Abstract
Background: Preservation of cultural heritage in the digital era confronts unprecedented complexity, demanding methods that can reliably document, analyze, and interpret ever-growing corpora of heritage sites. Conventional workflows—built for smaller, more uniform collections—struggle with today’s large-scale, heterogeneous datasets and often fail to surface robust, generalizable patterns. Methods: We designed an integrated computational workflow that fuses Geographic Information Systems (GIS), Natural Language Processing (NLP), and machine learning to interrogate 25,705 cultural heritage sites across Zhejiang Province, China. The pipeline combines kernel density estimation (bandwidth = 15.7 km), hierarchical spatial clustering, topic modeling via Latent Dirichlet Allocation (50 topics), and an ensemble classification stage incorporating Support Vector Machines, Random Forest, and Gradient Boosting. Results: Global spatial structure exhibits strong positive autocorrelation (Moran’s I = 0.847, p < 0.001). We further resolve five well-defined clusters that differ systematically in both heritage-site density and cultural attribute profiles. The ensemble classifier achieves 89.3% accuracy for heritage-type prediction (evaluated on a held-out set, n = 750), while the spatial prediction models prioritize 1,247 high-probability candidate locations for undocumented sites with a validation accuracy of 78.4%. Conclusions: Taken together, these results demonstrate the utility of a computational pipeline for province-scale cultural-heritage analysis, delivering decision-ready evidence to inform preservation planning and to target archaeological survey. The methodology is readily transferable to other regions and cultural settings, enabling rigorously evidence-based heritage management.
Introduction: The Challenge of Scale in Digital Archaeology
The "digital turn" in archaeology has generated a massive influx of data, yet heritage management remains largely tethered to manual, small-scale workflows. Traditional methods struggle to process the volume and heterogeneity of modern administrative registries—datasets often plagued by inconsistent metadata and varying descriptive granularity. This research addresses the "scale vs. nuance" trade-off: how to analyze an entire province’s worth of history (spanning millennia and tens of thousands of sites) without losing the specific cultural context of individual locations.
Methods: A Multimodal Computational Workflow
Our framework treats spatial and textual data as "equal citizens" through an integrated pipeline:
- Spatial Analysis (GIS): We utilized Kernel Density Estimation (KDE) to identify site concentrations and Hierarchical Clustering (Ward’s linkage) to define distinct cultural regions. Spatial autocorrelation was quantified using Moran’s I.
- Textual Analysis (NLP): To process complex site descriptions, we applied Latent Dirichlet Allocation (LDA) to extract 50 thematic topics. Furthermore, a Large Language Model (GPT-4-turbo) served as a semantic parser to harmonize terminology and classify sites by function and historical period.
- Machine Learning (ML): An ensemble model (SVM, Random Forest, and Gradient Boosting) was trained on combined spatial and textual features to predict heritage types. Additionally, a two-stage "Environmental Suitability Model" utilized topographic, hydrological, and land-use data to identify potential locations for undocumented sites.
Dataset Composition
The primary dataset comprises 25,705 cultural heritage sites collected from official government registries across 11 prefecture-level cities in Zhejiang Province. Each record includes WGS84 coordinates, descriptive text (historical background and significance), and categorical attributes such as protection level and historical period. To protect site confidentiality, coordinates were "jittered" with approximately 1km of Gaussian noise.
Key Insights: Cultural Fingerprinting and Settlement Logic
A significant finding of this study is the "Multimodal Advantage." Textual features (TF-IDF) emerged as the single most important predictor for site classification (34.2% importance), outperforming spatial coordinates. This suggests that the narrative "story" of a site is as analytically valuable as its geographic location.
Furthermore, the strong spatial clustering (Moran’s I = 0.847) indicates that heritage distribution is not merely an artifact of modern administrative boundaries. Instead, it reflects "deep-time" settlement logic—the long-term interaction between human populations, terrain, and resources. The framework resolved five distinct clusters representing unique "historical ecologies."
Results and Field Validation
The ensemble model achieved 89.3% accuracy in predicting heritage types (e.g., religious vs. industrial). Beyond classification, the model flagged 1,247 high-probability candidate locations for undocumented sites. In a pilot field survey of 25 predicted locations, 19 sites (76%) were confirmed to contain actual heritage, demonstrating the model's real-world utility for archaeological planning.
Limitations and Future Directions
While robust, the study relies on official registries, which may favor state-recognized monumental sites over vernacular heritage. Additionally, the current model often assigns a single "establishment" date, potentially flattening the "layered biographies" of sites that have been repurposed over centuries. Future work will integrate LiDAR and multispectral satellite imagery to refine site detection in low-visibility areas and incorporate crowdsourced oral histories to enrich the dataset.
Summary updated