Abstract
Recent advances in deep learning have achieved impressive accuracy in various building analysis tasks using street view imagery). However, a major challenge lies in the large-scale, labeled datasets typically required—an obstacle driven by limited raw data access and labor-intensive annotations. To overcome this, the present study introduces a domain adaptation (DA) framework for classifying exterior cladding materials. Six categories are targeted: Brick, Concrete, Glass, Stone, Mixed, and Others. A fully labeled dataset from Scotland and a partially labeled dataset from London form the basis of the approach, which leverages transformer-based architectures, data augmentation, and hyperparameter optimization to boost accuracy. In evaluations on unseen data, an axial transformer trained with augmented data and optimized hyperparameters emerged as most effective, achieving class-specific accuracies of 88.43% (Brick), 73.71% (Concrete), 68.67% (Glass), 91.33% (Stone), 86.65% (Mixed), and 83.46% (Others), culminating in an overall accuracy of 82.04%. These findings illustrate the potential of the DA-based method to maintain strong performance, with further refinements suggested for future work. The paper subsequently explores additional applications of this proposed strategy.