General remarks

This paper presents a multifaceted approach for modelling and upscaling spot data to landscape levelIt is a floow-up of a paper (Wangari et al. 2022, JGR: Biogeosciences, 127, e2022JG006901) that showed the measured GHG fluxes in more detail. It is noteworthy that the data from the spring campaign (March 14-15) was left out of this modelling exercise.

I generally like the setup and believe it brings useful information to landscape and lan use type assessments. There is not much to space for critics. The methods seem sound, text is well written and easily readable. However, it could benefit of more clarity in showing the improvement in upscaling prediction performance and restrictions of the measured GHG data in seasonal and spatial scales in Ch. 4.1. That would result in moderate changes only.

Upscaling of GHG fluxes measured from micro to macro scale is hampered by spatial and temporal uncertainties of varying biological and physical origins. At the same time, an adequate increase of the frequency of chamber measurements is hard. This paper uses flux data, soil physical and chemical characteristics from different types of ground vegetation-soil systems for high resolution upscaling to landscape level with help of remote sensed parameters and indices, and DEM for Random Forest modelling by different land use types separately.

Measurements of soil characteristics and GHG fluxes included two rather short campaigns in 2020. GHG’s were measured daytime using opaque chamber and “fast box” techniques during late June-early July and September 8-17. Those probably yielded estimates relative non-winter emisssion strengths rather than annual fluxes for the sites. In forest land the trees may contribute to the measured soil CO2 emissions through root respiration and add to uncertainties. For CO2 the opaque chamber flux represents a somewhat artificial sum of heterotrophic and autotrophic gas release, but not ecosystem net CO2 exchange that could be measured using e.g. using transparent chamber or eddy covariance over a landscape or within separate land use types. Thus the CO2 fluxes could be hard to compare with literature data. The authors should elaborate in discussion how their results could be applicable e.g. in land use planning or mitigation efforts given the representativeness of their data.

The paper claims an improved prediction performance compared to other approaches in upscaling the mosaic of landscape GHG fluxes. Table 3 shows the approach of this study compared to that of other published studies using RF. It is however difficult to evaluate the performance differences there of with other types of approaches. Please explain clearly why the present approach is an improvement over others. Are there any relative qualitative or quantitative indices for such evaluation?

Hot and cold landscape spots of emissions were identified and their contribution to overall GHG fluxes was evaluated. This is very useful for using the results in GHG mitigation.

Specific remarks

Lines 24-25 Please complete the comparison sentence, since you make comparisons between approaches in Ch. 3.5 (Fig. 6) and in discussion Table 3 and elsewhere. Be more specific.

In the submission by Wangari et al., the authors present the development of a random forest model to predict GHG fluxes at high spatial resolution for a study area in central Germany. The authors apply state of the art methods on a comprehensive and interesting dataset. The topic is of interesting to the readership of Biogeosciences. I have a few concerns regarding the RF model development that I wish to see addressed before the article can be considered for publication.

In my opinion, the data does not substantiate the development of a model at 1 m spatial resolution. The only variables that truly convey information at that scale are the ones derived from the DEM and they are not dominant in the important predictor variables in the CD models. The soil properties are interpolated using a simple interpolation routine. There exist a large body of literature on soil mapping and interpolation (also using ML based approaches like RF or geostatistics) and I find the applied approach to simplified to support a 1 m resolution. I would recommend to conduct the modelling at 10 m spatial resolution instead; meaning applying the IDW interpolation of the soil properties at 10 m and aggregating the DEM to 10 m as well.

The RF models are trained individually for the three land use classes while the summer and autumn data are treated jointly. I would expect that a RF model could easily utilize information from a land use map as additional predictor variable. Especially if the GHG fluxes show significantly different distributions across the three land use classes. For each of the three fluxes one model could be trained using data jointly from the three land use classes as well as both seasons. In line with the argument of the authors that joining summer and autumn trains more robust models, I would expect the same for including data from diverse land use classes.   

It is unclear whether the data plotted in Fig 3 are the 10-fold cv data or the 30% test data. It should be the 30% test data that is being evaluated here. Also, it would be very interesting to see the model’s performance for the 30% test data reported in similar way as the 70% used in the 10-fold cv evaluation in table 2. In this way the model’s robustness can be evaluated. Also, please discuss the limitations of a simple random split sample strategy taking inspiration in the following articles:

Bjerre, E., Fienen, M. N., Schneider, R., Koch, J., & Højberg, A. L. (2022). Assessing spatial transferability of a random forest metamodel for predicting drainage fraction. Journal of Hydrology, 612, 128177.

Meyer, H., & Pebesma, E. (2022). Machine learning-based global maps of ecological variables and the challenge of assessing them. Nature Communications, 13(1), 2208.

It is unclear how random forest hyperparameters where set and if a sensitivity analysis or tuning has been carried out.  

I lack a discussion on how many chamber measurements are needed for the proposed upscaling approach. This would be interesting for future design of upscaling experiments. Moreover, the measurement campaigns were carried out over a little more than a week. How does day to day and diurnal variability introduce uncertainty to the dataset? I also assume that the predictor variable soil temperature can be affected by temporal variability. How did the authors account for that in their analysis?

I think the authors should broaden their discussion up for alternative upscaling methods. Such a discussion should include process-based modelling and water table depth based upscaling using empirical response functions.

Tiemeyer, B., Freibauer, A., Borraz, E. A., Augustin, J., Bechtold, M., Beetz, S., ... & Drösler, M. (2020). A new methodology for organic soils in national greenhouse gas inventories: Data synthesis, derivation and application. Ecological Indicators, 109, 105838.

Koch, J., Elsgaard, L., Greve, M. H., Gyldenkærne, S., Hermansen, C., Levin, G., ... & Stisen, S. (2023). Water-table-driven greenhouse gas emission estimates guide peatland restoration at national scale. Biogeosciences, 20(12), 2387-2403.

Maybe I missed it, but  SR/ER_CO2 needs a clear definition.

I enjoyed reading the manuscript and look forward to seeing a revised version.