Improving predictions of rainfall-triggered debris flows using area-based probability intensity curves

Abstract Rainfall-triggered debris flows represent one of the greatest growing threats to co-located populations, infrastructure, and communities on or adjacent to steep slopes worldwide. Landslide-triggering rainfall thresholds are formulated to provide practical warnings of debris flow hazards. Statistical models, intensity-duration (I-D) curves or some adaptation thereof, account for about 70% of the methods used to define these thresholds (the number is higher if hybrid models are included). Despite widespread adoption, existing thresholds are beset by technical challenges including the almost ubiquitous occurrence of false positives and a general lack of clarity about the areal extent to which the threshold applies. A novel alternative is advanced over traditional methods to better predict debris flow occurrence. Area-based probability-intensity (ABPI) curves (rather than intensity-duration thresholds) are created by integrating a time-constrained landslide inventory with continuous gridded rainfall data. Landslide occurrence and non-occurrence are assembled at a given rainfall intensity to provide landslide probabilities by intensity and duration, per unit area. Rainfall corresponding to landslide occurrence and non-occurrence was compiled over several years. Though additional data is needed to create ABPI curves, that data is increasingly available, and the result is no harder to deploy than I-D curves. Performance testing demonstrates considerable improvements in precision, recall, and miss rate, and the critical success index. Improvements are demonstrated across multiple watersheds in British Columbia, Canada, followed by a discussion of remaining practical challenges that are not solved using either method. ABPI curves are intended to be rapid, robust, and accessible to the field of landslide scientists and practitioners.