基于机器学习方法构建的全球溶解氧格点数据(1960-2021)

我们提出了一种结合自组织映射网络与前馈神经网络（SOM-FFNN）的机器学习模型，用于重构全球溶解氧（DO）数据。首先，采用自组织映射（SOM）根据温度、盐度和溶解氧的分布特征将数据分类为12个生态化学省。随后，对于每个生态化学省，使用现有观测数据选择最优的前馈神经网络（FFNN）模型。接着，我们以温度、盐度、潜在密度异常、海底地形、深度、经度、纬度和时间的格点数据作为输入参数。通过选择最优模型，我们建立了协变量与溶解氧之间的非线性关系，并生成了每个生态化学省的格点数据。最后，我们将这些数据合并，形成全球溶解氧的格点数据集。该数据集涵盖了1960年至2021年的时间段，时间分辨率为月，空间分辨率为1°×1°，并覆盖从表层到2000米深度的41个标准层。

We propose a machine learning model combining Self-Organizing Map (SOM) and Feedforward Neural Network (FFNN) for reconstructing global dissolved oxygen (DO) data. First, the Self-Organizing Map (SOM) is employed to classify the data into 12 eco-chemical provinces based on the distribution characteristics of temperature, salinity, and dissolved oxygen. Subsequently, for each eco-chemical province, the optimal Feedforward Neural Network (FFNN) model is selected using existing observational data. Next, we take grid data of temperature, salinity, potential density anomaly, seabed topography, depth, longitude, latitude and time as input parameters. By selecting the optimal model, we establish the nonlinear relationship between covariates and dissolved oxygen, and generate grid data for each eco-chemical province. Finally, we merge these data to form a global dissolved oxygen grid dataset. This dataset covers the period from 1960 to 2021, with a temporal resolution of one month, a spatial resolution of 1°×1°, and encompasses 41 standard layers from the surface to a depth of 2000 meters.