Data-driven analysis of shallow landslide-to-stream connectivity

To improve freshwater quality and reduce sedimentation, detailed information on sediment delivery by rainfall-induced shallow landslides to stream channels is needed to better target mitigations. Data-driven landslide connectivity analysis can address this issue by identifying areas where future landslides are more likely to reach streams. We investigated landslide-to-stream connectivity using a comprehensive inventory of shallow landslides, focusing on a study area in Hawke's Bay, New Zealand, affected by intense rainfall in March 2022. We tested eleven logistic regression models, each with various parameter sets covering morphometry, land cover, lithology and landslide morphology. The most influential factor on connectivity was the overland flow distance to the stream network. A single-variable logistic regression model based on downslope distance achieved an AUROC of 0.84, demonstrating good predictive power. We assessed the models’ spatial transferability using an independent landslide inventory from another region with similar terrain, where the single-variable model maintained an AUROC of 0.84. Our analysis revealed no significant dependency of connectivity on land cover or near-surface lithology. However, landslides on pasture had significantly greater runout distances compared to native forests (with the same landslide source area), suggesting that land cover characteristics may influence connectivity. This study highlights the importance of detailed landslide connectivity analysis for targeting erosion control. We conclude that the single-variable model, based on downslope distance to the stream network, provides a robust and transferable tool for predicting landslide connectivity.