Publicação
Few-Shots Learning for Post-Earthquake Building Damage Assessment Using Metric-based and Transfer Learning Methods
| Resumo: | Earthquakes are among the most devastating natural disasters, often causing widespread destruction and loss of life. A rapid and accurate assessment of building damage is crucial for effective disaster response and recovery. However, traditional methods are time-consuming and require significant resources, while deep learning approaches struggle with challenges such as limited and imbalanced datasets. This study focuses on the Mexico City Earthquake (2017), and it utilizes satellite imagery to classify building damage into varying levels of severity. Moreover, this study explores the use of Few-Shot Learning (FSL) to overcome the challenges of limited and imbalanced data. In the realm of FSL, we use metric-based learning and transfer learning approaches to improve classification performance in scenarios where data is scarce and imbalanced. To evaluate the effectiveness of our approach, we evaluate and compare Prototypical Networks, EfficientNetB7, and ResNet50, analyzing key metrics such as precision, recall, F-score, and overall accuracy. Our findings reveal that Prototypical Networks outperform other models, particularly in identifying severely damaged structures. Additionally, data augmentation and oversampling are proven to be effective techniques for handling data imbalance. As the dataset is very limited in this study, the challenge of data scarcity, however, persists. To provide deeper insights, we integrate our damage predictions with geospatial mapping, revealing a decent correlation between predicted damage severity and actual impacted areas. Ultimately, this research highlights the potential of Few-Shot Learning in enabling rapid and scalable damage assessment in data-limited scenarios and contributes to improving emergency response efforts and optimizing resource allocation. |
|---|---|
| Autores principais: | Meskinyaar, Enayatullah |
| Assunto: | Building Damage Assessment Deep Learning Few Shots Learning Data Balancing Remote Sensing SDG 11 - Sustainable cities and communities |
| Ano: | 2025 |
| 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: | Earthquakes are among the most devastating natural disasters, often causing widespread destruction and loss of life. A rapid and accurate assessment of building damage is crucial for effective disaster response and recovery. However, traditional methods are time-consuming and require significant resources, while deep learning approaches struggle with challenges such as limited and imbalanced datasets. This study focuses on the Mexico City Earthquake (2017), and it utilizes satellite imagery to classify building damage into varying levels of severity. Moreover, this study explores the use of Few-Shot Learning (FSL) to overcome the challenges of limited and imbalanced data. In the realm of FSL, we use metric-based learning and transfer learning approaches to improve classification performance in scenarios where data is scarce and imbalanced. To evaluate the effectiveness of our approach, we evaluate and compare Prototypical Networks, EfficientNetB7, and ResNet50, analyzing key metrics such as precision, recall, F-score, and overall accuracy. Our findings reveal that Prototypical Networks outperform other models, particularly in identifying severely damaged structures. Additionally, data augmentation and oversampling are proven to be effective techniques for handling data imbalance. As the dataset is very limited in this study, the challenge of data scarcity, however, persists. To provide deeper insights, we integrate our damage predictions with geospatial mapping, revealing a decent correlation between predicted damage severity and actual impacted areas. Ultimately, this research highlights the potential of Few-Shot Learning in enabling rapid and scalable damage assessment in data-limited scenarios and contributes to improving emergency response efforts and optimizing resource allocation. |
|---|