Seminar: Xin Yu

Gross primary productivity (GPP) is the largest carbon flux between land and atmosphere and is fundamental to understanding terrestrial carbon cycling. While eddy covariance measurements provide in situ estimates of GPP, their sparse spatial distribution limits the accuracy of global upscaling approaches based on machine learning, meteorological data, and remote sensing. Satellite-based Solar-Induced Fluorescence (SIF), which is strongly correlated with GPP, offers globally consistent coverage that can help address this limitation. Here, we develop a transfer learning framework that integrates eddy covariance observations with TROPOMI SIF data. A feed-forward neural network is first pre-trained at the global scale to predict SIF from meteorological variables, vegetation indices, and plant functional types. It is then fine-tuned at the site level to predict GPP (TF model). The resulting model is subsequently applied to estimate GPP at the global scale. For comparison, two baselines, i.e., feedforward neural network and extreme gradient boosting, are trained only on site-level GPP. At the site scale, the TF model achieves comparable performance in predicting GPP with a Nash-Sutcliffe efficiency of 0.742 at the hourly scale. At the global scale, the TF model shows spatiotemporal agreement with SIF comparable to that of the baselines, but with notable improvements in tropical rainforests and croplands. This study demonstrates that the benefits of pre-training on SIF for GPP prediction are detectable, suggesting a potential pathway for alleviating the challenges of extrapolating sparse data to the globe.