Gross primary productivity and the predictability of CO₂: more uncertainty in what we predict than how well we predict it

Abstract

Abstract. The prediction of atmospheric CO2 concentrations is limited by the high interannual variability (IAV) in terrestrial gross primary productivity (GPP). However, there are large uncertainties in the drivers of GPP IAV among Earth system models (ESMs). Here, we evaluate the impact of these uncertainties on the predictability of atmospheric CO2 in six ESMs. We use regression analysis to determine the role of environmental drivers in (i) the patterns of GPP IAV and (ii) the predictability of GPP. There are large uncertainties in the spatial distribution of GPP IAV. Although all ESMs agree on the high IAV in the tropics, several ESMs have unique hotspots of GPP IAV. The main driver of GPP IAV is temperature in the ESMs using the Community Land Model, whereas it is soil moisture in the ESM developed by the Institute Pierre Simon Laplace (IPSL-CM6A-LR) and in the low-resolution configuration of the Max Planck Earth System Model (MPI-ESM-LR), revealing underlying differences in the source of GPP IAV among ESMs. Between 13 % and 24 % of the GPP IAV is predictable 1 year ahead, with four out of six ESMs showing values of between 19 % and 24 %. Up to 32 % of the GPP IAV induced by soil moisture is predictable, whereas only 7 % to 13 % of the GPP IAV induced by radiation is predictable. The results show that, while ESMs are fairly similar in their ability to predict their own carbon flux variability, these predicted contributions to the atmospheric CO2 variability originate from different regions and are caused by different drivers. A higher coherence in atmospheric CO2 predictability could be achieved by reducing uncertainties in the GPP sensitivity to soil moisture and by accurate observational products for GPP IAV.

Regional estimates of gross primary production applying the Process-Based Model 3D-CMCC-FEM vs. multiple Remote-Sensing datasets

“…Differences between 3D- CMCC-FEM and CFIX may thus more probably rely on different models’ parameterization and not just on the different approach used to simulate photosynthesis. Indeed, as outlined in other studies carried at global scales there is a higher uncertainties in interannual variability even across different RS-based datasets, (Butterfield et al, 2020; Zhang & Ye, 2021) and even in its correct (of GPP) calculation across different models (Dunkl et al, 2023).…”

“…2003). Yet interannual variability is the most uncertain signal event when comparing different data and often considered as an ‘acid test’ in vegetation modeling (Keenan et al, 2012; Collalti et al, 2016; Dunkl et al, 2023).…”

“…However, when applying the 3D-CMCC-FEM at a different spatial scales model parameters, structural, and input-related uncertainties might amplify or even buffer the error in simulating GPP as observed for other models (e.g. Dalmonech et al, 2015; Zheng et al, 2022; Dunkl et al, 2023). In any case, the overall, 3D-CMCC-FEM GPP (∼600 - ∼2200 gC m −2 yr −1 ) is well in the bounds of the ∼600 - ∼2500 gC m −2 yr −1 GPP values described in Collalti and Prentice (2019) for temperate deciduous and coniferous forests in Europe.…”

A Machine Learning Framework to Evaluate Vegetation Modeling in Earth System Models

“…Examples are seen in work done for simulating GPP using observations of meteorological data or satellite data (Sarkar et al, 2022;Z. Zhang et al, 2021), upscaling GPP estimates from eddy covariance sites (Yu et al, 2021), to constrain uncertainty in GPP projections from models (Schlund et al, 2020) and for evaluating GPP representation in models (Dunkl et al, 2023;Z. Zhang et al, 2021).…”

“…Machine learning (ML) approaches have recently been used extensively in the study as well as generation of more accurate GPP data sets. Examples are seen in work done for simulating GPP using observations of meteorological data or satellite data (Sarkar et al., 2022; Z. Zhang et al., 2021), upscaling GPP estimates from eddy covariance sites (Yu et al., 2021), to constrain uncertainty in GPP projections from models (Schlund et al., 2020) and for evaluating GPP representation in models (Dunkl et al., 2023; Z. Zhang et al., 2021). Our goal in this study is to use interpretable ML approaches (Doshi‐Velez & Kim, 2017; Molnar, 2020) to better understand the sources of differences in GPP estimates between ESMs.…”