地学前缘 ›› 2026, Vol. 33 ›› Issue (1): 511-522.DOI: 10.13745/j.esf.sf.2025.10.6

• 水文地质新技术新方法 • 上一篇    下一篇

融合数值模拟和机器学习的民勤盆地地下水潜力评价与主控因素识别

周斐然1(), 尹子悦1, 孙晓敏2, 宋健3, 杨蕴3, 吴剑锋1,*()   

  1. 1.南京大学 地球科学与工程学院, 江苏 南京 210023
    2.南京水利科学研究院, 江苏 南京 210029
    3.河海大学 地球科学与工程学院, 江苏 南京 211100
  • 收稿日期:2025-06-20 修回日期:2025-10-20 出版日期:2026-01-25 发布日期:2025-11-10
  • 通信作者: *吴剑锋(1971—),男,博士,教授,博士生导师,主要从事地下水数值模拟与优化研究。E-mail:jfwu@nju.edu.cn
  • 作者简介:周斐然(2001—),女,硕士研究生,主要从事地下水数值模拟研究。E-mail:feiranzhou@smail.nju.edu.cn
  • 基金资助:
    国家自然科学基金项目(42272277);博士后创新人才支持计划(BX20240165);江苏省卓越博士后计划(2024ZB125)

Integrating numerical simulation and machine learning for identification of groundwater potential zone and its governing factors in the Minqin Basin, Northwest China

ZHOU Feiran1(), YIN Ziyue1, SUN Xiaomin2, SONG Jian3, YANG Yun3, WU Jianfeng1,*()   

  1. 1. School of Earth Sciences and Engineering, Nanjing University, Nanjing 210023, China
    2. Nanjing Hydraulic Research Institute, Nanjing 210029, China
    3. School of Earth Science and Engineering, Hohai University, Nanjing 211100, China
  • Received:2025-06-20 Revised:2025-10-20 Online:2026-01-25 Published:2025-11-10

摘要:

绿洲作为干旱区生态安全的核心载体,其稳定性高度依赖地下水资源的可持续性,地下水潜力区的精准识别能为干旱区水资源优化配置提供决策支撑。本研究将民勤盆地作为西北干旱区内陆河流域的典型绿洲系统,提出了一种融合数值模拟和机器学习的干旱区地下水潜力综合评价方法。通过地下水数值模拟获取高精度的地下水埋深空间分布以及渗透系数、给水度两个关键参数,综合考虑气象、水文、土地利用、地形和地质5大类共18个影响因素,分别采用6种机器学习模型,系统评估了地下水潜力空间分布特征。研究结果表明:绿洲区地下水潜力呈现南高北低的空间格局,LightGBM模型(准确率为87.87%,F1分数为0.716,AUC为0.943)预测地下水潜力表现最优,XGBoost和随机森林次之,支持向量机、K近邻和BP神经网络的预测性能则相对较弱。在此基础上,通过随机森林、XGBoost和LightGBM 3个树模型计算特征重要性,结果显示地下水埋深(权重17.1%~18.5%)是影响民勤绿洲地下水潜力的关键主控因素,其次是潜在蒸散发(12.5%~14.2%)、大气降水(8.6%~12.5%)、NDVI(6.2%~12.8%)及地表高程(6.7%~11.4%)。本文提出的研究方法适用于干旱区的地下水潜力多参数评估体系,为干旱区绿洲地下水资源评价提供了科学依据。

关键词: 地下水潜力, 机器学习, 数值模拟, 民勤盆地, 干旱区

Abstract:

Oases are critical for maintaining ecological security in arid regions, and their stability is heavily dependent on sustainable groundwater resources. The delineation of groundwater potential zones (GPZs) offers decision-making support for optimal water resource allocation in arid areas. This study focuses on the Minqin Basin, a typical inland river oasis in northwest China, and proposes an integrated approach that combines numerical simulation with machine learning to assess groundwater potential in arid regions. Precise spatial distributions of groundwater depth and two key parameters (hydraulic conductivity and specific yield) were derived from groundwater numerical simulation. Considering 18 influencing factors across five categories (meteorology, hydrology, land use, topography, and geology), six machine learning models were employed to evaluate the spatial characteristics of groundwater potential. Results indicate that the LightGBM (Accuracy: 87.87%; F1 score: 0.716; AUC: 0.943) achieved the best performance, followed by XGBoost and Random Forest, whereas the Support Vector Machine, K-Nearest Neighbors, and BP Neural Network models demonstrated inferior performance. Subsequently, feature importance analysis using tree-based models (Random Forest, XGBoost, and LightGBM) revealed that groundwater depth (weight: 17.1%-18.5%) was the primary controlling factor. Secondary factors include potential evapotranspiration (12.5%-14.2%), precipitation (8.6%-12.5%), NDVI (6.2%-12.8%), and surface elevation (6.7%-11.4%). The proposed methodology establishes a multi-parameter assessment framework for groundwater potential in arid oasis areas, offering a scientific basis for sustainable groundwater management in water-scarce regions.

Key words: groundwater potential, machine learning, numerical simulation, Minqin Basin, arid regions

中图分类号: