This study develops a framework for decoupling and investigating urban and non-urban landslide mechanisms, focusing on Constantine, Algeria, a city with complex topography and high landslide susceptibility. The region presents a heterogeneous landscape, where dense urban zones coexist with bare rural areas, influencing slope stability differently. A landslide inventory of 184 events was compiled and classified into urban and non-urban categories. Using geospatial data (topography, hydrology, landcover, lithology) and machine learning models (Random Forest, XGBoost, LightGBM, Multi-Layer Perceptron, and Logistic Regression), landslide susceptibility maps were generated for three datasets: urban, non-urban, and mixed. Model performance was assessed using cross-validation and evaluation metrics (ROC-AUC, F1-score, precision, recall), while SHAP analysis provided insights into factor importance. The results reveal distinct landslide drivers across environments. In urban areas, landslides are primarily influenced by aspect, slope, and proximity to streams, while distance to roads plays a lesser role, likely due to engineered slopes and drainage infrastructure. In non-urban areas, distance to roads is the most critical factor, highlighting the destabilising effects of road cuts in rural landscapes. Slope and proximity to streams remain key determinants, with lithology playing a more significant role in naturally driven failures. This study underscores the importance of context-specific landslide modelling and the potential biases of using mixed urban and non-urban inventories. The findings provide actionable insights for targeted mitigation, land-use planning, and infrastructure design. By distinguishing between urban and non-urban landslides, this research bridges critical gaps in understanding landslide dynamics across diverse landscapes.

This study develops a framework for decoupling and investigating urban and non-urban landslide mechanisms, focusing on Constantine, Algeria, a city with complex topography and high landslide susceptibility. The region presents a heterogeneous landscape, where dense urban zones coexist with bare rural areas, influencing slope stability differently. A landslide inventory of 184 events was compiled and classified into urban and non-urban categories. Using geospatial data (topography, hydrology, landcover, lithology) and machine learning models (Random Forest, XGBoost, LightGBM, Multi-Layer Perceptron, and Logistic Regression), landslide susceptibility maps were generated for three datasets: urban, non-urban, and mixed. Model performance was assessed using cross-validation and evaluation metrics (ROC-AUC, F1-score, precision, recall), while SHAP analysis provided insights into factor importance. The results reveal distinct landslide drivers across environments. In urban areas, landslides are primarily influenced by aspect, slope, and proximity to streams, while distance to roads plays a lesser role, likely due to engineered slopes and drainage infrastructure. In non-urban areas, distance to roads is the most critical factor, highlighting the destabilising effects of road cuts in rural landscapes. Slope and proximity to streams remain key determinants, with lithology playing a more significant role in naturally driven failures. This study underscores the importance of context-specific landslide modelling and the potential biases of using mixed urban and non-urban inventories. The findings provide actionable insights for targeted mitigation, land-use planning, and infrastructure design. By distinguishing between urban and non-urban landslides, this research bridges critical gaps in understanding landslide dynamics across diverse landscapes.

This study develops a framework for decoupling and investigating urban and non-urban landslide mechanisms, focusing on Constantine, Algeria, a city with complex topography and high landslide susceptibility. The region presents a heterogeneous landscape, where dense urban zones coexist with bare rural areas, influencing slope stability differently. A landslide inventory of 184 events was compiled and classified into urban and non-urban categories. Using geospatial data (topography, hydrology, landcover, lithology) and machine learning models (Random Forest, XGBoost, LightGBM, Multi-Layer Perceptron, and Logistic Regression), landslide susceptibility maps were generated for three datasets: urban, non-urban, and mixed. Model performance was assessed using cross-validation and evaluation metrics (ROC-AUC, F1-score, precision, recall), while SHAP analysis provided insights into factor importance. The results reveal distinct landslide drivers across environments. In urban areas, landslides are primarily influenced by aspect, slope, and proximity to streams, while distance to roads plays a lesser role, likely due to engineered slopes and drainage infrastructure. In non-urban areas, distance to roads is the most critical factor, highlighting the destabilising effects of road cuts in rural landscapes. Slope and proximity to streams remain key determinants, with lithology playing a more significant role in naturally driven failures. This study underscores the importance of context-specific landslide modelling and the potential biases of using mixed urban and non-urban inventories. The findings provide actionable insights for targeted mitigation, land-use planning, and infrastructure design. By distinguishing between urban and non-urban landslides, this research bridges critical gaps in understanding landslide dynamics across diverse landscapes.