Modeling gross primary productivity across different European ecosystem types

Abstract

Predicting Gross Primary Productivity (GPP) is key for understanding ecosystem health and quantifying the global carbon cycle. While data-driven models have shown strong performance in capturing GPP dynamics at specific sites, their ability to generalize across ecosystems without site-specific recalibration remains largely untested. This study addresses this gap by evaluating the applicability of XGBoost and LSTM models in estimating GPP across different European ecosystems. We developed a unified (cross-site) modeling framework that integrates in-situ eddy covariance observations and Sentinel-2–derived vegetation indices using incremental learning. Models’ performance was assessed via: (i) site-specific models, developed to capture individual site characteristics, and (ii) cross-site generalization, including evaluation on an independent dataset of unseen ecosystems. SARIMAX is included as a site-specific statistical benchmark for comparison. Our findings indicate that XGBoost consistently outperformed the other models, achieving site-specific R² values above 0.90 in forest and grassland ecosystems and an average R² of 0.72 across unseen sites (range 0.66–0.78). LSTM exhibited better accuracy in predicting GPP peaks at site-specific level, particularly in cropland and forest ecosystems. At site-level, SARIMAX showed comparable performance to XGBoost but struggled in capturing the rapid temporal variation of GPP. These findings demonstrate the feasibility of a data-driven framework for cross-site GPP monitoring within European flux-tower networks, making a first step toward transferable GPP prediction without site-specific recalibration.