An Integration of Linear Model and ‘Random Forest’ Techniques for Prediction of Norway Spruce Vitality: A Case Study of the Hemiboreal Forest, Latvia

Round 1

Reviewer 1 Report

Drought is indeed the significant factor which stressing the forest growing. Compared with
traditional methods, remote sensing has various disadvantages, such as quick, efficient, etc.
The author conducted series of experiments about Norway spruce tree status monitoring, and
has shown some satisfied monitoring results. However, I have some confusions about this
study.
Firstly, vegetation index indeed close related with the growing status of vegetation, and
numbers of literatures have got great satisfied results, especially with high R
2. However, in this
study, the value of R2 was not more than 0.65. thus, I am wondering if there are something
wrong or some un-satisfactions for readers and foresters.
Secondly, the author said the RF was not sensitive when six classes were used at the same
time, whether there were some problems about the characteristic selection as the inputs for
the RF model? Is there some other intelligent algorithms can be used to improve the
robustness of the model? Such as SVM, AI or something else. By the way, the title of this study
used ‘decision tree’ but RF method was used in the text, although the RF was a method
developed from decision tree, they are different after all, so I suggest replaced decision tree
with RF.
Meanwhile, I want to know how to get the tree samples in the field, I see there were 800
samples was collected, what is the principle of selecting these samples? As said above, the
selection characteristics are important for RF model, especially the inputs in this study over a
long growing season.
Finally, RTK can be affected with some factors which result in some errors for the real
coordinate, the images match with UAS maybe not accurate enough. I cannot find some
information about the accuracy of RTK for coordinate.
Besides, Tables in this study have not an uniform format.

Author Response

Drought is indeed the significant factor which stressing the forest growing. Compared with
traditional methods, remote sensing has various disadvantages, such as quick, efficient, etc.
The author conducted series of experiments about Norway spruce tree status monitoring, and
has shown some satisfied monitoring results. However, I have some confusions about this
study.

 

• Firstly, vegetation index indeed close related with the growing status of vegetation, and numbers of literatures have got great satisfied results, especially with high R². However, in this study, the value of R² was not more than 0.65. thus, I am wondering if there are something wrong or some un-satisfactions for readers and foresters.

 

ANSWER:

We agree to Reviewer 1 that R² values in our study are lower than in it can be found in other studies of vegetation indices. Although, our relatively lower R² result may be related to various factors, while very detailed images and specific classes used in this study could be as main reasons for such differences.

 

Lines 326-332 are now complemented: “These results are lower than it was found in other studies for detecting stress induced changes [69, 70]. Mostly, studies with good model prediction power are based on lower resolution imagery (approx. 1.25-2.4 m). In our study a very detailed imagery (0.10 m) was used, but the drawback of such data is the high noise, that can occur from the background (due to the conical form of spruce crown) and/or partial crown damage, which had a negative effect on the outcame of the models.”

 

• Secondly, the author said the RF was not sensitive when six classes were used at the same time, whether there were some problems about the characteristic selection as the inputs for the RF model? Is there some other intelligent algorithms can be used to improve the robustness of the model? Such as SVM, AI or something else. By the way, the title of this study used ‘decision tree’ but RF method was used in the text, although the RF was a method developed from decision tree, they are different after all, so I suggest replaced decision tree with RF.

 

ANSWER:

We agree to Reviewer1, that it is possible that sensitivity of RF in our study could be affected by the selection of RF characteristics. Although, the RF fitting was performed by the default resampling implementation and parameter selection classes, therefore, we think that RF sensitivity issue could be related with problematic distinction between similar classes, namely the overlapping of the values of vegetation indices for so specific classes that were tested in this study (Stem damage, Root rot). While using simplified classes (Healthy vs Crown damages) with clear differences between classes a pronounced the classification also showing the potential of these data using for intermediate discoloration classes. Hence, modeling in our study was challenging when Stem damage and Root rot classes were included due to the Stem damage class which may be the initial stage of Root rot as the stem wounds are known as most common entry for fungal infection. Besides the infection may progress over time during longer observation periods therefore it is difficult to determine when the wound was infected. This requires to be constantly monitored and could be in the context for another study in the future. These specific classes were also less represented, resulting in an uneven distribution of trees by class, which also had some effect on the outcome. Furthermore, RF can accommodate noisy data and it is one of the outperforming decision tree classifiers in terms of classification accuracy, but others like SVM algorithm, which is known to be more sensitive to feature selection, although is less-user friendly due to lot of critical parameters that have to be set. Therefore, we decide to use the RF. We also added more

We thank Reviewer 1 for the suggestion to improve title, and now in the title decision tree is replaced with random forest.

 

Lines 444-450 are now complemented: “Similarly, in other studies [37, 80] when predicting the discolouration, the model ac-curacy was the greatest when distinct classes of physiological classes were used. Our results are consistent with these findings, suggesting that the RF sensitivity might be affected due to the problematic distinction between similar classes (Stem damage, Root rot). In general, the main difficulties to distinguish between these two classes derives from the fact that the stem wounds are known as the most common entry for fungal infections, consequently, in some cases perceived as an initial stage of Root rot.”

 

• Meanwhile, I want to know how to get the tree samples in the field, I see there were 800 samples was collected, what is the principle of selecting these samples? As said above, the selection characteristics are important for RF model, especially the inputs in this study over a long growing season.

 

ANSWER:

 

Tree sampling was based on the method of the National Forest Inventory in Latvia. It is based on the round sample plots with radius 12.62 m. Within a sample plot usually all trees greater than 6 cm at dbh are measured, but in this study we excluded those trees that were suppressed by adjacent trees and were in the shadows by larger trees. Moreover, the plot center were located to ensure that we had a Dead trees represented in each plot.

Lines 140-149 are now clarified: “The field sampling was conducted in eight round sample plots (500 m2, radius= 12.62 m) in May 2021 (after the first flight campaign). All plots were distributed across stands ensuring that each plot contains dead trees or trees with visible crown damages and the distance between the plot centers was at least 26 m. The coordinates of plot centers were measured with the Leica GS16 Global Navigation Satellite System (GNSS) receiver. Precise coordinates of each individual tree within sample plot were recorded by using Leica TS06 total station. The total station was positioned in plot center and pulsed a laser beam to each stem within plot at 1.3 m height. We excluded smaller trees if they were suppressed by adjacent trees and/ or if they were in the shadow of the larger trees. Overall, 800 trees were measured, from which 239 trees were cored.”

 

• Finally, RTK can be affected with some factors which result in some errors for the real coordinate, the images match with UAS maybe not accurate enough. I cannot find some information about the accuracy of RTK for coordinate.

 

ANSWER:

We are thankful to the Reviewer for the suggestion. We added the accuracy of RTK for coordinates

 

Lines194 are now complemented: “…with the position accuracy of 1-2 cm.”

 

• Besides, Tables in this study have not an uniform format.

 

ANSWER:

Tables are now formatted in an uniform style

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript describes the possibilities to use remote sensing data for predicting tree health classes by mixed models regression and Random Forest algorithm. The objectives of this work are interesting and fit well into the scope of the journal.

There are however, a number of flaws in the presentation, some of which can easily be corrected, others however can only be recorded and may reduce the confidence in the results.

1^st The selected stands seem to be rather similar, although the random effect of the stands is reported to be significant? This may reduce the generality of the results.

2^nd, The damage classes seem to be confounded.

3^rd, The description of the way how the classes have been assessed in the field is very poor. How exactly was the vitality assessed? This is the weakest point, because all results relate to these classes and it is well known that their assessment in the field may have substantial variations depending on the observers and the light conditions during the assessment. Further on: How was root rot assessed? Had the sample trees been felled? How exactly was “stem damage” assessed?

4th, In the methods the “Random Forest algorithm” is just named but not described.

5th, For the mixed models only the conditional R-squares are reported, which means the results are only valid for these four stands. The main predictive power could come from the stand variation rather than from the fixed effects.

6th, How exactly did the authors define "accuracy"?

Finally, the paper may need some English editing.

Author Response

• 1^st The selected stands seem to be rather similar, although the random effect of the stands is reported to be significant? This may reduce the generality of the results.

 

ANSWER:

We now added the marginal R² and we describe the proportion of variance explained by a grouping (random) factor, in addition we added a Table 4 to better present results of LMER models

Lines 318-320 are now complemented: “The random effect (Stand) of the models explained 29 to 45% of variance in selected vegetation indices values for different tree health classes (Table 4).”

 

• 2^nd, The damage classes seem to be confounded.

 

ANSWER:

We clarify the damage classes used in this and added Table 3 with description for each tree health class used in this study

3^rd, The description of the way how the classes have been assessed in the field is very poor. How exactly was the vitality assessed? This is the weakest point, because all results relate to these classes and it is well known that their assessment in the field may have substantial variations depending on the observers and the light conditions during the assessment. Further on: How was root rot assessed? Had the sample trees been felled? How exactly was “stem damage” assessed?

 

ANSWER:

We agree to Reviewer 2 that there was incomplete description about classes, hence we created Table 3 where we described the classes used in this study and also clarify the process of how vitality was assessed. Root rot class was assigned for those trees with decayed increment cores and/or trees with proper visual characteristics of root rot (making sure by control drilling). While Stem damage included those trees, which had significant bark striping wounds and/or stem cracks. Stem damages were assessed by inspecting each tree for visible stem damages and recorded if tree had bark stripping wound (greater than 10 cm width and greater than 20 cm length) or stem cracks.

Lines 152-158 are now clarified: “Vitality of crowns was visually assessed for decolorization and needle density from the ground and divided in three groups – visually healthy tree (individuals with no outward signs of stresses (e.g. drought induced decline), trees with noticeable damage to crown (e.g. yellowish green, yellow needles and brownish tree tops) and dead trees (complete loss of green foliage). Each tree was assessed from different angles and to reduce the bias associated with assessment subjectivity the same person performed all assessments.”

 

4th, In the methods the “Random Forest algorithm” is just named but not described.

 

ANSWER:

We added the description of the Random Forest

Lines 285-288 are now complemented: “RF is a method based on inductive decision trees and  can effectively handle high-dimensional, noisy and multi-source datasets without over-fitting [62,63]. The main features of RF include speed and flexibility in creating the relationship between input and output functions [62].

 

5th, For the mixed models only the conditional R-squares are reported, which means the results are only valid for these four stands. The main predictive power could come from the stand variation rather than from the fixed effects.

 

ANSWER:

We agree to Reviewer 2 and to indicates the amount of variation explained by the fixed effects we added also the marginal R² . We also added Table 4 with model summaries.

Lines 320-326 are now complemented: “Our models showed a moderate marginal R2 and total explanatory power – for NDRE index marginal R² was 0.26 and conditional R² was 0.49 (p<0.001), for NDVI index marginal R² was 0.34 and conditional R² was 0.60 (p<0.001), for RGI index marginal R2 was 0.36 and conditional R² was 0.55 (p<0.001), while for CI index marginal R² was 0.27 and conditional R² was 0.49 (p<0.001) (Figure 4, Table 4).”

 

6th, How exactly did the authors define "accuracy"?

 

ANSWER:

We rephrased accuracy to distinguish between the LMER and the RF. For LMER the predictive accuracy was used as a measure to fit the model, and in RF to evaluate classification result we used classification accuracy

 

Finally, the paper may need some English editing.

 

ANSWER:

The language and style have been revised throughout the manuscript.

Author Response File: Author Response.docx

Reviewer 3 Report

In this paper, the authors present a classification system to identify the health status of Norway spruce using multispectral data obtained by unmanned aerial systems. In general, the manuscript is correctly organized, clear and well-written. The experiment design is scientifically sound, and the obtained results can be interesting for other authors. However, the paper should be improved to be acceptable for publications. These are my specific comments:

 

1. In the abstract, you should use superscript for the "2" in R2, since this is the square of R.

 

2. In the introduction, you define the term UAS, but them you use UAV without using it. You should either use always one term or another, or define both of them.

 

3. In figure 1, it would be interesting to show a wider map (including not only Latvia) to put in context the area of study.

 

4. In section 2.2, please explain the procedure to obtain "tree increment cores". What does it consist of? How is it done?

 

5. In section 2.4, you refer to SlantView software. Instead of giving the exact version (v.2.15.0.2509), you should provide a link to this software. Is it free or commercial software?

 

6. The paragraphs in section 2.5 are not justified.

 

7. In figure 2, I suggest you add a legend with a scale indicating the meaning of the color of the pixels. It could be interesting to add an RGB image of this same area in order to better interpret the data.

 

8. In section 2.6, you should briefly explain the technique of linear mixed effect model, and provide references to it.

 

9. In line 180, remove the underscores in 2000 and 2019.

 

10. It would be interesting to add some samples of the UAV images used in the experiments, for example an image of one tree for each health class. In figure 4 you show some sample images, but they are not the images obtained by the drone.

 

11. The exact classes used in the classification should be presented and defined more clearly in section 2.

 

12. In line 180, you say that you use "root mean square error (RMSE)". However, in this case, you have a classification problem, not a regression problem. So, how do you define RMSE in your problem?

 

13. In line 199, you should not use ";" in the title of a section.

 

14. In section 3, you say that: "Our models showed substantial total explanatory power...". However, the obtained R^2 values are not especially good: 0.49, 0.60, 0.55 and 0.49. How do you explain it? How do you interpret these low values?

 

15. The quality of figure 3 should be improved. You should use a line graphic instead of a grayscale image. The same for figure 5.

 

16. Line 248: begging -> beginning

 

17. In all the tables, you should use the format of the journal. In the current version, different formats are used for the tables.

 

18. In table 3, why there is no data for "Age" in "Dead trees" in Stand 1?

 

19. The classification results obtained are not very good. Of course, if you reduce the problem to only two classes, the results are unfairly improved. But this is not an acceptable solution. You should provide some future lines of research in the conclusions to deal with this difficulty.

 

Author Response

1. In the abstract, you should use superscript for the "2" in R2, since this is the square of R.

 

ANSWER: 

We corrected R2 to R²

 

1. In the introduction, you define the term UAS, but them you use UAV without using it. You should either use always one term or another, or define both of them.

 

ANSWER:

 Lines 63-64 are rephrased “Alternative to such platforms are unmanned aircraft vehicle (UAV) platforms [19],….”

 

1. In figure 1, it would be interesting to show a wider map (including not only Latvia) to put in context the area of study.

 

 ANSWER:

We added a map of Europe and location of the Research forests of Kalsnava to Figure 1

In section 2.2, please explain the procedure to obtain "tree increment cores". What does it consist of? How is it done?

 

ANSWER:

Lines 158-171 are now complemented: “The tree age was detected as follows: within each plot, increment cores were taken during the field campaign from 30 trees selected based on the mean NDRE index values for tree crowns after the first flight campaign, (10 trees with the lowest values of NDRE index, 10 trees with medium NDRE index values and 10 trees with the highest NDRE index values). In addition, increment cores from trees with possible root rot infection were also taken (based on visual inspection) Tree increment cores were obtained at breast height (1.3 m) with a Pressler increment borer. The increment cores were processed and measured in the laboratory. Increment cores were air dried, grinded using sandpaper and LINTAB 5 (RinnTECH, Germany) measurement system was used to measure tree ring widths with a precision of 0.01 mm.”

 

1. In section 2.4, you refer to SlantView software. Instead of giving the exact version (v.2.15.0.2509), you should provide a link to this software. Is it free or commercial software?

 

ANSWER: We removed the exact version of the software as it a commercial and could be problems if we would provide a link to it.

 Lines 196-197. lines state now as follows: “The acquired raw images were processed using the SlantView software (SlantRange, Inc., San Diego, CA, USA).”

 

1. The paragraphs in section 2.5 are not justified.

 

ANSWER:

 We justified the paragraphs in section 2.5.

 

1. In figure 2, I suggest you add a legend with a scale indicating the meaning of the color of the pixels. It could be interesting to add an RGB image of this same area in order to better interpret the data.

 

ANSWER:

 Thank you for your suggestion, however, with this image we wanted to highlight the difficulties during the crown extraction, exceptionally for trees which was adjacent to each other. We added to Figure 2 a legend showing the band composite, as well the scale was added.

 

1. In section 2.6, you should briefly explain the technique of linear mixed effect model, and provide references to it.

 

ANSWER:

According to the suggestion of the Reviewer 3 we explained the technique of linear mixed effect model.

Lines 245-251 are complemented: “To evaluate the early detection of tree stress we observed temporal changes in vegetation indices by using the linear mixed effect model (LME).The LME also known as the variance component model is a statistical method that is widely used to model dependent data structures such as clustered data and longitudinal data [55]. The LME incorporates two parameters: the fixed effects and random effects [56]. The fixed effects have a common linear relationship for all the data, while the random effects can be used to account for structure of the data [56].”

 

1. In line 180, remove the underscores in 2000 and 2019.

 

ANSWER:

 After the formatting references according to guidelines these underscores were removed

 

1. It would be interesting to add some samples of the UAV images used in the experiments, for example an image of one tree for each health class. In figure 4 you show some sample images, but they are not the images obtained by the drone.

 

ANSWER:

 We created a new figure (Figure 3) with examples of the SlantRange 3PX orthomosaics (from flight campaign on 19th October 2021), and also he RGB images from last field campaign on October 2021, when we updated the tree health status

 

1. The exact classes used in the classification should be presented and defined more clearly in section 2.

 

ANSWER:

 We created Table 3 to be better presented and described more clearly the classes used in this study

 

12. In line 180, you say that you use "root mean square error (RMSE)". However, in this case, you have a classification problem, not a regression problem. So, how do you define RMSE in your problem?

 

ANSWER:

We agree to Reviewer 3 that there is classification problem and using the RMSE are not fully informative and so we removed from text, but these were as the criteria by which we selected the models. Similar to AIC the RMSE was based on the principle that the lower the RMSE is the better the fitting ability of the model.

 

13. In line 199, you should not use ";" in the title of a section.

 

ANSWER:

 We corrected the title of section 3.1.

 

Line 306 Now state: “Spectral reflectance of Vegetation health and monitoring of one vegetation season”

 

1. In section 3, you say that: "Our models showed substantial total explanatory power...". However, the obtained R^2 values are not especially good: 0.49, 0.60, 0.55 and 0.49. How do you explain it? How do you interpret these low values?

 

ANSWER:

We agree to Reviewer3 that R2 values in our study are lower than in it can be found in other studies of vegetation indices. Although, our relatively lower R² result may be related to various factors, while very detailed images and specific classes used in this study could be as main reasons for such differences. We also agree that using word “substantial” for our R² are too strong, therefore we rephrased to “moderate”.

 

Lines 326-332 are now complemented: “These results are lower than it was found in other studies for detecting stress induced changes [69, 70]. Mostly, studies with good model prediction power are based on lower resolution imagery (approx. 1.25-2.4 m). In our study a very detailed imagery (0.10 m) was used, but the drawback of such data is the high noise, that can occur from the background (due to the conical form of spruce crown) and/or partial crown damage, which had a negative effect on the outcome of the models.”

 

1. The quality of figure 3 should be improved. You should use a line graphic instead of a grayscale image. The same for figure 5.

 

ANSWER:

 We converted Figures to PNG format

 

1. Line 248: begging -> beginning

 

ANSWER:

 Line 313 we corrected a word

 

1. In all the tables, you should use the format of the journal. In the current version, different formats are used for the tables.

 

ANSWER:

 We formatted all tables according to journal guidelines

 

1. In table 3, why there is no data for "Age" in "Dead trees" in Stand 1?

 

ANSWER:

 Unfortunately, the increment cores from Dead trees were not collected in Stand 1, due to the decay. We added a footer to Table 3 with explanation

 

19.The classification results obtained are not very good. Of course, if you reduce the problem to only two classes, the results are unfairly improved. But this is not an acceptable solution. You should provide some future lines of research in the conclusions to deal with this difficulty.

 

We agree that such way to increase the accuracy of classification was not entirely correct. Although, we wanted to highlight the potential of these multispectral remote sensing data in classification by using simplified classes. We are thankful to Reviewer 3 for this suggestion, we identified possible future challenges that could be considered in future research

 

ANSWER:

Lines 486-491 are complemented: “Therefore, future research should address the main challenges to solve limitations of the current study, focusing on how to increase the stability of the RF classifier by reducing the imbalance of the representation between classes. Also, how to deal with the damage progression over time in long term observations. And uneven distribution of the affected needles within crowns are another set of challenges for future re-searches.”

 

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors responded sufficiently to my comments, except the very narrow stand variation wich limits the generality of the results.

Reviewer 3 Report

The authors have satisfactorily answered all my questions, making the necessary changes. In my opinion, the article has improved in quality and is now acceptable for publication.