Few Shot Learning on the Swiss River Network

Co-Supervised by: Benjamin Fankhauser

If you are interested in this topic or have further questions, do not hesitate to contact benjamin.fankhauser@unibe.ch.

Background / Context

Water temperature is an important variable for the health of our water ecosystems. An increase of a few degrees has drastic effects. Further, many factors influence the water temperature: sun radiation, power production, industrial cooling, etc. The unknown contribution of these factors makes it unfeasible to model. Therefore, we rely on statistical models.

But long-term temperature monitoring is very costly and labor-intensive. An alternative is battery-powered sensors, which only last for up to three months and produce reliable measurements during this period. The goal of this project is to explore if few shot learning techniques can be applied to such short-term temperature measurements.

Research Question(s) / Goals

  • What are suitable few-shot techniques for time series, especially water temperature?
  • What are the minimal data requirements to achieve an RMSE<0.9?
  • What can we do with three months of data?

Approach / Methods

Begin with a literature review on few-shot learning in general, with a focus on its application to time series. The goal is to establish a baseline. You can then implement one promising existing method or develop your own. It is expected that both a baseline and one additional model of your choice are implemented and evaluated on the dataset from at least one station.

The modeling performance (RMSE) will be discussed, although achieving high accuracy is not the primary objective of this work.

Expected Contributions / Outcomes

A good outcome would be determining which methods are available and the expected modeling performance of baseline models on this dataset. The main research question is twofold: what are the data requirements to achieve a RMSE<0.9 and what can we do with three months of data.

Required Skills / Prerequisites

We use Python and PyTorch. Some basic knowledge of deep learning will be beneficial.

Further Reading / Starting Literature

  • “Graph-Based Deep Learning on the Swiss River Network” by Fankhauser B. Bigler V. Kaspar R (2023)
  • “Meta-Learning with Memory-Augmented Neural Networks” by Santoro et al (2016)