the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 18 Dec 2023
Similar importance of inter-tree and intra-tree variations in wood density observations in Central Europe
Abstract. Wood density is a crucial variable linked to mechanical, physiological, and ecological properties. In this study, we analyzed an extensive dataset of over 48,000 wood density samples collected from 2,920 trees. Our aim was to explore variations in wood density, at both inter-tree and intra-tree levels, along with the factors contributing to these variations. Inter-tree variations reveal significant differences in wood density among eight dominant species, highlighting their role in shaping wood density. As tree species exhibit specific spatial distributions associated with microhabitats, we anticipated a link between wood density distribution and microhabitat. Using a feature selection approach and random forest model, we identified six predictors, including satellite-based vegetation indexes, topographic variables, and soil sand content, capable of predicting 91 % of spatial wood density variations. The Normalized Difference Vegetation Index (NDVI) positively representing the amount of carbon within trees correlated with wood density, while the Normalized Difference Water Index (NDWI), reflecting water content, and soil sand content showed negative associations. Geomorphons and soil sand context provided insights into wood density variations and specific landforms. Lower wood density values were linked to landforms with low geomorphons (summit, ridge, or shoulder), whereas higher wood density was found in landforms with high geomorphons (valley, depression, or hollow areas). Furthermore, our study highlighted the importance of considering intra-tree variation, a facet often overlooked in previous research. Interestingly, the magnitude of intra-tree variation is comparable to, and in some species even exceeds, that of inter-tree variations. The intra-tree wood density samples display significant differences both vertically along the height and radially from the center to the bark zones of trees. These variations are influenced by tree growing strategy, living conditions, and physiological structure. In summary, our research delved into the multifaceted features of wood density, shedding light on critical aspects of this fundamental variable.
- Preprint
(1196 KB) - Metadata XML
-
Supplement
(105 KB) - BibTeX
- EndNote
Status: closed (peer review stopped)
-
RC1: 'Comment on egusphere-2023-2691', Anonymous Referee #1, 16 May 2024
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2023/egusphere-2023-2691/egusphere-2023-2691-RC1-supplement.pdf
-
RC2: 'Comment on egusphere-2023-2691', David M. Barnard, 24 Jun 2024
This paper utilizes an impressive dataset of 48,000 wood density samples from forest in Poland to investigate within and among tree variations in wood density. Taxonomic and landscape factors are correlated with tree density, then a feature selection and random forest approach is used to model spatial variation in wood density using remote sensing metrics.
While I think this dataset and findings are interesting, there is some disconnect between the objectives and analyses, and the analytical approach (and interpretation therein) needs improvement. Moreover, the introduction and discussion are both lacking in depth of narrative development and interpretation and lacks a thorough review of the wide body of literature that has focused on wood density variability.
For this review, however, I will focus primarily on the analytical approach, as I think corrections must be made here first before the interpretation of results in the discussion can be evaluated.
- Were density samples collected to minimize inclusion of compression or tension wood? Xylem cells can develop strong differences in cell wall thickness (and thus density) due to directional effects of topography/wind speed which could bias density measurements and possibly inflate within-tree variability. It’s mentioned that rings were sampled for radial profiles in north and south directions, so it seems likely that some compression or tension wood has been included and possibly influencing results. This needs clarification.
- A better description of the predictor data is needed. Line 133 references Table 1, but there are no tables present in the version of this manuscript that I reviewed or in the supplemental. What is the spatial resolution of the remote sensing data or the DEMs used to calculate geomorphons? Why were the specific spectral indices selected and from what satellite products were they computed?
- A better description of the spatial sampling design is needed. The authors explain different height and age classes, and the total number of samples/trees/plots in Fig 1, and that 30+ samples were collected per tree, but the number of trees in each plot is not provided. This is quite important as it defines the data’s hierarchical structure and should drive the analytical approach.
- With hierarchical data such as this, ANOVA is not an appropriate analytical approach. These data should be analyzed with a generalized linear mixed effects model so that the appropriate variance can be partitioned to the random effects (e.g. sample, tree, plot) rather than being fully attributed to the fixed effects (leaf type, family, species, age, dbh, etc). Without a proper hierarchical analysis, the results are challenging to interpret with confidence. Using a GLMM would also allow for all density samples to be pooled into a single analysis (with random effects for sample nested in tree nested in site), implicitly accounting for within tree variation in density, rather than averaging all density measurements per tree and then analyzing at the tree-level. This approach also increases degrees of freedom substantially, thereby increasing interpretive power.
- It’s not clear in the objectives why remote sensing data were used for the random forest modeling when they weren’t used for any of the other analyses. Was this an attempt to be able to predict wood density in areas where samples were not collected but remote sensing data is available? If so, the authors need to explicitly state this objective.
- I don’t understand why a feature selection procedure was used prior to random forest modeling (lines 130-131). Random forest has a built in feature selection process and is, for the most part, robust to a large number of predictors (assuming hyperparameters are appropriately defined). This two-part process may be omitting predictors that RF may have otherwise identified as important.
- More information is needed on the random forest modeling methods, which package/library was used. How were hyperparameters selected or tuned? Were data centered or scaled in any way? Are random forests run at the sample-, tree-, or plot-level? If at tree- or plot-level how were density measurements aggregated? If averaged, were the individual density measurements normally distributed or was a median reported?
- Out of bag estimates from random forest models are not “predictions,” (e.g. lines 193, 265, 324) they report a variance explained only for the model fit to the data i.e. OOB estimates are for variance explained on data subsets withheld when each tree is defined. If the authors would like to report model “predictions” then the model needs to be tuned to a subset (70-80%) of the data (preferably with entire sites/plots omitted) and tested on another subset that has been entirely withheld from model tuning process. This also means that the 91% reported accuracy is likely overestimated in terms of predicting wood density estimates from “new” or withheld data. In simpler language: as reported, the random forest model explains how well the model captures variation in the data, but that has little bearing on the ability of the model to predict (i.e. extrapolate) density estimates using spectral data from trees/areas not included in the dataset used to tune the model.
- Technically, spatial data inherently violate model assumptions of data independence for random forests. While it is increasingly common to use RF for modeling spatial data, the authors need to acknowledge potential biases in model performance and partial effects of predictors due to spatial autocorrelation. There are several packages available to run random forests on spatial data such as the ‘spatialRF’ package in R.
Minor comments:
Line 17 – Typo: “representing” should be “represented”
Line 19 – Define or describe geomorphons at first mention rather than in following sentences.
Lines 64 and 76 – This study looks at variation *among* trees (i.e. comparison among many) rather than between trees – "between" implies a comparison between two individuals.
Line 80 – What does it mean that trees were aged 5 years?
Citation: https://doi.org/10.5194/egusphere-2023-2691-RC2
Status: closed (peer review stopped)
-
RC1: 'Comment on egusphere-2023-2691', Anonymous Referee #1, 16 May 2024
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2023/egusphere-2023-2691/egusphere-2023-2691-RC1-supplement.pdf
-
RC2: 'Comment on egusphere-2023-2691', David M. Barnard, 24 Jun 2024
This paper utilizes an impressive dataset of 48,000 wood density samples from forest in Poland to investigate within and among tree variations in wood density. Taxonomic and landscape factors are correlated with tree density, then a feature selection and random forest approach is used to model spatial variation in wood density using remote sensing metrics.
While I think this dataset and findings are interesting, there is some disconnect between the objectives and analyses, and the analytical approach (and interpretation therein) needs improvement. Moreover, the introduction and discussion are both lacking in depth of narrative development and interpretation and lacks a thorough review of the wide body of literature that has focused on wood density variability.
For this review, however, I will focus primarily on the analytical approach, as I think corrections must be made here first before the interpretation of results in the discussion can be evaluated.
- Were density samples collected to minimize inclusion of compression or tension wood? Xylem cells can develop strong differences in cell wall thickness (and thus density) due to directional effects of topography/wind speed which could bias density measurements and possibly inflate within-tree variability. It’s mentioned that rings were sampled for radial profiles in north and south directions, so it seems likely that some compression or tension wood has been included and possibly influencing results. This needs clarification.
- A better description of the predictor data is needed. Line 133 references Table 1, but there are no tables present in the version of this manuscript that I reviewed or in the supplemental. What is the spatial resolution of the remote sensing data or the DEMs used to calculate geomorphons? Why were the specific spectral indices selected and from what satellite products were they computed?
- A better description of the spatial sampling design is needed. The authors explain different height and age classes, and the total number of samples/trees/plots in Fig 1, and that 30+ samples were collected per tree, but the number of trees in each plot is not provided. This is quite important as it defines the data’s hierarchical structure and should drive the analytical approach.
- With hierarchical data such as this, ANOVA is not an appropriate analytical approach. These data should be analyzed with a generalized linear mixed effects model so that the appropriate variance can be partitioned to the random effects (e.g. sample, tree, plot) rather than being fully attributed to the fixed effects (leaf type, family, species, age, dbh, etc). Without a proper hierarchical analysis, the results are challenging to interpret with confidence. Using a GLMM would also allow for all density samples to be pooled into a single analysis (with random effects for sample nested in tree nested in site), implicitly accounting for within tree variation in density, rather than averaging all density measurements per tree and then analyzing at the tree-level. This approach also increases degrees of freedom substantially, thereby increasing interpretive power.
- It’s not clear in the objectives why remote sensing data were used for the random forest modeling when they weren’t used for any of the other analyses. Was this an attempt to be able to predict wood density in areas where samples were not collected but remote sensing data is available? If so, the authors need to explicitly state this objective.
- I don’t understand why a feature selection procedure was used prior to random forest modeling (lines 130-131). Random forest has a built in feature selection process and is, for the most part, robust to a large number of predictors (assuming hyperparameters are appropriately defined). This two-part process may be omitting predictors that RF may have otherwise identified as important.
- More information is needed on the random forest modeling methods, which package/library was used. How were hyperparameters selected or tuned? Were data centered or scaled in any way? Are random forests run at the sample-, tree-, or plot-level? If at tree- or plot-level how were density measurements aggregated? If averaged, were the individual density measurements normally distributed or was a median reported?
- Out of bag estimates from random forest models are not “predictions,” (e.g. lines 193, 265, 324) they report a variance explained only for the model fit to the data i.e. OOB estimates are for variance explained on data subsets withheld when each tree is defined. If the authors would like to report model “predictions” then the model needs to be tuned to a subset (70-80%) of the data (preferably with entire sites/plots omitted) and tested on another subset that has been entirely withheld from model tuning process. This also means that the 91% reported accuracy is likely overestimated in terms of predicting wood density estimates from “new” or withheld data. In simpler language: as reported, the random forest model explains how well the model captures variation in the data, but that has little bearing on the ability of the model to predict (i.e. extrapolate) density estimates using spectral data from trees/areas not included in the dataset used to tune the model.
- Technically, spatial data inherently violate model assumptions of data independence for random forests. While it is increasingly common to use RF for modeling spatial data, the authors need to acknowledge potential biases in model performance and partial effects of predictors due to spatial autocorrelation. There are several packages available to run random forests on spatial data such as the ‘spatialRF’ package in R.
Minor comments:
Line 17 – Typo: “representing” should be “represented”
Line 19 – Define or describe geomorphons at first mention rather than in following sentences.
Lines 64 and 76 – This study looks at variation *among* trees (i.e. comparison among many) rather than between trees – "between" implies a comparison between two individuals.
Line 80 – What does it mean that trees were aged 5 years?
Citation: https://doi.org/10.5194/egusphere-2023-2691-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 334 | 98 | 35 | 467 | 48 | 25 | 21 |
- HTML: 334
- PDF: 98
- XML: 35
- Total: 467
- Supplement: 48
- BibTeX: 25
- EndNote: 21
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1