Anh Phan, Hiromichi Fukui

Accurate terrestrial gross primary productivity (GPP) estimates are crucial for developing effective climate change policies. However, quantifying GPP is challenging due to sparse ground observations and the complexity of plant functional types (PFTs). In this study, we address these challenges by evaluating various aspects of a data-driven model, including the architecture of time series deep learning models, the optimal sequence length for input data, and the selection of an appropriate PFT dataset to improve GPP prediction accuracy. We introduce FluxFormer, a comprehensive framework and global dataset designed to optimize GPP estimates from 2001 to 2020 at a 0.1-degree spatial resolution. FluxFormer leverages the updated global PFT dataset v2.0.8 from the ESA Land Cover Climate Change Initiative (ESA-CCI) and combines this with time series remote sensing and climate data using a Multivariate Time Series (MVTS) Transformer model. Our comprehensive evaluations show that FluxFormer’s model architecture and optimal sequence length selection significantly improve monthly GPP predictions and their mean seasonal cycle, especially in tropical regions. We also demonstrate that incorporating the ESA-CCI PFT dataset v2.0.8 yields a more reliable GPP dataset compared to using the Moderate Resolution Imaging Spectroradiometer 1D PFT dataset. Additionally, FluxFormer exhibits reduced interannual variability in arid regions and captures a positive long-term GPP trend (2001-2021) consistent with carbon dioxide (CO2) fertilization effects, an aspect missing in some existing datasets. FluxFormer can thus serve as a tool for refining carbon flux estimates and for cross-verifying datasets.