1992-2020全球大洋二氧化碳分压格点数据产品

通过将逐步回归法和前反馈神经网络结合起来，我们构建了一个逐步前反馈神经网络拟合自算法，来在全球大洋不同区域挑选与表层海水二氧化碳分压紧密相关的预测参数。我们用自组织映射神经网络将全球大洋划分成了11个区域，在每个区域挑选出来了使二氧化碳分压预测的平均误差最低的预测参数组合。基于这些预测参数，再利用前反馈神经网络构建了从1992年1月到2020年12月每月的全球大洋表层海水二氧化碳分压1° × 1°格点数据。与原始数据集SOCAT间的平均误差为12.44 μatm，标准误差为19.41 μatm.

By combining the stepwise regression method and feedforward neural network, we constructed a stepwise feedforward neural network self-fitting algorithm to select predictive parameters closely related to the partial pressure of carbon dioxide in surface seawater across different regions of the global ocean. Using a self-organizing map (SOM) neural network, we divided the global ocean into 11 regions. In each region, we identified the combination of predictive parameters that minimizes the mean error of carbon dioxide partial pressure prediction. Based on these selected parameters, we further employed a feedforward neural network to generate monthly 1° × 1° gridded data of the partial pressure of carbon dioxide in surface seawater of the global ocean from January 1992 to December 2020. The mean error relative to the original SOCAT dataset is 12.44 μatm, with a standard error of 19.41 μatm.