PREDICTING SURFACE SEAWATER DIMETHYLSULFIDE (DMS) CONCENTRATIONS USING A MACHINE LEARNING ALGORITHM (TREENET)

In order to investigate ecological complexities of surface ocean dimethylsulfide (DMS) distribution, a machine learning algorithm (TreeNet) was combined with ArcGIS to create informative predictions of marine DMS concentrations on a global scale. Monthly climatologies of marine DMS concentrations were calculated from 17 environmental predictor variables. We present a use of spatial modeling for predicting DMS concentrations at the sea surface using a machine learning algorithm. Mean squared error (MSE) and r2 values were applied to evaluate model performance among a series of random data subsets extracted from NOAA's Pacific Marine Environmental Laboratory DMS database. Monthly r2 values ranged from 0.29 to 0.60. Well known areas of high DMS concentrations were closely reproduced. We found that photosynthetically active radiation, solar radiation dose, and standard deviation of sea surface temperature played the most important roles in determining seawater DMS concentrations. These concepts, tools and data layers may be used for further hypothesis testing, and to objectively predict the spatial distribution of ocean compounds. This will enable improved global understanding of ocean biogeochemical systems.

为探究表层海洋二甲基硫（Dimethylsulfide, DMS）分布的生态复杂性，本研究将机器学习算法TreeNet与ArcGIS相结合，以实现全球尺度下海洋DMS浓度的有效预测。研究基于17项环境预测变量，计算得到海洋DMS浓度的月气候态数据集。本文展示了依托机器学习算法开展海面DMS浓度空间预测的建模应用。采用均方误差（Mean Squared Error, MSE）与决定系数r²作为模型性能评价指标，评估所用数据子集提取自美国国家海洋和大气管理局（National Oceanic and Atmospheric Administration, NOAA）太平洋海洋环境实验室（Pacific Marine Environmental Laboratory）的DMS数据库。各月的决定系数r²取值介于0.29至0.60之间，且能够精准复现已知的高DMS浓度海域分布特征。研究发现，光合有效辐射（photosynthetically active radiation）、太阳辐射剂量（solar radiation dose）以及海表温度（sea surface temperature）标准差是调控海水DMS浓度的核心影响因素。本研究所用的研究思路、分析工具与数据图层可用于后续的假说验证，以及客观预测海洋化合物的空间分布格局，这将有助于深化对全球海洋生物地球化学系统的认知。