img

官方微信

高级检索

中国沙漠 ›› 2026, Vol. 46 ›› Issue (3): 175-186.DOI: 10.7522/j.issn.1000-694X.2026.00032

• • 上一篇    

基于XGBoost-SHAP方法的黄土高原生态质量演变及驱动因素分析

李晓鹏1(), 李康1, 徐静2, 贾富贵1, 雷双1   

  1. 1.兰州财经大学,农林经济管理学院 /,甘肃 兰州 730020
    2.兰州财经大学,会计学院,甘肃 兰州 730020
  • 收稿日期:2026-01-09 修回日期:2026-03-02 出版日期:2026-05-20 发布日期:2026-06-11
  • 作者简介:李晓鹏(1981—),女,甘肃天水人,博士,教授,研究方向为生态环境与区域发展。E-mail: lixiaopeng1396@163.com
  • 基金资助:
    兰州财经大学科研专项经费;兰州财经大学2025年度高等教育研究项目(LJY202513);兰州财经大学数字化赋能交叉融合课程

Eco-environment evolution and driving factors of the Loess Plateau based on the XGBoost-SHAP approach

Xiaopeng Li1(), Kang Li1, Jing Xu2, Fugui Jia1, Shuang Lei1   

  1. 1.College of Agricultural,Forestry and Economic Management /, Lanzhou University of Finance and Economics,Lanzhou 730020,China
    2.School of Accounting, Lanzhou University of Finance and Economics,Lanzhou 730020,China
  • Received:2026-01-09 Revised:2026-03-02 Online:2026-05-20 Published:2026-06-11

摘要:

为研究黄土高原生态质量的长期演变并揭示其驱动机制,本文以2000—2025年为研究期,基于遥感生态指数(RSEI)构建生态质量时空序列,开展等级结构与转移特征分析,并结合Sen趋势估计与Mann-Kendall检验识别变化方向及显著性;在驱动机制方面,选取2000、2010、2020年为代表年份,整合气候、地形、土地利用及人类活动等因子,构建XGBoost模型并引入SHAP实现贡献分解与非线性响应识别。结果表明:(1)黄土高原多年RSEI均值为0.449,生态质量空间格局呈东南高、西北低的梯度分异,等级结构以一般与较差为主体,高等级区域呈带状或斑块状集聚;(2)2000—2025年生态质量整体改善,RSEI年均值由0.395升至0.474,低等级面积占比由53.86%降至35.40%,等级结构向中高等级方向调整;(3)5 a尺度等级转移以稳定保持为主,稳定比例为65.99%~80.32%,转移以相邻等级间的渐进转换为主要形式,不同阶段改善与退化强度存在差异;(4)驱动因素中水分相关因子稳定主导,年降水量与水分亏缺贡献居前,关键因子存在阈值与分段响应特征,表现为坡度约在10°附近由弱负贡献转为正贡献、年降水量约在400 mm附近由抑制转为促进、水分亏缺约在50 mm附近由正向或近零贡献跃迁为稳定负贡献。研究结果可为黄土高原生态质量的长期监测评估、分区治理与生态工程优化提供定量依据与机制参考。

关键词: 黄土高原, RSEI, 时空演变, 等级转移, XGBoost, SHAP, 驱动因素

Abstract:

To characterize long-term changes in eco-environment quality on the Loess Plateau and to reveal the underlying drivers, we constructed a spatiotemporal series of eco-environment quality for 2000-2025 using the Remote Sensing Ecological Index (RSEI). Class structure and transitions were analyzed, and Sen's slope estimation and the Mann-Kendall test were applied to identify change direction and significance. For mechanism interpretation, three representative years (2000, 2010, and 2020) were selected. Climate, topography, land use, and human-activity factors were integrated to develop XGBoost models, and SHAP was employed to decompose contributions and identify nonlinear responses. The results show that: (1) The multi-year mean RSEI was 0.449, with a distinct southeast-northwest gradient (higher in the southeast and lower in the northwest); the class composition was dominated by the “moderate” and “poor” categories, while higher-quality classes exhibited banded or patchy clustering. (2) Eco-environment quality improved overall from 2000 to 2025, with the annual mean RSEI increasing from 0.395 to 0.474 and the areal proportion of low-quality classes decreasing from 53.86% to 35.40%, indicating a structural shift toward medium-to-high classes. (3) Five-year class transitions were dominated by persistence, with stable proportions of 65.99%-80.32%, and changes mainly occurred as gradual shifts between adjacent classes, although the relative strength of improvement and degradation varied among periods. (4) Water-related factors consistently played a dominant role, with annual precipitation and water deficit ranking highest in contribution, and key drivers exhibited threshold and segmented responses, including a transition of slope effects from weakly negative to positive at approximately 10°, a shift of precipitation effects from inhibiting to promoting at around 400 mm, and a shift of water-deficit effects from positive or near-zero to persistently negative at about 50 mm. These findings provide quantitative evidence and mechanistic insights to support long-term monitoring and assessment, zoned management, and optimization of ecological restoration projects on the Loess Plateau.

Key words: Loess Plateau, RSEI, spatiotemporal dynamics, class transition, XGBoost, SHAP, driving factors

中图分类号: