Towards a universal hydrologic metric for predicting rainfall-triggered landslide timing
Abstract. Metrics that approximate hillslope hydrologic response to rainfall are fundamental for informing landslide risk reduction efforts, such as early warning systems and hazard models. Notwithstanding the numerous publications using different wetness metrics that underpin and largely control the accuracy of landslide risk reduction products, a robust comparison of a broad array of different wetness metrics for regional landslide analysis is currently lacking in the literature. In this study, we statistically compare common wetness metrics for predicting the temporal occurrence of rainfall-triggered landslides at regional scales (> 1000 km2) using a landslide inventory covering the contiguous United States. We find that representations of hillslope wetting and drainage using parsimonious leaky bucket models that only require rainfall input and an estimated drainage factor can identify landslide-triggering hydrologic conditions across disparate ecological regions more accurately than unmodified precipitation metrics or more complex hydrological models. Due to the proliferation of global precipitation datasets and the limited input data needed for the parsimonious leaky bucket model, this model could be used to improve tools for landslide risk reduction.
This manuscript compares multiple wetness metrics for identifying the timing of rainfall-triggered landslides across the contiguous United States and concludes that a simple antecedent wetness index may provide broadly transferable predictive capability. However, I do not believe that the underlying landslide data or the evaluation framework can support this conclusion. The inventory combines heterogeneous sources and mapping procedures, converts polygon inventories to centroid points, accepts failure times with only daily precision, and identifies rainfall-triggered cases largely by excluding earthquake inventories and requiring non-zero precipitation on the reported failure day. This procedure does not establish rainfall causality, while essential attributes such as landslide type, size, depth, and movement mechanism are unavailable or inconsistent. The inclusion of post-fire debris flows and other mechanistically distinct phenomena further compromises sample homogeneity. More importantly, the percentile-based evaluation relies on the unverifiable assumption that the incomplete inventory is unbiased with respect to rainfall extremity, although the manuscript itself acknowledges substantial reporting biases toward particular regions and widespread landslide events. Multiple landslides generated by the same storm and the repeated metric values calculated for each landslide also introduce strong dependence and pseudoreplication, yet the Bayesian model does not appear to include landslide-, storm-, inventory-, or region-level dependence structures. No independent temporal, spatial, or event-based validation is presented, and the analysis therefore represents a retrospective ranking at known landslide locations rather than a robust assessment of predictive skill or false-alarm performance. In addition, the statistical formulation in the supplement contains serious inconsistencies: Equation S6 reverses the percentile scale as written, Equation S7 uses a standard deviation where a variance is required for beta-distribution parameterization, and the upper bound in Equation S15 reduces to zero. These problems make the analysis difficult to reproduce and cast doubt on the reported statistical comparisons. Because the principal conclusions depend directly on uncertain event labels, non-independent samples, an inadequately validated performance measure, and an internally inconsistent statistical model, these are foundational issues that cannot be addressed through routine revision. I therefore recommend rejection.