the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 Apr 2022
Modelling the impact of wood density dependent tree mortality on the spatial distribution of Amazonian vegetation carbon
Abstract. Spatially heterogeneous plant mortality rates are an important predictor of the distribution of vegetation carbon in Amazonia. Reproducing the spatial gradients of vegetation carbon in Amazonia and the observed decline in the intact Amazonian carbon sink since 1990 is a challenge faced by dynamic global vegetation models (DGVMs). In this paper, we implement spatially variable mortality rates in TRIFFID, the DGVM currently coupled to the Joint UK Land Environment Simulator (JULES), and compare with the standard model which assumes a homogeneous mortality rate. Spatially variable gridded fields of Amazonian tree mortality are created using a well-known relationship between mortality and wood density, and three independent wood density maps. The diversified mortality scheme substantially improves the representation of vegetation carbon in TRIFFID when compared to observations, with a 90 % reduction in model bias and an increase in the Pearson correlation coefficient with observed biomass. JULES now captures the observed variability of both mortality and vegetation carbon to a greater extent, demonstrating the potential of using easily-measured traits, like wood density, to add spatial and functional diversity into DGVMs. Despite this, the spatial variation of vegetation carbon simulated with the new mortality fields (with standard deviation 15 MgCha-1) is still less than half of the variation in the observed data (standard deviation 35 MgCha-1). Future work should consider the effects of additional processes, like fire, drought and the phosphorus cycle, on the simulated distribution of vegetation carbon in the Amazon.
This preprint has been withdrawn.
-
Withdrawal notice
This preprint has been withdrawn.
-
Preprint
(4285 KB)
Interactive discussion
Status: closed
-
RC1: 'Comment on bg-2022-87', Izabela Aleixo, 15 May 2022
This is a very well-written paper on a topic that should be of wide-interest across biogeosciences discussions. The authors included the tree mortality rate through the wood density-mortality relationship as a predictor of the carbon distribution of vegetation in the Amazon. Different dynamic global vegetation models (DGVMs) and four different mortality equations were used to compare with the standard models, which assume homogeneous mortality rates throughout the Amazon basin. This approach brought important improvements in the representation of the spatial dynamics of Carbon in the vegetation of the Amazon, showing a greater correlation between the model with variable mortality and the observed biomass.
Wood density is an important trait in the determination of mortality, and it is relatively easy to obtain, but even so, it does not fully represent the mortality patterns of trees. This is demonstrated by the low variation of the mortality data estimated in this paper, in relation to the actual values ââobserved. Although there is a well-known relationship between wood density and mortality rate, we know that tree mortality results from the interaction between extrinsic environmental conditions, such as climate and other tree ecological traits. Local soil conditions, topography, occurrence of lightning, drought, fire and other environmental and climatic factors also affect mortality patterns, making prediction for such a heterogeneous Amazon basin difficult (e.g. New Phytologist (2019) doi: 10.1111/nph. 16260). As well as extrinsic conditions affect mortality rates, other functional traits also play a key role in determining mortality rates across the basin, such as the phenological behavior of species (e.g. Nat. Clim. Chang. (2019) https: //doi.org/10.1038/s41558-019-0458-0). These factors add even more complexity to the modeling of mortality, making this variable difficult to represent in DGVMs.
Likewise, mortality is an important process that determines the stock of biomass in Amazonian forests, but it is not sufficient by itself to explain the distribution of vegetation along the basin.
Despite these challenges, the authors did a great job of testing different mortality models and equations, carefully explaining the effects that each variable had on biomass estimates. It is an important advance that can be very useful when applied to the science of climate change and effects on Amazonian biomass. The methodology used opens the way for the use of other traits (such as phenology) in mortality estimates, as well as the use of processes other than mortality in modeling the spatial distribution of Amazon Carbon.
The article brings a detailed and very rich discussion about the main points of interest of the scientific community, increasing even more the importance of this manuscript. The methods are presented very clearly, despite the complexity of the subject. It also shows where there are some data gaps where researchers should focus efforts to increase our ability to understand the carbon dynamics of this important Amazon forest ecosystem.
Combining mortality variations as a result of wood density is a path that proved to be very useful and easy to implement to improve biomass stock estimates, although it needs special care in obtaining data and equations used.
Citation: https://doi.org/10.5194/bg-2022-87-RC1 -
RC2: 'Comment on bg-2022-87', Anonymous Referee #2, 05 Jun 2022
Hancock et al. derive spatial maps of wood densities using three independent approaches based on (1) species distribution models, (2) a random forest model and (3) a spatial interpolation method based on inventories. They then apply an ensemble of four different empirical wood density- mortality relationships to derive location-specific mortality rates. Finally, they apply these mortality rates in the DGVM TRIFFID to improve the spatial patterns of aboveground biomass in the Amazon rainforest. The authors find that specific combinations of the wood density maps and mortality-wood density relationships improve spatial patterns of vegetation aboveground biomass in TRIFFID.
The manuscript is for the most part very well written and each section is logically structured and easy to follow. I very much appreciate the effort the authors took to derive the three different maps of wood density which makes their analysis more robust. Furthermore, I also like the fact that (on top of the wood density maps) the authors also test several mortality functions.
However, I am not convinced by their finding/conclusion that these new spatially-explicit mortalities improve the models' performance in reproducing the aboveground biomass patterns in the Amazon rainforest. The spatial patterns in Figure 3 do not look very much different compared to control simulations despite just showing lower (M2, M4) absolute aboveground biomass (AGB) throughout the whole Amazon basin.
The authors apply several indices (absolute bias, CRMSD and Pearson correlation coefficient, Table 3) to support their results. However, most of the indices CRMSD and Pearson’s R do not change much (or consistently) between the simulated combinations and the observations. Only the absolute bias is consistently lower for most of the simulations and the authors mainly argue around this index and its improvement.
If I understood it correctly this index compares mean observed AGB across all grid cells with mean simulated AGB across all grid cells. If that is the case, I do not agree that this measure should be used here at all. The lower values of this measure do not support an improvement in spatial patterns, but rather a canceling/averaging out of over-and underestimation of AGB as seen in Figure 5.
I am puzzled why this index was chosen at all, as I think a more ‘fair’ comparison here is to apply indices such RMSE that do pixel-by-pixel comparison and only then aggregate across the whole basin. I think it would be very helpful to also see RMSE between observed AGB and modelled AGB.
The authors do apply a similar indicator CRMSD which does not show lead to pronounced differences across all simulations.
Furthermore, the authors argue that they find a 40% improvement in Pearsons’ R which is only true for one of the simulations and lower for most of the other simulations. I think it is not fair to just pick the best candidate out of all the simulations and argue that there is an improvement.
Finally, three of the four mortality wood-density relationships are similar (Figure A1), but lead to very different overall aboveground biomass distributions (Figure 3) which shows that AGB is very sensitive to the mortality rates. I am wondering if the relationship between mortality and wood density is necessary overall to improve the spatial patterns or if just an adjustment of the standard constant mortality rate would lead to similar improvements.
Overall, I am not convinced by the authors conclusions that the spatial explicit mortality rates based on wood density lead to an improvement in the AGB patterns of the Amazon basin and would therefore not recommend it for publication in Biogeosciences in its current state.
However, if the authors would show (1) that the RMSE is consistently improving between simulations and observations and (2) that a similar improvement cannot be achieved by simply adjusting the standard constant mortality rates, I would be happy to review the manuscript again.
Other comments
Lines 56-59: I think this part should be rather moved toward the discussion. Ending a paragraph with ‘more data are needed to support this hypothesis’ and then starting the next paragraph ‘Given the evidence above…’ does disturb the flow of reading.
Line 104 and equation (1): what are the indices i and j?
Line 112: I would reframe ‘It was only through…’ to ‘We modified gamma in Eq (1) to …’. Otherwise, this would mean that there is literally no other way/process/function how wood density could affect vegetation carbon in JULES.
Line 117: What are noncompetitive mortality values? Low values? This deserves one more sentence for clarification.
Lines 126 – 129: It is not fully clear to me why there are two different \rho_max values for M1 and M4. Why not simply including an if-statement to avoid mortalities getting to zero when wood densities are too large?
Line 139: If M2 is an improvement on M1 why not skip M1 entirely from this analysis?
Line 167: Some more information about the RF model setup would be helpful. What were the exact predictors? How many train and test samples? I think it would also be useful to show some validation metrics.
Line 172: I very much like that the authors included spatial autocorrelation.
Line 180: I am also missing some information about the setup of the Kriging? Ordinary or universal Kriging? What about the other parameters such as the nugget?
Line 186: What does a coefficient of determination = 0.35 say? Is that a good value? This should be also mentioned somewhere.
Line 198: I wonder how robust the results are when using another forcing e.g. GLDAS 2.0. I know that this might be a lot of effort, but I think it is worth testing if the results of this study are robust against the choice of climate forcing.
Line 220: I think M1 should only be mentioned in the supporting information as it leads to a dieback and is only shown in Figures 1 and 2 and not in the other figures.
Lines 225 to 235: I find the declarations of vegetation carbon with hats confusing. I think the veg subscript is useless as C is always used in combination with ‘veg’. Why not drop the hat and the ‘veg’ in the subscript in favor of another index? For example, simulated vegetation carbon could be C_sim and observed vegetation carbon C_obs.
Equation 3: As I see no index x here, is that the difference between average simulated vegetation carbon and modelled vegetation carbon? If yes, I do not think this is an adequate measure for evaluating spatial patterns (See my main point).
Equation 4: Why not simply use RMSE to compare to the observations? Why CRMSD?
Figure 1: It is hard to compare the plots when all color legends have different ranges. I think they all should be the same range e.g. from 0.42 to 0.81. In the description text kriged map should also have quotes similar to ‘SDM-based’ and ‘RF-based’.
Figure 2: Again, I think equal color legend ranges would be very helpful when comparing the plots. Furthermore, I would drop M1 from this analysis and move it the to the appendix. Similar to M1-M4 I would also use consistent naming of the three maps in the plot and the description: ‘SDM-based’ map, ‘RF-based’ map and ‘Kriged’ map.
Table 2: What are Occurrence and Observation? SDM and RF based? If yes, I would use consistent naming here.
Figure 3: The legend says vegetation carbon? Is it that or AGB? I thought vegetation carbon also includes BGB.
Line 288: I do not see a reduction in CRMSD. These values are only very slightly different compared to the control simulation.
Line 290: What does it mean that they are less effective?
Figure 4: This is a nice figure, but apart from g and i the additional spatial variation is very small that M2-M4 have compared to the absolute differences in AGB in Figure 3.
Figure 5: This figure is also nice, but I am questioning if we really need the maps of wood density and mortality to get to such an improvement. I would like to see if simply adjusting the constant value of gamma could lead to quite similar results.
Figure 6: I do not understand the purpose of this figure. Why is it so clumped, e.g why are there no values around 0.008? Some more sentences explaining this figure would be useful.
Line 329: ‘will be discussed later’. Where? Please refer to the subsection where this is discussed.
Line 371: What does relatively low mean?
Line 375: This sounds very interesting, and I think it should definitely be tried out here.
Line 402: I do not think that the bias statement is useful. I think this statement is misleading and only shows that in Figure 5 b, e, h the red and blue parts cancel out. I agree that this is better than a complete under or overestimation, but I do not agree that it is improving the spatial patterns.
Line 446: Good point. How certain are the authors that the wood density – mortality relationship is needed to improve the spatial patterns. Couldn’t a height-mortality relationship also be reasonable?
Line 457: I think phosphorus availability and its gradient inverse to AGB could also deserve some sentences here (section 4.3.1).
Citation: https://doi.org/10.5194/bg-2022-87-RC2
Interactive discussion
Status: closed
-
RC1: 'Comment on bg-2022-87', Izabela Aleixo, 15 May 2022
This is a very well-written paper on a topic that should be of wide-interest across biogeosciences discussions. The authors included the tree mortality rate through the wood density-mortality relationship as a predictor of the carbon distribution of vegetation in the Amazon. Different dynamic global vegetation models (DGVMs) and four different mortality equations were used to compare with the standard models, which assume homogeneous mortality rates throughout the Amazon basin. This approach brought important improvements in the representation of the spatial dynamics of Carbon in the vegetation of the Amazon, showing a greater correlation between the model with variable mortality and the observed biomass.
Wood density is an important trait in the determination of mortality, and it is relatively easy to obtain, but even so, it does not fully represent the mortality patterns of trees. This is demonstrated by the low variation of the mortality data estimated in this paper, in relation to the actual values ââobserved. Although there is a well-known relationship between wood density and mortality rate, we know that tree mortality results from the interaction between extrinsic environmental conditions, such as climate and other tree ecological traits. Local soil conditions, topography, occurrence of lightning, drought, fire and other environmental and climatic factors also affect mortality patterns, making prediction for such a heterogeneous Amazon basin difficult (e.g. New Phytologist (2019) doi: 10.1111/nph. 16260). As well as extrinsic conditions affect mortality rates, other functional traits also play a key role in determining mortality rates across the basin, such as the phenological behavior of species (e.g. Nat. Clim. Chang. (2019) https: //doi.org/10.1038/s41558-019-0458-0). These factors add even more complexity to the modeling of mortality, making this variable difficult to represent in DGVMs.
Likewise, mortality is an important process that determines the stock of biomass in Amazonian forests, but it is not sufficient by itself to explain the distribution of vegetation along the basin.
Despite these challenges, the authors did a great job of testing different mortality models and equations, carefully explaining the effects that each variable had on biomass estimates. It is an important advance that can be very useful when applied to the science of climate change and effects on Amazonian biomass. The methodology used opens the way for the use of other traits (such as phenology) in mortality estimates, as well as the use of processes other than mortality in modeling the spatial distribution of Amazon Carbon.
The article brings a detailed and very rich discussion about the main points of interest of the scientific community, increasing even more the importance of this manuscript. The methods are presented very clearly, despite the complexity of the subject. It also shows where there are some data gaps where researchers should focus efforts to increase our ability to understand the carbon dynamics of this important Amazon forest ecosystem.
Combining mortality variations as a result of wood density is a path that proved to be very useful and easy to implement to improve biomass stock estimates, although it needs special care in obtaining data and equations used.
Citation: https://doi.org/10.5194/bg-2022-87-RC1 -
RC2: 'Comment on bg-2022-87', Anonymous Referee #2, 05 Jun 2022
Hancock et al. derive spatial maps of wood densities using three independent approaches based on (1) species distribution models, (2) a random forest model and (3) a spatial interpolation method based on inventories. They then apply an ensemble of four different empirical wood density- mortality relationships to derive location-specific mortality rates. Finally, they apply these mortality rates in the DGVM TRIFFID to improve the spatial patterns of aboveground biomass in the Amazon rainforest. The authors find that specific combinations of the wood density maps and mortality-wood density relationships improve spatial patterns of vegetation aboveground biomass in TRIFFID.
The manuscript is for the most part very well written and each section is logically structured and easy to follow. I very much appreciate the effort the authors took to derive the three different maps of wood density which makes their analysis more robust. Furthermore, I also like the fact that (on top of the wood density maps) the authors also test several mortality functions.
However, I am not convinced by their finding/conclusion that these new spatially-explicit mortalities improve the models' performance in reproducing the aboveground biomass patterns in the Amazon rainforest. The spatial patterns in Figure 3 do not look very much different compared to control simulations despite just showing lower (M2, M4) absolute aboveground biomass (AGB) throughout the whole Amazon basin.
The authors apply several indices (absolute bias, CRMSD and Pearson correlation coefficient, Table 3) to support their results. However, most of the indices CRMSD and Pearson’s R do not change much (or consistently) between the simulated combinations and the observations. Only the absolute bias is consistently lower for most of the simulations and the authors mainly argue around this index and its improvement.
If I understood it correctly this index compares mean observed AGB across all grid cells with mean simulated AGB across all grid cells. If that is the case, I do not agree that this measure should be used here at all. The lower values of this measure do not support an improvement in spatial patterns, but rather a canceling/averaging out of over-and underestimation of AGB as seen in Figure 5.
I am puzzled why this index was chosen at all, as I think a more ‘fair’ comparison here is to apply indices such RMSE that do pixel-by-pixel comparison and only then aggregate across the whole basin. I think it would be very helpful to also see RMSE between observed AGB and modelled AGB.
The authors do apply a similar indicator CRMSD which does not show lead to pronounced differences across all simulations.
Furthermore, the authors argue that they find a 40% improvement in Pearsons’ R which is only true for one of the simulations and lower for most of the other simulations. I think it is not fair to just pick the best candidate out of all the simulations and argue that there is an improvement.
Finally, three of the four mortality wood-density relationships are similar (Figure A1), but lead to very different overall aboveground biomass distributions (Figure 3) which shows that AGB is very sensitive to the mortality rates. I am wondering if the relationship between mortality and wood density is necessary overall to improve the spatial patterns or if just an adjustment of the standard constant mortality rate would lead to similar improvements.
Overall, I am not convinced by the authors conclusions that the spatial explicit mortality rates based on wood density lead to an improvement in the AGB patterns of the Amazon basin and would therefore not recommend it for publication in Biogeosciences in its current state.
However, if the authors would show (1) that the RMSE is consistently improving between simulations and observations and (2) that a similar improvement cannot be achieved by simply adjusting the standard constant mortality rates, I would be happy to review the manuscript again.
Other comments
Lines 56-59: I think this part should be rather moved toward the discussion. Ending a paragraph with ‘more data are needed to support this hypothesis’ and then starting the next paragraph ‘Given the evidence above…’ does disturb the flow of reading.
Line 104 and equation (1): what are the indices i and j?
Line 112: I would reframe ‘It was only through…’ to ‘We modified gamma in Eq (1) to …’. Otherwise, this would mean that there is literally no other way/process/function how wood density could affect vegetation carbon in JULES.
Line 117: What are noncompetitive mortality values? Low values? This deserves one more sentence for clarification.
Lines 126 – 129: It is not fully clear to me why there are two different \rho_max values for M1 and M4. Why not simply including an if-statement to avoid mortalities getting to zero when wood densities are too large?
Line 139: If M2 is an improvement on M1 why not skip M1 entirely from this analysis?
Line 167: Some more information about the RF model setup would be helpful. What were the exact predictors? How many train and test samples? I think it would also be useful to show some validation metrics.
Line 172: I very much like that the authors included spatial autocorrelation.
Line 180: I am also missing some information about the setup of the Kriging? Ordinary or universal Kriging? What about the other parameters such as the nugget?
Line 186: What does a coefficient of determination = 0.35 say? Is that a good value? This should be also mentioned somewhere.
Line 198: I wonder how robust the results are when using another forcing e.g. GLDAS 2.0. I know that this might be a lot of effort, but I think it is worth testing if the results of this study are robust against the choice of climate forcing.
Line 220: I think M1 should only be mentioned in the supporting information as it leads to a dieback and is only shown in Figures 1 and 2 and not in the other figures.
Lines 225 to 235: I find the declarations of vegetation carbon with hats confusing. I think the veg subscript is useless as C is always used in combination with ‘veg’. Why not drop the hat and the ‘veg’ in the subscript in favor of another index? For example, simulated vegetation carbon could be C_sim and observed vegetation carbon C_obs.
Equation 3: As I see no index x here, is that the difference between average simulated vegetation carbon and modelled vegetation carbon? If yes, I do not think this is an adequate measure for evaluating spatial patterns (See my main point).
Equation 4: Why not simply use RMSE to compare to the observations? Why CRMSD?
Figure 1: It is hard to compare the plots when all color legends have different ranges. I think they all should be the same range e.g. from 0.42 to 0.81. In the description text kriged map should also have quotes similar to ‘SDM-based’ and ‘RF-based’.
Figure 2: Again, I think equal color legend ranges would be very helpful when comparing the plots. Furthermore, I would drop M1 from this analysis and move it the to the appendix. Similar to M1-M4 I would also use consistent naming of the three maps in the plot and the description: ‘SDM-based’ map, ‘RF-based’ map and ‘Kriged’ map.
Table 2: What are Occurrence and Observation? SDM and RF based? If yes, I would use consistent naming here.
Figure 3: The legend says vegetation carbon? Is it that or AGB? I thought vegetation carbon also includes BGB.
Line 288: I do not see a reduction in CRMSD. These values are only very slightly different compared to the control simulation.
Line 290: What does it mean that they are less effective?
Figure 4: This is a nice figure, but apart from g and i the additional spatial variation is very small that M2-M4 have compared to the absolute differences in AGB in Figure 3.
Figure 5: This figure is also nice, but I am questioning if we really need the maps of wood density and mortality to get to such an improvement. I would like to see if simply adjusting the constant value of gamma could lead to quite similar results.
Figure 6: I do not understand the purpose of this figure. Why is it so clumped, e.g why are there no values around 0.008? Some more sentences explaining this figure would be useful.
Line 329: ‘will be discussed later’. Where? Please refer to the subsection where this is discussed.
Line 371: What does relatively low mean?
Line 375: This sounds very interesting, and I think it should definitely be tried out here.
Line 402: I do not think that the bias statement is useful. I think this statement is misleading and only shows that in Figure 5 b, e, h the red and blue parts cancel out. I agree that this is better than a complete under or overestimation, but I do not agree that it is improving the spatial patterns.
Line 446: Good point. How certain are the authors that the wood density – mortality relationship is needed to improve the spatial patterns. Couldn’t a height-mortality relationship also be reasonable?
Line 457: I think phosphorus availability and its gradient inverse to AGB could also deserve some sentences here (section 4.3.1).
Citation: https://doi.org/10.5194/bg-2022-87-RC2
Data sets
Data for: Modelling the impact of wood density dependent tree mortality on the spatial distribution of Amazonian vegetation carbon Mathilda Hancock, Stephen Sitch, Fabian J. Fischer, Jérôme Chave, Michael O'Sullivan, Dominic Fawcett, Lina M. Mercado https://doi.org/10.5281/zenodo.6388019
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 718 | 233 | 44 | 995 | 35 | 42 |
- HTML: 718
- PDF: 233
- XML: 44
- Total: 995
- BibTeX: 35
- EndNote: 42
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Cited
2 citations as recorded by crossref.
- Modeling the topographic influence on aboveground biomass using a coupled model of hillslope hydrology and ecosystem dynamics Y. Fang et al. 10.5194/gmd-15-7879-2022
- Vicarious calibration of GEDI biomass with Landsat age data for understanding secondary forest carbon dynamics N. Jha et al. 10.1088/1748-9326/ad3661