|The paper presents the inverse modeling system developed at NIES, based on the NIES-TM and FLEXPART transport models. The aim of the paper, as I understand it, is to present the system, as well as introduce some technical and developments, and in particular a new method to compute flux covariance matrices. The stated aim of the development is eventually to go towards optimizing anthropogenic CO2 emissions at a high resolution. |
The authors have addressed several of the comments from the initial round of review, and the paper has overall improved. However, I remain unconvinced that it should be published in ACP: The main unique features of the inverse model are the construction of flux covariance matrices using implicit diffusion instead of classical Gaussian correlation functions, and the high-resolution transport achieved via a coupling between NIES-TM and FLEXPART. I think that both are good ideas, which deserve to be explored and eventually published. But they are also very technical, rather out of the scope of ACP (it would fit better in a journal like GMD). A publication in ACP could still be justified if it was demonstrated that these developments lead to considerable changes in our comprehension of CO2 fluxes, or if they enabled inversions that would not otherwise be possible. Unfortunately, neither is done in the paper:
- There is no real really focus on the CO2 fluxes. Compared to the initial submission, they added a section presenting some optimized fluxes, but this is clearly just illustrative, the authors themselves state that “more detailed comparisons […] should be made after improving the inversion setup”. I understand that choice of not focusing on the fluxes during the development phase of the system, but it then it only leaves the quality and novelty of the inversion setup as argument for publication.
- On that second point, the authors unfortunately chose a setup that doesn’t put their model at an advantage compared to existing inverse models: the approach is supposed to facilitate the optimization of CO2 fluxes at a very high resolution, but they used correlation distances of 500 km, which, concretely, means that adjacent pixels are not resolved independently at all by the inversion, and the fluxes are only adjusted by large patches of a few hundred kms (as illustrated in Fig 5, left). What is then the advantage, in terms of scientific results and in terms of computational efficiency, compared to an inversion that would optimize fluxes only at a resolution of just 1°? Would their system still perform well in conditions where others would struggle? If the correlation distances were set small enough so that the inversion can adjust some of the fine-scale spatial structure of the fluxes, would the system still perform well (intuitively, I imagine that the number of iterations required will increase exponentially when the correlation distances are reduced …)? Maybe I am excessively pessimistic, but I think it should be relatively easy for the authors to prove me wrong!
- The use of low-resolution meteorology is understandable for a model development paper, but it makes it even more complicated to judge the interest of the work presented. Furthermore, the authors use a comparison to CarbonTracker as a sort of benchmark, but there are so many differences between the two systems that it’s hard to retain anything from that comparison: the transport models are different, the meteo data driving them is different, the inversion scheme is different, the prior constraints are different, if anything, it shows the NIES-TM-FLEXPART system in an un-necessarily unfavourable view.
My recommendation in the initial submission oscillated between rejection (with some encouragement to improve and resubmit the manuscript to another journal, maybe GMD), and a major revision. This was based pretty much on the same reasons as listed above, and since the authors only offered a minor revision and did not resolve these issues, I cannot change my recommendation.
If the authors want to persist with publishing this paper in ACP, then I would recommend either enriching the paper with a comprehensive set of sensitivity inversions, to assess the use of each of the unique features of their setup, or alternatively, produce a “reference” inversion, using their best possible setup (I.e. including the needed improvements listed in the last sentence of Section 5.2).
Besides the major comments above, the paper is overall in a very good shape (and given my overall review, I did not spend too much time in looking for very minor issues). One issue though is the way FLEXPART footprints are computed: depending on the release time, the footprints will span a period of 2 to 3 days. I don’t know what the consequences are in practice, but I guess that it can introduce some form of systematic bias in the daily cycle, since the length of the footprints will be a function of their longitude and release time. Since the footprints ingested by the system are eventually only 2D, this may not matter too much, but this should be verified and justified in the paper.