Rapid mapping of volcanic eruption building damage: A model based on prior knowledge and few-shot fine-tuning

Large-scale volcanic eruptions always inflict severe damage upon facilities and cause significant environmental pollution. Building damage caused by lava flow and volcanic ash coverage can reflect the infrastructure devastation and victim scope within the affected region. The application of machine learning methods for building damage automated identification from remote sensing imagery typically relies on a large number of training samples. However, labeled data scarcity is a common issue in the field of disasters, particularly in volcanic eruptions. To address this, we propose a two-stage building damage quick mapping workflow, which combines a building localization model trained on prior knowledge and a damage classification model fine-tuned on few-shot volcanic eruption-related samples. The classification model utilizes a CNN-based Siamese network for bi-temporal image feature extraction and comparison, with the backbone initialized with pre-trained weights from ImageNet. We conducted building damage classification tasks for single-disaster scenarios and cross-disaster domain scenarios in the eruptions of Mount Semeru, Tonga, and ST. Vincent; the visual damage level of each building was used as ground truth. The results demonstrate that our model can identify building damage efficiently and accurately in different volcanic eruption scenarios, with an over 93% F1-score on the 2-way 20-shot tasks. Furthermore, though building samples from different volcanic eruption regions present cross-domain challenges, our model can adapt to different feature domains by being supplemented with a few samples of another volcanic eruption disaster. Additionally, in the case of Mount Semeru Eruption, we gain insights into the potential of building damage statistics in post-eruption environmental assessments. To further enhance the model robustness on mixed-domain samples and multi-level damage classification tasks, issues including sample bias of certain disaster sources should be addressed.