R&D highlights edition 2019

PROJECTS Flood risk 16 MACHINE LEARNING FOR GLOBAL FORECASTING OF RAINFALL-INDUCED LANDSLIDES Rainfall-induced landslides are the most common type of landslide across the world. They are complex in nature and forecasting them is therefore very challenging. In this project, we used machine learning to predict rainfall-induced landslides with a combination of several controlling factors and rainfall data as the main triggering factor. L andslides are catastrophic geo-hazards that threaten urbanisation worldwide. Population growth, in combination with the construction of critical infrastructure such as roads and pipelines, in landslide- prone areas increases the risk associated with landslides. Of the multiple factors that trigger landslides, rainfall is the most common: it has caused thousands of landslides in the past decade only, and some of the deadliest (an example being the debris flow event in August 2017 in and around Freetown in Sierra Leone, which caused 1141 fatalities). Rainfall-induced landslides are normally triggered by intense and/or prolonged precipitation and they take the form of shallow slides and debris flows. Given the complex nature of rainfall-induced landslides, a single source of data such as rainfall or terrain features, or the geotechnical properties of slopes, will not be enough to forecast them. In order to prevent or minimise the catastrophic consequences of these events, it is therefore important to combine multiple sources of data to forecast more accurately when and where they will occur. The present study therefore set up a data-driven framework for forecasting rainfall-induced landslides. A database was created for nearly 11,000 landslides, including the date and location of events across the world between 2007 and 2018, as well as the triggering and pre- disposing factors that caused these landslides. The database was then used to train supervised Machine Learning (ML) classification algorithms. Binary classification methods such as logistic regression were used to distinguish between landslides and non-landslide cases. To train the ML model, we built eleven sample sets (E0 to E10) with different combinations of triggering and controlling factors (model features). The sample sets were assigned to training (67%) and test (33%) sets which were then used for the training and assessment of the logistic regression model respectively. The accuracy of the logistic regression model was estimated with the Receiver Operating Characteristic (ROC) curves and the associated Area Under Cut. The outcome of this study is being implemented in a Landslide EarlyWarning System (LEWS) as part of the Delft-FEWS platform. The forecasting tool developed in this project can be used for regional landslide forecasting after regional adaptation. Landslide forecasting framework for this study Features used to train the machine learning algorithms Accuracy of logistic regressionmodel for classifying landslides and non-landslides Contact: Faraz S. Tehrani, Faraz.Tehrani@deltares.nl , t +31 (0)6 4691 1742 Giorgio Santinelli, Giorgio.Santinelli@deltares.nl, t +31 (0)6 1582 1809 Further reading : Tehrani et al. (2019). “A framework for predicting rainfall-induced landslides using machine learning”, Proceedings of the XVII ECSMGE-2019 Geotechnical Engineering foundation of the future, September 2019, Iceland.