The devastating impacts of the widespread flooding and landsliding in Puerto Rico following the September 2017 landfall of Hurricane Maria highlight the enhanced hazard potential from increasingly extreme storms in mountainous humid-tropical climate zones. Long-standing conceptual models for hydrologically driven hazards in Puerto Rico posit that hillslope soils remain wet throughout the year and antecedent soil wetness imposes a negligible effect on hazard potential. Our post-Maria in situ hillslope hydrologic observations indicate that while some slopes remain wet throughout the year, others exhibit appreciable seasonal and intra-storm subsurface drainage. Therefore, we used receiver-operating characteristic analysis and the Threat Score (TS) skill statistic to evaluate the performance of hydro-meteorological (soil wetness and rainfall) versus intensity-duration (rainfall only) hillslope hydrologic response thresholds that identify the onset of positive pore-water pressure, a predisposing factor for widespread slope instability in this region. We found that the hydro-meteorological thresholds outperformed intensity-duration thresholds for a seasonally wet, coarse-grained soil (TS = 0.8 vs. 0.6, respectively), although they did not outperform intensity-duration thresholds for a perennially wet, fine-grained soil (TS = 0.2 vs. 0.2, respectively). These soils types may also produce radically different stormflow responses, with subsurface flow being more common for the coarse-grained soils underlain by intrusive rocks versus infiltration excess and/or saturation excess for the fine-grained soils underlain by volcaniclastic rocks. We conclude that variability in soil-hydraulic properties, as opposed to the humid-tropical climate zone, is the dominant factor that controls runoff generation and modulates the importance of antecedent soil wetness for our hillslope hydrologic response thresholds. Our findings encourage further deployment of continuous in situ hydrologic monitoring to facilitate the development of empirical hillslope hydrologic response and landslide thresholds for regional-scale hazard warning systems that must account for spatially variable soil types.

We apply deep learning to a synthetic near-surface hydrological response dataset of 4.4 million infiltration scenarios to determine conditions for the onset of positive pore-water pressures. This provides a rapid assessment of hydrologic conditions of potentially hazardous hillslopes where mass wasting is prevalent, and sidesteps the computationally expensive process of solving complex, highly non-linear equations. Each scenario considers antecedent soil moisture and storm depth with varying soil properties based on those measured at a USGS site in the East Bay Hills, CA, USA. Our model combines antecedent soil wetness and storm conditions with soil-hydraulic properties and predicts a binary output of whether or not positive pore pressures were generated. After parameterization, pore-water pressure conditions can be returned for any combination of antecedent soil moisture content and storm depth values. Similar to previous work, a deep learning model reduces computational cost: processing time is decreased by more than an order of magnitude for 1D simulated infiltration scenarios while maintaining high levels of accuracy. While the physical relevance and utility behind process-based numerical modeling cannot be replaced, the comparatively reduced computational cost of deep learning allows for rapid modeling of pore-water pressure conditions where solving complex, highly non-linear equations would otherwise be required. Furthermore, comparing the solution of a deep learning model with a hydrological model exemplifies how similar results can be produced through highly divergent mathematical relationships. This provides a unique opportunity to understand which variables are most relevant for the prediction of positive pore-water pressures on hillslopes, and can represent landslide-relevant hydrologic conditions for hillslopes where rapid analysis is imperative for informing potential hazard mitigation efforts. Ultimately, a calibrated deep learning model may reduce the need for computationally expensive physics-based modeling, which are often time and resource intensive, while providing critical statistical insight for the onset of hazardous conditions in landslide-prone areas.