Globally- vs Locally-trained Machine Learning Models for Land-slide Detection: A Case Study of a Glacial Landscape

Abstract

Landslide risk mitigation is limited by data scarcity. This could be improved using continuous landslide detection systems. In order to investigate which image types and machine learning (ML) models are most useful for landslide detection in a Norwegian setting, we compared the performance of five different ML models, for the Jølster case study (30-July-2019), in Western Norway. These included three globally pre-trained models; i) the Continuous Change Detection and Classification (CCDC) algorithm, ii) a combined k-means clustering and Random Forest classification model, and iii) a convolutional neural network (CNN), and two locally-trained models, including; iv) Classification and Regression Trees and v) a U-net CNN model. Images used included Sentinel-1, Sentinel-2, digital elevation model (DEM) and slope. The globally-trained models performed poorly in shadowed areas, and were all outperformed by the locally-trained models. A maximum Matthew’s correlation coefficient (MCC) score of 89% was achieved with model v, using combined Sentinel-1 and -2 images as input. This is one of the first attempts to apply deep-learning to detect landslides with both Sentinel-1 and -2 images. Using Sentinel-1 images only, the locally-trained deep-learning model significantly outperformed the conventional ML model. These findings contribute towards developing a national continuous monitoring system for landslides.

Understanding Landslide Expression in SAR Backscatter Data: A Global Study

“…The dataset produced in this study provides a diverse training dataset that can be used for developing generalised automatic detection models. Previous studies have shown that locally-trained deeplearning models such as U-Net can detect landslides with good accuracy from Sentinel-1 images due to their ability to differentiate random speckle noise from clusters of pixels related to changes to the ground surface (Ganerød et al, 2023). The results of this study can be used to improve the design of deep-learning models through understanding how to ensure representative training cases and relevant input data are included.…”

“…Preliminary studies have yielded promising results at a local scale particularly with deep-learning methods. However, further development is needed to achieve automatic landslide detection models that perform well in diverse environments (Ganerød et al, 2023). Eventually, operational real-time monitoring and alert systems, such as presently exist for detection of illegal deforestation (Reiche et al, 2021), could be developed for landslides.…”

“…(Burrows et al, 2022) used Sentinel-1 time-series to backdate monsoon-triggered landslides and explored the performance of different bands. There have also been a few attempts to automate this process using locally-trained deep-learning models (Ganerød et al, 2023;Nava et al, 2022).…”