rear edge, aligning more closely with current observations. Real-time updates of the internal friction angle parameters are achieved, gradually aligning the parameters with the observed values. As a result, the root mean square deviation of the model decreases from 0.17 to 0.04, bringing the model's prediction closer to the observed values. Consequently, landslide risk assessment based on the particle filter assimilation method more accurately reflects the current landslide situation and exhibits higher prediction accuracy.

The urbanization in mountainous areas will cause land use changes, thereby potentially altering the landslide susceptibility. However, there is a lack of research on landslide susceptibility under future land use dynamics in landslide hazard management at present. In this study, taking the towns along the Yangtze River in Wanzhou District, Chongqing City as an example, we first used the Patch-generating Land Use Simulation (PLUS) model to predict the dynamic factors of land use in 2025 and 2030. Then, combined with the static factors collected and landslide data in the study area, we constructed a database for susceptibility assessment. The Particle Swarm Optimization (PSO) algorithm was used to adjust the hyperparameters of the ensemble learning models based on the GBDT (Gradient Boosting Decision Tree) framework, including XGBoost, LightGBM, and CatBoost, and conduct dynamic susceptibility modeling. Finally, we mapped the landslide susceptibility under scenarios of future land use dynamics based on the optimal evaluation model. The results show that： (1) the land use change results based on the PLUS model can provide accurate basic data for predicting future landslide-prone areas; (2) by comparing the AUC results before and after the involvement of the PSO algorithm in the modeling, we found that the PSO-CatBoost model was able to better fit the nonlinear relationship between the landslide inventory and the triggering factor in the study area, and achieved an optimal modeling accuracy; (3) our research framework that integrates future land use dynamics and landslide susceptibility suggested that future development of building lands in townships along the Yangtze River will increase landslide susceptibility. The research approach proposed in this paper is of great practical significance for building a resilient and effective disaster prevention and reduction system in mountainous urban areas and can provide scientific guidance for the future planning of mountainous urban areas.

The causes of landslide disasters are complex, and landslide susceptibility assessment is of great significance for disaster warning, prevention, and control management. In the previous mapping studies on landslide susceptibility assessment, land use change factor was not considered. This paper proposed a combination of factors for landslide susceptibility assessment by considering land use dynamic change factor. The landslide frequency ratio was used to quantitatively measure the correlation between land use change and landslide development. And Logistic Regression (LR) model was used to compare the prediction ability of the model before and after the introduction of land use change factor. We constructed three machine learning models: Decision Tree (DT), Gradient Boosting Decision Tree (GBDT), and Random Forest (RF). We used AUC and other indicators to compare model performance. Finally, we took Sanming City of Fujian Province as the study area and the whole Fujian Province as the verification area to conduct the landslide susceptibility assessment research. The results show that there is a strong correlation between land use change factor and landslide development. The inclusion of land use change factor improves model prediction accuracy, which indicates that it is necessary to introduce dynamic factor in the assessment of landslide susceptibility. The verification results show that RF model has higher prediction accuracy than DT and GBDT. The high landslide prone areas are mainly distributed in the west and central of Sanming City, where the land use change degree is high, and the impact of human activities is great. The low landslide prone areas basically locate in the high-altitude areas with little influence of human activities. This study provides a new research perspective for landslide susceptibility assessment and helps to explore the impact of human activities on disaster formation.

[Objectives] To investigate the use of temporal deformation fractal features for identifying landslides in alpine glaciated areas and analyze their applicability. [Methods] The deformation time series of the Chamoli landslide and its neighboring glacier were characterized using slope (average deformation rate) and fractal features (fractal dimension and fractal goodness of fit). Cluster analysis was used to distinguish landslide areas from glaciers and analyze influencing factors. [Results] The deformation time series of landslides exhibited higher fractal dimensions and lower fractal goodness of fit compared to glaciers. While significant differences in the slope of deformation time series (average deformation rate) were observed between landslides and glaciers, clustering analysis based solely on deformation rate achieved an accuracy of only 61.70%. In contrast, using fractal indexes of the deformation time series (fractal dimension and fractal goodness of fit) significantly improved clustering accuracy to nearly 84.00%. The applicability of this method is attributed to intrinsic differences in material composition, influencing factors, and developmental evolution between landslides and glaciers. Compared to glaciers, landslides are more complex in material composition, influenced by multiple factors, and exhibit greater variability in their deformation time series. [Conclusions] The study demonstrates the feasibility of identifying landslides in alpine glaciated areas using fractal features of deformation time series. In the context of global warming, this method has the potential to support landslide identification and contribute to disaster prevention and mitigation efforts in alpine glacier regions.

[Objectives] Using deep learning methods for landslide identification can significantly improve efficiency and is of great importance for landslide disaster prevention and mitigation. The DeepLabV3+ algorithm effectively captures multi-scale features, thereby improving image segmentation accuracy, and has been widely used in the segmentation and recognition of remote sensing images. [Methods] We propose an improved model based on DeepLabV3+. First, the Coordinate Attention (CA) mechanism is incorporated into the original model to enhance its feature extraction capabilities. Second, the Atrous Spatial Pyramid Pooling (ASPP) module is replaced with the Dense Atrous Spatial Pyramid Pooling (DenseASPP) module, which helps the network capture more detailed features and expands the receptive field, effectively addressing the limitations of inefficient or ineffective dilated convolution. A Strip Pooling (SP) branch module is added in parallel to allow the backbone network to better leverage long-range dependencies. Finally, the Cascade Feature Fusion (CFF) module is introduced to hierarchically fuse multi-scale features, further improving segmentation accuracy. [Results] Experiments on the Bijie landslide dataset show that, compared with the original model, the improved model achieves a 2.2% increase in MIoU and a 1.2% increase in the F1 score. Compared with other mainstream deep learning models, the proposed model demonstrates higher extraction accuracy. In terms of segmentation quality, it significantly improves the overall accuracy in identifying landslide areas, reduces misclassification and omission, and yields more precise delineation of landslide boundaries. [Conclusions] Based on experiments using the landslide debris flow disaster dataset in Sichuan and surrounding areas, along with practical application verification, the proposed method demonstrates strong recognition capability across landslide images in diverse scenarios and levels of complexity. It performs particularly well in challenging environments such as areas with dense vegetation or proximity to rivers, showing strong generalization ability and broad applicability.

[Objectives] The quality of training samples significantly impacts model performance and prediction accuracy. In regions with limited sample data, the small number of samples and their uneven spatial distribution may prevent the model from effectively learning the features of disaster-inducing factors. This increases the risk of overfitting and ultimately affects the accuracy of model predictions. Therefore, it is crucial to collect and optimize training samples based on regional characteristics. [Methods] To address this issue, this study proposes a sampling optimization method for training samples. The method combines the Prototype Sampling (PBS) approach for selecting landslide-positive samples with an unsupervised clustering model for training sample selection. This results in a screened and expanded positive sample dataset and an objectively extracted negative sample dataset, forming an optimized training sample dataset. Subsequently, the Random Forest (RF) and Support Vector Machine (SVM) models, which are well suited for handling small sample data, were employed to construct a landslide susceptibility evaluation model. Comparative experiments were conducted using Raw Data (RD), a dataset with only Data Augmentation (DA), and the optimized dataset. Model prediction performance was assessed using metrics such as the Area Under the Curve (AUC). Additionally, the frequency ratio method was applied to optimize the results of landslide susceptibility zoning. Finally, a case study was conducted in Putian City, where landslide sample data is relatively scarce, to verify the effectiveness and generalization capability of the proposed sampling optimization method. [Results] The results indicate that models trained on the SO dataset achieved AUC improvements of 10.69% and 18.23% compared to those trained on the RD and DA datasets, respectively, demonstrating a significant enhancement in predictive performance. This suggests that selecting and expanding positive samples while objectively extracting negative samples can improve model accuracy and mitigate the overfitting problem during training. Furthermore, the frequency ratio analysis revealed that the SO-RF model achieved higher frequency ratios in regions with extremely high and high susceptibility than the SO-SVM model, indicating that SO-RF is more suitable for evaluating landslide susceptibility in regions with limited landslide sample data, such as Putian City. [Conclusions] The proposed training sample optimization approach, combined with machine learning evaluation methods, demonstrates high applicability and accuracy. Therefore, the findings of this study provide valuable insights into machine learning-based sampling strategies for landslide susceptibility assessment.

[Objectives] The influence of negative sample selection and machine learning models on landslide susceptibility evaluation cannot be overlooked. [Methods] To investigate the impact of these two factors on landslide susceptibility assessment, this study examines the Nujiang Valley section of the Nujiang River Basin. A weighted information quantity model was proposed to optimize negative sample selection. Thirteen influencing factors, including topography, land use, and average annual rainfall, were selected. Three machine learning models were employed: Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Gradient Boosting Decision Tree (GBDT). A comparative analysis of landslide susceptibility was conducted against traditional random sample selection methods. Additionally, the effect of rainfall factors on susceptibility classification was analyzed. [Results] The results indicate that: (1) The optimized negative sample selection improved landslide density by 0.0103, 0.0639, and 0.004 0, respectively, for the three models. The AUC values increased by 0.033, 0.018, and 0.008, respectively. (2) Among the susceptibility evaluation models, the GBDT model performed best, improving accuracy by 3.8% and 1.7% compared to the SVM and CNN models, respectively. (3) Incorporating average monthly rainfall data for summer and winter (2019—2020) into the GBDT model revealed an increase in high and relatively high susceptibility zones during summer, particularly in the southern regions of Liuku Town and Shangjiang Town. [Conclusions] The optimization of negative samples based on the weighted information quantity model is reasonable and effective. As a landslide susceptibility evaluation model, the GBDT model is the most suitable for the disaster-prone environment of the Nujiang Valley, where precipitation significantly impacts landslide susceptibility.

[Objectives] This study aims to investigate the spatiotemporal evolution characteristics and driving mechanisms of landslide evolution in the Three Gorges Reservoir Area (TGRA) under compound extreme weather events in 2022, characterized by severe drought in the Yangtze River Basin and localized heavy rainfall within the reservoir area. It also seeks to address the knowledge gap in understanding landslide evolution under extreme climatic conditions. [Methods] Focusing on the Zigui-Fengjie section, this study utilized Sentinel-1 SAR data and the Small Baseline Subset (SBAS) InSAR technique to monitor surface deformation and identify active landslides. A dynamic evaluation of landslide susceptibility was conducted combining the information value model and SBAS-InSAR results. Based on reservoir water level fluctuations, the study period was divided into two intervals: normal weather (July 2020 to July 2022) and extreme weather (July 2022 to September 2023), to comparatively analyze landslide evolution patterns and driving mechanisms. [Results] The results are as follows: (1) A total of 136 active landslides were identified. The most favorable geomorphic conditions for landslide development included slopes of 10°-30°, southeasterly to northwesterly orientations, elevations of 100~400 m, distances to rivers less than 200 m, and distributions mainly in clastic and mixed clastic-carbonate rock areas. Many landslides were located in rainfed farmland within 100 m of roads. (2) The combination of InSAR technology with traditional landslide susceptibility assessment models enabled dynamic assessment. The method can be updated synchronously with InSAR deformation data to reflect the current state of landslide evolution in a timely manner. (3) Under extreme weather conditions, landslide risks in the study area increased significantly, while the relatively low water level of the Three Gorges Reservoir had no significant negative impact on reservoir bank stability. (4) Precipitation was identified as the primary driver of dynamic landslide susceptibility evolution, with the susceptibility-precipitation response varying considerably across regions with different lithologies. [Conclusions] By integrating SBAS-InSAR time-series deformation monitoring with the information value model, this study reveals the spatiotemporal variability of landslide risk under extreme weather conditions. It addresses critical gaps in understanding landslide evolution mechanisms in the TGRA and provides a scientific foundation for landslide monitoring, early warning, and risk management.