Deep Learning-Driven Green Building Facade Segmentation: Enhancing Sustainable Urban Development Through U-Net Architectures, Edge Detection, and Attention Mechanisms

This work aims to redefine the function of green buildings (GB) in sustainable urban development via the use of deep learning (DL) algorithms for precise segmentation of building facades. This method utilizes U-Net models to examine architectural characteristics, improving accuracy in facade segmentation and promoting ecologically friendly design and urban planning. A dataset of 606 open-source photos, tagged with 51,731 architectural elements, was used, and class imbalances were mitigated by six data augmentation methods. The research started with a foundational U-Net model (Model I) trained on the original dataset, followed by an augmented U-Net model (Model II) using Canny Edge Detection (CED) to increase edge definition. Model III was subsequently created by integrating an attention mechanism into Model II. Model assessments indicated that Model III had the maximum performance, providing comprehensive facade forecasts with an accuracy of 0.99. This study illustrates that the amalgamation of deep learning approaches with edge detection, data augmentation, and attention processes significantly enhances GB segmentation accuracy, providing a beneficial resource for architects and urban planners in facade evaluation. Improved precision in identifying architectural features facilitates more sustainable urban planning, promoting energy-efficient, low-carbon, and environmentally friendly building methods. Ultimately, these strategies facilitate urban infrastructure development while minimizing environmental impact, so significantly enhancing environmental sustainability.