Journal of Tropical Oceanography >
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)
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).
ZHANG Xianqing , LI Cai , Zhou Wen , LIU Cong , XU Zhantang , CAO Wenxi , YANG Yuezhong . Studying on diffuse attenuation coefficient in the South China Sea based on volume scattering function and absorption coefficient*[J]. Journal of Tropical Oceanography, 2023 , 42(3) : 86 -95 . DOI: 10.11978/2022146
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] |
|
/
〈 |
|
〉 |