Deep Learning Classification and Feature  Analysis for Ganjiang River Basin Traditional Villages Multicultural Types

This study tackles the challenge of identifying regional cultural features in traditional villages across China's Ganjiang River Basin by proposing a deep learning-based image classification method. Focusing on five cultural types (Hakka, Linchuan, Luling, Yuanzhou, and Yuzhang) as classification labels, we constructed an image dataset and innovatively adapted image recognition architectures for village feature extraction. The research advances three key aspects: (1) A novel framework integrating feature focusing, knowledge transfer, and environmental perturbation optimization, developed through comparative analysis of nine model architectures; (2) Evaluation using Accuracy, Precision, Recall, F1 Score, ROC-AUC, and PRC curves to verify model consistency; (3) Visual interpretation through Grad-CAM, RPN, and UMAP methods. Results demonstrate the modified ConvNeXt model achieves optimal performance, effectively capturing cultural elements in village imagery. This approach enables efficient extraction of regional cultural characteristics from traditional settlements, offering an interpretable technical solution for digitally preserving cultural heritage.