Analysis of the Diurnal, Weekly, and Seasonal Cycles and Annual Trends in Atmospheric CO₂ and CH₄ at Tower Network in Siberia from 2005 to 2016

Round 1

Reviewer 1 Report

This study focuses on data analysis of measured CO2 and CH4 data from tower network over Siberia. The analysis is clear and straight-forward, and the results are interesting for data users for future analyses and modeling studies. Given the significant values of this dataset, detail statistical analysis like this work is necessary to inform future data use. Thus I recommend for publications.

Minor issues:

Line 36-37, this assertion are not well-supported. If these events are short-lived, then their influences to large temporal scales maybe limited. Please consider toning down this assertion.

Fig. 1, 'measured by CarbonTracker-CH4' is not an accurate expression. CT-CH4 is a inversion model. I would suggests to use 'estimated' instead. When showing the CT-CH4 flux, it is also important to mention that whether CT-CH4 has assimilated these tower data or not.

Fig.2, what's the grey dots stand for?

Line 155, 'VGN showed showed a faster increase after 2012', why is that? Do we need to worry about end-effect of model curve fitting?

Table 2, indicate units of CO2 and CH4, and 'detrended' in table caption.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This study uses a recently introduced time-series model (Prophet) to extract salient features of the in-situ Siberian tower network-based GHG data (CO2 and CH4). Then this study aims at comparing all of these components (trend, seasonal, sub-seasonal and diurnal) tying the findings with meteorological observations. These findings include multi-periodic seasonality, the prevalence of high-concentration events and correlation with temperature inversion, wind speed, and direction.

The major concerns of this study are:

1) Introduction of the paper doesn't provide sufficient motivation to understand the signatures of these time-series in greater detail. Authors mentioned that Trend, seasonality and few other sub-seasonal behaviors have already been studied before for this data. It is however not clear that why would this study be necessary? How is this work different and unique from previous studies mentioned? In short, it lacks to depict the impact of this work w.r.t. scientific innovations or to the particular scientific community (for example the inverse modeling community)

2) In line number 34, authors started with the fact that "most crucial shortcoming of the studies is their failure to use advanced statistical methods to conduct time-series analysis", the question is how can this be a failure of the previous studies? Does it mean previous studies couldn’t extract meaningful information from the tower data? If a simple trend, seasonality decomposition method serves the purpose, why do we need an advanced method? Does this mean that the method adopted in this paper performs better? If yes, under what metrics? What are the implications? This paper doesn't elaborate on any of this.

3) In the Results and discussion section, authors need to provide a paragraph on how the Prophet model has been run. This should include but not limited to specific trend models, the number of change points, change point estimation strategy used, any tuning parameters specified, the number of periods in the Fourier decomposition, setup of simulated historical forecasts (SHFs), etc. This is crucial for the manuscript that uses a particular model.

4) Figure 8; Is this a 1-sigma, 2-sigma or 3-sigma shading? Why is the uncertainty similar between no-data time-periods compared to data-dense periods? Shouldn’t the uncertainty be larger in these regions? How did the authors assume such a strong systematic structure when there is no data (see plot (b) and (f) )? This is very important for this manuscript as high concentration days depend on this.

Other particular comments are as follows:

- In line number 5, authors mentioned that "inspired by the nature of time-series from the Facebook social network"; what does it mean by “nature of time-series from Facebook Social Network?”
- In line number 14, it says "For some sites, the wind direction indicated the location of strong local sources of CO2 and CH4"; clarify this point, as the reader might be confused and think that wind direction indicates the presence of the sources!
- In line number 35, authors mentioned that usage of the averaging kernel will lead to loss of information. Yes, but there are other well established statistical methods that can also extract long-term trends without applying averaging kernel. So, what does it tell the reader?
- In line number 61, what does it mean by “natural emissions of CH4”? Which version of Carbon tracker is used here?
- In line number 85, provide the reference of the paper that has described the averaging procedure as collection flips between inlets within an hour and multiple inlets mean smaller averaging window within an hour and hence larger uncertainty.
- In line number 97, unless authors show that in an inset with a zoomed-in plot, it is not clear how daily variations are stronger than seasonal variation.
- Instead of table 1, a much easier and more informative way of showing data coverage is to show a heatmap of towers (y-axis) vs time (x-axis). Data dense periods can also be easily identified.
- In line number 105, what does it mean by "The parametric assumption of [20] shows that e(t) is normally distributed"?
- Line number 107 Why only compare with “traditional exponential smoothing models”? There are other techniques like
Dynamic Linear Models (aka Kalman Filters), Singular Spectrum Analysis (SSA), etc. In particular, Kalman Filter doesn’t even require the data to be equally spaced either as
it depends on previous time points (which may or may not be equally spaced)
- In line number 112, authors mentioned that "The large number of gaps that exist in the JR-STATION measurements make the Prophet model particularly suitable to simultaneously reveal the daily, weekly, and monthly seasonality from the hourly time-series."; Why is Prophet model a great candidate under missing values? There is no discussion on how missing values from the tower data have been modeled or tackled.
- Line number 120 talks about change points. How did the authors estimate the change-points? How many change points were there? — There was no discussion on this in the methods or in the discussion section. This is crucial as one main point of this study is handling missing data.
- In line number 139, what does it mean by "reducing trend flexibility"? Perhaps the author can mention this in the Prophet model run discussion paragraph.
- In 141, how small are the differences? Are the differences discernible?
- In line 161, authors mentioned "unlike other statistical methods..."; which statistical models are being referred here? definitely not Dynamic Linear Models mentioned above. DLMs also handle missing value quite nicely using predicted states as the posterior states where data are missing.
- In line 165, What does it mean? authors need to specifically explain the impact of studying “multiyear seasonal variation” and how multiyear seasonal distinguishes itself from studies like [5].
- Table 2 4th column is Diff(%), percentage differences w.r.t what? Are the CO2 working day enhancements statistically significantly different from the Holiday enhancements?
- Line 193-198, what about sensitivity w.r.t. the transportation sector? The transportation sector is one of the largest contributors to anthropogenic CO2. The author should focus more on that for CO2 in comparison to industrial activity in Siberia.
- General question for section 3.3, What about the distributional change in CO2 w.r.t. weekday vs. weekend? Authors can plot average diurnal patterns for different days of the week on the same plot to visually see any differences. The authors may want to use two periods (5 and 7) in the Prophet model to see if there is discernible differences.
- Change figure 6 Y-axis label from CO2, ppm and CH4, ppb to CO2 (ppm) and CH4 (ppb)
- Table 3 doesn't seem to be absolutely relevant to be included in the main manuscript.
- In line 243, authors need to specify that this is approximate cross-validation, not the traditional k-fold cross-validation where random subsets of data in the training set are held out for prediction.
- Table 4; It would be informative to also provide a plot of the forecasted lines along with time points.
How did the authors assume such a systematic structure when there is no data (example: plot (b) and (f) ). This is very important for this manuscript as high-concentration days depend on this.
- In line number 313, authors mention that "this study shows that it is possible to obtain information on the variations of the concentrations of these gases"; I don’t see enough rationale behind the assertion here. It is always possible to obtain information on variations. Why is it not obvious?
- General comments for line number 322. Wind speed, direction, period of high concentration, time of the day and temperature inversion are all correlated. However, it is not obvious from this study. The authors can make pairwise scatter plots or overlayed time-plots to facilitate these correlations and their shapes. These would also be useful to the inversion modeling community.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

This study uses a recently introduced time-series model (Prophet) to extract salient features of the in-situ Siberian tower network-based GHG data (CO2 and CH4). Then this study aims at comparing all of these components (trend, seasonal, sub-seasonal and diurnal) tying the findings with meteorological observations. These findings include multi-periodic seasonality, the prevalence of high-concentration events and correlation with temperature inversion, wind speed, and direction.

One standing concern of the paper is the following:

4) Figure 8; The authors have mentioned that this is an 80% confidence interval i.e. 1.28-sigma and also explained how the periodic structures have been estimated leveraging other year's data. However, the paper still doesn't explain why is the uncertainty similar between no-data time-periods compared to data-dense periods? Shouldn’t the uncertainty be larger in these regions? Statistically, data-sparse regions should have higher uncertainty and thus wider confidence interval.

 

Author Response

We thank the reviewer for constructive and helpful suggestions. We have provided our responses to the reviewers’ comments and believe that our manuscript is much improved as a result.

The reviewers’ specific comments (shown in blue) are addressed below.

4. The following statement should be added to Figure 8 description:
"By default Prophet will only return uncertainty in the trend and observation noise. To get uncertainty in seasonality, it needs to do full Bayesian sampling. We tried different options of the model tunning, but couldn't found out how to convert data-density into uncertainty calculation."

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.