Enhancing landslide susceptibility mapping using a positive-unlabeled machine learning approach: a case study in Chamoli, India

The Indian Himalayas' susceptibility to landslides, particularly as a location where climate change effects may be event catalysts, necessitates the development of dependable landslide susceptibility maps (LSM). This study diverges from traditional binary classification models, framing LSM as a positive-unlabeled learning problem. This approach acknowledges that regions without recorded landslides are not necessarily at low risk but could simply have not experienced landslides yet. The study utilizes novel positive-unlabeled learning-enhanced algorithms—Random Forest, K-Nearest Neighbor, and Decision Tree—to create LSM for Chamoli district, India. Eleven causative factors for landslides are identified, including elevation, aspect, slope, geology, geomorphology, distance to lineament, lithology, NDVI, distance to river, distance to road and residential land use. To address spatial correlation biases, instead of randomly splitting the dataset, the study adopts spatial splitting to get the training and testing datasets. The study reveals that positive-unlabeled learning substantially improves the Area Under Curve and recall, leading to a more conservative LSM compared to binary classification methods. Analysis shows that the southern region of Chamoli exhibits high recall but lower accuracy, suggesting a latent high landslide susceptibility despite a lack of historical landslides in this region. The study also quantifies the impact of human activity on landslide risk, indicating an elevated threat to life and the local economy, especially in Chamoli's southwestern areas.