Abstract
This work contributes a new framework for establishing data-driven rainfall thresholds in high-risk, data-limited contexts. Rainfall thresholds are commonly used to characterise the precipitation needed to trigger landslides in a region. However, these empirical relationships are sensitive to the exact definition of a “rainfall event”, especially how the minimum inter-event time (MIT) and triggering event (TE) are defined. Using Bayesian inference (BI) and nonlinear least-squares (NLS) techniques, this study evaluates how variations in MIT and TE definitions affect rainfall threshold estimation, considering both Event Rainfall–Duration $$\left( {E{-}D} \right)$$ and Intensity–Duration $$\left( {I{-}D} \right)$$ spaces. The dataset includes 15-min rainfall measurements from 52 gauges recorded from 2005 to 2023, as well as a regional landslide dataset compiled from British Geological Survey records covering the South Wales coalfields. Findings reveal that BI -derived thresholds are more stable than NLS -based thresholds, showing smaller parameter changes and fewer unrealistic curves, particularly in I–D space, where NLS often produces near-flat thresholds. Overall, both BI and NLS approaches demonstrate their strongest performance at MIT = 48 h, emphasising the role of extended antecedent rainfall in triggering spoil tip failures. This study demonstrates how the integration of robust Bayesian methods facilitates the downscaling of global thresholds to data-scarce regions and how careful event delineation practices can improve landslide prediction.