A 2001–2022 global gross primary productivity dataset using an ensemble model based on the random forest method

作者：Xin Chen, Tiexi Chen, Xiaodong Li, Yuanfang Chai, Shengjie Zhou, Renjie Guo, Jie Dai · 发表于：Biogeosciences · 年份：2024 · DOI：10.5194/bg-21-4285-2024 · 被引用次数：15 · 研究领域：Remote Sensing in Agriculture、Solar Radiation and Photovoltaics

Abstract. Advancements in remote sensing technology have significantly contributed to the improvement of models for estimating terrestrial gross primary productivity (GPP). However, discrepancies in the spatial distribution and interannual variability within GPP datasets pose challenges to a comprehensive understanding of the terrestrial carbon cycle. In contrast to previous models that rely on remote sensing and environmental variables, we developed an ensemble model based on the random forest method (denoted ERF model). This model used GPP outputs from established models: Eddy Covariance Light Use Efficiency (EC-LUE), GPP estimate model based on Kernel Normalized Difference Vegetation Index (GPP-kNDVI), GPP estimate model based on Near-Infrared Reflectance of Vegetation (GPP-NIRv), Revised-EC-LUE, Vegetation Photosynthesis Model (VPM), and GPP estimate model based on the Moderate Resolution Imaging Spectroradiometer (MODIS). These outputs were used as inputs to estimate GPP. The ERF model demonstrated superior performance, explaining 85.1 % of the monthly GPP variations at 170 sites and surpassing the performance of selected GPP estimate models (67.7 %–77.5 %) and an independent random forest model using remote sensing and environmental variables (81.5 %). Additionally, the ERF model improved accuracy across each month and with various subranges, mitigating the issue of “high-value underestimation and low-value overestimation” in GPP estimates. Over the period from 2001 to ...