Optimal Survey Design for Forest Carbon Monitoring in Remote Regions Using Multi-Objective Mathematical Programming

Round 1

Reviewer 1 Report

The estimation of carbon storage and scientific decision-making of sampling strategy in remote areas is of special ecological significance. This study present a multi-objective mathematical programming model which help identify the set of Pareto-optimal combinations of field- and remote sensing sampling strategies for boreal forest carbon, using data collected from the Tanana Valley of interior Alaska (USA), and demonstrate how multi-objective mathematical programming help identify the best compromise sampling intensity combinations. The structure of the paper is basically clear, the discussion is detailed, and the expression needs to be modified to improve the clarity. Some detailed comments are listed below:

 

1）Line 193, suggest the superscripts a, b, s be changed to t, w, s, corresponding to tree, wood and soil, which is more intuitive；

2）Line 203-204,’the total cost of remote sensing analyses for j remote sensing plots and the field plot measurement cost for i field plots’. Change to …for i field plots and for j remote sensing plots, as same as other places putting I before j.

3)Eq(7) on page 6, Can i and j be interchanged here? Why?

4)Line 215,how to calculate the variance?

5) Line 216-217, correlation coefficient are calculated for two series numbers, what are the two series here?

6) Line 218,Z is the maximum number of remote sensing plots, why not field plots?

Comments for author File: Comments.pdf

Author Response

Reviewer #1:

The estimation of carbon storage and scientific decision-making of sampling strategy in remote areas is of special ecological significance. This study present a multi-objective mathematical programming model which help identify the set of Pareto-optimal combinations of field- and remote sensing sampling strategies for boreal forest carbon, using data collected from the Tanana Valley of interior Alaska (USA), and demonstrate how multi-objective mathematical programming help identify the best compromise sampling intensity combinations. The structure of the paper is basically clear, the discussion is detailed, and the expression needs to be modified to improve the clarity. Some detailed comments are listed below: 

1. Line 193, suggest the superscriptsa, b, s be changed to t, w, s, corresponding to tree, wood and soil, which is more intuitive

Author response: The superscripts have been replaced

2. Line 203-204,’the total cost of remote sensing analyses for j remote sensing plots and the field plot measurement cost for i field plots’. Change to …for i field plots and for j remote sensing plots, as same as other places putting I before j.

Author response: This change was made.

3. Eq(7) on page 6, Can i and jbe interchanged here? Why?

Author response: We actually noticed an error in lines 214-217:  denotes the variance and  denotes the correlation coefficient. This error was corrected. Equation 7, i refers to the number of field plots, j refers to the number of lidar plots, and  refers to the correlation coefficient between lidar metrics and carbon pool k.  They are not interchangeable. This is perhaps best understood if you look at the behavior of this variance estimator as the lidar-carbon correlation goes to 0 (reduces to the variance using only field plots – lidar provides no additional information) and as the lidar-carbon correlation goes to 1 (reduces to the variance using only lidar plots – lidar plots provide same information as the field plots). This makes sense, but also illustrates that i and j are not interchangeable.  

4. Line 215, how to calculate the variance?

Author response: As stated on line 218, we calculated the carbon variances and lidar correlations based on a preliminary analysis of tree-, down woody material- and soil carbon using a pilot field and lidar dataset in the Tanana valley of interior Alaska (Pattison et al., 2018)

5. Line 216-217, correlation coefficient are calculated for two series numbers, what are the two series here?

Author response: We aren’t sure if we completely understand the question, but  is the correlation coefficient from the regression relationship between lidar metric(s) and the plot-level carbon in pool k. Since we are interested in carbon in three different carbon pools, there are three different correlation coefficients.

6. Line 218,Z is the maximum number of remote sensing plots, why not field plots?

Author response: This formula was adopted from Cochrane’s “Sampling Techniques” (1977): see Section 12 about Double Sampling. Z is the number of elements in the population. While it is realistic to obtain a full census of lidar plots (i.e. wall-to-wall lidar coverage, a common situation in other regions), it is never realistic to obtain a full census of field plots. Therefore, Z represents the maximum number of lidar plots (full lidar coverage), not field plots.   

Reviewer 2 Report

The manuscript developed the evaluation methods of optimal survey design for forest carbon monitoring in Alaska region of the United States. It’s an interesting topic for the design of field work. The manuscript was well written and organized. Some minor concerns were listed below:

 

The abstract is too long, and it could be shortened within 200–250 words.

 

More references should be cited to enrich the introduction part.

 

Figure 1 is too obscure for readers. The labels need to improve resolution. The information of compass, lat/log, and image were missed.

 

The parameters and data used in the proposed method were depended on USFS forest inventory. If the method is intended to use in other countries or regions, how to estimate/obtain these parameters and data?

Author Response

Reviewer #2:

The manuscript developed the evaluation methods of optimal survey design for forest carbon monitoring in Alaska region of the United States. It’s an interesting topic for the design of field work. The manuscript was well written and organized. Some minor concerns were listed below: 

1. The abstract is too long, and it could be shortened within 200–250 words.

Author response: The abstract has been shortened to 236 words.

2. More references should be cited to enrich the introduction part.

Author response: We feel that the Introduction provides motivation/context for the study and adequately acknowledges the previous work done in this area through text and citations. As stated in lines 90-94, there has been relatively little work done in this area (with the notable exceptions of Kohl et al., 2011 and Barrett et al., 2009) which might explain why there are fewer citations in this part of the introduction.

3. Figure 1 is too obscure for readers. The labels need to improve resolution. The information of compass, lat/log, and image were missed.

Author response: Figure 1 was revised and now is higher resolution, has lat/lon graticules, scale bar, and inset showing location of Alaska and study area within North America.

4. The parameters and data used in the proposed method were depended on USFS forest inventory. If the method is intended to use in other countries or regions, how to estimate/obtain these parameters and data?

Author response: The carbon variances and lidar correlations were based on a preliminary analysis of tree-, down woody material- and soil carbon using a pilot field and lidar dataset in the Tanana valley of interior Alaska. We agree that it is ideal to use estimates based on a pilot study conducted in a representative portion of the region of interest; however, it is possible to implement this optimization approach using estimates based on either expert knowledge or data collected in a similar forest region/country. While conducting a dedicated pilot study within a particular country may not be feasible, at this point many regions of the world have field measurements and lidar (or other airborne remote sensing) collected in over a smaller area that could be used to parameterize the optimization model.

Reviewer 3 Report

This paper presents an approach to identify the optimal number of two types of samples - ground plots, and airborne lidar, when sampling 3 pools which combine 4 of the 5 IPCC carbon pools. It presents several optimal scenarios based on different budgets, and variances in each pool. These optimal scenarios aim for the lowest standard error in one or more pools. 

It is a well written an interesting paper. Relevant for those planning data collection, or looking to validate biomass maps for example?

Some feedback;

On terms, I also see that sometimes 'tree' biomass is used, and sometimes AGB is, so just check that the correct term is used (see figure 2 where the figure has one term and caption the other).

I would use the term REDD+ rather than REDD.

Abstract line 24 - In the abstract, you mention it is "less practical" to monitor soil carbon and DWM. I think it just has not been done/demonstrated yet, which is what you imply later (but it would be very practical if it could be done).

In the intro - could very briefly mention the need to monitor due to international requirements such as for the FAO FRA, and ultimately for NDCs, global stocktake. Just thinking beyond REDD+.

On a broader note, I like the focus on field plots and airborne plots, but since satellite derived lidar measurements are now more and more available, would it be worth mentioning the relevance of this work to those - for example gedi? Or national biomass maps which could be used as input data too as a wall-to-wall source. Might be interesting to say something briefly about this if you think it is not outside the scope of the paper...

ON the findings, I wonder how do these standard errors in C/ha compare to what is required in practice - I'm thinking of of monitoring for REDD+ / other voluntary carbon market accuracy requirements. FCPF for example has quite strict accuracy requirements (where payments are cut if these are not met). Your paper can be relevant to these monitoring needs. There may be other uses for this data which requires certain accuracy of data.

Line 356-357 - OK this suprised me, so would it mean that in situations where field measurements are much more uncertain/ prone to errors (I'm thinking of difficult to monitor ecosystems etc), then your findings/approach would apply there too?

For the discussion, I wonder if you can clarify the link of your results to these things, if any are relevant and within scope: in practice, selecting the number and location of plots in tropical areas is not straightforward, since some plots are not able to be visited as are inaccessible, or areas become unsafe, so they are incomplete /not sampled or moved to another location. Does this ability to get random plots, and being left with plots in one biome/veg type which sometimes happens will really mean that more plots are needed. Sampling in more heterogeneous countries vs large homogenous areas?

GFOI MGD reference should be year 2020: https://www.reddcompass.org/mgd/resources/GFOI-MGD-3.1-en.pdf I suppose Box 26 is of relevance - perhaps? 

Author Response

Reviewer #3:

This paper presents an approach to identify the optimal number of two types of samples - ground plots, and airborne lidar, when sampling 3 pools which combine 4 of the 5 IPCC carbon pools. It presents several optimal scenarios based on different budgets, and variances in each pool. These optimal scenarios aim for the lowest standard error in one or more pools. 

It is a well written an interesting paper. Relevant for those planning data collection, or looking to validate biomass maps for example?

Some feedback;

1. On terms, I also see that sometimes 'tree' biomass is used, and sometimes AGB is, so just check that the correct term is used (see figure 2 where the figure has one term and caption the other).

Author response: Tree carbon was used throughout the paper, and refers to both live and dead standing trees.  The one place AGB was used it was referring to the specific response variable in a previous study (Naesset and Gobakken, 2008).

2. I would use the term REDD+ rather than REDD.

Author response: This change was made throughout the document.

 

3. Abstract line 24 - In the abstract, you mention it is "less practical" to monitor soil carbon and DWM. I think it just has not been done/demonstrated yet, which is what you imply later (but it would be very practical if it could be done).

Author response: The abstract was revised significantly and this sentence was changed to a more direct statement: “Lidar measurements are typically highly correlated with aboveground tree carbon, but are less-strongly correlated with other carbon pools, such as down woody materials (DWM) and soil”

4. In the intro - could very briefly mention the need to monitor due to international requirements such as for the FAO FRA, and ultimately for NDCs, global stocktake. Just thinking beyond REDD+.

Author response: The authors agree and have added the following sentence to this paragraph: “While the requirements of REDD+ programs are often a primary driver for implemen-tation of forest carbon monitoring systems, there are other national-scale reporting ac-tivities that depend upon forest monitoring, including the FAO Global Forest Resources Assessment (FAO, 2020) and nationally determined contributions (NDCs) to reduce national emissions and adapt to the impacts of climate change as required by the Paris Agreement (UNFCCC, 2015).”

5. On a broader note, I like the focus on field plots and airborne plots, but since satellite derived lidar measurements are now more and more available, would it be worth mentioning the relevance of this work to those - for example gedi? Or national biomass maps which could be used as input data too as a wall-to-wall source. Might be interesting to say something briefly about this if you think it is not outside the scope of the paper...

Author response: We agree, and have revised a sentence in the Discussions and Conclusions section to indicate that these methods are applicable to other types of remote sensing: “Thus, it would also be applicable to other remote sensing technologies that might become available in the future as long as information about their error and correlation profiles is given.”

6. ON the findings, I wonder how do these standard errors in C/ha compare to what is required in practice - I'm thinking of of monitoring for REDD+ / other voluntary carbon market accuracy requirements. FCPF for example has quite strict accuracy requirements (where payments are cut if these are not met). Your paper can be relevant to these monitoring needs. There may be other uses for this data which requires certain accuracy of data.

Author response: We agree that the target levels of precision (standard error) will vary depending on the requirements of the monitoring program. For example, in the US FIA program the specified target precision for volume is 10% (per billon cubic feet of growing stock trees on timberland), while forest projects enrolled in the California Air Resources Board carbon offset program need to meet a precision threshold for sampling error of 20% of the mean.  In this manuscript we have tried to emphasize that the target levels for precision (standard error) in the optimization algorithm should be based on the specific application (using FIA in Alaska as an example).  

7. Line 356-357 - OK this suprised me, so would it mean that in situations where field measurements are much more uncertain/ prone to errors (I'm thinking of difficult to monitor ecosystems etc), then your findings/approach would apply there too?

 

Author response: Yes, our approach would apply wherever tradeoffs exist behind minimizing expected errors for different pools of forest carbon due to budget constraints. In our case study, carbon variance in the three carbon pools we examined did not affect Pareto-optimality. The surveyor, however, would still have to pick from the 4 field-remote sampling intensity combinations found in this study if they wish to have an optimal sampling design for a given budget. It is possible of course that in other regions or ecosystems, the effect of carbon variance on Pareto-optimality would be different.  

8. For the discussion, I wonder if you can clarify the link of your results to these things, if any are relevant and within scope: in practice, selecting the number and location of plots in tropical areas is not straightforward, since some plots are not able to be visited as are inaccessible, or areas become unsafe, so they are incomplete /not sampled or moved to another location. Does this ability to get random plots, and being left with plots in one biome/veg type which sometimes happens will really mean that more plots are needed. Sampling in more heterogeneous countries vs large homogenous areas?

Author response: The referee is absolutely right about the importance of practical constraints on sampling such as lack of access or safety. However, random (probability) sampling is required for a design-unbiased estimator. In practice, such issues could be addressed by having a back-up protocol in place should access to a randomly selected plot proves to be problematic. We feel that this issue, as important as it is, is beyond the scope of this paper as it applies to all sampling designs, optimal or not.   

9. GFOI MGD reference should be year 2020: https://www.reddcompass.org/mgd/resources/GFOI-MGD-3.1-en.pdf I suppose Box 26 is of relevance - perhaps?

Author response: This revision was made.