Assessing the insensitivity of machine learning-based GPP estimation to data expansion and the effectiveness of vegetation partitioning strategies

Abstract Gross Primary Production (GPP) is a fundamental component of the ecosystem carbon cycle, and its accurate estimation is critical for understanding the global carbon budget and the ecosystem’s response to climate change. Theoretically, increasing the number of observation sites and extending the observational period should improve the predictive accuracy of machine learning models. However, in practice, model performance improvements are not strictly linear and may exhibit saturation, where additional data provide diminishing returns. Based on GPP flux observations across Europe from 2001 to 2020, this study conducted two comparative experiments to systematically assess the impact of data expansion on GPP estimation accuracy. A plant functional types (PFTs) classification strategy was further introduced to explore its potential for improving model performance. The results show that extending the temporal range from 2001 to 2020 only slightly increased R 2 from 0.722 to 0.725. Spatial expansion further reduced accuracy, with R 2 dropping from 0.740 to 0.734. In contrast, incorporating PFT-based modeling significantly enhanced estimation performance, raising R 2 to approximately 0.77 (p < 0.01) and reducing prediction errors. However, changes in data volume have an impact on the spatiotemporal pattern of GPP estimates by the model, with significant inconsistencies in trends and seasonal dynamics across regions and vegetation types. These findings highlight the limitations of current machine learning models and the fact that, as data availability increases, improving the accuracy of GPP estimates will rely more on optimized model structure and ecological stratification rather than simply on data volume.