IAP-24-095

Representing runout in landslide nowcasts to improve disaster information and response

When earthquakes shake steep terrain they often trigger hundreds to thousands of ‘coseismic’ landslides, damaging buildings and infrastructure, killing and injuring people, and delaying response efforts. Understanding where these coseismic landslides have happened or will happen is a vital tool in mitigating their risk whether in the years before or days after the earthquake. Landslide models have been developed to provide predictions for both situations, with the major difference being the extent to which the shaking intensity can be constrained (i.e., much easier after the earthquake than before).

Empirical or statistical models of landslide initiation consistently provide the most skilful predictions, but suffer from overfitting and are highly dependent on training data. Physics-based models provide some insight but are over-constrained and hard to parameterise. Both physics-based and empirical models typically ignore runout and focus on initiation, yet runout is important for determining landslide impact, often makes up the majority of the disturbed area, and is governed by fundamentally different physical processes to initiation. Empirical models currently neglect runout or conflate it with initiation, while physics-based models capable of treating the coupled initiation-runout process are infeasible to run at landscape scale. Incorporating runout into landslide hazard modelling transforms the problem from local to non-local; this is a harder modelling problem necessitating a new modelling approach that remains data-driven but with physics-based constraints.

Aim: improve the predictive skill of coseismic landslide models both by improving initiation modelling and introducing runout representation.

The project will deliver two distinct and improved types of model driven by user needs (D1 and D2): D1) a post-event model which must: use only predictors that are widely available within the area of interest, have short runtimes because outputs are time critical, and produce outputs relevant to emergency response; and D2) a pre-event model that can: use a wider range of predictors because location is more tightly constrained and there is more time to gather input data, and have longer run times because it is not ‘time critical’. Because shaking intensity is poorly constrained, any pre-event model will need to considermultiple intensity scenarios and user needs will determine output resolution. This PhD will focus initially on D1 but improvements in coupled prediction of initiation and runout will provide valuable insights to inform D2.

M1) Gathering landslide data. Empirical models need to be trained on a wide constituency of landslide datasets to avoid overfitting. Landslide maps have been generated for many of the large earthquakes over the last 20 years including for the Gorkha and Kashmir earthquakes, and the supervisory team already hold >30 suitable landslide inventories. The student will supplement the global set of inventories with new landslide datasets from past earthquakes mapped using an automated detection algorithm, ALDI, designed specifically to identify coseismic landslides from archival Landsat data and that has been shown to identify landslides with comparable accuracy to manual mapping (Milledge et al., 2021). These automated maps will be manually checked, clustered and classified, and then calibrated with and tested against manual mapping for subsets of the same events.

M2) Gathering predictor datasets. Landslide initiation is primarily controlled by shaking intensity, topographic slope, and material strength. The starting model will use only predictors from global datasets as used in the current USGS global landslide model (Nowicki Jessee et al., 2018) which releases rapid predictions available in the USGS Shakemap environment after every significant earthquake (Allstadt et al., 2022). This will enable easy adoption of improvements and implementation of the model within the USGS real time information services. However, there is also the opportunity to test the performance improvement from more accurate locally available predictors (e.g. 1:250k or 1:50k maps for lithology; higher-resolution elevation data (e.g. 2-10 m cell size), historic shaking intensity, and recent rainfall (e.g. PERSIANN-CDR and APHRODITE). There is also the opportunity to test a suite of alternative metrics of shaking intensity hypothesised to capture landslide relevant shaking properties (using USGS or other published outputs at suitable sites).

M3) Developing, training, and testing empirical-physical models. The student will draw on existing theory on non-local hillslope processes (Foufoula-Georgiou et al., 2010; Furbish et al., 2020), to test and extend existing runout representations capable of aggregating initiation probability downslope with path-dependent physically-based constraints (e.g. Milledge et al., 2019; Geng et al., 2022). The supervisory team already hold both working versions and source code for one such model. The student can then compare these against more complete runout representations such as RAMMS (e.g., Zimmerman et al., 2020) and MassWastingRunout (Keck et al., 2024) in a nested set of test cases. These will range from individual landslides where the physical properties are well constrained and thus physical models are well parameterised, to valley and event scale cases where the physical model is treated as a competitor rather than the truth. The student will train the model on landslide data collected in M1 with covariates from M2 using widely different landscapes to avoid overfitting and local focus to test hypotheses about the value of better local data, pre-event displacement, and post-event coherence. Holdback testing can be undertaken both within training events and using new events, measuring performance relative to the current state of the art (physically-based treatments at small scale, and USGS models at event scale).

Year 1

Reading and literature review, determination of key research questions; compilation of landslide inventories and application of automated mapping techniques. Visit to project partners Nowicki Jessee and/or Allstadt to understand the current USGS model(s).

Year 2

Implement existing landslide models and gather global and locally higher resolution/accuracy datasets (M2). Begin model development for coupled initiation-runout model. Attend UK geomorphology conference.

Year 3

Train and test coupled initiation-runout models using globally available predictors. Present outcomes at international conference. Begin to draft papers and thesis.

Year 3.5

Write up and submit thesis. Finalise publications. Present results at international conference.

Techniques in topographic analysis, spatial/statistical modelling and machine learning, Matlab, and numerical modelling will form the core of this project. Development of the necessary skills will be facilitated through in-house expertise in Newcastle and Durham, via NERC researcher training events, and via internationally recognised summer schools, such as the LARAM landslide summer school or the Community Surface Dynamics Modelling System (CSDMS) Spring School which provide training in the process understanding and the application to hazard and risk analysis and the modelling and programming and high-performance computing expertise respectively.

Further training in transferable skills, including project management, scientific writing, oral and written communication, and media and public engagement, is available via the Newcastle doctoral training program and via the cross-disciplinary training provided as part of the IAPETUS2 DTP.

The student will be encouraged to present their work at departmental conferences and seminars at Newcastle and Durham and will develop an ability to communicate across disciplinary boundaries with a particular focus on Natural Science and Engineering. They will also be encouraged to attend relevant national and international conferences throughout their PhD.