Forecasting landslides from ground motion

Predicting a landslide needs more than the motion itself. It needs the water cycle, erosion, freeze and thaw, land cover and the soil material, and how they combine in different conditions. We pretrain one model on these signals without labels, then check whether it forecasts better, transfers to regions with few recorded events, and adapts with fewer parameters than the classical statistical and physical models.

Radar ground motion is the central signal. It shows where to look, and how it changes over time shows whether the ground is speeding up.

The first test uses recorded Norwegian landslides from the NVE inventory, a Swedish quick-clay case, a quiet control site with no recorded event, and the mining towns of Kiruna and Malmberget. Every dot on a site map is a 10 m cell seen by both radar tracks. Click it to see its vertical, east and line-of-sight time series.

These maps are training data for the model, not a monitoring service for any asset. Each one covers a window around a known incident, from about two years before to two months after, so the model can learn what ground motion looks like before a landslide. They do not run up to today.

1. 1

   Find usable past events

   Audit the NVE and SGI landslide inventories for events with a date and a place good enough to test a forecast.

2. 2

   Define the test sites

   Pick incidents, draw a box around each, and find the Sentinel-1 tracks that cover it.

3. 3

   Process the radar

   Download the images and process one stack per track on GPU nodes. Winter images are dropped because snow breaks the radar phase.

4. 4

   Add the drivers

   Daily rain, temperature, snow and melt from seNorge, and atmospheric delay from ERA5.

5. 5

   Correct and cut the cubes

   Mask unreliable pixels, remove height-related phase, choose a stable reference, and cut cubes for the model.

6. 6

   Process only what was known on the day

   For a forecast on a given date the radar is reprocessed with images up to that date, so later images cannot leak into earlier values.

7. 7

   Freeze the split

   Fix which events are for training, validation and testing before any model sees them, with withheld regions and later years.

8. 8

   Pretrain and test

   Train the foundation model and compare it with a persistence baseline and a weather-only classifier.

• A wider processing window gave more usable ground. At Lærdal, a 150 m phase-linking window raised the share of cells with vertical and east motion from 13% to 44%.
• Snow breaks multi-year series. Series that cross a snowy winter jumped by 8 to 19 mm. Using only June to October images cut the jumps to about 1.3 to 2.2 mm at four sites, and all sites now use that rule.
• Our rates do not yet agree with EGMS. Within a summer the difference is 2.0 to 3.6 mm, but the spatial correlation of whole-period rates is still low.
• Weather alone forecasts weakly. Water reaching the ground gives a ROC of 0.62 at 7 days and 0.56 at 30 days on later years.
• The model has not beaten persistence yet. Trained on one site and tested on another, it reconstructs hidden dates better than persistence but forecasts future motion worse. Persistence is the bar to beat.

• Rates are fits, not forecasts. Rates in mm per year are straight-line fits over the processed window. The changes in mm are the direct measurement.
• Gaps in winter are intended. We use only snow-free images: at most 20 % of the area may hold more than 5 mm of snow water equivalent. Radar phase through snow is unreliable, and this rule cut winter jumps from 8–19 mm to about 2 mm. The gaps of several months each year are the result. Stenungsund lies outside the snow data and keeps all its images.
• Coverage is patchy in steep, wooded and snowy ground. Radar loses coherence there, so empty areas are missing data, not stable ground.
• East and vertical assume no north motion. Each dot combines an ascending and a descending track.
• Kiruna and Malmberget show one summer. They are processed over a single season and serve as a check on the radar quality, not as forecast cases.
• Giljabekken is not on the map yet. Its ascending track still gives empty output, so it is being rerun.

Processed from Copernicus Sentinel-1 data. Weather from seNorge and ERA5. Landslide records from NVE and SGI. Base map tiles © Esri.