GeoHUM Internship
Building a flood catalogue, night-time light research and deep learning for flood masks
Executive summary
This report presents the outcomes of my internship that aimed to explore and enhance disaster response strategies through the utilization of data-driven approaches. The core tasks undertaken during this internship were meticulously designed to contribute significantly to the field of disaster response and preparedness. The initial task focused on the development of a comprehensive floods catalogue by exploring existing global datasets. This catalogue provides a vital resource, compiling essential information and parameters related to flooding incidents. The organized and easily accessible information within this catalogue has the potential to inform decision-making and disaster response strategies in regions prone to floods. The second task delved into the exploration of night-time light datasets, aiming to extract valuable insights into human activity patterns in disaster-affected regions. By analyzing these datasets, a deeper understanding of nocturnal movement, infrastructure utilization, and activity patterns was gained. A substantial portion of the internship was dedicated to investigating flood mapping techniques using advanced deep learning methodologies. This research aimed to leverage cutting-edge algorithms to produce more automated flood maps from Sentinel-1 images. The outcomes of this research were presented at the AI4Copernicus 2023 conference, a significant platform for disseminating advancements at the intersection of Artificial Intelligence and Copernicus data.
Task description
From March to May 2023, I engaged in a three-month internship opportunity, funded by the Christian Doppler Society. This enriching experience took place within the GeoHUM CDL laboratory. A notable aspect of this internship was its collaborative nature, as I had the privilege of working alongside the dedicated professionals of Doctors Without Borders/Médecins Sans Frontières (MSF) Austria. I committed 20 hours per week to delve into various tasks extensively described in the next sections. My primary responsibility encompassed providing essential support in the execution of tests on datasets with potential applicability in emergency scenarios. This role became especially pertinent during two significant events in which Médecins Sans Frontières (MSF) participated: the Turkey-Syria earthquake and the flooding crisis in Mozambique.
Task 1: Development of a flood catalogue
The initial task involved a systematic exploration of existing datasets with the aim of constructing a comprehensive catalogue focused on floods. This effort aimed to collate pertinent information and parameters associated with flooding incidents. The emphasis was placed on curating global datasets that could significantly contribute to informed decision-making and disaster response strategies at an event of a flood.
Task 2: Exploration of NTL dataset
The next task encompassed an in-depth investigation of datasets capturing night-time light emissions. This was an experiment in trying to establish whether satellite NTL datasets can aid in the mapping of new settlements after a disaster.
Task 3: Flood mapping using deep learning techniques
A substantial portion of my internship was dedicated to a study of flood mapping using deep learning algorithms. This research aimed to leverage automated methods to create flood maps from Sentinel-1 images after a flood event. The findings of this research were presented at the AI4Copernicus 2023 conference.
AI4Copernicus Conference
On May 25, 2023, I presented my research titled Enhancing flood mapping from Sentinel-1 SAR data using deep learning at the AI4Copernicus 2023 Conference in Luxembourg. The conference was attended by scholars and AI experts all round Europe and I had the unique opportunity to share the insights garnered from my research.
Reccommendation for future interns
Future inters could consider developing an automated system to create flood masks using Sentinel-1 images, rather than relying on existing online platforms. This could improve the efficiency and accuracy of flood monitoring. In the process offered a chance to enhance skills in image processing and data analysis while contributing to more effective disaster management techniques.