2:45 PM - 3:00 PM
[ACG39-05] Comparison of Machine Learning, Remote Sensing, and Process-Based Models in GPP Estimation: Insights from Multi-Model Evaluation
Keywords:Gross Primary Production, Machine Learning, Remote Sensing, Process-Based Models, Carbon Cycle, Model Comparison
Gross Primary Production (GPP) is a key indicator of terrestrial ecosystem carbon uptake, and its accurate estimation is crucial for understanding carbon cycle dynamics. Various methods have been developed for GPP estimation, including Machine Learning (ML)-based models, Remote Sensing (RS)-based models, and process-based models (e.g., TREDNY). However, significant discrepancies exist among these approaches due to differences in data sources, assumptions, and model sensitivities to environmental drivers.
This study compares GPP estimates from three approaches: Machine Learning (ML), Remote Sensing (RS), and process-based models (e.g., TRENDY) at a spatial resolution of 0.25°. By analyzing seasonal variations, spatial distributions, and interannual trends, we assess model consistency and uncertainty. Correlation coefficients, root mean square error (RMSE), and bias are computed to evaluate their performance.
Results indicate that the variation range of GPP estimates from the Trendy models is relatively large, with some models producing significantly higher values than ML and RS models. Compared to ML and RS, TRENDY models exhibit stronger fluctuations, particularly in high-latitude regions (e.g., Siberia and Northern Canada), where cloud cover and satellite data limitations contribute to estimation discrepancies. In tropical rainforest areas (e.g., the Amazon), some models yield lower GPP estimates, likely due to vegetation canopy cover affecting ML and RS model accuracy. Among the evaluated models, NIES estimates are close to the mean value, whereas E3SM produces higher GPP estimates than other models.
This comparative analysis highlights the strengths and limitations of each approach, providing insights into their applicability for carbon cycle research. The findings contribute to improving multi-model GPP assessments and enhancing our understanding of terrestrial carbon fluxes under changing climate conditions.
This study compares GPP estimates from three approaches: Machine Learning (ML), Remote Sensing (RS), and process-based models (e.g., TRENDY) at a spatial resolution of 0.25°. By analyzing seasonal variations, spatial distributions, and interannual trends, we assess model consistency and uncertainty. Correlation coefficients, root mean square error (RMSE), and bias are computed to evaluate their performance.
Results indicate that the variation range of GPP estimates from the Trendy models is relatively large, with some models producing significantly higher values than ML and RS models. Compared to ML and RS, TRENDY models exhibit stronger fluctuations, particularly in high-latitude regions (e.g., Siberia and Northern Canada), where cloud cover and satellite data limitations contribute to estimation discrepancies. In tropical rainforest areas (e.g., the Amazon), some models yield lower GPP estimates, likely due to vegetation canopy cover affecting ML and RS model accuracy. Among the evaluated models, NIES estimates are close to the mean value, whereas E3SM produces higher GPP estimates than other models.
This comparative analysis highlights the strengths and limitations of each approach, providing insights into their applicability for carbon cycle research. The findings contribute to improving multi-model GPP assessments and enhancing our understanding of terrestrial carbon fluxes under changing climate conditions.