Tianbao Huo’s research while affiliated with Lanzhou Jiaotong University and other places

The complex structures and intricate hyperparameters of existing deep learning models make achieving higher accuracy in landslide susceptibility assessment timeconsuming and labor-intensive. Deep forest (DF) is a decision treebased deep learning framework that uses a cascade structure to process features, with model depth adapting to the input data. To explore a more ideal landslide susceptibility model, this study designed a landslide susceptibility model combining convolutional neural networks (CNN) and DF, referred to as CNN-DF. The Bailong River Basin, a region severely affected by landslides, was chosen as the study area. Firstly, the landslide inventory and influencing factors of the study area were obtained. Secondly, an equal number of landslide and non-landslide samples were selected under similar environmental constraints to establish the dataset. Thirdly, CNN was used to extract high-level features from the raw data, which were then input into the DF model for training and testing. Finally, the trained model was used to predict landslide susceptibility. The results showed that the CNN-DF model achieved high prediction accuracy, with an AUC of 0.9061 on the testing set, outperforming DF, CNN, and other commonly used machine learning models. In landslide susceptibility maps, the proportion of historical landslides in the very high susceptibility category of CNN-DF was also higher than that of other models. CNN-DF is feasible for landslide susceptibility assessment, offering higher efficiency and more accurate results. Additionally, the SHAP algorithm was used to quantify the contribution of features to the prediction results both globally and locally, further explaining the model. The landslide susceptibility map based on CNN-DF can provide a scientific basis for landslide prevention and disaster management in the target area.

Global warming is accelerating the permafrost degradation along the Qinghai-Tibet Railway (QTR), causing the surface deformation (SD) of the railway subgrade. Especially in the Salt Lake to Wuli section of the QTR, the permafrost is widely distributed, and the SD has been the most serious. However, the spatio-temporal characteristics and mechanism of SD are still unclear. In addition, it is very important to predict the future trend of SD. Therefore, we acquired time series SD results from 2019 to 2022 based on Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) and analyzed the spatio-temporal characteristics and mechanism of SD in the Salt Lake to Wuli section. Subsequently, the EnvCA-GRU model for SD prediction was developed, integrating the Multi-Head Cross-Attention (MHCA) mechanism and Gated Recurrent Unit (GRU) to account for changes in environmental factors (EFs). The model was then employed to forecast SD trends over the next two years. Our results showed that the SD was uneven in the Salt Lake to Wuli section of the QTR from 2019 to 2022, there were six typical deformation areas, and the maximum cumulative ground subsidence reached 126.79 mm. The SD velocity of the sunny slope was higher than that of the shady slope, and the closer to the QTR, the greater the ground subsidence. Land surface temperature (LST), normalized difference vegetation index (NDVI), and precipitation are the main factors affecting SD. Our proposed EnvCA-GRU prediction model fusing NDVI, LST, and precipitation showed an RMSE of 0.153 and an R² of 0.991, the proposed model was reliable. The maximum cumulative ground subsidence of six typical areas by July 2024 reached 177.52, 268.08, 287.73, 270.99, 190.70, and 211.89 mm, respectively. The results of this study can play a guiding role in the early warning and mitigation of ground subsidence disasters along the QTR.

Thaw slump susceptibility mapping (TSSM) of Qinghai-Tibet Railway corridor (QTRC) is the prerequisite and basis for disaster assessment and prevention of permafrost projects. The objective of this study is to construct ensemble learning models based on single classifier models to generate the TSSM of the QTRC, compare and verify the performance of the models, and further explore the relationship between the high susceptibility area and environmental factors of the QTRC. The collinearity analysis was carried out by selecting 14 thaw slump conditioning factors (TSCFs). We used the balance bagging method for sample optimization, and the data set was divided into 70% training set and 30% verification set. Convolutional neural network (CNN), multilayer perceptron (MLP), support vector regression (SVR), random forest (RF) single classifiers were selected to construct blending and stacking ensemble learning models for the TSSM. The results showed that there was no collinearity among the 14 TSCFS. The comparison of model performance revealed that all models had good performance, but the constructed stacking and blending ensemble learning models had stable performance and high prediction accuracy for TSSM. The stacking ensemble learning model had the best effect, and the area under curve (AUC) value of receiver operating characteristic (ROC) curve reached 0.9607. It showed that the generated TSSM of QTRC based on stacking ensemble learning model had the highest reliability. The QTRC has local areas with high thaw slump susceptibility, mainly concentrated in the permafrost areas with high altitude, high slope, adjacent faults, sparse vegetation, ice and snow and the more cumulative precipitation.

The ancient Suoertou landslide seriously threatens the surrounding population's lives and property. Monitoring and predicting this landslide is crucial to ensure the affected areas’ safety. The previous research on the landslide's displacement characteristics and mechanisms has lacked detailed analyses. Additionally, its future development trends must be understood. Therefore, we conducted a detailed analysis of the ancient Suoertou landslide's displacement characteristics and mechanisms using small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) monitoring results from 2018 to 2023. Furthermore, by applying a gated recurrent unit (GRU) prediction mode that incorporates the refined displacement characteristics and mechanisms, we forecasted the landslide's displacement trends. The results show that this landslide is currently undergoing overall slow displacement with violent fluctuations in localized areas. Tectonic movement, precipitation, human activities, river erosion, and other factors interact, forming a vicious development displacement mechanism. According to our prediction, the displacement of this landslide will be in a trend of fluctuating increase from June 2023 to June 2024. In particular, the local area will undergo abnormal displacement acceleration. The results of the current research provide a scientific basis upon which to monitor landslides, promote their management, and reduce the risk of losses due to landslide disasters.

Accurate surface deformation (SD) predictions are critical for early warning and timely remediation of infrastructure damage. However, the current SD prediction models do not integrate the multi-scale features of InSAR SD and environmental factors (EFs), which makes their prediction results inaccurate. To address this limitations, we proposed a bi-directional gated recurrent unit (BiGRU) multi-output SD prediction network (GLER-BiGRUnet), which mainly included global-local feature extraction (GLFE), multi-factor cross-attention residual (MCAR) and local residual module embedded in self-attention mechanism (RCSA) modules. Specifically, Dense and Conv1D were concatenated in the GLFE module to extract global-local SD features. The long time-series dependence between EFs and SD was learned in the MCAR module using the Multi-head cross attention (MHCA) mechanism to obtain the corresponding attention weight feature matrix. The residual connection and self-attention (SA) mechanisms were used in the RCSA module to merge the multi-scale features and enhance the model fitting ability. We chose four typical regions in the permafrost area of Qinghai-Tibet Railway (QTR) as the scene for the experiment. The spatial distribution and local profile exhibited relatively small discrepancies between the prediction results of the GLER-BiGRUnet model and the InSAR SD. Meanwhile, the average RMSE of the GLER-BiGRUnet model in the four typical regions was 0.19 mm, and the proposed model had the best evaluation index compared with other SD prediction models. Additionally, the prediction trend of SD of the proposed GLER-BiGRUnet model was consistent with the original InSAR SD, and the prediction results were more stable than those of other prediction models. The SD prediction model proposed in this paper contributes to early warning of surface deformation.

Landslide susceptibility mapping (LSM) is of great significance for regional land resource planning and disaster prevention and reduction. The machine learning (ML) method has been widely used in the field of LSM. However, the existing LSM model fails to consider the correlation between landslide and disaster-prone environment (DPE) and lacks global information, resulting in a high false alarm rate of LSM. Therefore, we propose an LSM method with GraphTransformer that considers the DPE characteristics and global information. Firstly, correlation analysis and importance analysis are employed on nine landslide contributing factors (LCFs), and the landslide dataset is generated by combining remote sensing image interpretation and field verification. Secondly, a graph constrained by environment similarity relationship is constructed to realize the correlation between landslide and DPE. Then, the Transformer module is introduced to construct a Graph-Transformer model that considers the global information. Finally, the LSM is generated and analyzed, and the accuracy of the proposed model is compared and evaluated. The experimental results show that the environment similarity relationship graph effectively improves the accuracy of the models and weakens the influence of environmental differences on the models. Compared with graph convolutional network (GCN), graph sample and aggregate (GraphSAGE), and graph attention network (GAT) models, the AUC value of the proposed model is more than 2.05% higher under the environment similarity relationship. In addition, the AUC value of the proposed model is more than 8.8% higher than that of traditional ML models. In conclusion, our proposed model framework can get better evaluation results than most existing methods, and its results can provide effective ways and key technical support for landslide disaster investigation and control