This repository contains all resources for the tutorial "Efficient Deep Learning for Earth Observation" held at the International Geoscience and Remote Sensing Symposium 2023.
The goal of this tutorial is introduce and showcase methods for a more efficient training of Deep Learning models for Earth observation applications. The following topics will be introduced in our practical lab sessions:
- Data Fusion
- Multitask Learning
- Self-supervised Learning
In this tutorial we make heavy use of ben-ge dataset, which is available here. The ben-ge paper is available on arxiv.
Course materials are provided in the form of Jupyter Notebooks that use Pytorch to implement Deep Learning models. The Notebooks can be accessed via the launchers provided below, or they can be download by cloning this repository. If you are running the Notebooks in the cloud, we recommend to use Google Colab as it provides access to GPUs for more efficient training.
| Time slot | Lab | Content | CoLab Notebook Launchers | MyBinder Notebook Launchers |
|---|---|---|---|---|
| 9:00 - 10:20 | Lab 1 | Intro (slides) / Deep Learning Recap + Data Fusion (slides) |  |
 |
| 10:40 - 12:00 | Lab 2 | Multi-task Learning (slides) | |
|
| 15:40 - 17:00 | Lab 3 | Self-supervised Learning (slides) | |
|
The tutorial will be presented by Michael Mommert, Joelle Hanna, Linus Scheibenreif and Damian Borth, who are all researchers at the AIML Lab at the University of St. Gallen.
All contents of this tutorial are provided under the BSD-3-Clause license.