Rainfall-induced landslides and debris flows in Chieng So commune, Son La Province, Vietnam

On 26 July 2025, a short-duration, high-intensity rainfall event occurred in Chieng So Commune and its surrounding areas of Son La Province, Vietnam. The rainfall was mainly concentrated between 14:00 and 22:00 local time, with a total accumulation of 145.8 mm, accounting for more than 63.8% of the total rainfall recorded during the event period (19–26 July). This event triggered 1885 landslides within an area of approximately 408 km 2 . Using the landslide inventory, analysis of the landslide mobility shows that most landslides have a height-to-travel-distance ratio below 1, indicating high mobility. Landslides are mainly located near ridgelines or streams, which is different from some regions worldwide where they mostly occur close to streams. The probability of landslides exceeding 500 m 2 , 2000 m 2 , and 10,000 m 2 is 89.1%, 39.2%, and 8%, respectively. To evaluate the influence of antecedent rainfall variables on landslide occurrence, antecedent rainfall predictors were derived by identifying rainfall events, alongside topographic and geological predictors, to train an Extreme Gradient Boosting (XGB) model using spatial cross-validation with different block sizes. The results indicate that a 2000 m block size provides an appropriate compromise, balancing the reduction of spatial dependence with the need to maintain stable and reliable model performance. Lithology, dominated by strongly weathered granite or volcanic rock formations, is the most influential factor for landslide prediction, with a mean SHAP value of 0.25, followed by 8-h antecedent rainfall (0.18), slope (0.08) and 87-h antecedent rainfall (0.04). Furthermore, centered ALE values increase significantly when antecedent rainfall accumulations over 8 h and 87 h reach approximately 75 mm and 125 mm, respectively. These values are specific to the analyzed events and reflect model-derived statistical thresholds rather than absolute physical triggering limits. By integrating rainfall analysis, landslide geometric characteristics, and machine learning, this study offers important insights for rainfall-induced landslides, improving hazard reduction strategies and early warning systems.