Constructing an Extensible Building Damage Dataset via Semi-supervised Fine-Tuning across 12 Natural Disasters

Post-disaster building damage assessment (BDA) is vital for emergency response. Deep learning (DL) models are increasingly being applied to achieve quick and automatic BDA on disaster remote sensing imagery, and their performance largely relies on the knowledge base offered by the dataset. However, constructing a BDA dataset requires intensive expert labeling work and a massive time, leading to a substantial lag in dataset enrichment and model development in the current research field. To address this, this paper introduces a new multidisaster BDA benchmark, the extensible building damage (EBD) dataset, which includes over 18,000 pre- and post-disaster image pairs from 12 recent disaster events, covering over 175,000 building annotations with 4-level damage labels. Unlike previous BDA datasets, EBD follows a semiautomatic labeling workflow and has reduced construction time by 80% compared to full manual labeling. In this process, the DL model served as the machine expert to perform automatic labeling. It was pretrained on the xView2 building damage dataset and then transferred to each new disaster scenario via semi-supervised fine-tuning (SS-FT). SS-FT not only leverages a few labeled samples for supervised fine-tuning but also incorporates both labeled and unlabeled samples into pixel-level contrastive learning. Results demonstrate that the DL model has considerably improved annotation performance under SS-FT. A series of analyses have proven EBD’s building damage feature diversity, practical value in emergency mapping, and knowledge enrichment to the existing benchmark. EBD advances data renewal for natural disaster scenarios and supports the application of artificial intelligence in emergency response efforts.