Post-event information in annual satellite embeddings inflates the apparent performance of landslide susceptibility models: a paired audit of 18 events

Yifan Du , Tianhe Zheng, Peiyan Yang, Yibo Li, Junhang Zhou, Kun Yang

Landslide susceptibility models are often evaluated by spatial cross-validation on the inventory of a single event. When predictors come from annual satellite products, the layer for the event year can also record the event's aftermath. We audited 18 earthquake- and rainfall-triggered events with 61,402 landslide locations, changing only the input year of AlphaEarth Foundations embeddings while samples, spatial folds and models stayed fixed. Event-year inputs raised the area under the receiver operating characteristic curve (AUC) by 0.0874 (95% confidence interval 0.0649–0.1098) for a point model and 0.0705 (0.0475–0.0934) for a patch model. The increase persisted after correction for pre-event drift and appeared in point inventories and in inventories with mapped extents. Sentinel-2 inputs with a fixed number of dates gained discrimination when the observation window moved past the event. Across two municipalities affected by the May 2023 Emilia-Romagna event, the highest-scoring 10% of the area captured 30.0% of the mapped landslide area with prior-year inputs and 49.3% with event-year inputs. The event-year layer replaced 46.4% of that top-ranked area. These are within-event validation results, not forecasting skill. In retrospective tests trained on earlier landslides with pre-target inputs, embeddings added 0.059 AUC beyond conventional factors and landslide history in North Carolina. In Lushan they added 0.008, and the interval included zero. Susceptibility assessments that use annual satellite products should report the observation window of each predictor, the chronology of the inventory and the evaluation domain.