基于梯度提升机的喀斯特高原湖泊叶绿素a浓度遥感反演研究（英文）

曹卫堂; 周忠发; 孔杰; 王艳碧; 解茹凯

doi:10.13928/j.cnki.wrahe.2026.02.003

水利水电技术（中英文） ›› 2026, Vol. 57 ›› Issue (2) : 32 -53. DOI: 10.13928/j.cnki.wrahe.2026.02.003

水环境治理与水生态修复专栏

基于梯度提升机的喀斯特高原湖泊叶绿素a浓度遥感反演研究（英文）

曹卫堂 ¹^,² ,
周忠发 ¹^,²^,³^,⁴ ,
孔杰 ¹^,² ,
王艳碧 ¹^,² ,
解茹凯 ¹^,²

作者信息 +

Remote sensing inversion of chlorophyll-a concentration in karst plateau lakes based on gradient boosting machine

Weitang CAO ¹^,² ,
Zhongfa ZHOU ¹^,²^,³^,⁴ ,
Jie KONG ¹^,² ,
Yanbi WANG ¹^,² ,
Rukai XIE ¹^,²

Author information +

文章历史 +

PDF (1208K)

摘要

【目的】针对喀斯特高原湖泊因高pH(平均值> 8.2)、高悬浮颗粒物(SPM>50 mg/L)及季节性水文波动导致传统遥感模型面临光谱信号混合与非线性关系拟合不足的问题，以典型喀斯特高原湖泊平寨水库为对象，旨在实现该类水体叶绿素a(Chla)浓度的高精度遥感反演，【方法】研究采用Sentinel-2 MSI Level 2A影像(空间分辨率为10~60 m)及40个野外采样点数据，提出“短波敏感单波段+跨波段线性组合”特征工程策略，筛选出高敏感性光谱特征(包括单波段B3、B1、B2、B5、B4及线性组合B1+B3、B2+B3、B3+B5等),并利用梯度提升机(GBM)模型对非线性关系的高效拟合能力构建Chla浓度遥感反演框架，通过数据预处理与超参数优化提升模型对光谱特征与Chla浓度间非线性关系的拟合能力，【结果】结果显示，构建的GBM模型反演精度达决定系数(R²)=0.908、均方根误差(RMSE)=0.731μg/L、平均绝对误差(MAE)=0.529μg/L,较传统单波段线性模型(B3波段，R²=0.560 7)精度提升62%,平寨水库Chla浓度呈显著季节特征，夏季平均值为10.22μg/L,冬季为2.46μg/L,春季和秋季分别为6.01μg/L和5.88μg/L,其变化主要受水温(相关系数r=0.730)和总有机碳(TOC,相关系数r=0.783)驱动，且高pH环境下总氮生物可利用性的负反馈机制体现了喀斯特水体的特殊性，【结论】为喀斯特高原湖泊Chla浓度高精度遥感监测提供“敏感波段组合+机器学习”的技术方案，同时为水库水质管理与生态保护提供科学支撑。

Abstract

[Objective] For karst plateau lakes,traditional remote sensing models face challenges of spectral signal mixing and insufficient fitting of nonlinear relationships due to high p H values (average>8.2),high suspended particulate matter (SPM>50 mg/L),and seasonal hydrological fluctuations.The Pingzhai Reservoir,a typical karst plateau lake,is taken as the study area,and the aim is to achieve high-precision remote sensing inversion of chlorophyll-a (Chla) concentration in such water bodies. [Methods] Sentinel-2 MSI Level 2A imagery (with a spatial resolution of 10~60 m) and data from 40 field sampling points were used.A “shortwave-sensitive single-band+cross-band linear combination” feature engineering strategy was proposed to screen highly sensitive spectral features (including single bands B3,B1,B2,B5,B4 and linear combinations such as B1+B3,B2+B3,B3+B5).Additionally,a remote sensing inversion framework for Chla concentration was constructed by leveraging the efficient fitting capability of the Gradient Boosting Machine (GBM) model for nonlinear relationships.The fitting capability of the model for the nonlinear relationship between spectral features and Chla concentration was enhanced through data preprocessing and hyperparameter optimization. [Results] The result showed that the constructed GBM model achieved an inversion accuracy with a coefficient of determination (R²) of 0.908,root mean square error (RMSE) of 0.731μg/L,and mean absolute error(MAE) of 0.529μg/L,representing a 62% improvement in accuracy compared to the traditional single-band linear model (B3 band,R²=0.560 7).The Chla concentration in Pingzhai Reservoir showed significant seasonal characteristics,with average values of 10.22μg/L in summer,2.46μg/L in winter,6.01μg/L in spring,and 5.88μg/L in autumn.Its variation was primarily driven by water temperature (correlation coefficient r=0.730) and total organic carbon (TOC,correlation coefficient r=0.783),and the negative feedback mechanism of total nitrogen bioavailability in high p H environments reflected the distinctive characteristics of karst water bodies. [Conclusion] The findings provide a technical solution of“sensitive band combination+machine learning”for high-precision remote sensing monitoring of Chla concentration in karst plateau lakes,while also offering scientific support for reservoir water quality management and ecological protection.

关键词

喀斯特高原湖泊 / 叶绿素a浓度 / 遥感反演 / 哨兵-2号影像 / 梯度提升机模型 / 影响因素

Key words

karst plateau lakes / chlorophyll-a concentration / remote sensing inversion / Sentinel-2 imagery / gradient boosting machine model / influencing factors