VODCA2GPPv2 - An updated global model for estimating GPP from microwave satellite observations with enhanced cross-biome consistency

Abstract

The monitoring of Gross Primary Production (GPP) on a global scale is essential for understanding the role of terrestrial ecosystems in the carbon cycle. Over the past few decades, significant progress has been made in the ability to globally monitor GPP using process-based models and remote sensing techniques. Despite these advancements, there are still substantial differences between GPP products and large uncertainties in GPP estimates. Recently, Vegetation Optical Depth (VOD) has emerged as a useful indicator for deriving GPP from microwave satellite observations. The carbon-sink driven approach developed by Teubner et al. (2019) utilizes VOD as a proxy for the carbon-sink strength of terrestrial ecosystems to derive GPP. Wild et al. (2022) further adapted this approach, creating a global long-term GPP dataset called VODCA2GPP, based on VOD observations from the Vegetation Optical Depth Climate Archive (VODCA). This approach has shown promising results with good agreements with in-situ GPP observations and independent GPP datasets. However, the model still exhibits limited performance in certain regions and biomes, particularly in arid regions and the tropics, where in-situ data is scarce.This study builds on the VODCA2GPPv1 model by Wild et al. (2022) and tries to make it more consistent across biomes. This was done by employing a new random forest machine learning model, by merging three different eddy covariance datasets to more than double the training data in comparison with VODCA2GPPv1 and by adding two new predictors: Land Cover and low frequency VOD. Validation with in-situ GPP observations showed significant improvements in comparison with VODCA2GPPv1. Median correlations increased from 0.67 to 0.78 r, RMSE decreased from 2.81 to 2.25 gC/m2/d, and bias decreased from 0.25 to -0.04 gC/m2/d. Analyzing the cross-validation results based on land cover demonstrateda more consistent performance of the model, making it better suited for diverse regions. Comparisons with the independent FLUXCOM, MODIS and TRENDY GPP datasets revealed good temporal agreement with mean global correlations of 0.56, 0.62 and 0.42 r respectively, which could mostly be improved in comparisonto VODCA2GPPv1 (+0.06, -0.02 and +0.03 r). Furthermore, the new model reduced global overestimation with respect to these datasets (bias to FLUXCOM and MODIS could be reduced by 0.44 and 0.45 gC/m2/d respectively).However, the new model still has limitations. It still tends to globally overestimate GPP, particularly in tropical regions. Additionally, it exhibits limited performance in arid environments, highlighting the importance of accounting for water limitation in future models. Overall, the inclusion of new predictors and additional in-situ data has resulted in a model that aligns better with in-situ GPP observations and independent GPP datasets. It also demonstrates improved consistencyacross different biomes and land cover classes. VODCA2GPPv2 complements existing GPP products and its long temporal availability makes it a valuable tool for studying the carbon cycle over extended time periods.