Publicação
AI-Based Flood Detection Using TerraMind and Sentinel Data
| Resumo: | Floods are among the most frequent and damaging natural hazards. Timely flood extent maps support emergency response, especially in high-risk regions such as South Sudan. Satellite flood mapping is often limited by data gaps, particularly the lack of cloud-free Sentinel-2 imagery during flood events. This thesis presents a deep learning framework that uses Sentinel-1 SAR and Sentinel-2 optical data when both are available and remain operational when one modality is missing. The framework combines a pretrained multimodal encoder (TerraMind) with a UPerNet decoder. Thinking in Modalities is used to enable inference with incomplete inputs. The model is fine-tuned on the Sen1Floods11 dataset and evaluated using Sentinel-1 only, Sentinel-2 only, and Sentinel-1 + Sentinel-2 configurations. Flood mapping is derived from pre- and post-event predictions and produces flood probability maps, binary flood masks, and uncertainty outputs. Benchmark results show strong segmentation performance on Sen1Floods11. In the South Sudan transfer case, the fused Sentinel-1 + Sentinel-2 configuration produces the most coherent flood patterns. Sentinel-1 only fails to produce reliable flood maps in this study area. It overestimates flooding and introduces scattered false detections. Sentinel-2 outputs are cleaner, but clouds and cloud shadows reduce the observable flood signal and lead to underestimation. Overall, multimodal fusion reduces single modality errors and supports more consistent flood extent mapping under varying data availability. |
|---|---|
| Autores principais: | Tariq, Ayesha |
| Assunto: | Flood mapping Remote sensing Deep learning Multimodal data fusion Sentinel-1 and Sentinel-2 |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Floods are among the most frequent and damaging natural hazards. Timely flood extent maps support emergency response, especially in high-risk regions such as South Sudan. Satellite flood mapping is often limited by data gaps, particularly the lack of cloud-free Sentinel-2 imagery during flood events. This thesis presents a deep learning framework that uses Sentinel-1 SAR and Sentinel-2 optical data when both are available and remain operational when one modality is missing. The framework combines a pretrained multimodal encoder (TerraMind) with a UPerNet decoder. Thinking in Modalities is used to enable inference with incomplete inputs. The model is fine-tuned on the Sen1Floods11 dataset and evaluated using Sentinel-1 only, Sentinel-2 only, and Sentinel-1 + Sentinel-2 configurations. Flood mapping is derived from pre- and post-event predictions and produces flood probability maps, binary flood masks, and uncertainty outputs. Benchmark results show strong segmentation performance on Sen1Floods11. In the South Sudan transfer case, the fused Sentinel-1 + Sentinel-2 configuration produces the most coherent flood patterns. Sentinel-1 only fails to produce reliable flood maps in this study area. It overestimates flooding and introduces scattered false detections. Sentinel-2 outputs are cleaner, but clouds and cloud shadows reduce the observable flood signal and lead to underestimation. Overall, multimodal fusion reduces single modality errors and supports more consistent flood extent mapping under varying data availability. |
|---|