Toward Sustainable Heritage Documentation: A Deep Transfer Learning Approach to Classifying Saudi Architectural Façade Styles
TL;DR
A deep-learning classification framework for building façade imagery considering the Saudi architectural styles and indicates that ConvNeXtBase with two-stage transfer learning is a promising step toward automated documentation of Saudi architectural styles.
Abstract
The Kingdom of Saudi Arabia launches the Saudi Architecture Map, which includes 19 architectural styles for the country’s main regions. These architectural styles are shaped by distinct climatic, material, and cultural determinants across the country. A deep dive into the architectural elements that define these styles reveals many similarities and nuances, making manual classification of modern buildings challenging. Despite the kingdom’s rich heritage, and since the launch of the Saudi Architecture Map and its contained styles and corresponding design guidelines, no operational digital pipeline currently exists for automated, quantitative, and reproducible classification of façade imagery against these styles. This limits heritage documentation and compliance monitoring at scale. This study attempts to fill this gap. We developed a deep-learning classification framework for building façade imagery considering the Saudi architectural styles. We curated a dataset of 3750 façade images, comprising 2320 full-building façade images and 1430 architectural-element images. We collected the data from online heritage repositories. We benchmarked five convolutional backbones (ResNet50, EfficientNetB0, EfficientNetV2M, ConvNeXtBase, and InceptionV3) under a unified two-stage transfer learning strategy. The first stage trained classifier heads on architectural-element images with frozen backbone weights, and the second stage fine-tuned the top 30% of backbone layers on full-building façade images. Among the five CNN backbones, ConvNeXtBase achieved the strongest performance across all four metrics, achieving 90.75% in overall accuracy, 88.06% in F1-Score, 92.05% in producer accuracy, and 86.65% in user accuracy. Ten of the eighteen styles exceeded 90% F1, and two styles—Bisha Desert and Tabuk Coast, both with very small corrected test counts (n = 6 and n = 1, respectively)—fell below 70%. These results indicate that ConvNeXtBase with two-stage transfer learning is a promising step toward automated documentation of Saudi architectural styles. Yet, further validation through ablation testing or calibration is needed before deployment for compliance monitoring.