Refining marine net primary production estimates: Advanced uncertainty quantification through probability prediction models

Abstract

In marine ecosystems, net primary production (NPP) is important, not merely as a critical indicator of ecosystem health, but also as an essential component in the global carbon cycling process. Despite its significance, the accurate estimation of NPP is plagued by uncertainty stemming from multiple sources, including measurement challenges in the field, errors in satellite-based inversion methods, and inherent variability in ecosystem dynamics. This study focuses on the aquatic environs of Weizhou Island, located off the coast of Guangxi, China, and introduces an advanced probability prediction model aimed at improving NPP estimation accuracy while partially addressing its associated uncertainties within the current modeling framework. The dataset comprises eight distinct sets of monitoring data spanning January 2007 to February 2018. NPP values were derived using three widely recognized estimation methods - the Vertically Generalized Production Model (VGPM); the Carbon, Absorption, and Fluorescence Euphotic-resolving (CAFE) model; and the Carbon-based Productivity Model (CbPM) - serving as model outputs for further analysis. The study evaluates two probability prediction approaches: a Bayesian probability prediction model based on empirical distribution and a deep-learning-based probability prediction model. These methods are employed to meticulously quantify the uncertainty in NPP. The results highlight the effectiveness of probability prediction models in capturing the dynamic trends and uncertainties in marine NPP. Notably, the neural-network-based model demonstrates superior accuracy and reliability compared to the Bayesian approach. Furthermore, the models are applied to prognosticate NPP variations in specific marine regions, efficaciously elucidating interannual trends. This research advances the methodological precision in partially quantifying NPP uncertainty related to parameter and input data variability while highlighting the need for future structural uncertainty assessments through multi-model comparisons. Copyright: