Self-supervised Pretraining with Edge Guidance for Building Damage Assessment

Songxi Yang,Bo Peng,Tang Sui,Qunying Huang

DOI: https://doi.org/10.1145/3687123.3698288

2024-01-01

Abstract:Building damage assessment is a critical subtask within GeoAI-driven remote sensing semantic segmentation, where deep neural networks have been widely applied. Most existing works typically use pre- and post-disaster images as input of a siamese deep neural network under supervised learning, which requires a large amount of labeled data. However, in real-world scenarios, acquiring massive labeled datasets is often difficult, making fully supervised methods less practical. To overcome this, we propose a self-supervised pretraining framework based on pre-post remote sensing image pairs. In the first stage, a dual denoising autoencoder with Vision Transformer backbone is proposed for image representation learning. In the second stage, two downstream tasks-building localization and building damage severity-are performed. Additionally, we incorporate an edge guidance module and an edge detection loss to further enhance performance in downstream tasks. On the xBD dataset, the largest building damage assessment dataset, the proposed method achieves an F1 score of 0.895 for building localization, outperforming state-of-the-art image segmentation techniques. Additionally, it receives an F1 score of 0.704 in building damage severity compared to state-of-the-art in all self-supervised learning methods.