Effects of Bark Beetle Outbreaks on Forest Landscape Pattern in the Southern Rocky Mountains, U.S.A.

Round 1

Reviewer 1 Report

Dear Authors

Random Forest and severity models appeared robust to many investigated factors, though they demonstrated that severe defoliation can lead to overestimates of beetle-caused tree mortality, thus, non-lethal defoliation events should be considered in future development of maps characterizing tree mortality. One time complete defoliation of conifers causes their death, while deciduous trees can be defoliated multiple times and they survive. From this reason application of shortwave-infrared bands and spectral indices that incorporate them are particularly useful as sensitive to tree mortality. Could chlorophyll fluorescence, as inversely proportional to photosynthesis efficiency, could also help detect physiological changes in trees that do not yet show visible changes in the assimilation apparatus?

Also, please consider the below remarks:

L53-57 „Improved mapping efforts that combined tailed field data and remotely sensed products are critical to improve understanding of the extent, severity, and spatial patterns of outbreaks.” That statement can be justified with the following reference:  Nowakowska, J. A., Hsiang, T., Patynek, P., Stereńczak, K., Olejarski, I., & Oszako, T. (2020). Health assessment and genetic structure of monumental Norway spruce trees during a bark beetle (Ips typographus L.) outbreak in the Białowieża Forest District, Poland. Forests, 11(6), 647.

Authors compared field and remote sensing data and found a high agreement of the results (accuracies). It supports also further statement L67“Given  a  suitable  landscape, outbreaks  may  be  incited  by  warm  temperatures  and drought  conditions”

Figure 2. what is LandTrends? Explenation appears much further in L213

L224 and L 303 “Non-Iterative Clustering” – probably Interactive?

L276 this dot “.” Cannot stand alone

Author Response

Thanks for the helpful feedback! All author responses are listed in italics below.

Random Forest and severity models appeared robust to many investigated factors, though they demonstrated that severe defoliation can lead to overestimates of beetle-caused tree mortality, thus, non-lethal defoliation events should be considered in future development of maps characterizing tree mortality. One time complete defoliation of conifers causes their death, while deciduous trees can be defoliated multiple times and they survive.

Insect defoliators in this study area are affecting both deciduous and coniferous trees, and do not typically cause tree mortality in a single defoliation event. We have attempted to clarify this in the text by adding “(e.g., non-lethal defoliation of conifers),” on L250 and “Non-lethal defoliation by other insect species” on L283.

From this reason application of shortwave-infrared bands and spectral indices that incorporate them are particularly useful as sensitive to tree mortality. Could chlorophyll fluorescence, as inversely proportional to photosynthesis efficiency, could also help detect physiological changes in trees that do not yet show visible changes in the assimilation apparatus?

This is an interesting idea, and this type of approach (e.g., using solar‐induced chlorophyll fluorescence (SIF)) would likely improve the initial detection of tree mortality due to bark beetle attack. However, to our knowledge, these data, which are primarily derived from GOME and OCO-2 sensors, are not available at a sufficient spatial resolution or temporal extent for a comparison with the Landsat-based approaches presented here. Though we are aware of downscaled products that use MODIS data, even these data sources are substantially coarser (0.05 x 0.05 degrees) and available for a much shorter period than Landsat products. This seems like a great opportunity for future study, and we are aware of other researchers developing sophisticated methods such as this to improve the ability to detect plant hydraulic stress and initial infestation.

Also, please consider the below remarks:

L53-57 Improved mapping efforts that combined tailed field data and remotely sensed products are critical to improve understanding of the extent, severity, and spatial patterns of outbreaks.” That statement can be justified with the following reference:  Nowakowska, J. A., Hsiang, T., Patynek, P., Stereńczak, K., Olejarski, I., & Oszako, T. (2020). Health assessment and genetic structure of monumental Norway spruce trees during a bark beetle (Ips typographus L.) outbreak in the Białowieża Forest District, Poland. Forests, 11(6), 647.

Authors compared field and remote sensing data and found a high agreement of the results (accuracies). It supports also further statement L67“Given  a  suitable  landscape, outbreaks  may  be  incited  by  warm  temperatures  and drought  conditions”

Thanks for this suggestion. The reference that you have recommended is an interesting study and we agree that it is relevant to the topic of our paper. However, we focused primarily on studies of North American species of Dendroctonus and Dryocoetes bark beetles in the introductory section to allow for a more thorough review of the most closely related literature. While European spruce beetle (Ips typographus) has somewhat similar dynamics to North American mountain pine beetle and spruce beetle, it is outside the focal genera and geographic region of this study. Because we have already included a small number of references that relate to this species of Ips (e.g., [1,2]), we decided not to include another one here.

Figure 2. what is LandTrends? Explanation appears much further in L213

Good point. We have added the following statement to the caption for Figure 2: “LandTrendr (Landsat-based detection of trends in disturbance and recovery) is a pixel-based temporal segmentation algorithm used to identify homogeneous periods of spectral increase, stability, and decline [62,63].”

L224 and L 303 “Non-Iterative Clustering” – probably Interactive?

SNIC stands for Simple Non-Iterative Clustering, as opposed to SLIC (Simple Linear It-

erative Clustering). It is our understanding that the “iterative” part of the name refers to multiple iterations to reach cluster convergence during the segmentation process. In comparison, SNIC converges in a single iteration, making it better for use in parallel processing and cloud-based platforms such as Google Earth Engine.

L276 this dot “.” Cannot stand alone

We have changed the text in this section to “Field data, spatial data, statistical code, and model outputs are available through Dryad Digital Repository <https://doi.org/10.5061/dryad.1rn8pk0sn> [79].” This DOI link is not yet active, but will be shortly after article publication.

Reviewer 2 Report

This manuscript is very well written, informative and easy to understand. It presents the spatial and temporal information combined with extensive field data outputting maps which present severity of bark beetle- tree mortality 1997-2019 in subalpine forests throughout the Southern Rocky Mountains, USA. Remote sensing and field surveys are combined in order to study the effects of bark beetle outbreaks on forest pattern.

They are just few errors in the text:

Lines 133-139 please give scientific names of species listed

Lines 640, 696, 715, 764: Use italic for scientific names of species

Author Response

Author responses are listed below in italics

This manuscript is very well written, informative and easy to understand. It presents the spatial and temporal information combined with extensive field data outputting maps which present severity of bark beetle- tree mortality 1997-2019 in subalpine forests throughout the Southern Rocky Mountains, USA. Remote sensing and field surveys are combined in order to study the effects of bark beetle outbreaks on forest pattern.

Thank you! We appreciate the kind words.

They are just few errors in the text:

Lines 133-139 please give scientific names of species listed

As suggested, we have now added the describer/authority (i.e., full scientific name) for each species in this section (L123-131) and at first mention of other species throughout.

Lines 640, 696, 715, 764: Use italic for scientific names of species

Good point. This was an issue with our citation manager which we have corrected in the updated draft.

Reviewer 3 Report

The study used Landsat time series images and random forest machine learning to map bark beetle outbreaks in the southern Rocky mountains, US. Both the presence and severity of bark beetle caused tree mortality are mapped and validated with field samples with classification accuracy 80% (presence) and R2 = 0.7 (severity), respectively. The spatial pattern of the cumulative tree mortality over 1997-2019 was also qualified. The study is important for forest disturbance study and its environment.

This study claimed using extensive field plot validation points, which are seldom used in previous literatures.

This paper is well written. I support its publication after a minor review.

In the 10-fold validation, maximum 245*90% training samples for training. Is it too less for random forest?

Abstract: Please report both the RMSE and R2 for the regression based severity.

Lines 170-171, what are the accuracies of the severity of the 245 field plot?

It is good to know the authors did a bias correction of random forest output.

Author Response

All author responses are italicized below

The study used Landsat time series images and random forest machine learning to map bark beetle outbreaks in the southern Rocky mountains, US. Both the presence and severity of bark beetle caused tree mortality are mapped and validated with field samples with classification accuracy 80% (presence) and R2 = 0.7 (severity), respectively. The spatial pattern of the cumulative tree mortality over 1997-2019 was also qualified. The study is important for forest disturbance study and its environment. This study claimed using extensive field plot validation points, which are seldom used in previous literatures. This paper is well written. I support its publication after a minor review.

Thank you for the kind words!

In the 10-fold validation, maximum 245*90% training samples for training. Is it too less for random forest?

This is a good question. Our understanding is that Random Forest is one of the best approaches in “high p, low n” situations in which there are large number of predictors and a relatively small number of samples. We feel that the results of cross-validation, variable importance plots, and partial dependence plots together indicate that the relationships between the RF models and the training set are strong and intuitive (greater spectral change is indicative of greater tree mortality). But, as with any statistical algorithm, extending predictions from a relatively small training set (in cross-validation or in the full model) to a much larger sample should be done cautiously. Thus, we also graphically compared RF-predicted maps to aerial imagery throughout the region to confirm that predictions were logical (e.g., Fig. 4e-g). A larger field dataset would certainly be helpful in future studies, though we are confident in the results of this study despite the potential limitations of our sample size.

Abstract: Please report both the RMSE and R2 for the regression based severity.

We have now added the Cohen’s Kappa coefficient (for the classification model) and the RMSE (for the regression model) to the abstract section, for additional accuracy metrics.

Lines 170-171, what are the accuracies of the severity of the 245 field plot?

The field surveys recorded information about all live and dead trees larger than 4cm in diameter (at 1.37 m above the ground) within the sampled area, typically to the closest 0.1 cm, which was then used to calculate live and dead tree basal area and percent loss c. 1990s-2010s. Because we have this information, the field-derived estimates give a very accurate representation of total forest density and disturbance severity in each location. The reviewer may also be asking us to include OOB error from the full models (rather than the 10-fold cross-validation. However, these values are presented in Appendix 4 and are very similar to those obtained from cross-validation. We decided to exclude these from the main text as it may have felt like excessive detail in the results (because the accuracy statistics from cross-validation are already presented).

It is good to know the authors did a bias correction of random forest output.

Thanks! Note that based on the feedback from another reviewer, we have added an additional section to Appendix 3 that further clarifies the rationale behind bias correction. We felt this procedure was necessary because the low- and high-severity pixels are very important ecologically, and we wanted to ensure that models were effectively characterizing these areas.

Reviewer 4 Report

The authors used Landsat time series images to map bark beetle outbreaks. The study is well designed, and the description is clear. Below are my comments:

Line160: Some plots are smaller than Landsat pixels (e.g., 20 by 20m). How do you match the plots with pixels, for example, if a plot across multiple pixels but not dominate any of them? 
Appendix 2: Why most of plots says "Plots included in final analysis had no visible tree mortality in 1-m aerial imagery and Landsat time series." But the manuscript says "all field plots had evidence of bark beetle presence" at line 165.
Table 1: These Tasseled Cap coefficients were for Landsat 5 top-of-atmosphere reflectance, but the research used GEE multi-source surface reflectance.
Line 240: How many plots were used for independent validation? The model reported accuracy from OOB is usually overestimated.
Line 248: This is suspicious. RF might predict wrong on extreme values, but not the overall distribution. It would be interesting to see the comparison of mean and standard deviations of RF-predicted and observed pixel values and the comparison of accuracy and tree mortality map between corrected and uncorrected predictions.

Author Response

All author responses are given in italics below

The authors used Landsat time series images to map bark beetle outbreaks. The study is well designed, and the description is clear. Below are my comments:

Thank you!

Line160: Some plots are smaller than Landsat pixels (e.g., 20 by 20m). How do you match the plots with pixels, for example, if a plot across multiple pixels but not dominate any of them? 

This is a good question and something that we should have clarified in the text. We added the following text at L223-226: “For comparison with field data, we extracted values of each LTS predictor at each plot center location. Though the footprints of our field plots sometimes intersected multiple pixels, this approach maintains raw values in the LTS data.” As the reviewer is alluding to, this means that surveys could sometimes span multiple Landsat pixels and smaller plots might not span an entire pixel. We selected the pixel located at the center of each field plot, with the assumption that was the pixel that would be most representative of the field plot. Altogether, we found very little variation in model accuracy among data contributors that utilized slightly different plot sizes and study designs, indicating that these results were relatively robust to variations in these factors (graphical comparison of model residuals by data contributor is presented in Appendix 4).

Appendix 2: Why most of plots says "Plots included in final analysis had no visible tree mortality in 1-m aerial imagery and Landsat time series." But the manuscript says "all field plots had evidence of bark beetle presence" at line 165.

We agree that this text was unclear as originally worded. Because some of the field data was collected prior to 2019, there was the potential that LandTrendr-derived variables of total spectral decline 1997-2019 could include tree mortality that happened after the time of field surveys but prior to 2019. To assess this, we used aerial imagery and Landsat time series at each field plot location to visually confirm that there was no visible tree mortality from the time of field data collection to the end of the study period (2019). We have attempted to clarify this point by removing this text in the referenced Appendix table and adding the following footnote:

“1: Visual interpretations of Landsat time series and high-resolution imagery were used at each field plot location to ensure that there was no visible mortality that occurred after the time of field surveys but prior to 2019, confirming that subsequent disturbances were not present in LandTrendr-derived predictors (i.e., total spectral decline 1997-2019) that would have been excluded in reference data.”

Table 1: These Tasseled Cap coefficients were for Landsat 5 top-of-atmosphere reflectance, but the research used GEE multi-source surface reflectance.

This is a good point. The Tasseled Cap coefficients included here were developed for top of atmosphere reflectance rather than surface reflectance data. However, a wide range of studies have used these coefficients with surface reflectance data as well (e.g., [3–5]). In fact, these values are included in the source code for LandTrendr on Google Earth Engine, so any study using Tasseled Cap variables with LandTrendr is likely using these coefficients with surface reflectance data. Fundamentally, these coefficients are acting to create weighted combinations of different bands, and these combinations (TCB, TCG, TCW) should illustrate relatively similar concepts whether applied to TOA or SR data. Our intention in using these coefficients was not to precisely recreate the first three principal components in the original transformation of Crist (1985), but rather to develop three indices that are commonly used in other similar studies.

Line 240: How many plots were used for independent validation? The model reported accuracy from OOB is usually overestimated.

This is a good question. We did not use a split sample approach in which a subset of plots was held out of all RF models. Instead, the accuracy metrics presented in the text are based on 10-fold cross-validation, in which 10 different models were fit using the same hyperparameters and LTS predictors, but different subsets of the data. Prior to this, hyperparameters were optimized using a different 10-fold cross-validation (rather than OOB error). For each of the 10 models, 90% of the data was used as a training set and 10% was withheld as a test set, such that each field plot was used as a part of the test set in one of the runs and used as a part of the training set in the remaining nine models. Predicted and observed values were then compared by merging the results from all runs. Thus, we do not present OOB error for either of the Random Forest models, though model accuracy from cross-validation was fairly similar to the OOB error rates – both are presented in Appendix 4. We have also modified the wording of L268-273 of the results to read: “Based on 10-fold cross-validation, the RF presence model had a classification accuracy of 80.3% and Cohen’s Kappa of 0.61; the RF severity model had an R² value of 0.68 and a root-mean-squared-error (RMSE) of 17.3.”

Line 248: This is suspicious. RF might predict wrong on extreme values, but not the overall distribution. It would be interesting to see the comparison of mean and standard deviations of RF-predicted and observed pixel values and the comparison of accuracy and tree mortality map between corrected and uncorrected predictions.

Thanks for this comment. We agree that it is important to be cautious in applying these sorts of corrections that modify model predictions. However, we base this correction on the fact that several previous studies have recognized this as a problem (for example, see the following quote from [6])

“Initial explorations also indicated that the final [Random Forest regression] model was biased, in that CBI was overpredicted at low CBI values and underpredicted at high CBI values (Figure 3). This is because the model fits the mean response, and data extremes rarely represent the mean response”

And another quote from [7]

“On the other hand, because extreme observations are estimated using averages of response values that are closer to those observations, large values of the regression function are underestimated and small values of the regression function are overestimated. Consequently, bias is not negligible and bias correction is necessary”

This bias toward the mean is a logical result of how tree-based algorithms such as Random Forests make predictions. In this study, we found that our RF model of severity was underpredicting high values and overpredicting low values when comparing predicted and observed values from cross-validation (though the mean value was quite accurate), similar to issues brought up by other studies using RF regression. In our case, the extreme low and high values are quite important because they represent places with very little impact from the beetles (potential disturbance refugia) and the most severe impacts (near-total canopy replacing disturbances). Therefore, we applied a correction factor to match the statistical distributions of predicted and observed values and improve the predictions of extreme values. This same type of correction (i.e., “Z-score matching”) is commonly applied with gridded climate data [8–10]. The reviewer also suggested using a figure to compare the statistical distributions of observed, predicted, and corrected values, which was included in the prior version of Appendix 4, but lost among the many plots and statistical code. Therefore, we added a section to Appendix 3, entitled “Correcting Bias in Random Forest Predictions of Outbreak Severity” that includes this figure (Fig. A3.4) as well as a better description of why we performed these corrections. In this section of the appendix, we included the suggested terms (mean and SD of the predicted and observed values) that were used in the bias correction. These terms and Fig. A3.4 also indicate that the RF model was effectively predicting the mean of the distribution, but underestimating the variance. Lastly, we confirmed that RMSE and R² of the RF severity model (presented in the results) were being calculated using the bias-corrected values rather than uncorrected values, which led to slight changes in the text. We feel that all of these changes have helped to improve the manuscript and are very useful additions. Much appreciated!

References

1. Seidl, R.; Müller, J.; Hothorn, T.; Bässler, C.; Heurich, M.; Kautz, M. Small beetle, large-scale drivers: How regional and landscape factors affect outbreaks of the European spruce bark beetle. J. Appl. Ecol. 2016, 53, 530–540, doi:10.1111/1365-2664.12540.
2. Biedermann, P.H.W.; Müller, J.; Grégoire, J.C.; Gruppe, A.; Hagge, J.; Hammerbacher, A.; Hofstetter, R.W.; Kandasamy, D.; Kolarik, M.; Kostovcik, M.; et al. Bark Beetle Population Dynamics in the Anthropocene: Challenges and Solutions. Trends Ecol. Evol. 2019, 34, 914–924, doi:10.1016/j.tree.2019.06.002.
3. Cohen, W.B.; Yang, Z.; Healey, S.P.; Kennedy, R.E.; Gorelick, N. A LandTrendr multispectral ensemble for forest disturbance detection. Remote Sens. Environ. 2018, 205, 131–140, doi:10.1016/j.rse.2017.11.015.
4. Kennedy, R.E.; Yang, Z.; Gorelick, N.; Braaten, J.; Cavalcante, L.; Cohen, W.B.; Healey, S. Implementation of the LandTrendr Algorithm on Google Earth Engine. Remote Sens. 2018, 10, 691.
5. Schroeder, T.A.; Wulder, M.A.; Healey, S.P.; Moisen, G.G. Mapping wildfire and clearcut harvest disturbances in boreal forests with Landsat time series data. Remote Sens. Environ. 2011, 115, 1421–1433, doi:10.1016/j.rse.2011.01.022.
6. Parks, S.A.; Holsinger, L.M.; Koontz, M.J.; Collins, L.; Whitman, E.; Parisien, M.A.; Loehman, R.A.; Barnes, J.L.; Bourdon, J.F.; Boucher, J.; et al. Giving ecological meaning to satellite-derived fire severity metrics across North American forests. Remote Sens. 2019, 11, 1–19, doi:10.3390/rs11141735.
7. Zhang, G.; Lu, Y. Bias-corrected random forests in regression. J. Appl. Stat. 2012, 39, 151–160, doi:10.1080/02664763.2011.578621.
8. Flint, L.E.; Flint, A.L. Downscaling Future Climate Scenarios to Fine Scales for Hydrologic and Ecological Modeling and Analysis. Ecol. Process. 2012, 1, 1–15.
9. Rodman, K.C.; Veblen, T.T.; Battaglia, M.A.; Chambers, M.E.; Fornwalt, P.J.; Holden, Z.A.; Kolb, T.E.; Ouzts, J.R.; Rother, M.T. A Changing Climate is Snuffing Out Post-Fire Recovery in Montane Forests. Glob. Ecol. Biogeogr. 2020, 29, 2039–2051, doi:10.1111/GEB.13174.
10. Bouwer, L.; Aerts, J.; Van de Coterlet, G.; Van de Giesen, N.; Gieske, A.; Mannaerts, C. Evaluating Downscaling Methods for Preparing Global Circulation Model (GCM) Data for Hydrological Impact Modeling. In Climate Change in Contrasting Biver basins: Adaptation Strategies for Water, Food and Environment; CAB International Publishing: London, 2004; pp. 25–47 ISBN 0851998356.