Towards automated registration of climate-related landslides in Norway by combining Google Earth Engine, global precipitation datasets and AI

The Norwegian Mass Movements inventory is crucial for producing landslide susceptibility maps and early warning thresholds. However, it has significant sampling and spatial bias, with approximately 90% of registered landslides found within 100 m of a road. Applying AI, and the computing power of Google Earth Engine, to extract information from earth observation data, has great potential to improve our understanding of the true spatial distribution of landslides in Norway. Recently, globally-trained generalised ML algorithms have been developed, aiming to detect landslides from satellite images in regions where they have not been previously trained. Here we investigate how these tools can be applied in Norwegian conditions.

This study consists of two parts; 1) to evaluate how well existing generalised ML landslide detection algorithms perform in Norwegian conditions, and 2) to investigate methods for automatically back-dating and extracting trigger information for newly detected landslides using the Google Earth Engine platform. Two generalised ML methods using Sentinel-2 images, proposed by Prakash et. al (2021) and Tehrani et. al (2021), were tested on the Jølster case study (30.07.2019) from western Norway. This case study is a very well documented example of a multiple landslide event, triggered by extreme rainfall, and represents some of the ‘unique’ fjord- and mountainous-environments in Norway. In part two; backdating and extracting trigger information with Google Earth Engine - the investigated methods were tested on specific debris flow at Vassenden, using Sentinel-2 satellite images and global precipitation datasets (GSMaP and GPM).

Preliminary detection results were relatively poor. The Prakash algorithm vastly overestimated landslide activity, and the Tehrani algorithm did not detect any landslides. We found that snow cover, seasonal vegetation and lighting changes in the input images - factors that greatly affect detectability of landslides in Norway - were not sufficiently accounted for in the two methods tested.  In the second part; extracting the date and trigger information -  a mean-NDVI time-series of Sentinel-2 images within a one-year window was produced for the landslide area, and the date range of vegetation loss determined. The precipitation datasets were filtered to identify the magnitude and time of maximum precipitation at the landslide point, within the previously determined date range.

To conclude, these early, generalised ML landslide detection models show good potential to be applied in Norway, however they do require retraining and further development to perform well in the local conditions. Due to the strong seasonal effects, a more suitable approach for improving landslide inventories could be to conduct annual regional surveys, then backdate the newly detected landslides and assign most-likely-trigger information. Modifications to the preparation of input images are recommended to account for the seasonal conditions, including a) widening the time window for image selection to one year, b) creating a cloud-free composite based on a modified greenest-pixel approach, and c) filtering for snow. We plan to expand this study to include case studies from a diverse range of locations and seasonal conditions in Norway, and to retrain and modify the machine learning pipelines to further improve detection results.