The Application of Transformer Model in Building Information Modeling

Abstract

Building information modeling leverages high-resolution satellite imagery and change detection for urban planning and disaster monitoring. This study enhances building information modeling accuracy by integrating coordinate attention with a Transformer hybrid architecture. Local feature extraction by convolutional neural network and global context modeling by Transformer are combined. Feature exchange techniques and a hollow space pyramid pooling module improve multi-scale change detection. Lightweight designs, including depthwise separable convolutions and Ghost modules, reduce computational costs. Experimental results show the model stabilizes after 80 iterations, achieving 95% accuracy and a 1.37% mIoU improvement. With a Kappa value of 0.795 and minimal parameters, the framework enables efficient synthetic aperture radar-based building change detection, suitable for real-time urban monitoring. The raised model can achieve the task of building information modeling, laying the foundation for large-scale automatic recognition and classification of building images.