1 No.4 Geological Brigade, Jiangsu Geological Bureau, Suzhou 215129, China
2 School of Rail Transportation, Soochow University, Suzhou 215131, China
3 Suzhou Institute of Archaeology, Suzhou 215005, China
Show less
文章历史+
Received
Published
Revised
2025-09-11
2026-08-28
2026-01-12
Online Date
2026-09-04
PDF (15110K)
摘要
太湖流域作为中国东南沿海地区的重要文化发源地,自新石器时代以来即孕育了丰富的人类活动遗存。以苏州市域为代表,综合利用遥感影像、地形数据及主要水文地貌因子,构建基于频率比(frequency ratio, FR)模型与最大熵(maximum entropy,MaxEnt)模型的考古遗址空间预测模型。为克服传统FR模型在因子加权方面的局限性,引入条件注意力机制构建改进型加权模型(AM_FR),对各环境因子敏感性进行量化赋权,并对其预测性能进行系统对比。结果表明:高程、地表起伏度和粗糙度是影响遗址分布的主要环境因子;AM_FR模型在训练与测试数据上的AUC(area under curve)值优于传统FR与MaxEnt模型,表现出更优的拟合度与预测能力。重分类与栅格叠加分析显示,研究区内高概率区域主要分布于地势微高、近水资源丰富的滨湖缓坡地带,符合“择水而居、因地制宜”的聚落选址逻辑,并在构建模型指引下发现了新的遗址。研究成果不仅验证了机器学习方法在考古预测建模中的适用性,也为提升区域考古调查效率与文化遗产保护提供了技术支撑与方法借鉴。
Abstract
The Taihu Lake Basin in southeastern China represents a crucial cradle of prehistoric civilization, having sustained extensive human occupation since the Neolithic period. An archaeological site predictive model was developed for the Suzhou area by integrating multisource geospatial data—including remote sensing imagery, topographic parameters, and hydro-geomorphic factors—within a combined frequency ratio (FR) and maximum entropy (MaxEnt) modeling framework. To address limitations in factor weighting inherent in conventional FR approaches, an innovative conditional attention mechanism was incorporated to construct an enhanced weighted model (AM_FR), which quantified the sensitivity of various environmental predictors and optimized their contributions. Model performance was systematically evaluated and compared using training and testing datasets. The results show that elevation, surface roughness, and relief intensity are the primary environmental determinants influencing site location. The AM_FR model achieves higher area under curve (AUC) values on both training and testing datasets compared to standard FR and MaxEnt models, demonstrating superior fitting accuracy and predictive capability. Spatial overlay and reclustering analysis reveal that high-probability areas are predominantly distributed on gentle slopes near lakeshores with moderate elevations and reliable water access, consistent with the settlement principle of “adapting to habitat and leveraging natural advantages”. Guided by the established model, a new site is discovered. The research results validate the effectiveness of machine learning methods in archaeological predictive modeling. The proposed AM_FR approach offers a scientifically sound methodology and technical framework for improving the efficiency of regional archaeological survey and cultural heritage conservation practices.
JI Jia-hui, LI Xue-wen, WANG Kai-jie, NIU Yu-long, WANG Yu-hai, TANG Qiang.
Prediction of Neolithic Sites Based on an Improved Frequency Ratio Model: Taihu Lake Basin[J].
Science Technology and Engineering, 2026, 26(24): 10286-10296 DOI:10.12404/j.issn.1671-1815.2506556
新石器时代是人类社会由狩猎采集向农耕定居转变的关键时期,聚落的空间分布特征在很大程度上反映了史前人类对自然环境的适应策略与文化发展路径[1-2]。考古学者长期以来致力于揭示史前聚落的选址机制及其与环境因子的耦合关系,期望以此深化对先民社会结构、生业形态与人地关系演变的理解[3]。近年来,伴随地理信息系统(geographic information system,GIS)与空间建模技术的快速发展,考古遗址预测模型(archaeological predictive models,APM)逐步应用于文化遗产保护与考古调查规划之中[4],成为提升调查效率与科学性的重要工具[5-6]。通过定量分析环境因子与已知遗址之间的空间关系,APM能够识别潜在遗址高适宜性区域[7],为未知遗址发现提供方向,也为研究史前聚落的分布机制奠定基础[8]。
在遗址预测研究方法方面,中外学者已发展出多种模型方法,如逻辑回归、变异系数分析、证据权重(weight of evidence,WoE)模型以及基于机器学习的最大熵(maximum entropy,MaxEnt)模型等[12]。其中,频率比(frequency ratio,FR)模型因其方法简便、逻辑清晰,适合小样本条件下的环境因子解释,已被广泛应用于地质灾害预测与考古选址研究中。然而,传统FR模型存在因子权重难以区分、分类阈值主观、空间依赖性缺失等问题,难以应对复杂地貌背景下的多因子交互关系。相比之下,MaxEnt模型是一种基于“存在数据”的机器学习方法,能够在缺乏负样本的条件下,通过环境约束下的最大熵分布估计遗址适宜性,其在生态学与考古学中的应用成果显示出良好的非线性拟合能力与泛化性能[13-14]。此外,随着集成学习思想的发展,学界亦尝试引入多模型融合策略(如XGBoost、逻辑回归叠加等)以进一步提升模型稳定性和预测精度[15]。
在最大熵模型中,为弥补缺乏真实“非遗址”数据的局限,需引入背景点模拟区域环境特征的总体分布。采用地理信息系统(geographic information system,GIS)平台中随机点生成工具,在研究区域范围内等概率抽取不小于已知遗址数量10倍的背景点,作为MaxEnt模型的负样本近似[33]。为降低潜在的空间自相关与信息泄露风险,背景点生成过程中设置缓冲区,以剔除已知遗址及其周边1 km范围内的区域。
MaxEnt模型在整体预测性能上略优于FR模型,在训练集上其接收者操作特征曲线下面积(area under curve,AUC)达到0.705,而FR模型为0.671,两者均处于良好水平,而基于注意力机制的频率比模型在训练集上的AUC值则达到了0.802,优于前二者。这表明模型在遗址与环境因子关系建模方面具备较强区分能力。进一步引入注意力机制后,空间图像更为平滑,边界识别更自然,说明注意力机制有助于增强预测的整体稳定性与适用性。
DaiXiangming. The phased changes of Chinese prehistoric society and the formation of early states[J]. Acta Archaeologica Sinica, 2020(3): 309-336.
[5]
ClarkA E, RanlettS, StinerM C. Domestic spaces as crucibles of Paleolithic culture: an archaeological perspective[J]. Journal of Human Evolution, 2022, 172: 103266.
CaoYaowen. A predictive model for pre-Qin period sites in the Conghua Liuxi River Basin based on Logistic regression analysis[J]. Geomatics & Spatial Information Technology, 2021, 44(5): 124-127, 131.
[8]
HodderI, OrtonC. Spatial analysis in archaeology[M]. Cambridge: Cambridge University Press, 1976.
[9]
SchifferM B. Advances in archaeological method and theory[M]. New York: Elsevier Academic Press, 1986: 397-452.
[10]
YaworskyP M, VernonK B, SpanglerJ D, et al. Advancing predictive modeling in archaeology: an evaluation of regression and machine learning methods on the Grand Staircase-Escalante National Monument[J]. PLoS One, 2020, 15(10): e0239424.
ZhangPeng, YangYishi, ZhangShanjia, et al. Settlement location selection and prediction in the Late Neolithic Age in the northeast of the Tibetan Plateau: a case study in Zhuanglang River Basin[J]. Quaternary Sciences, 2025, 45(1): 267-282.
YuLijun. Spatial analysis and remote sensing detection research of archaeological sites in the Taihu Lake Basin during the Neolithic period[D]. Beijing: University of Chinese Academy of Sciences (Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences), 2018.
[15]
YangL, ZhaoY, YuanW, et al. GIS-based analysis of the regional typologyof neolithic archaeological cultures in the Taihu Lake region of China[J]. Land, 2024, 13(2): 244.
[16]
ZhangH, ChengH, SinhaA, et al. Collapse of the Liangzhu and other Neolithic cultures in the lower Yangtze region in response to climate change[J]. Science Advances, 2021, 7(48): eabl7287.
[17]
NovielloM, CafarelliB, CalculliC, et al. Investigating the distribution of archaeological sites: multiparametric vs probability models and potentials for remote sensing data[J]. Applied Geography, 2018, 95: 34-44.
[18]
ZhangY C, HeJie. Maxent modeling for distribution of rock art landscape in Tongtian River Basin[J]. Landscape Architecture, 2021, 28(11): 44-50.
WuRuichan, GanShu, LiQiangqiang. Predictive distribution research of Neolithic sites in Yulin area based on MaxEnt model[J]. Software, 2019, 40(9): 148-152.
[21]
WachtelI, ZidonR, GartiS, et al. Predictive modeling for archaeological site locations: comparing logistic regression and maximal entropy in north Israel and northeast China[J]. Journal of Archaeological Science, 2018, 92: 28-36.
[22]
ShaoB, MonteithF, YouZ, et al. The role of environmental factors in the spatiotemporal distribution of millet in late Neolithic to Bronze Ages sites in the Tibetan Plateau and surrounding regions[J]. Journal of Archaeological Science, 2024, 166: 105976.
[23]
WangY, ShiX, OguchiT. Archaeological predictive modeling using machine learning and statistical methods for Japan and China[J]. ISPRSInternational Journal of Geo-Information, 2023, 12(6): 244.
[24]
NaT, KawamuraY, KangS, et al. Hazard mapping of ground subsidence in east area of Sapporo using frequency ratio model and GIS[J]. Geomatics, Natural Hazards and Risk, 2021, 12(1): 347-362.
[25]
NicuI, Mihu-PintilieA, WilliamsonJ. GIS-based and statistical approaches in archaeological predictive modelling (NE Romania)[J]. Sustainability, 2019, 11(21): 5969.
ChuYongbin, ZhuLidong, YangWenguang, et al. Method for extracting area-elevation contour curves using ArcGIS and Excel[J]. Science Technology and Engineering, 2021, 21(36): 15372-15377.
ZhangChenjia, ZhuLei, YuLu. A survey of attention mechanisms in convolutional neural networks[J]. Computer Engineering and Applications, 2021, 57(20): 64-72.
[30]
WarrenR, AschD. A predictive model of archaeological site location in the eastern prairie peninsula[M]. Boca Raton, FL: CRC Press, 1999: 5-32.
[31]
DormannC F, ElithJ, BacherS, et al. Collinearity: a review of methods to deal with it and a simulation study evaluating their performance[J]. Ecography, 2013, 36(1): 27-46.
[32]
WangJ, ChenX, ZhuX H, et al. Taihu Lake, lower Yangtze drainage basin: evolution, sedimentation rate and the sea level[J]. Geomorphology, 2001, 41(2): 183-193.
[33]
ChenX, LiuL. The archaeology of China: from the late Paleolithic to the early Bronze Age[M]. Cambridge: Cambridge University Press, 2012.
LanTingcheng, ZhaoDachuan. A study on pig domestication in the Qiantang River Basin during the Neolithic Age: Majiabang Culture, Songze Culture, and Liangzhu Culture[J]. Swine Industry Science, 2019, 36(5): 126-128.
WuJingda, GuoLixin. Site selection and subsistence of ancient people in the Taihu Lake Basin during the Neolithic period: focusing on Majiabang Culture and Songze Culture[J]. Agricultural Archaeology, 2024(3): 5-12.
[38]
JiangL, LiuL. New evidence for the origins of sedentism and rice domestication in the lower Yangzi River, China[J]. Antiquity, 2006, 80(308): 355-361.
[39]
WuY, WangC, ZhangZ, et al. Subsistence, environment, and society in the Taihu Lake area during the Neolithic era from a dietary perspective[J]. Land, 2022, 11(8): 1229.
[40]
LiX. Determination of the temporal-spatial distribution patterns of ancient heritage sites in China and their influencing factors via GIS[J]. Heritage Science, 2024, 12(1): 143.
ZhangXiaojing, YangZaiqiang, WangXuelin, et al. Evaluation of climate comfort in the Jiangsu-Zhejiang-Shanghai region of the middle and lower reaches of the Yangtze River[J]. Science Technology and Engineering, 2014, 14(27): 122-127.
TianJie, ZhuYouchen, LiLinzhi, et al. Predictive modeling of Neolithic archaeological sites in Shandong Province based on GIS and XGBoost algorithm[J]. Journal of Beijing Normal University (Natural Science Edition), 2025, 61(3): 394-404.
[45]
KempfM. The application of GIS and satellite imagery in archaeological land-use reconstruction: a predictive model?[J]. Journal of Archaeological Science: Reports, 2019, 25: 116-128.
[46]
Mesa-MingoranceJ L, Ariza-LópezF J. Accuracy assessment of digital elevation models (DEMs): a critical review of practices of the past three decades[J]. Remote Sensing, 2020, 12(16): 2630.
[47]
PhillipsS J, AndersonR P, DudíkM, et al. Opening the black box: an open-source release of MaxEnt[J]. Ecography, 2017, 40(7): 887-893.
[48]
NogueiraS, SechidisK, BrownG. On the stability of feature selection algorithms[J]. Journal of Machine Learning Research, 2018, 18: 1-54.
SunBaoyan, YangZhengyang, ChenKuan, et al. Application of integrated aerial and ground photograph 3D reconstruction technology in archaeology[J]. Science Technology and Engineering, 2019, 19(17): 262-266.
[51]
PradhanB, LeeS. Delineation of landslide hazard areas on Penang Island, Malaysia, by using frequency ratio, logistic regression, and artificial neural network models[J]. Environmental Earth Sciences, 2010, 60(5): 1037-1054.
[52]
BuiD T, TuanT A, KlempeH, et al. Spatial prediction models for shallow landslide hazards: a comparative assessment of the efficacy of support vector machines, artificial neural networks, kernel logistic regression, and logistic model tree[J]. Landslides, 2016, 13(2): 361-378.
[53]
BengioY, SerdyukD, ChorowskiJ, et al. Attention-based models for speech recognition[J]. Neural Information Processing Systems, 2015, 28: 577-585.
[54]
JanochaK, CzarneckiW M. On loss functions for deep neural networks in classification[J]. Schedae Informaticae, 2017, 25: 49-59.
[55]
PhillipsS J, AndersonR P, SchapireR E. Maximum entropy modeling of species geographic distributions[J]. Ecological Modelling, 2006, 190(3): 231-259.
[56]
LiZ, LiuY, ZengH. Application of the MaxEnt model in improving the accuracy of ecological red line identification: a case study of Zhanjiang, China[J]. Ecological Indicators, 2022, 137: 108767.
QinLing, FuDaolian, ZhangHai. A study on the wild food resource catchments of early agricultural settlements: cases from the Lower Yangtze River and the Central Plains[J]. Quaternary Sciences, 2010, 30(2): 245-261.
[59]
LiG, DongJ, CheM, et al. GIS and machine learning models target dynamic settlement patterns and their driving mechanisms from the Neolithic to Bronze Age in the northeastern Tibetan Plateau[J]. Remote Sensing, 2024, 16(8): 1454.
[60]
ZhangP, YangY, ZhangS, et al. Spatial predictive modeling of Late Neolithic settlements in the northeastern Tibetan Plateau: a case study in Zhuanglang River basin[J]. Quaternary Sciences, 2025, 45(1): 267-282.
JinZhixin, SunHaohang, LiLin, et al. Detection of moving targets in sports scenes using multi-attention mechanism fusion[J]. Science Technology and Engineering, 2025, 25(29): 12570-12578.
BiShuoben, JiHan, YangHongru. Clustering analysis of the neolithic settlement sites in Zhengzhou-Luoyang area based on DBSCAN[J]. Science Technology and Engineering, 2014, 14(32): 266-270.