Automated UAV-based bio-colonization detection of buildings and infrastructure using multispectral fusion network

Bio-colonization on civil infrastructure poses long-term deterioration problems and maintenance challenges. Traditional inspection methods, which primarily rely on manual visual surveys, are labor-intensive, subjective, and limited in spatial coverage. Recently, UAV-based monitoring method has been rapidly developed and widely applied across various domains. However, the accuracy and robustness of commonly deployed bio-colonization detection methods could not fulfill the practical applications due to uncomplete data from single modality. To address these issues, this research presents an automated integrated framework that combines multispectral and RGB imagery with a deep semantic segmentation network for accurate and large-scale detection of bio-colonization on buildings and infrastructure. Firstly, a multispectral Unmanned Aerial Vehicle (UAV) data acquisition workflow has been developed, followed by a three-step calibration process to precisely align multispectral and RGB images. Subsequently, a lightweight semantic segmentation network was proposed based on a U-Net architecture with MobileNetV3 backbone, integrated with a novel Band Context Attention (BCA) module that enhances attention to spectral-channel features. Then, a comprehensive ablation experiment was conducted on 15 channel fusion configurations, revealing that RGB combined with the Normalized Difference Vegetation Index (NDVI) yields the best performance among all configurations. Further comparisons among attention mechanisms demonstrated that the BCA-integrated model achieved the highest performance, reaching a mean intersection over union (mIoU) of 81.61 %, outperforming other attention modules evaluated in our experiment. The proposed framework provides an efficient and generalizable solution for bio-colonization monitoring and contributes to intelligent maintenance of civil infrastructure under real-world conditions.