A method for protecting architectural heritage in smart cities based on deep learning models: taking real-time weed monitoring as an example

Timely and accurate detection of invasive weeds on the facades of historic buildings is crucial for the preservation of architectural heritage in smart cities. However, challenges such as complex surface textures, low-contrast areas, and weed occlusion often hinder effective detection by traditional computer vision methods. In this study, YOLOv11-SWDS is proposed, which is specifically designed for real-time weed detection in heritage conservation scenarios. The backbone network is redesigned with a multi-scale feature extraction architecture (BLRA), which effectively improves the representation ability of fine-grained features in visually cluttered environments. In addition, we integrate channel-level (CBAM) and spatial-level (C2PSA) attention modules to enable the model to better distinguish weeds from building textures and shadows. To support actual deployment, we optimize the model using quantization and knowledge distillation, significantly reducing GFLOPs and parameter size. Experiments on a custom annotated facade weed dataset demonstrate the superiority of the proposed model, achieving an F1 score of 86.0% and a mAP@50 of 89.7%, outperforming the baseline model while maintaining a low inference latency (<200 ms) suitable for edge devices. This research contributes to the development of interpretable and deployable deep learning models for smart heritage conservation and provides a scalable solution for urban facade maintenance.