Item: Avalanche detection in Sentinel-1 radar images using convolutional neural networks
-
-
Title: Avalanche detection in Sentinel-1 radar images using convolutional neural networks
Proceedings: International Snow Science Workshop Proceedings 2018, Innsbruck, Austria
Authors:
- Per Egil Kummervold [ Norut Northern Research Institute, Tromsø, Norway ]
- Eirik Malnes [ Norut Northern Research Institute, Tromsø, Norway ]
- Markus Eckerstorfer [ Norut Northern Research Institute, Tromsø, Norway ]
- Ingar M. Arntzen [ Norut Northern Research Institute, Tromsø, Norway ]
- Fillipo Bianchi [ Norut Northern Research Institute, Tromsø, Norway ]
Date: 2018-10-07
Abstract: Knowledge about frequency and location of avalanche activity is important for avalanche forecasting and hazard mapping. Traditional field monitoring has limitations especially when surveying large remote areas. Thus avalanche detection in Sentinel-1 radar satellite imagery has been developed in recent years as an alternative. Current state-of-the-art automatic signal processing results in an accuracy of roughly 80%, but has in problematic cases (snow turning from wet to dry) an accuracy below 50% when compared to manual interpretation. We thus explored the use of convolutional neural networks (VGG-19 and AConvNets) in detecting avalanches in radar images, and evaluated if these networks were able to outperform currently used radar image classification. The CNN’s produced consistently accuracies above 90%. While conventional signal processing seems to fail on images that are easily categorised by human experts, the neural networks seem to have problems with the same images that are also considered borderline cases by a human expert. It is likely that enlarged and improved datasets, as well as transferred learning, can increase the accuracy even more.
Object ID: ISSW2018_P04.9.pdf
Language of Article: English
Presenter(s):
Keywords: Sentinel-1, avalanche detection, convolutional neural network, artificial intelligence.
Page Number(s): 377-381
-