Architectural heritage assessment increasingly relies on automated visual analysis, yet existing deep learning approaches often lack interpretability and provide limited insight into how cultural value judgments are formed. To address this gap, this study proposes an interpretable multi-task framework—YOLOv11-CVHP—for architectural heritage image recognition and integrated value classification. The model incorporates a lightweight backbone network (RepGhostNet), an enhanced attention module (ArchDetectAttn), and the WIoU loss function to improve detection accuracy and robustness. Based on the architectural components and semantic attributes detected by YOLOv11-CVHP, seven visual–cultural variables were constructed to quantify heritage characteristics. A Random Forest classifier was then applied to predict four-level integrated value grades. Although Random Forest is commonly regarded as a black-box model, interpretability is achieved through the incorporation of SHAP, which attributes the contribution of each visual–cultural feature to the final value grade, allowing transparent analysis of the decision process. Results indicate that Cultural Value (Intellectual) consistently serves as the dominant discriminative factor across all levels, while Historical Period and Structural Integrity play critical roles in differentiating between higher value categories. The classifier demonstrates strong generalization, with five-fold Precision–Recall curves showing stable performance and ROC–AUC scores exceeding 0.90 on both training and test sets. In conclusion, the integrated YOLOv11-CVHP and SHAP-enhanced Random Forest framework provides both high accuracy and clear interpretability, offering a practical and explainable solution for automated architectural heritage identification and value assessment.
Hu et al. (Fri,) studied this question.