EDB-HSTEU-Net : Earthquake-damaged building detection using a novel hybrid swin transformer efficient U-Net (HSTEU-Net) and transfer learning techniques from post-event VHR remote sensing data
Khankeshizadeh, Ehsan LU ; Mohammadzadeh, Ali and Jamali, Sadegh LU
(2025)
In Journal of Building Engineering
108.
- Abstract
Rapid and accurate generation of building damage maps (BDMs) following earthquakes is crucial for effective disaster response and rescue operations. With the increasing availability of optical very high-resolution remote sensing (OVHR-RS) images, two considerable deep learning challenges arise: (1) developing robust models for data-rich regions and (2) adapting pre-trained models for areas with limited labeled data, where ground-truth collection is often infeasible. This study introduces the Hybrid Swin Transformer Efficient U-Net (HSTEU-Net), a novel deep-learning-based model designed to generate BDMs exclusively using post-event OVHR-RS data, eliminating dependence on pre-disaster imagery. The model incorporates a dual-encoder U-Net... (More)
Rapid and accurate generation of building damage maps (BDMs) following earthquakes is crucial for effective disaster response and rescue operations. With the increasing availability of optical very high-resolution remote sensing (OVHR-RS) images, two considerable deep learning challenges arise: (1) developing robust models for data-rich regions and (2) adapting pre-trained models for areas with limited labeled data, where ground-truth collection is often infeasible. This study introduces the Hybrid Swin Transformer Efficient U-Net (HSTEU-Net), a novel deep-learning-based model designed to generate BDMs exclusively using post-event OVHR-RS data, eliminating dependence on pre-disaster imagery. The model incorporates a dual-encoder U-Net architecture, integrating ImageNet pre-trained Swin Transformer and EfficientNet components to enhance local and global damage feature extraction. A key challenge addressed is the application of transfer learning in disaster scenarios with sparse labeled data. As a solution to this problem, the study proposes a data-driven transfer learning (DTL) technique, which selectively integrates relevant samples from the source domain to refine the model in the target region. The research follows a four-phase methodology: data preprocessing, model training, BDM generation and evaluation, and transfer learning implementation. Experimental validation using OVHR-RS data from Haiti-Port-au-Prince (source) and Iran-Bam (target) demonstrates the superior performance of HSTEU-Net, achieving an 85.59 % damage detection rate (DDR)—outperforming state-of-the-art models. The proposed DTL technique further boosts DDR to 87.02 %, even in data-limited environments. These findings suggest that HSTEU-Net, in combination with the DTL strategy, can contribute toward addressing some of the key challenges in scalable, accurate, and resource-efficient BDM for disaster management applications.
(Less)
- author
- Khankeshizadeh, Ehsan
LU
; Mohammadzadeh, Ali
and Jamali, Sadegh
LU
- organization
- publishing date
- 2025-08
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- Building damage map, Data-based transfer learning, Dual-encoder U-Net, SwinTransformer, Transfer learning
- in
- Journal of Building Engineering
- volume
- 108
- article number
- 112889
- publisher
- Elsevier
- external identifiers
-
- scopus:105004878522
- ISSN
- 2352-7102
- DOI
- 10.1016/j.jobe.2025.112889
- language
- English
- LU publication?
- yes
- id
- 840943e4-241e-4318-b19f-b4596ccd0c12
- date added to LUP
- 2025-07-17 11:26:26
- date last changed
- 2025-07-17 11:27:46
@article{840943e4-241e-4318-b19f-b4596ccd0c12,
abstract = {{<p>Rapid and accurate generation of building damage maps (BDMs) following earthquakes is crucial for effective disaster response and rescue operations. With the increasing availability of optical very high-resolution remote sensing (OVHR-RS) images, two considerable deep learning challenges arise: (1) developing robust models for data-rich regions and (2) adapting pre-trained models for areas with limited labeled data, where ground-truth collection is often infeasible. This study introduces the Hybrid Swin Transformer Efficient U-Net (HSTEU-Net), a novel deep-learning-based model designed to generate BDMs exclusively using post-event OVHR-RS data, eliminating dependence on pre-disaster imagery. The model incorporates a dual-encoder U-Net architecture, integrating ImageNet pre-trained Swin Transformer and EfficientNet components to enhance local and global damage feature extraction. A key challenge addressed is the application of transfer learning in disaster scenarios with sparse labeled data. As a solution to this problem, the study proposes a data-driven transfer learning (DTL) technique, which selectively integrates relevant samples from the source domain to refine the model in the target region. The research follows a four-phase methodology: data preprocessing, model training, BDM generation and evaluation, and transfer learning implementation. Experimental validation using OVHR-RS data from Haiti-Port-au-Prince (source) and Iran-Bam (target) demonstrates the superior performance of HSTEU-Net, achieving an 85.59 % damage detection rate (DDR)—outperforming state-of-the-art models. The proposed DTL technique further boosts DDR to 87.02 %, even in data-limited environments. These findings suggest that HSTEU-Net, in combination with the DTL strategy, can contribute toward addressing some of the key challenges in scalable, accurate, and resource-efficient BDM for disaster management applications.</p>}},
author = {{Khankeshizadeh, Ehsan and Mohammadzadeh, Ali and Jamali, Sadegh}},
issn = {{2352-7102}},
keywords = {{Building damage map; Data-based transfer learning; Dual-encoder U-Net; SwinTransformer; Transfer learning}},
language = {{eng}},
publisher = {{Elsevier}},
series = {{Journal of Building Engineering}},
title = {{EDB-HSTEU-Net : Earthquake-damaged building detection using a novel hybrid swin transformer efficient U-Net (HSTEU-Net) and transfer learning techniques from post-event VHR remote sensing data}},
url = {{http://dx.doi.org/10.1016/j.jobe.2025.112889}},
doi = {{10.1016/j.jobe.2025.112889}},
volume = {{108}},
year = {{2025}},
}