Using Location Embeddings for Improving Transferability of Flood Mapping Models

Suggested By: Bruno Menini Matosak, Getachew Gella (CDL GEOHUM)

Keywords: Deep Learning; Flood Mapping; Transferability; Foundation Models

Objective: To quantitatively evaluate how the utilization of location embeddings during training and inference may improve the transferability of flood mapping models.

Short Description: Floods are hazardous events that impact millions of people annually. In this context, it is essential that humanitarian and hazard response are well informed about the extent of affected areas in a timely manner. To achieve this, reducing the time needed to select and train a model is crucial. If a flood mapping model is highly transferable, it can be used successfully in a wider variety of regions, making the process of mapping floods in these areas more straightforward. This approach could be particularly important for regions where there is little or no reference data available for training.

In this proposed master thesis, the student will study how the inclusion of location embeddings generated from the SatCLIP foundation model affect the transferability of common flood mapping models based on convolutional neural networks and image transformers.

Start: Anytime

Relevant Studies:

1. Klemmer, Konstantin, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. 2024. “SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery.” arXiv:2311.17179. Preprint, arXiv, April 12. https://doi.org/10.48550/arXiv.2311.17179.
2.  Bentivoglio, Roberto, Elvin Isufi, Sebastian Nicolaas Jonkman, and Riccardo Taormina. 2022. “Deep Learning Methods for Flood Mapping: A Review of Existing Applications and Future Research Directions.” Hydrology and Earth System Sciences 26 (16): 4345–78. https://doi.org/10.5194/hess-26-4345-2022.