Open access
Author
Date
2021Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Global warming is a defining challenge of our time with devastating consequences for local habitats. High mountain areas are particularly affected by global warming leading to a decline of their cryosphere (glaciers, snow cover and permafrost).
In high-alpine steep bedrock, permafrost thaw decreases the stability of mountain slopes leading to an increase of rockfalls and landslides and thereby putting life and built infrastructure at risk.
Monitoring these environmental changes is important for natural hazard warning and understanding the geophysical processes leading to such hazards. Moreover, by providing evidence from large-scale, long-term measurements, environmental monitoring helps to bolster scientific findings and can call attention to the immediate impacts of climate change. The rise of wireless sensor networks offers a range of possibilities for environmental monitoring enabling large-scale deployments with high spatial-temporal resolution using many different sensor types. The cheap and diverse sensors can be installed at hard to reach places with little available networking or power infrastructure. However, the resulting datasets (often heterogenous and long-term measurements) require a complex data analysis. Moreover, networking or power failures often lead to an error-prone data collection and a fragmented and noisy datasets. Analyzing these datasets typically requires dedicated domain-expert knowledge which can not be scaled to long-term monitoring datasets. Machine learning provides options to extract information automatically but these techniques usually require a clean dataset for training and their performance is strongly affected by differences in the distribution of training and test data.
In this dissertation, we consequently develop tools and methods applicable to heterogeneous, long-term, noisy datasets originating in wireless sensor network deployments. The main contributions of the dissertation are
- A methodology to work with fragmented and noisy data from a real-world sensor network deployment at Matterhorn, Switzerland. The methodology uses active learning with human-in-the-loop and a heterogeneous set of sensors to systematically filter out unwanted influences from seismic signals.
- The development and installation of an array of low-power, event-triggered micro-seismic sensors for the purpose of rockfall early warning. In addition, a machine-learning based human footstep classifier is designed and optimized for computation on memory-constraint embedded devices to detect humans in the hazard zone.
- Unsupervised and semi-supervised methods designed to bridge machine learning technology and domain-expert knowledge by providing experts with automated information extraction and machine-learning algorithms with crucial information such as information about the system context.
- foReal, a data analytics and visualization platform which allows to combine data from different sources. It is designed for long-term and large-scale environmental datasets and focuses on robustness against data corruption, missing data and misconfigurations during data processing as well as misinterpretations during experiment design and analysis. The tooling developed enables fast and easy exchange between experts of various domains and offers the public access to scientific data. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000536692Publication status
publishedExternal links
Search print copy at ETH Library
Journal / series
TIK-SchriftenreiheVolume
Publisher
ETH ZurichOrganisational unit
03429 - Thiele, Lothar (emeritus) / Thiele, Lothar (emeritus)
More
Show all metadata
ETH Bibliography
yes
Altmetrics