A new dataset, the Extensible Building Damage (EBD) dataset, offers significant improvements in disaster response mapping by combining satellite imagery and deep learning techniques. This dataset, covering 12 natural disasters, uses semi-supervised fine-tuning (SS-FT) to reduce the time and effort traditionally required for manual damage labeling, speeding up disaster recovery efforts globally.
Building damage assessments (BDA) are crucial for post-disaster recovery, as they help in identifying areas most in need of urgent assistance. However, current BDA methods suffer from slow dataset development, largely due to manual labeling requirements. The new Extensible Building Damage (EBD) dataset addresses this by leveraging deep learning for semi-automated labeling, improving the speed and accuracy of damage assessment in disaster zones. Based on these challenges, or due to these problems, there is a need for further research into semi-automated disaster response technologies.
Researchers from Zhejiang University and the RIKEN Center for Advanced Intelligence in Japan, with collaboration from various international institutions, have introduced the EBD dataset, published (DOI: 10.34133/remotesensing.0733) in J ournal of Remote Sensing . This dataset represents a leap in disaster mapping by using machine-driven annotation to assist human experts in quickly categorizing building damage post-disaster. The SS-FT method it uses provides an innovative solution to the traditionally slow and labor-intensive task of damage classification.
The EBD dataset includes over 18,000 image pairs from 12 major natural disasters, with labels for over 175,000 buildings. Unlike earlier efforts, the dataset uses a semi-automatic annotation process, drastically reducing the manual workload by 80%. The SS-FT method not only utilizes a small amount of manually labeled data but also incorporates large sets of unlabeled samples for improved accuracy. This breakthrough provides faster, more reliable damage assessment results, particularly in areas with limited human resources.
The process begins with a pre-trained model using a historical dataset, which is then fine-tuned on disaster-specific data through the SS-FT method. By comparing pre- and post-disaster images, the model automatically classifies damage into four categories: No Damage, Minor Damage, Major Damage, and Destroyed. The SS-FT method has proven to improve model accuracy especially in situations with limited labeled samples. This capability is demonstrated through disaster events such as Hurricane Ian and the Turkey Earthquake, where the model showed significant improvements over pre-trained only setting and supervised fine-tuning setting.
"By reducing the reliance on manual labeling, the EBD dataset represents a major step forward in how we can use artificial intelligence in disaster response," said Dr. Zeyu Wang, a leading researcher on the project. "This system not only accelerates post-disaster recovery but also makes it more scalable, meaning it can be used globally to address future disaster events."
The research used high-resolution satellite imagery from the Maxar Open-Data Program, processing bi-temporal images to assess building damage. The SS-FT method was implemented using the PyTorch framework, with the model optimized on NVIDIA GPUs. The process involved multiple rounds of fine-tuning, using both labeled and unlabeled data to improve damage classification accuracy.
The EBD dataset has the potential to transform emergency response by providing rapid, accurate damage assessments. As this dataset continues to grow, it could be integrated into broader global disaster monitoring systems, offering valuable insights for climate change-related disasters. Additionally, its semi-automated labeling system can be applied to new disaster scenarios, making it an indispensable tool for disaster management worldwide. The future of disaster response relies on datasets like EBD, offering more timely and precise interventions to save lives.
###
References
DOI
Original Source URL
https://spj.science.org/doi/10.34133/remotesensing.0733
Funding Information
This work was supported by the National Key Research and Development Program of China under495Grants 2019YFE0127400.
About Journal of Remote Sensing
The Journal of Remote Sensing , an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.
Journal of Remote Sensing
Not applicable
Constructing an Extensible Building Damage Dataset via Semi-supervised Fine-Tuning across 12 Natural Disasters
7-Aug-2025
The authors declare that they have no competing interests.