Automatic landslide detection using Sentinel-1 and -2 images - a glacial case study

Although Norway is a country with rough terrain and a high frequency instable steep slopes, there is a scarcity of landslide data available. This limits the accuracy of thresholds for early warning systems, and hazard maps, both of which rely on historic event data. There is great potential to supplement existing ground-based observations with automated landslide detection, using satellite imagery and deep learning. In working towards an automated system for landslide detection in Norway, we investigated which imagery types and machine-learning models performed best for detecting landslides in a formerly glaciated landscape.

We locally trained a deep learning model with the use of Keras, TensorFlow 2 and U-net architecture. As input data, we used multi temporal composites with Sentinel-1 and -2 image stacks of all available images from one month pre- and post-event. Processed bands included: dNDVI (difference in maximum normalised difference vegetation index) from Sentinel-2, and pre- and post-event Synthetic Aperture Radar (SAR) data (terrain-corrected, mean of multi-temporal ascending descending images, in VV polarisation) from Sentinel-1. Training and evaluation were performed with a well-verified landslide inventory of 120 manually mapped rainfall-triggered landslides from Jølster (30-July-2019), in Western Norway. We tested the model with four input data settings using different bands and various polarization for the pre- and post-event SAR data, including: 1) full version (all 13 bands) 2) dNDVI (Sentinel-2), preVV, postVV (Sentinel-1), 3) preVV, postVV (Sentinel-1), and 4) post-R, post-G, post-B, post-NIR, dNDVI (Sentinel-2). The results were compared to the results of a pixel-based conventional machine learning model (Classification and Regression Tree) using the same input data. The second input data setting provides the best results. The performance scores show precision results for all four input data settings between 80-85%, with Matthews corelation coefficient values from 51-89%. Moreover, the deep-learning model significantly outperforms the conventional machine learning model in the input data setting #3. We see that the patch-based classification method far out-performs the pixel-classification due to the ability to differentiate the landslide signal from random noise produced from speckle in undisturbed areas. In addition, this represents one of the first attempts to fuse SAR and optical data for landslide detection, and we show there is an advantage in doing so in this case.