Enhancing Smart Tourism: EfficientNet-Based Recognition of Saudi Arabia’s Cultural Landmarks

Abstract

Advances in computer vision and deep learning are reshaping tourism by enabling automated landmark recognition and enhancing visitor experience. This study evaluates EfficientNetB0 for recognizing 16 Saudi landmarks spanning cultural, historical, and modern sites, and integrates predictions with a bilingual (Arabic/English) metadata repository containing names, locations, historical context, and visiting recommendations. A curated dataset of 480 images (balanced via offline augmentation) was used, and the models were trained using a two-phase transfer learning procedure. Under 5-fold stratified cross-validation, EfficientNetB0 achieved a mean accuracy of 0.994 and an F1-score of 0.992, outperforming ResNet50 and MobileNetV3Small. On an external test set of unseen images, the model reached 92.5% top-1 accuracy, with errors concentrated among architecturally similar sites (e.g., Diriyah vs. Ibrahim Palace). A Gradio-based web interface provided near real-time recognition (177 ms) and enriched predictions with bilingual cultural metadata, achieving high usability (SUS 80–100). These findings demonstrate not only technical robustness but also practical contributions to smart tourism, aligning with Saudi Vision 2030 in cultural preservation, digital transformation, and sustainable tourism development.