基于体散射函数及吸收系数的南海水体漫射衰减系数研究*
张现清(1998—), 女, 山东省济宁市人, 硕士研究生, 主要从事海洋光学研究。email: |
*感谢中国科学院南海海洋研究所海洋光学团队的所有人员, 他们对本次研究实验数据的获取作出了贡献, 感谢国家自然科学基金组织的南海北部调查航次。 |
Copy editor: 林强
收稿日期: 2022-06-27
修回日期: 2022-08-12
网络出版日期: 2022-08-24
基金资助
国家自然科学基金(41976181)
国家自然科学基金(41976172)
国家自然科学基金(41976170)
广州市科技计划重点项目(201707020023)
南方海洋科学与工程广东省实验室(广州)人才团队引进重大专项项目(GML2019ZD0305)
热带海洋环境国家重点实验室自主研究项目(LTOZZ1602)
Studying on diffuse attenuation coefficient in the South China Sea based on volume scattering function and absorption coefficient*
Copy editor: LIN Qiang
Received date: 2022-06-27
Revised date: 2022-08-12
Online published: 2022-08-24
Supported by
National Natural Science Foundation of China(41976181)
National Natural Science Foundation of China(41976172)
National Natural Science Foundation of China(41976170)
Science and Technology Planning Project of Guangzhou City China(201707020023)
Southern Marine Science and Engineering Guangdong Laboratory (Guangzhou)(GML2019ZD0305)
Open Project Program of the State Key Laboratory of Tropical Oceanography(LTOZZ1602)
漫射衰减系数Kd(z, λ)是估算水下光场及水色要素剖面分布、研究浮游植物光合作用及赤潮灾害预警方法的重要参数, 它是一个准固有光学特性参数, 是波长 和剖面深度z的函数, 除与水体吸收、散射或后向散射有关外, 对归一化的水体体散射函数即散射相函数的角度分布极为敏感。本文基于广角体散射函数测量仪(volume scattering and attenuation meter, VSAM)、吸收衰减系数测量仪ac-9和ac-s以及海洋光学剖面仪Profiler Ⅱ OCI/R-200I和HyperPro Ⅱ在南海海域实测数据, 利用LightGBM、随机森林(random forest, RF)、CatBoost三种高效机器学习方法, 首次构建了基于体散射函数β(ψ)、吸收系数a及对应剖面深度z的漫射衰减系数Kd(650)剖面分布估算模型, 并综合R2、RMSE、MAPE以及估算与实测数据的对比进行模型评价, 结果表明, 三种机器学习模型中, CatBoost模型的R2和RMSE分别为0.8534和0.0472m-1, 均优于RF和LightGBM; CatBoost模型的MAPE为11.0585%, 低于RF模型但略高于LightGBM模型; 通过对比估算和实测结果发现, CatBoost模型估算结果与实测结果最为相近, 是Kd(650)最优估算模型。利用CatBoost模型, 结合实测体散射函数β(ψ)、吸收系数a及其相应剖面深度z, 对南海北部多个站点15m以浅Kd(650)的剖面分布估算表明, 上述站点Kd(650)在5、10、15m三个水层变化范围为0.275~0.7m-1, 5m水层的Kd(650)较为平稳, 10与15m水层Kd(650)跨度较大。本研究方法考虑了多角度体散射函数分布对漫射衰减系数的贡献, 为基于固有光学特性参数估算Kd(z, λ)提供了新方法思路。
关键词: 漫射衰减系数Kd(650); 体散射函数; 吸收系数; 机器学习
张现清 , 李彩 , 周雯 , 刘聪 , 许占堂 , 曹文熙 , 杨跃忠 . 基于体散射函数及吸收系数的南海水体漫射衰减系数研究*[J]. 热带海洋学报, 2023 , 42(3) : 86 -95 . DOI: 10.11978/2022146
Diffuse attenuation coefficient of downwelling irradiance Kd(z, λ) is an important parameter for estimating the profile distribution of underwater light filed and water constituents, and studying the photosynthesis of the phytoplankton and warning method of harmful algae bloom. Kd(z, λ) is a “quasi-inherent” optical property as a function of wavelength and depth . Not only is it sensitive to absorption and scattering/backscattering coefficient, but also sensitive to the angular distribution of the normalized volume scattering function (i.e., scattering phase function). In this study, based on the volume scattering function [VSF, β(ψ, z)] in seven directions determined with a custom in situ device called VSAM (volume scattering and attenuation meter), the absorption coefficient a(z) determined with the WET Labs ac9 and ac-s, and the downwelling irradiance Ed(z) determined with the Satlantic Profiler Ⅱ OCI/R-200 and HyperPro Ⅱ in the north South China Sea (SCS) with a broad range, using LightGBM, Random Forest (RF) and CatBoost, three machine learning models for estimating the profile distribution of Kd(z, 650) were developed at first, and they were then evaluated by the key indicators including R2、RMSE、MAPE, as well as the comparison between in situ measured Kd(650) and estimated Kd(650). The evaluation indicated that the CatBoost model performed the best with R2 of 0.8534, RMSE of 0.0472 m-1, MAPE of 11.0585%, and the estimated Kd(650) was also closest to the measured Kd(650). Using the established CatBoost model, input inherent optical properties (IOPs) were the absorption coefficient, the volume scattering function (VSF), and their profile depth, the Kd(650) profile distribution among 15 m in the north SCS was estimated. The result shows that Kd(650) varies from 0.275 to 0.7 m-1 at 5, 10 and 15 m underwater. At 5 m, Kd(650) is relatively stable while it varies greatly at 10 and 15 m. The contribution of volume scattering function distribution to Kd(z, λ) is considered in this study, which provides a new idea and method for accurate estimation and acquisition of Kd(z, λ) based on inherent optical properties (IOPs).
图1 站点分布图基于国家测绘地理信息局标准地图服务网站下载的审图号为GS(2016)1665号的标准地图制作, 底图无修改 Fig 1 Location of stations in the north South China Sea. This map is produced based on the standard map on GS (2016) 1665 download from http://bzdt.ch.mnr.gov.cn, without modification on the base map |
表1 LightGBM算法主要超参数Tab. 1 The main hyperparameters of Light GBM |
名称 | 含义 |
---|---|
learning_rate | 学习率 |
max_depth | 树的最大深度 |
num_leaves | 树的叶子数量 |
min_data_in_leaf | 一个叶子上最小数据量 |
min_sum_hessian_in_leaf | 一个叶子上的最小海森值之和 |
feature_fraction | 随机选取的参数比例 |
bagging_fraction | 训练样本的采样比例 |
reg_alpha | L1正则化 |
reg_lambda | L2正则化 |
表2 机器学习超参数调优结果Tab. 2 The hyperparameters results of CatBoost/LightGBM/RF |
CatBoost | LightGBM | RF |
---|---|---|
n_estimators =800 | learning rate = 0.008,max_depth = 2,num_leaves= 10, min_data_in_leaf=18, min_sum_hessian_in_leaf =0.001, feature_fraction=0.3,bagging_fraction=0.8,reg_alpha=0.001,reg_lambda=0.03 | n_estimators=300 max_features=6 |
图4 14°N断面Kd(650)实测值与三种模型估算值剖面对比图 (站点E74、E73、E69、E66及E63所在断面)Fig. 4 Comparison of in situ measured Kd(650) profile and estimated Kd(650) profile by CatBoost/LightGBM/RF at latitude of 14°N (including stations of E74, E73, E69, E66 and E63) |
图6 CatBoost模型估算17°N Kd(650)剖面图(站点S40、S43、S44以及S45所在断面)Fig. 6 The estimated Kd(650) profile by CatBoost at latitude of 17°N. (including stations S40, S43, S44 and S45) |
[1] |
曹渝昆, 朱萌, 2019. 基于主成分分析和LightGBM的风电场发电功率超短期预测[J]. 上海电力学院学报, 35(6): 562-566.
|
[2] |
宫鹏, 王德兴, 袁红春, 等, 2021. 基于LightGBM的南太平洋长鳍金枪鱼渔场预报模型研究[J]. 水产科学, 40(5): 762-767.
|
[3] |
卢雪梅, 苏华, 2020. 基于OLCI数据的福建近海悬浮物浓度遥感反演[J]. 环境科学学报, 40(8): 2819-2827.
|
[4] |
徐磊, 吴鹏, 徐明生, 等, 2021. 基于卷积神经网络与LightGBM的短期风电功率预测方法[J]. 水电能源科学, 39(2): 209-212+199.
|
[5] |
张天一, 苏华, 杨欣, 等, 2020. 基于LightGBM的全球海洋次表层温盐遥感预测[J]. 遥感学报, 24(10): 1255-1269.
|
[6] |
赵文静, 曹文熙, 胡水波, 等, 2018. MODIS-Aqua漫射衰减产品Kd(490)在南海海域的精度对比[J]. 光学精密工程, 26(1): 14-24.
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
|
/
〈 | 〉 |