地球信息科学学报 ›› 2023, Vol. 25 ›› Issue (10): 2026-2038.doi: 10.12082/dqxxkx.2023.230152

• 地理空间分析综合应用 • 上一篇    下一篇

基于随机森林模型的“网格-月”尺度武装冲突风险预测及影响因素分析——以中南半岛为例

杜树坤1(), 张晶1,2,*(), 韩志军1,2, 公茂玉1,2   

  1. 1.信息工程大学地理空间信息学院,郑州 450001
    2.智慧中原地理信息技术河南省协同创新中心,郑州 450001
  • 收稿日期:2023-03-27 修回日期:2023-06-02 出版日期:2023-10-25 发布日期:2023-09-22
  • 通讯作者: * 张晶(1974—),女,辽宁西丰人,博士,教授,主要从事人文地理学理论与应用研究。E-mail: gybzj@163.com
  • 作者简介:杜树坤(1998—),男,山东东营人,硕士生,主要从事人文地理学与地理大数据研究。E-mail: a412309072@sina.com
  • 基金资助:
    国家自然科学基金项目(41301125);国家社科基金重大项目(20&ZD138)

Armed Conflict Risk Prediction and Influencing Factors Analysis Based on the Random Forest Model at the Grid-month Scale: A Case Study of Indochina Peninsula

DU Shukun1(), ZHANG Jing1,2,*(), HAN Zhijun1,2, GONG Maoyu1,2   

  1. 1. College of Geospatial Information, Information Engineering University, Zhengzhou 450001, China
    2. Collaborative Innovation Center of Geo-information Technology for Smart Central Plains, Zhengzhou 450001, China
  • Received:2023-03-27 Revised:2023-06-02 Online:2023-10-25 Published:2023-09-22
  • Contact: * ZHANG Jing, E-mail: gybzj@163.com
  • Supported by:
    National Natural Science Foundation of China(41301125);National Social Science Foundation of China(20&ZD138)

摘要:

掌握周边地区武装冲突风险形势对我国“一带一路”倡议推进和海外投资建设具有十分重要的意义。由于武装冲突风险涉及的因素众多,很多数据时空精度有限,以往研究的尺度大多集中于“国家-年”层面,未能从次国家尺度上预测武装冲突风险。通过将武装冲突与政治、经济、社会和地理等专题的多源数据匹配到统一的“网格-月”尺度的时空框架中,构建了多个基于随机森林模型的武装冲突风险预测模型,以中南半岛为例,对比各专题模型和集成模型的预测精度,将预测结果与实际的武装冲突风险时空分布情况进行比较与分析,计算各影响因素的权重并分析其影响作用。研究结果表明:① 基于随机森林模型的冲突预测模型比传统的逻辑回归模型预测精度更高,其中集成模型的准确率、ROC曲线下面积和PR曲线下面积分别提高了0.017 7、0.436 2和0.171 2;② 中南半岛武装冲突风险受政治、经济和社会要素影响较高,地理要素的相关性较弱,但随着风险水平发生变化,影响因素的作用程度也在改变;③ 在基础专题数据的支撑下,顾及冲突的时空依赖性可以明显提高模型的预测精度;④ 与大尺度研究相比,“网格-月”尺度的冲突预测结果精度更高,可解释性也更强。本研究可为我国海外投资与当地冲突风险防控与治理等提供参考和依据。

关键词: 武装冲突, 随机森林模型, 风险预测, 影响因素, 中南半岛

Abstract:

A better understanding of the threat of armed conflict in a region is essential to advance the Belt and Road Initiative and overseas investment and construction. Most existing studies have concentrated on a "country-year" level and have limited accuracy in predicting armed conflict risk at the sub-national level, because the armed conflict risk involves numerous influencing factors and many data have limited spatiotemporal accuracy. In our study, we built several models based on random forest methods for armed conflict risk prediction by integrating multi-source armed conflict data with political, economic, social, and geographic thematic information into a unified spatiotemporal framework at "grid-month" scale. Taking the Indochinese Peninsula as an example, we compared the prediction accuracy of each thematic model and the integrated model for armed conflict risk. Then we compared the prediction results against the actual spatiotemporal distribution of armed conflicts, and the weights of each influencing factor were calculated and analyzed. The results show that: (1) compared to the traditional logistic regression model's performance, the accuracy, area under the ROC curve, and area under the PR curve of the integrated random forest model increased by 0.017 7, 0.436 2, and 0.171 2, respectively; (2) the political, economic, and social factors had a significant impact on the risk of armed conflict in the Indochina Peninsula, while geographic factors were less important. However, as the risk level changes, the degree of influence of these factors also changed; (3) the model prediction accuracy of armed conflict risk can be greatly increased by taking into account the spatiotemporal dependence of conflicts, which was supported by the underlying thematic data; (4) the conflict results predicted at the "grid-month" scale were more precise and interpretable compared to large-scale prediction results. This study provides a reference and basis for China's overseas investment as well as local conflict risk prevention, control, and governance.

Key words: armed conflict, Random Forest Model, risk prediction, influence factor, Indochina Peninsula