中国东部海域浮游植物类群遥感反演研究

浮游植物类群遥感反演能够为全面认识浮游植物在海洋生态系统中的作用提供重要的数据资料。但由于复杂的水体光学特性，近海浮游植物类群遥感反演存在着巨大挑战。本研究以复杂光学二类水体—中国东部海域为研究区，通过使用3种建模方法，即波段组合法、基于奇异值分解的多元线性回归法、基于奇异值分解的XGBoost回归法，利用遥感反射率数据反演浮游植物类群。经原位实测数据集验证，基于奇异值分解的XGBoost回归法构建的8类浮游植物叶绿素a浓度反演模型的精度最高，其中硅藻、甲藻的叶绿素a浓度反演模型在验证集上的决定系数均大于0.7。相比之下，3种建模方法估算得到的绿藻、蓝藻和金藻的叶绿素a浓度精度较低（验证结果的决定系数小于0.45）。同时，研究评估了OLCI影像的3种大气校正方法（C2RCC、POLYMER、MUMM）在中国东部海域的适用性。结果显示，相对于其他两种大气校正算法，C2RCC在各波段有较好的表现（均方根误差小于0.004 8 sr?1）。将3种浮游植物类群反演模型应用到大气校正后的OLCI影像，验证结果显示，利用基于奇异值分解的多元线性回归法建立的硅藻叶绿素a浓度模型有较好的反演精度（决定系数为0.56）。

Remote sensing retrieval of phytoplankton group in the eastern China seas

Remote sensing retrieval of phytoplankton group can provide important data for a comprehensive understanding of the role of phytoplankton in marine ecosystem. However, due to the complex optical characteristics, there are still great challenges in the remote sensing retrieval of phytoplankton group in offshore waters. In this study, the eastern China seas region, a complex optical class II water body, is taken as the research area. By using three modeling methods, namely band combination method, multiple linear regression method based on singular value decomposition (SVD+MLR) and XGBoost regression method based on singular value decomposition (SVD+XGBoost), the phytoplankton group is retrieved from remote sensing reflectance (R_rs) data. Verified by the in-situ measured data set, the chlorophyll a (Chl a) concentration retrieval model of eight phytoplankton groups by SVD+XGBoost has the highest accuracy, and the determination coefficient (R2) of Chl a concentration inversion model of diatoms and dinoflagellates in the validation set is greater than 0.7. In contrast, the accuracy of Chl a concentration of chlorophytes, cyanobacteria and chrysophytes estimated by the three modeling methods is low (the R2 of the validation results is less than 0.45). At the same time, the applicability of three atmospheric correction methods of OLCI images (C2RCC, POLYMER and MUMM) in the eastern China seas is evaluated. The results show that compared with the other two atmospheric correction algorithms, C2RCC has better performance in each band (root mean square error is less than 0.0048 sr?1). Finally, the performance of the retrieval model on satellite images is verified by the in-situ data. The validation results show that the diatoms Chl a concentration model established by SVD+MLR has better accuracy (the R2 is 0.56), while the Chl a concentration inversion models of other phytoplankton groups have poor results.