Probabilistic rainfall threshold of landslides in Data-Scarce mountainous Areas: A case study of the Bailong River Basin, China

The damage caused by rainfall-induced landslides has increased globally. The development of urbanisation has led to the expansion of residential areas in mountainous areas, and human activities have accelerated slope instability. However, the limited data records in mountainous zones have resulted in low-precision landslide rainfall thresholds. Establishing a landslide early warning model in areas with scarce data is an unresolved problem. This study uses the Bailong River Basin in western China as an example. By downscaling Tropical Rainfall Measuring Mission (TRMM) data to a daily resolution of 1 km and conditionally merging it with ground station rainfall, we obtain high-precision rainfall data to establish rainfall threshold curves based on Bayes' theorem and the frequency method. Subsequently, a landslide susceptibility map based on deep learning was integrated to construct a regional landslide early warning model. The results are as follows: First, compared with downscaling rainfall and rain gauge interpolation data, the daily rainfall data with a resolution of 1 km that combines the two has better overall performance and higher accuracy. Second, we generated the event-rainfall and-duration threshold curves of 5%, 20%, and 50% probability by combining the frequency method and Bayesian theorem, and show that with a probability of landslide occurrence of 50%, while the intercept of Bayes' theorem is greater than that of the frequency method, the overall trend is consistent. Finally, by using a typical mass rainfall event to test the performance, the results show that the early warning capability of the model integrated with the rainfall threshold and landslide susceptibility map is more accurate than that of using a separate rainfall threshold. The outcomes of this research are expected to provide efficient support for early warning and risk management of geological disasters in mountainous areas with scarce data.