Spatial dynamic rainfall thresholds for landslide in the Aburrá Valley, Colombia
Abstract. Rainfall-triggered landslides are frequent in the Aburrá Valley, a densely populated Andean valley in Colombia, where warnings must tell where as well as when slopes become critical. We derived rainfall thresholds for each slope unit from 492 dated landslides (2011–2025), four terrain variables, rainfall and the El Niño-Southern Oscillation (ENSO) phase. Non-landslide observations were drawn under three designs that share the landslide's slope unit, its date, or neither, each resampled 100 times. For each design we compared logistic regression, generalized additive models and random forests. A one-day window best described the trigger; for antecedent rainfall, 30 days was chosen among several equivalent windows. Median areas under the receiver operating characteristic (ROC) curve ranged from 0.955 to 0.988. The algorithms differed little within a design, whereas the design changed which predictors carried the signal. Three cut-offs on the ROC curve of the additive model define four alert levels, whose rainfall thresholds change with land cover, slope, soil, ENSO phase and antecedent rainfall. In the larger of two landslide events of 2025 kept out of training, 78 % of the evaluated slope units with landslides reached High or Very high before the report, while 11 % of the valley did.