FD-SC: focus-diffusion self-contrastive framework empowers efficient remote sensing building damage detection

doi:10.1016/j.isprsjprs.2026.04.032

FD-SC: focus-diffusion self-contrastive framework empowers efficient remote sensing building damage detection

2026-04-20

0

PRE

AI

DOI：10.1016/j.isprsjprs.2026.04.032

Original

Original request for help

Save

Abstract

En 中文

Efficient building damage detection (BDD) based on remote sensing images is crucial for rapid and accurate disaster assessment. However, existing methods struggle to balance accuracy and inference efficiency, and depend heavily on large amounts of annotated data. This conflicts with the demands for high precision, high time sensitivity, and limited resources in actual disaster scenarios. To address these challenges, this study proposes the Focus-Diffusion Self-Contrastive (FD-SC) framework, which enhances damage feature representation through the fusion of supervised and self-supervised signals, while improving inference efficiency and data utilization efficiency. The framework consists of the Focus-Diffusion Network (FDN) and Self-Contrastive Learning (SCL). FDN simulates the selective attention mechanism of the human visual system to enable dynamic sparse computation and feature diffusion, thereby improving inference efficiency. Meanwhile, lightweight difference perception and multi-scale interaction mechanisms are proposed to further enhance damage feature representation. And SCL avoids complex data augmentation and pretraining. It exploits semantic correlations within a single feature to construct contrastive samples for self-contrastive learning, thereby improving data utilization efficiency. Furthermore, this study constructs a high-resolution Gaza dataset with 35,939 damaged buildings. Experiments on the GE dataset (Global Earthquake dataset) and the Gaza dataset show that the F1 of FD-SC improves by at least 1.34% and 1.59%, respectively. And FD-SC contains only 3.34 M parameters and 4.25G FLOPs, achieving an inference speed of 39.92 FPS. Moreover, few-shot experiments confirm its robustness under data-scarce conditions, and transfer experiments on the Hawaii wildfire verify its strong generalization capability.

Keywords:

Building damage detection

Remote sensing

Self-contrastive learning

Efficient inference

Sparse computation

FD-SC: focus-diffusion self-contrastive framework empowers efficient remote sensing building damage detection

Abstract

Journal

Organization