Traditional vernacular architecture is an important component of historical and cultural heritage, and the accurate identification of its construction period is of great significance for architectural heritage conservation, historical research, and urban–rural planning. However, traditional methods for period identification are labor-intensive, potentially damaging to buildings, and lack sufficient accuracy. To address these issues, this study proposes a deep learning-based method for classifying the construction periods of traditional vernacular architecture. A dataset of traditional vernacular architecture images from the Longzhong region of Gansu Province was constructed, covering four periods: before 1911, 1912–1949, 1950–1980, and from 1981 to the present, with a total of 1181 images. Through comparative analysis of three mainstream models—ResNet50, EfficientNet-b4, and Vision Transformer—we found that EfficientNet demonstrated optimal performance in the classification task, achieving Accuracy, Precision, Recall, and F1-scores of 85.1%, 81.6%, 81.0%, and 81.1%, respectively. These metrics surpassed ResNet50 by 1.4%, 1.3%, 0.5%, and 1.2%, and outperformed Vision Transformer by 8.1%, 9.1%, 9.5%, and 9.1%, respectively. To further improve feature extraction and classification accuracy, we propose the “local–global feature joint learning network architecture” (DualBranchEfficientNet). This dual-branch design, comprising a global feature branch and a local feature branch, effectively integrates global structure with local details and significantly enhances classification performance. The proposed architecture achieved Accuracy, Precision, Recall, and F1-scores of 89.6%, 87.7%, 86.0%, and 86.7%, respectively, with DualBranchEfficientNet exhibiting a 2.0% higher Accuracy than DualBranchResNet. To address sample imbalance, a hybrid triplet loss function (Focal Loss + Triplet Loss) was introduced, and its effectiveness in identifying minority class samples was validated through ablation experiments. Experimental results show that the DualBranchEfficientNet model with the hybrid triplet loss outperforms traditional models across all evaluation metrics, particularly in the data-scarce 1950–1980 period, where Recall increased by 7.3% and F1-score by 4.1%. Finally, interpretability analysis via Grad-CAM heat maps demonstrates that the DualBranchEfficientNet model incorporating hybrid triplet loss accurately pinpoints the key discriminative regions of traditional dwellings across different eras, and its focus closely aligns with those identified by conventional methods. This study provides an efficient, accurate, and scalable deep learning solution for the period identification of traditional vernacular architecture.

The extraction of features and classification of traditional dwellings plays significant roles in preserving and ensuring the sustainable development of these structures. Currently, challenges persist in subjective classification and the accuracy of feature extraction. This study focuses on traditional dwellings in Gansu Province, China, employing a novel model named Improved Swin Transformer. This model, based on the Swin Transformer and parallel grouped Convolutional Neural Networks (CNN) branches, aims to enhance the accuracy of feature extraction and classification precision. Furthermore, to validate the accuracy of feature extraction during the prediction process and foster trust in AI systems, explainability research was conducted using Grad-CAM-generated heatmaps. Initially, the Gansu Province Traditional Dwelling Dataset (GTDD) is established. On the constructed GTDD dataset, the Improved Swin Transformer attains an accuracy of 90.03% and an F1 score of 87.44%. Comparative analysis with ResNet-50, ResNeXt-50, and Swin Transformer highlights the outstanding performance of the improved model. The confusion matrix of the Improved Swin Transformer model reveals the classification results across different regions, indicating that the primary influencing factors are attributed to terrain, climate, and cultural aspects. Finally, using Grad-CAM-generated heatmaps for explaining classifications, it is observed that the Improved Swin Transformer model exhibits more accurate localization and focuses on features compared to the other three models. The model demonstrates exceptional feature extraction ability with minimal influence from the surrounding environment. Simultaneously, through the heatmaps generated by the Improved Swin Transformer for traditional residential areas in five regions of Gansu, it is evident that the model accurately extracts architectural features such as roofs, facades, materials, windows, etc. This validates the consistency of features extracted by the Improved Swin Transformer with traditional methods and enhances trust in the model and decision-making. In summary, the Improved Swin Transformer demonstrates outstanding feature extraction ability and accurate classification, providing valuable insights for the protection and style control of traditional residential areas.