Archaeological Predictive Modeling Using Machine Learning and Statistical Methods for Japan and China

Yuan Wang, Xiaodan Shi, Takashi OguchiIsprs Int. J. geo inf 2023 · DOI

TL;DR

This paper introduces a novel hybrid archaeological predictive modeling (APM) approach, AM_FR, that combines deep learning with a frequency ratio model. This method aims to improve the accuracy of predicting archaeological site locations by better capturing the complex relationships between environmental factors and human settlement patterns, outperforming traditional methods in their study areas of Japan and China.

What Problem Does This Paper Address?

Archaeological Predictive Modeling (APM) is crucial for understanding past human behavior, managing cultural heritage, and guiding archaeological fieldwork. However, existing APM methods often struggle to accurately represent the intricate interplay between various environmental factors and the spatial distribution of archaeological sites. Many models focus on a limited set of topographic features or overlook the importance of how these factors interact. This study addresses the need for more sophisticated APM techniques that can effectively integrate diverse environmental data and learn complex relationships to provide more reliable predictions, particularly in regions with distinct geographical characteristics like Japan and China.

Methods

The authors propose a novel AM_FR (Attention Mechanism - Frequency Ratio) model. This hybrid approach leverages:

  • Frequency Ratio (FR): A statistical method to quantify the relationship between individual environmental factors (e.g., elevation, slope) and the presence of archaeological sites. It calculates the ratio of site density within a specific factor class to the overall site density.
  • Conditional Attention Mechanism (AM): A deep learning component that learns the relative importance (weights) of different input factors and their interactions for predicting archaeological site locations. This mechanism allows the model to focus on the most relevant environmental features, mimicking how humans might prioritize certain environmental cues.

The AM_FR model integrates these by using the attention mechanism to determine weights for each factor, which are then applied to the FR values. This weighted combination is used to calculate a site probability (SP). The study also compares this novel approach against two established methods:

  • Frequency Ratio (FR) model: A standalone statistical approach.
  • Maximum Entropy (MaxEnt): A widely used machine learning model for presence-only data.

Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC) and Kvamme's Gain.

Datasets

The study utilized archaeological site data from two distinct regions:

  • Japan: Archaeological site data (n=1367) from the Kofun database, focusing on burial tombs from the Kofun period (3rd-6th centuries CE).
  • Shaanxi Province, China: Archaeological site data (n=200) from the Chinese Archaeology website, encompassing monuments and tombs spanning over 2000 years (1046 BC–907 AD).

Environmental data included:

  • Topographic Factors: Elevation, slope, roughness, relief degree of the land surface (RDLS), plan curvature, profile curvature, and cutting depth. These were derived from Digital Elevation Models (DEMs) for Japan (10m resolution) and China (30m resolution).
  • Hydrological Factor: Distance from major rivers, obtained from national hydrological databases.

Key Insights and Ideas

  • Environmental Factors are Key, but Their Interactions Matter: The study reinforces that topographic and hydrological factors are critical for predicting archaeological site locations. However, the novel AM_FR model highlights that understanding how these factors interact (learned by the attention mechanism) is crucial for improved predictive accuracy.
  • Topography Shapes Human Activity: The research strongly suggests that ancient human settlement preferences were heavily influenced by relatively flat, gentle terrain. This is attributed to the needs for transportation, agriculture, and easier construction of settlements, especially for the earthen tombs prevalent in Japan.
  • Regional Differences in Factor Importance: While general trends exist, the specific environmental factors most influential for site prediction differ between Japan and China. This underscores the need for localized APM approaches. For instance, RDLS and cutting depth were highly important in Japan, while slope and elevation were more dominant in Shaanxi.
  • Deep Learning Enhances Traditional Methods: The integration of a deep learning attention mechanism with the statistical FR model proved more effective than using either method in isolation, suggesting a powerful synergy for complex spatial prediction tasks.

Key Results

  • AM_FR Superiority: The AM_FR model consistently outperformed both the standalone FR model and the MaxEnt model in terms of AUC values for predicting archaeological sites in both Japan and Shaanxi, China. The mean AUC for AM_FR was 0.901 (Japan) and 0.816 (Shaanxi), compared to MaxEnt's 0.872 and 0.785, and FR's 0.851 and 0.791, respectively.
  • Kvamme's Gain Nuance: While AM_FR showed higher overall accuracy (AUC), the MaxEnt model demonstrated a higher efficiency in identifying "very high probability" regions, as indicated by Kvamme's Gain values (0.92 for Japan, 0.89 for China with MaxEnt vs. 0.78 and 0.84 for AM_FR). However, this came at the cost of classifying more areas as moderate or low probability, potentially missing some sites.
  • Factor Importance: In Japan, RDLS, cutting depth, and roughness were the most significant factors. In Shaanxi, slope, RDLS, and elevation were most influential. All these factors generally showed negative correlations with site occurrence, favoring flatter terrains.

Limitations and Future Work

  • Site Type Specificity: The authors acknowledge that their models, focused on tombs and monuments, might not be as effective for predicting other site types like villages or hunting camps, which have different locational requirements.
  • Marginal Areas: The predictive maps for areas like Hokkaido in Japan showed high probability zones, which the authors attribute to the models considering only natural factors and not socio-cultural aspects that historically influenced settlement patterns in these regions.
  • Future Work: The study suggests evaluating the AM_FR model's effectiveness on other site types and in different geographical contexts. Further research could also incorporate socio-cultural factors to refine predictions, especially in marginal or culturally distinct areas.

Summary updated

BibTeX
@article{wang2023archaeological,
  author = {Yuan Wang and Xiaodan Shi and Takashi Oguchi},
  title = {{Archaeological Predictive Modeling Using Machine Learning and Statistical Methods for Japan and China}},
  journal = {Isprs Int. J. geo inf},
  year = {2023},
  doi = {10.3390/ijgi12060238},
}
Figure