Oceanic Primary Production Estimation Based On Machine Learning

Abstract

AbstractOceanic primary production (OPP) is crucial for ecosystem services and global carbon cycle. However, sensitivity to geographic and environmental characteristics limits the application of semi‐empirical OPP estimate models, such as the vertically generalized productivity model (VGPM) and its modified version, particularly in coastal regions. In addition, the difficulty in collecting necessary parameters also hampers long‐term OPP estimates. Data‐driven machine learning (ML) methods can automatically capture the relationships between the input parameters and the objective; hence, they may become new methods for global OPP estimates. In this study, the effectiveness of ML methods to estimate OPP and the key attributes influencing ML performance in different regions and seasons are discussed. First, the ML models obtain a lower root mean square error than the VGPM and Eppley‐VGPM. In addition, the random forest (RF) model achieves the best performance among the four selected ML models. The enhancement in the accuracy of OPP estimates based on the RF model is more obvious in coastal regions than in the open ocean. In the four seasons, the RF model obtains better estimates of OPP than the Eppley‐VGPM, especially for summer. Moreover, input attributes, including sea surface temperature (SST), photosynthetic active radiation (PAR), and chlorophyll‐a concentration (Chlor‐a), achieve the best performance. The suitable alternative input attributes are SST/Chlor‐a in the coastal regions, and single Chlor‐a, SST/Chlor‐a, and PAR/Chlor‐a in the open ocean. Except for the SST/PAR/Chlor‐a combination, input with Chlor‐a, that is, SST/Chlor‐a and PAR/Chlor‐a, result a relatively acceptable performance in the four seasons.