Abstract: Satellite remote sensing technology, leveraging its high spatial resolution and short revisit cycle, offers an effective approach for water quality monitoring. This study constructed a machine learning model based on Extreme Gradient Boosting (XGBoost) researching on reservoirs in Gu'an Hot Spring Park. Using domestic remote sensing images from the Gaofen (GF), Ziyuan (ZY), Beijing (BJ), and Jilin (JL) satellite series, the model retrieved key water quality parameters including chlorophyll-a, pH, ammonia nitrogen, and dissolved oxygen. The reflectance was derived from remote sensing images after preprocessing steps such as radiometric calibration and atmospheric correction, while synchronized measured data were obtained from buoy-based online monitoring stations. Band combinations with high correlation (coefficient>0.6) selected through Pearson correlation analysis were used as model inputs. The dataset was split into training, validation, and test sets in a 7:1:2 ratio, and model performance was evaluated using R², RMSE, and MAPE. The results demonstrated that the XGBoost model outperformed comparative models—Random Forest, AdaBoost, Deep Neural Networks, Convolutional Neural Networks, and Ordinary Linear Regression—in retrieving all four water quality parameters, exhibiting excellent fitting capability on the training set and strong generalization ability on the test set. High-spatial-resolution water quality distribution maps generated by the model revealed a spatial pattern of “poor at the edges and good in the center” within the study area, with water quality in the southern reservoir being superior to that in the northern reservoir. This research provides effective technical support for precise monitoring and management of water quality in small and medium-sized water bodies.

Key words: remote sensing images, water quality monitoring, XGBoost model, inversion, spatial distribution

摘要： 卫星遥感技术凭借高空间分辨率和低重访周期,为水质监测提供了有效途径。本研究以固安县温泉园区调蓄水池为研究对象,构建了基于极端梯度提升树(XGBoost)的机器学习模型。研究采用国产高分、资源、北京、吉林系列卫星的遥感影像,反演了叶绿素a、pH值、氨氮、溶解氧等关键水质参数。遥感影像经辐射定标、大气校正等预处理后计算反射率,实测数据来源于浮标在线监测站获取的同步水质参数。通过皮尔逊相关分析筛选高相关性波段组合(相关系数>0.6)作为模型输入,将数据集按7:1:2划分为训练集、验证集和测试集,采用R²,RMSE,MAPE评估模型性能。结果表明,与随机森林、AdaBoost、深度神经网络、卷积神经网络及普通线性回归模型相比,XGBoost模型在4项水质参数的反演中均表现出最优的性能,其在训练集上具有卓越的拟合能力,在测试集上展现了强大的泛化能力。基于该模型生成的高空间分辨率水质分布图揭示,研究区域水质参数空间分布呈现“边缘差、中央好”特征,且南库水质优于北库。该研究为中小型水体水质的精准监测与管理提供了有效技术支撑。

关键词: 遥感影像, 水质监测, XGBoost模型, 反演, 空间分布