Bridging the Gap in Disaster Response: Cross-Resolution Domain Adaptation for Rapid Damage Assessment

This thesis explores the use of free, high-revisit Sentinel-1/2 data for building damage assessment by aligning their features with high-resolution xBD datasets through domain adaptation and state-of-the-art architectures.

Requirements

  • M.Sc. in Data Science, Computer Science, Artificial Intelligence, Mathematics, or similar
  • Strong knowledge of Python and deep learning frameworks (PyTorch, Lightning)
  • Knowledge of Computer Vision (Semantic Segmentation, Change Detection)
  • Familiarity with Foundation Models (ViTs) and emerging architectures like Mamba
  • Interest in geospatial data processing (Rasterio, QGIS)

Description

Remote Sensing Damage Assessment is the process of identifying how natural or man-made disasters like floods, earthquakes, or wildfires affect structures using satellite images. By looking at images from before and after an event, algorithms can find hit areas and label how bad the damage is for houses and infrastructure. This helps groups like the civil protection know where to send help first.

Right now, the best models use Very High Resolution (VHR) images, but VHR data is expensive and often arrives with delays. On the other side, the European Copernicus Sentinel-1 (SAR) and Sentinel-2 (Multispectral) satellites are free and take pictures of the same spot every few days, but the images are grainier (10-20m per pixel).

The goal of this thesis is to build a way to do damage assessment on Sentinel data by using the labels found in high-res datasets. Since standard models fail because the resolution is too different, this research focuses on techniques like transfer learning, domain adaptation, cross-modal distillation. The idea is to teach a model to find damage in low resolution Sentinel pixels by using the features from VHR images to guide it during training.

The work will rely on multiple label sources to train and test the models. This includes the xBD (xView2) dataset, which contains VHR images from disasters worldwide with labels classifying buildings as fine, slightly damaged, or destroyed. To expand the scope beyond buildings, the research also incorporates the SYSU-CD dataset for diverse change detection in roads, vegetation, and water bodies. We will also use labels from the Copernicus CEMS Rapid Mapping portal, which provides grading maps of real-world damage derived from satellite activations. Examples of models used include ResNet, Vision Transformers (ViT), and Mamba.

Main Activities

  • Data Retrieval: Downloading and rasterizing CEMS vector data for ground-truth labels or creating low-res labels from the xBD dataset to match Sentinel resolutions.
  • Baselines: Training and testing a model using the obtained labels and Sentinel-2 data to set a benchmark without extra cross-resolution help.
  • Methodology Development: Using tricks like Teacher-Student setups to improve the Sentinel model’s performance.
  • Radar Integration: Checking if Sentinel-1 SAR or InSAR data improves the performance of damage detection or helps when it is too cloudy for optical sensors.

Contacts