Estimating primary production in the California Current System using machine learning methods

Primary Production (PP) is a key indicator for assessing the photosynthetic rate of marine phytoplankton. Over the past 40 years, models for estimating PP using remote sensing technology have been continuously developed. While these models have achieved high accuracy in open oceans, their performance in optically complex coastal regions remains limited. With an attempt to develop accurate and robust PP models for coastal environments from satellite measurements, this study aimed to explore machine learning (ML) methods in satellite retrieval of PP values. The California Current System (CCS), one of the world's four largest eastern boundary current systems, has abundant in-situ measurements of PP. Combining these data with remote sensing data, we developed multi-parameter fusion ML algorithms and conducted a comparative analysis with three other PP models. The results indicated that the ML model exhibited high applicability in the remote sensing inversion of PP. The inversion accuracy of the ML model (average RMSE: 266.3 mgC·m⁻²·d⁻¹, average MAPD: 49.9%, average Bias: 3.2 mgC·m⁻²·d⁻¹) outperformed PP models (average RMSE: 1127.0 mgC·m⁻²·d⁻¹, average MAPD: 151.6%, average Bias: 471.6 mgC·m⁻²·d⁻¹). The XGBoost model improves the inversion accuracy of PP in coastal waters more accurately than other models. Based on this model, we analyzed the spatio-temporal distribution characteristics of PP in the CCS from 2012 to 2022. The findings showed distinct monthly distribution patterns of PP on spatial scales, with a decrease from nearshore to offshore areas. On temporal scales, there was an increase trend from February to August, followed by a decline trend until the next February. Additionally, this study further explored the relationship between variations in PP within the CCS and climatic phenomena, specifically the El Niño-Southern Oscillation (ENSO) and the Pacific Decadal Oscillation (PDO). The results showed that abnormal changes in sea surface temperature (SST) were negatively correlated with PP. These findings enhance the methodologies for remote sensing observations of PP and provide innovative perspectives on understanding the dynamics of marine phytoplankton.