[H-TT15] Environmental Remote Sensing

[HTT15-P01] Remote estimation of phytoplankton primary production in clear to turbid waters by integrating a semi-analytical model with a machine learning algorithm

*Zhaoxin Li¹, Wei Yang², Bunkei Matsushita³, Akihiko Kondoh² (1.Graduate School of Science and Engineering, Chiba University, 2.Center for Environmental Remote Sensing, Chiba University, 3.Graduate School of Life and Environmental Sciences, University of Tsukuba)

Keywords:Depth-integrated primary production, Photosynthetic parameters, Ocean color remote sensing, Enhanced random forest regression

Remote estimation of phytoplankton primary production has long been recognized as an important method for investigating the responses of aquatic ecosystems to global climate change. The theory-based primary production model (TPM), one of the earlier proposed models, is potentially applicable to a variety of water bodies because of its semi-analytical scheme. Its accuracy is highly dependent on whether the photophysiological response of phytoplankton is adequately parameterized by a suite of photosynthetic parameters, two of which are the assimilation number (P^B_max) and the light saturation parameter (E_k). The remote assignment of P^B_max and E_k is acknowledged to be a challenging task, and the limited progress that has been made in the assignment of these parameters has impeded extensive use of the TPM. In this study, we proposed a machine learning algorithm, the enhanced random forest regression (ERFR), to retrieve P^B_max and E_k from satellite observations. The ERFR were then integrated with the TPM (together termed as TPM_ERFR) to estimate daily depth-integrated primary production (IPP) in clear to turbid waters. The ERFR was trained and validated using in situ datasets from representative water bodies throughout the Earth. Evaluations with independent in situ data and matchup data showed that the ERFR outperformed conventional empirical and semi-analytical algorithms, and it could better capture the variability of P^B_max and E_k than look-up-table methods. The root mean square difference (RMSD) of the satellite-based IPP estimates from the TPM_ERFR was lower than 0.27. In contrast, the benchmark models generally yielded IPP estimates with RMSDs of 0.27–0.62. The TPM_ERFR was then implemented to climatological satellite products (2010–2019) to reassess global IPP. Reasonable spatial distributions of IPP were preliminarily demonstrated, especially in polar, coastal, and inland waters. These results indicated the potential utility of the TPM_ERFR to generate seamless IPP distributions in clear to turbid waters.