Predicting the Recorded Impact of Glacial Lake Outburst Floods in Gilgit-Baltistan: What a Small Event Database Can and Cannot Teach a Machine Learning Model

Glacial Lake Outburst Floods (GLOFs) occur when a natural dam fails, or a glacial lake abruptly empties, triggering catastrophic downstream consequences. These events destroy farmland, bridges, and homes and cause deaths in mountain communities, but records of earlier events are limited and irregular. This study investigates the predictive power of primary archival data, specifically a GLOF’s location, year, lake type, and trigger mechanism, in determining whether a given event resulted in a documented impact. I compared four classifiers on 137 events from the HMAGLOFDB database, using leave-one-out cross-validation. The Gradient Boosting classifier achieved an accuracy of 70.1%, outperforming the baseline of 56.9%. The underlying dataset reveals historical biases, as the proportion of events with documented impact shifts significantly across different eras. Instead of mapping physical hazards, the models simply map where and when events happened to be documented. Random 80/20 splits ranged from 50.0% to 89.3%, which shows how easily a single split can mislead. Physical predictors and spatial validation are the next steps.