基于超参数优化CatBoost算法的河流悬浮物浓度遥感反演

悬浮物浓度（TSM）是水生态环境评价的重要参数之一,及时掌握河流悬浮物浓度动态变化信息对于内陆水质监测、水环境治理是十分必要的。本研究基于野外实测光谱和悬浮物浓度数据,筛选与悬浮物浓度高度相关的波段组合反射率作为自变量,基于CatBoost、随机森林和多元线性回归算法构建悬浮物浓度遥感反演模型,采用带交叉验证的网格搜索法分别对CatBoost和随机森林2种机器学习模型进行超参数调优,确定模型最优参数配置,并对比不同模型反演精度,确定最优模型。基于最优模型,利用2019—2020年多时相Sentinel-2 MSI遥感影像,反演闽江下游悬浮物浓度,并分析其时空变化特征。结果表明：① b4/b3、(b6-b3)/(b6+b3)、(b4+b8)/b3、(1/b3-1/b4)×b5是MSI反演闽江下游TSM浓度的最佳波段组合反射率; ② 对比其他2种模型,基于超参数优化的CatBoost算法建立的悬浮物反演模型精度最高,其决定系数R²为0.95,均方根误差RMSE和平均绝对百分比误差MAPE分别为15.32 mg/L和19.68%; ③ 2019—2020年闽江下游悬浮物浓度分布“西低东高”,白沙至琅岐入海口呈升高趋势;④ 悬浮物浓度夏季最高,冬季和秋季次之,春季最低。本研究可为闽江下游悬浮物浓度监测及时空变化分析提供一种有效的技术手段和理论参考。

Retrieving Suspended Matter Concentration in Rivers based on Hyperparameter Optimized CatBoost Algorithm

Total Suspended Matter (TSM) is one of the significant parameters of aquatic ecological environment assessment. It is necessary to grasp the dynamic change information of river suspended solids concentration in time for inland water quality monitoring and water environment management. This paper is based on field measured spectra and suspended matter concentration data, the band combination reflectance that is highly correlated with the concentration of suspended solids is selected as the independent variable. The remote sensing inversion model of suspended solids concentration is constructed based on CatBoost, random forest, and multiple linear regression algorithms. In order to determine the optimal parameter configuration for the models, the grid search method with cross-validation is used for hyperparameter tuning of two machine learning models, i.e., CatBoost and Random Forest, respectively. And the inversion accuracy of different models is compared to determine the optimal model. Based on the optimal model, multi-temporal Sentinel-2 MSI remote sensing images from 2019 to 2020 are used to invert suspended matter concentrations in the lower reaches of the Minjiang River and analyse their spatial and temporal variation characteristics. The results indicate that: ① b4/b3, (b6-b3)/(b6+b3), (b4+b8)/b3, (1/b3-1/b4)×b5 are the best band combination reflectance for MSI inversion of TSM concentrations in the lower Minjiang River; ② Compared with the other two models, the suspended matter concentrations inversion model based on CatBoost algorithm with hyperparameter optimized has the highest accuracy, with a coefficient of determination R² of 0.95, Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) of 15.32 mg/L and 19.68%, respectively; ③ The distribution of suspended matter concentrations in the lower reaches of the Minjiang River from 2019 to 2020 is "low in the west and high in the east", with a rising trend from Baisha to the mouth of the Langqi inlet; ④ The suspended matter concentration is highest in summer, followed by winter and autumn, and lowest in spring. This study provides an effective technical means and theoretical reference for the monitoring and spatio-temporal variation analysis of suspended matter concentration in the lower reaches of Minjiang River.