A Computational Approach to Traditional Chinese Architectural Styles: Intelligent Identification and Spatial Pattern Analysis

The preservation of traditional architecture, which is a tangible carrier of regional culture and identity, requires precise and scalable methods for style classification. Here, we address the critical need for such methodologies by developing and applying an integrated computational framework that synergizes deep learning with spatial analytical techniques from cultural geography to document and analyze traditional architectural heritage. The primary research objectives were threefold: first, to construct a high-performance automated system for quantitatively classifying major Chinese traditional architectural styles(CTASs) from visual data; second, to leverage the system's outputs to analyze the multiscale spatial patterns of these styles; and third, to interpret these patterns to identify cultural boundaries and regional diversity, thereby providing a data-driven foundation for heritage conservation and planning. To achieve these aims, the research methodology began with the curation of a multimodal image database derived from 646 nationally designated traditional villages. An EfficientNet model, enhanced for multiscale feature fusion to capture both global and local details, was trained on this dataset to classify six typical styles: Jing, Jin, Chuan, Wan, Su, and Min. To ensure transparency and interpretability, Grad-CAM was used to visualize the key architectural components that informed the model's decisions. Subsequently, the geographically referenced classification results were subjected to a comprehensive spatial analysis suite, including spatial autocorrelation, standard deviational ellipse analysis, and calculation of Shannon diversity and Simpson evenness indices at the provincial level. The results and conclusions of the integrated analysis are detailed and multifaceted. The EfficientNet model achieved an overall classification accuracy of 90%, confirming the efficacy of the deep-learning approach. However, a misclassification rate of 10.7% was observed between the Jing and Chuan styles. Grad-CAM analysis provided critical insights into this phenomenon, revealing that the model's confusion stemmed from a shared, significant focus on similar wooden walls and colonnade features in both the Jing and Chuan styles, highlighting subtle inter-style visual overlaps. At the macroscale, the spatial distribution analysis confirmed that CTASs generally follow a cultural dissemination model that conforms to mountains, waters, and other natural barriers. For instance, the northern Jing and Jin styles cluster in the Yellow River Basin, the southern Su and Wan styles follow intricate water networks, and the Min and Chuan styles are dispersed across hilly and mountainous terrain. At the mesoscale, Local Indicators of Spatial Association (LISA) analysis identified statistically significant high-high clustering areas for each style. Synthesizing these with topographic and hydrological data allowed for the delineation of coherent cultural geography: five distinct cultural core areas (where a single style is intensely concentrated), three well-defined cultural transition belts (exhibiting high stylistic mixing), and two cultural fracture zones (areas of low style density and diversity). This regionalization provides a precise spatial template for designing cross-provincial heritage corridor protection systems. At the micro (provincial) scale, the diversity and evenness indices revealed three characteristic patterns: Multi-Style Fusion Zones (high diversity, high uniformity), Oligarchic Equilibrium Zones (low diversity, high uniformity), and Style-Pure Zones (low diversity, low uniformity). These patterns offer crucial insights for tailoring provincial-level conservation strategies and regional planning. This study makes significant strides on these two fronts. Methodologically, it pioneers a replicable and scalable framework that successfully bridges computer vision and cultural geography, offering a new and robust tool for automated architectural stylistic analysis. Substantively, it provides the first comprehensive, quantitative, and multiscale spatial dissection of CTASs based on large-scale empirical data. Collectively, these findings provide a rigorous scientific foundation for evidence-based policies in multiscale heritage conservation, cultural landscape management, and sustainable regional development.