Challenges for satellite-based deep-seated landslide nowcasting

Landslides are a major geohazard in hilly and mountainous environments. We focus on slow-moving, deep-seated landslides that are characterized by gradual, non-catastrophic deformations of millimeters to decimeters per year and cause extensive economic damage. Where landslide hazard mitigation is impossible, Early Warning Systems are a valuable alternative to reduce landslide risk. Recent studies have demonstrated the effective application of machine learning for deformation forecasting to specific cases of slow-moving, non-catastrophic, deep-seated landslides. To test to what extent a combination of data-driven machine learning techniques and remote sensing observations can be used for landslide deformation forecasting, we developed a machine learning based nowcasting model on the multi-sensor monitored, deep-seated, Vögelsberg landslide, near Innsbruck, Austria. Our goal was to link the landslide deformation pattern to the conditions on the slope, and to produce a four-day, short-term forecast, a nowcast, of deformation accelerations.

Precipitation, snowmelt, soil moisture, evaporation, and temperature were identified as hydro-meteorological variables with high potential for forecasting deformation acceleration. Time series of those variables were obtained from remote sensing sources where possible, and otherwise from reanalysis sources as surrogate for data that is likely to be available in the near future. Deformation, the result of slope instability, was monitored daily by a local, automated total station of the Division of Geoinformation of the Federal State of Tyrol.

The five years of daily deformation and hydro-meteorological observations at the Vögelsberg landslide is quite limited for a machine learning model. To limit the complexity of the model, and the number of parameters to be optimized, the model was designed to mimic a bucket model, a simple hydrological model. A shallow neural network based on Long-Short Term Memory, was implemented in TensorFlow, as custom sequence of existing building blocks. In addition, a traditional neural network and recurrent neural network were tested for comparison.

Thanks to the limited complexity of the model, the major contributors could be determined by trial-and-error of nearly 150 000 model variations. Models including soil moisture information are more likely to generate high quality nowcasts, followed by models based solely on precipitation or snowmelt. Although none of the shallow neural network configurations produced a convincing nowcast deformation, they provide important context for future attempts. The machine learning model was poorly constrained as only five years of observations were available in combination with the four acceleration events that occurred in these five years. Furthermore, standard error metrics, like mean squared error, are unsuitable for model optimization for landslide nowcasting.

We showed that landslide deformation nowcasting is not a straightforward application of machine learning. The complexity of the machine learning model formulation at the Vögelsberg illustrates the necessity of expert judgement in the design and evaluation of a data-driven nowcast of slow deforming slopes. A future, successful nowcasting system will require a simple, robust model and frequent, high quality and event-rich data to train upon.