This study provides excellent insight into spatial heterogeneity of GHG fluxes in a fairly unprecedented sampling extent of chamber-based fluxes. The authors compensate well for the lack of temporal coverage in fluxes by leveraging the spatial information available through a hierarchical Bayesian modeling approach to NEE. The remote sensing analysis is thorough. The comparison of scaling models provides useful insight into important multivariate influences on GHG fluxes.

 

I have two areas of significant improvement to address. The first is a lack of information provided about the Bayesian hierarchical modeling. The authors reference Williams et al. 2006 for their model structure. This should be explicitly provided, along with the parameters that are being estimated, at least in the supplement. The authors mention using vague prior information for these parameters. The prior distributions and initial values used for the MCMC chains should be provided (at least in the supplement). There is no mention of posterior predictive checks or tests for convergence. These are necessary to ensure the model is appropriate for the data and that parameters are estimated correctly (without need for a longer burn-in for example). There should be some presentation of the posterior distributions for parameter estimates. There could be some additional discussion as well related to how much the random effect of plot contributed predictions, or how variable the random effect was within vegetation types, etc.

 

The second area for improvement is related to using back-transformations of log-transformed predictions. From my understanding, soil C, biomass, and soil moisture were log-transformed during their upscaling. Then they were back-transformed and subsequently used as drivers to predict GHG fluxes. Back transforming a prediction (from a non-affine transformation) will introduce bias that needs to be corrected. For a useful explanation of the problem, see this blogpost: https://florianwilhelm.info/2020/05/honey_i_shrunk_the_target_variable/. There are multiple methods available for correcting back-transformation bias, some of which are analytical such as in the case of simple linear regression. See this paper for a comparison of several bias correction methods for GBM models: https://kdd-milets.github.io/milets2022/papers/MILETS_2022_paper_0925.pdf. Since all three of the back-transformed variables rank as fairly high predictors, and are especially important at high soil C, high soil moisture, etc where the back-transformation bias is larger, it is critical to correct this bias. The CH4 flux scaling similarly needs a back transformation bias correction, since a cube-root transformation is also non-affine, and these predicted fluxes are subsequently back transformed for comparing to in situ fluxes and calculating carbon budgets.

 

Minor comments:

There are numerous regressions demonstrating model performance (Fig 4, FigS3), with the r-squared reported. The slopes intercepts, and p-values should also be reported, as this would help assess performance and bias in the model predictions.

 

Font sizes for Fig 3 are too small.

 

It is unclear what the 'Agency 2017' reference is in Fig 1. It is also unclear what the colored vegetation boxes correspond to in panel (c) of Fig 1.