Te Pei

Landslide Susceptibility Assessment (LSA) is essential for disaster mitigation and spatial planning, especially in regions with complex terrain and increasing development pressure. This study focuses on the northeastern U.S., particularly the New York Tri-State Area, where glacially shaped hilly terrain, steep slopes, and variable elevations, combined with frequent extratropical cyclones and intense rainfall, contribute to spatially heterogeneous landslide risk. Given the area’s dense population and economic significance, high-resolution evaluation and management are therefore urgently needed. While machine learning (ML) shows significant potential in LSA, many existing models fail to account for spatial heterogeneity among geo-environmental variables and often lack explainability due to their black-box nature. To address these limitations, we propose a geographically explainable AI framework that combines advanced ensemble machine learning with interpretable techniques to better capture local variability and provide spatial insights into landslide susceptibility. In this study, landslide inventories were compiled from the United States Geological Survey and state agency databases. Multi-source geospatial data, including topographic, hydrological, geological, and anthropogenic factors, were used to develop a high-resolution landslide susceptibility map for the study area based on spatial cross-validation. Results show that the proposed framework outperforms several traditional ML models such as Random Forest and LightGBM in predictive performance. In addition to improved accuracy, the framework captures spatial heterogeneity by combining global and localized models through spatial weighting, allowing it to reflect locally varying patterns of landslide susceptibility. It also identifies the relative influence of landslide contributing factors across different geographic contexts, revealing nonlinear relationships and threshold effects between key variables and landslide occurrences. These capabilities enhance the overall interpretability of LSA. The proposed framework is applicable to regional-scale geohazard assessments and offers valuable insights into resilient land-use planning and adaptive disaster mitigation strategies in the Northeastern U.S.

Landslide risk is traditionally predicted by process‐based models with detailed assessments or point‐scale attribute‐based machine learning (ML) models with first‐ or second‐order features (e.g., slope and curvature) as inputs. One could hypothesize that terrain patterns might contain useful higher‐order information that could be extracted, via computer vision ML models, to elevate prediction performance beyond that achievable with attribute‐based models. We put this hypothesis to the test in the state of Oregon, where a large landslide data set is available. A Convolutional Neural Network (CNN) using 2D geospatial and terrain data (CNN2D) reached state‐of‐the‐art single‐model scores for Precision (0.90) and Recall (0.86), along with other metrics. CNN2D's Precision‐Recall Pareto front, formed by applying different hyperparameters, dominated attribute‐based models like Random Forest (RF1D) by a substantial margin, attesting to the value of fine‐scale terrain patterns. However, CNN2D's superiority required high‐resolution rainfall (∼800 m) and terrain (∼10 m) data sets: as the resolution coarsened, all models declined in performance but CNN2D's scores decreased more than RF1D's. Ensembling CNN2D and RF1D produced even better Recall (0.90), and this cross‐model‐type ensemble was also better than other ensembles. These models further showed robust results in cross‐regional validation. Rainfall, land cover, and elevation were the most important predictors, while prescribed Plan and Profile Curvature fields were also highly useful inputs (perhaps due to the size of the training data set). Based on the results of our analyses, we generated landslide susceptibility maps which provide insights into spatial patterns of landslide risk.

Landslide susceptibility measures the probability of landslides occurring under certain geo-environmental conditions and is essential in landslide hazard assessment. Landslide susceptibility mapping (LSM) using data-driven methods applies statistical models and geospatial data to show the relative propensity of slope failure in a given area. However, due to the rarity of multi-temporal landslide inventory, conventional data-driven LSMs are primarily generated by spatial causative factors, while the temporal factors remain limited. In this study, a spatiotemporal LSM is carried out using machine learning (ML) techniques to assess rainfall-induced landslide susceptibility. To achieve this, two landslide inventories are collected for southwestern Pennsylvania: a spatial inventory and a multi-temporal inventory, with 4543 and 223 historical landslide samples, respectively. The spatial inventory lacks the information to describe landslide temporal distribution; there are insufficient samples in the temporal inventory to represent landslide spatial distribution. A novel paradigm of data augmentation through non-landslide sampling based on domain knowledge is applied to leverage both spatial and temporal information for ML modeling. The results show that the spatiotemporal ML model using the proposed data augmentation predicts well rainfall-induced landslides in space and time across the study area, with a value of 0.86 of the area under the receiver operating characteristic curve (AUC), which makes it an effective tool in rainfall-induced landslide hazard mitigation and forecasting.