By combining machine-driven annotation with expert review, the EBD dataset makes disaster damage mapping faster, more accurate, and globally scalable – a breakthrough for saving lives in crisis response.

Research: Constructing an Extensible Building Damage Dataset via Semi-supervised Fine-Tuning across 12 Natural Disasters. Image Credit: Bell nipon / Shutterstock
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, mainly 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, there is a need for further research into semi-automated disaster response technologies.
Introducing the EBD dataset
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 in the Journal 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 employs 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.
How the SS-FT method works
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 was demonstrated through disaster events such as Hurricane Ian and the Turkey Earthquake, where the model showed significant improvements over both pre-trained-only and supervised fine-tuning settings.
"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."
Technical implementation and impact
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.
Source:
Journal reference: