Using Deep Learning to Backcast Hydrologic Response and Inform Landslide Early Warning Systems
Loading...
Date
2024-01-09
Authors
Sheppard, Jonathan
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oregon
Abstract
Landslides are difficult to predict due to the influence of variable geologic and environmental factors, such as geomechanical properties, rainfall, ground saturation, topography, and earthquakes, exert on the probability of a slope failure. Deep learning (DL) models can accurately predict the site-specific hydrologic response on hillslopes using soil moisture, pore pressure, and rainfall monitoring data. Landslide early warning systems can utilize empirical thresholds from deep learning-derived soil hydrology properties to improve landslide hazard prediction accuracy. We study the possibility of improving a logistical regression-based landslide early warning system being used in Sitka, AK by incorporating pore pressure responses that correspond to past known landslide events. Because pore pressure records for past known events are nonexistent, we must backcast soil hydrology timeseries from weather records, without including antecedent soil hydrology as initial conditions. We assess the accuracy of predictions at various rainfall intensity thresholds made by a Long Short-Term Memory (LSTM) DL model trained on weather features compared to a model that includes antecedent soil hydrology conditions. We find that the average accuracy of our model decreases by up to 20% for important, high-intensity rainfall events.
Description
Keywords
Backcasting, Deep Learning, Landslide Early Warning System, Landslides, Pore Pressure