Project

Landslide hazard assessment in mountainous areas

The research objective of this project is to combine existing scientific models, remote sensing and ground survey data with machine learning to assess landslide hazard in the mountains of Nepal. First, we will look at regional analyses to produce probability estimates of landslide susceptibility for co-seismic earthquakes. Second, we will attempt to predict the mass and displacement of co-seismic and rainfall-induced landslides at selected sites. In the longer term, research focus will shift to modelling landslide risk and the connection with climate change.

Every year, landslides in Nepal lead to loss of life and property. A reliable landslide forecasting and early warning system would minimise such losses. In fact, landslide early warning systems are quite rare around the world. Landslides are triggered by a multitude of factors including earthquakes, rainfall, slope instability and human activities. On the other hand, there is a lack of systematic data collection and monitoring of landslides in Nepal. Therefore, building reliable forecasting models is challenging. The intersection of process-based models, remote sensing and machine learning provides as of yet unexplored opportunities.

Project description

Machine learning methods can extract complex relationships between diverse predictors and indicators of landslide hazard, such as the probability of a landslide or its size and displacement. Machine learning requires data on predictor inputs and landslide labels (or inventories). This project will start with a regional analysis of landslide susceptibility, focusing on coseismic landslides. Previous studies have published inventories of landslides caused by the 2015 earthquake, and the largest inventory contains more than 24000 landslides. The main challenge for landslide hazard assessment in Nepal is the lack of high-quality predictor data. Commonly used predictors include peak ground acceleration (PGA) during earthquakes, topographical features (elevation, slope, aspect), geological features (rock type, lithologic strength, distance to fault lines), hydrological features (precipitation, soil moisture, distance to drainage), and anthropogenic factors (land cover and land use, distance to roads). Most of these inputs may be available, but their quality is a concern. Some inputs, such as PGA, can be obtained from existing models. Remote sensing provides data for some inputs, such as precipitation, land cover and land use. With the existing predictor data and landslide inventories, machine learning can output the probability of a landslide in a given place.

Site-specific landslide hazard assessment will require more detailed data on geo-technical parameters (including soil properties), weather, and geomorphology (including river sections and topography). The second step will target areas that have more detailed predictor data available, such as Kathmandu valleys or highways, and model rainfall-induced landslides as well. The data will come from government entities, including the Department of Hydrology and Meteorology. Neural networks can be pretrained with high-quality data from Italy and fine-tuned with data from Nepal, using a method called domain adaptation. Data for Italy will be obtained in collaboration with the University of Florence. Landslide hazard models will predict where and when a landslide will occur and its mass and displacement. Mass and displacement will act as proxies for risk of damaging infrastructure and settlements as well as damming rivers downstream. Irrespective of predictive performance, this step will produce a workflow for assessing landslide hazards in areas with high-quality predictor data. Collaborating partners will include NASA Goddard Space Flight Centre (GSFC). International Centre for Integrated Mountain Development (ICIMOD), Practical Action Nepal and government entities. Efforts will be made to build a database of predictors for selected river basins, e.g. Melamchi and Upper Tamakoshi. Ground-based data collection will require equipment and infrastructure, such as weather stations and drones fitted with LiDAR sensors. The third step will use collected data to recalibrate hydrological models for the selected sites. Outputs of hydrological models, along with other predictors, will be passed to landslide hazard assessment models. These models can also be pretrained with data from Italy and fine-tuned in Nepal. The models can be validated on the 2021 Melamchi landslide or areas around the Upper Tamakoshi hydropower station or on a held-out test set. In the long term, this project aims to provide near real-time monitoring of hydrological processes and imminent landslide risk in the mountains and flooding risk downstream. Once we have fairly accurate hydrological and landslide hazard models, we can study the impact of climate change on landslides, hydrology and agriculture.

Results

Actions completed so far include:

  • Review of landslide mapping and forecasting methods
  • Review of landslides susceptibility mapping in Nepal and landslide forecasting in nearby region called Sikkim (India)
  • Introductory meetings with organisations working in this field (Department of Hydrology and Meteorology, Practical Action, ICIMOD, NASA GSFC, University of Twente.