Knowledge-graph reasoning for temporal lag effects and mechanism inversion of heavy-rainfall-induced landslide events: case study on Hong Kong

Heavy rainfall can sustain landslide activity for several days after a storm, but daily event inventories often obscure this delayed response and provide little information on the evidence used to infer failure processes. We developed a knowledge-graph workflow that links 2,961 Hong Kong landslide records from 2001 to 2021 with daily rainfall, slope attributes, spatial relations, and 57 investigation reports. Lag analysis was possible for 2,950 records, mechanism analysis for 2,940 records, and 43 events had traceable report anchors. Daily incidence rates and threshold–window tests identified a high-hazard stage on days 0–3 and an extended-hazard stage on days 4–7. With a 50 mm threshold and a 7 days look-back window, 42.2% of events occurred on a qualifying rainfall day and 30.1% occurred during the following week. A nearest-station test using 21 gauges produced 89.4% agreement with the baseline stage classification. The graph combined storm, slope-form, neighborhood, and regional evidence to estimate five mechanism classes. Pore-pressure-dominated soil slides and channelized debris flows accounted for 78.9%–84.4% of classifications across lag stages, while posterior uncertainty increased with lag. Leakage-controlled leave-one-out testing yielded Top-1 and Top-2 accuracies of 0.163 and 0.349. These results support storm-cluster interpretation and investigation prioritization, while event-specific causes still require field evidence.