赵宗泽, 周钰瑶, 陈时雨.夜光遥感城市等级分类方法与鲁棒性特征分析[J].河南理工大学学报（自然科学版）,2025,44(6):156-164.

ZHAO Z Z, ZHOU Y Y, CHEN S Y. Classification methods for urban hierarchy and analysis of robust features using nighttime light remote sensing[J].Journal of Henan Polytechnic University(Natural Science) ,2025,44(6):156-164.

夜光遥感城市等级分类方法与鲁棒性特征分析

赵宗泽1, 周钰瑶1, 陈时雨2

1.河南理工大学 测绘与国土信息工程学院，河南 焦作  454000；2.信阳师范大学 地理科学学院，河南 信阳  464000

摘要: 目的 城市等级分类与城市规划和城市可持续发展密切相关。为评估机器学习分类方法在城市等级分类方面的适用性，分析夜间灯光遥感数据中可用于城市等级分类的特征的鲁棒性。  方法 利用2021年度NPP-VIIRS（national polar-orbiting partnership - visible infrared imaging radiometer）夜间灯光遥感影像提取中国297个城市的12个统计特征，基于机器学习方法对城市进行等级识别分类，并使用SHAP（shapley additive exPlanations）对模型解释归因、识别鲁棒性特征。 结果 结果表明，每种机器学习方法的ROC（receiver operating characteristic）曲线下面积AUC（area under the curve）均在0.90以上，表现出良好的分类性能，其中随机森林的分类性能最好，准确率为0.75，F₁值为0.764 6；针对不同等级城市，机器学习方法的识别能力不同，三、四线城市的分类精度明显低于一、二、五线城市，使用XGBoost与随机森林集成的组合模型方法，总体准确率达到了0.783 3，F₁值为0.805 1；不同模型中每个特征的重要性不同，但鲁棒的夜间灯光特征仍然是一致的。 结论 研究表明，基于夜间灯光遥感影像利用机器学习分类方法能够较好地识别城市等级，存在具有鲁棒性的夜光特征变量，降低了城市等级分类复杂度、提高了分类效率。

关键词:夜光遥感;机器学习;城市等级分类;特征重要性;精度评价

doi:10.16186/j.cnki.1673-9787.2023120025

基金项目:国家自然科学基金资助项目（U22A20566，42071405）

收稿日期:2023/12/11

修回日期:2024/02/12

出版日期:2025/10/14

Classification methods for urban hierarchy and analysis of robust features using nighttime light remote sensing

Zhao Zongze1, Zhou Yuyao1, Chen Shiyu2

1.School of Surveying and Land Information Engineering， Henan Polytechnic University， Jiaozuo  454000， Henan， China；2.School of Geographical Sciences， Xinyang Normal University， Xinyang  464000， Henan， China

Abstract: Objectives Urban hierarchy classification levels is closely related to urban planning and sustainable development. This study aims to evaluate the applicability of machine learning classification methods for urban hierarchy classification and analyze the robustness of nighttime light remote sensing features.  Methods Based on the 2021 NPP-VIIRS （National Polar-orbiting Partnership-Visible Infrared Imaging Radiometer） nighttime light remote sensing imagery， 12 statistical features were extracted from 297 cities in China. Machine learning methods were employed for urban hierarchy recognition and classification， and the SHAP （Shapley Additive exPlaneations） method was used for the model interpretation and robust feature identification. Results The results show that all machine learning methods achieved AUC values above 0.90 under ROC （Receiver Operating Characteristic） curves， demonstrating excellent classification performance. Among them， random forest exhibited the best performance. with an accuracy of 0.75 and an F₁ value of 0.764 6. The recognition ability of machine learning methods varied across different city tiers： The classification accuracy for third-tier and fourth-tier cities was significantly lower than that of first-tier， second-tier， and fifth-tier cities. Classification accuracy for third- and fourth-tier cities was significantly lower than for first-， second-， and fifth-tier cities. The ensemble model combining XGBoost and random forest achieved an overall accuracy of 0.783 3， with an F₁ value of 0.805 1. While feature importance varied across different models， the robust nighttime light features remained consistent.  Conclusions This study demonstrates that machine learning classification methods based on nighttime light remote sensing imagery can effectively identify urban hierarchies. The identification of robust nighttime light feature variables reduces the complexity of urban hierarchy classification and improves classification efficiency. This approach provides a scientific basis for urban classification.

Key words: nighttime light remote sensing; machine learning; urban hierarchy classification; feature importance; classification accuracy