Low-data & low-code geo-spatial Deep-Learning
Towards a Low-data, Low-code Geospatial Deep Learning Platform for Domain Experts: A first application to wildfire impact severity assessment.
Factsheet
- Schools involved School of Engineering and Computer Science
- Institute(s) Institute for Data Applications and Security (IDAS)
- Research unit(s) IDAS / Applied Machine Intelligence
- Funding organisation Others
- Duration (planned) 01.03.2022 - 28.02.2023
- Head of project Prof. Dr. Souhir Ben Souissi
- Project staff Céline Hüttenmoser
- Partner Hasler Stiftung
- Keywords Geo-Spatial, Wildfire Burn Severity, Deep Learning, Computer Science
Situation
The latest theoretical and practical advancements in data engineering and deep-learning have allowed domain-experts (such as environmental scientists & emergency management agencies) to perform prediction, classification, regression and segmentation tasks for their areas of interest (e.g. geo-spatial imaging) at a scale and accuracy that was previously impossible. Unfortunately, in these applied settings the benefits of AI & Deep-Learning are now showing diminishing returns, due to improper or noisy deployment environments as well as lack of data for domain-specific usage.
Course of action
In this project we aim to: (a) Construct more specialized data-sets that better mimic the real-world problems we have identified. (b) Experiment with DL (Deep-Learning) solutions that try to mitigate these problems in a specific context (such as Geo-spatial Deep Learning). Finally, (c) prototype a low-code solution for easily deploying and fine-tuning these type of systems in practice. We expect that the resulting work for what we describe as low-data & low-code Deep-Learning will be more amenable to specialized configuration by domain-experts who do not have an advanced AI or data engineering background. As a first application, the researchers are testing their deep-learning solutions to identify the severity of wildfire burns by using satellite images.
