Ensemble Machine Learning for Mapping Tree Species Alpha-Diversity Using Multi-Source Satellite Data in an Ecuadorian Seasonally Dry Forest

Round 1

Reviewer 1 Report

This paper combines multi-source remote sensing data with Model Ensembles for mapping tree species α-Diversity in Dry tropical forest, In my opinion, this study has certain practical significance for regional tree species α-Diversity acquisition by remote sensing methods. However, there are still many problems in the article, and I think the article will be more specific after the revision

 Keyword: The keyword here is not LIDAR but simulated LIDAR

 Line117： (mm mo^-1 ), Can you explain what this unit stands for?

 Figure1(b,c)：What's the little yellow box

 Line137: within a 9-ha plot on 225,,,the 225 is the numbei of subplots?

 Line142:: I don't think it's d.b.h., it's DBH

 Line144: Add space between 5 and power

 Table 2 evi calculation expression appears serial

 UTM zones are not mentioned in the article

Author Response

Comments and Suggestions for Authors

 

Reviewer 1

This paper combines multi-source remote sensing data with Model Ensembles for mapping tree species α-Diversity in Dry tropical forest, In my opinion, this study has certain practical significance for regional tree species α-Diversity acquisition by remote sensing methods. However, there are still many problems in the article, and I think the article will be more specific after the revision

 

We have reorganized the manuscript and revised sections to clarify methods and results according to reviewer input. Changes from line-by-line comments are below.

 

Keyword: The keyword here is not LIDAR but simulated LIDAR

We have changed the keyword from LiDAR to ‘simulated LiDAR’.

 

Line 117: (mm mo^-1 ), Can you explain what this unit stands for?

We have changed mm mo-1 to “mm per month” so that the unit measurements are more clearly stated.

 

Figure1(b,c)：What's the little yellow box

The smaller yellow box is the study area. The larger grey box containing the study area has been enlarged and inset over the landscape map to show a zoomed-in view during leaf-on and leaf-off periods. We have revised the figure to indicate which is the actual study area location and which is the enlarged area. We have re-written the figure captions to better explain what is shown in each map.

 

Line137: within a 9-ha plot on 225,,,the 225 is the numbei of subplots?

Yes, it is a little unclear as to how the 9-ha study site is set up for tree measurements. We revised this sentence to indicate that the 9-ha area was sub-divided into 225, 20m x 20m (400 m²) plots for making systematic tree measurements and orientation while inside the study area. We hope it is clearer with these changes.

 

Line142: I don't think it's d.b.h., it's DBH

We have changed d.b.h to DBH throughout. In addition, we removed t.p.h for trees per ha since it is only used once.

 

Line144: Add space between 5 and power

We have made this correction.

 

Table 2 evi calculation expression appears serial

This is the standard EVI calculation when NIR, Red and Blue bands are present according to Huete et a. 2002.

 

UTM zones are not mentioned in the article

We have added that UTM WGS84 Zone 17 was the projection system used for tree stem mapping in Methods and Materials. 

Author Response File: Author Response.docx

Reviewer 2 Report

In this manuscript, various remote sensing images are used for estimating tropical forest diversity with stacking ensemble learning. The experiment approach and results are extremely interested. However, it is hard to see any scheme in this piece of writing; it is very confused for readers, specially for the parts of introduction and results. I suggest that the structure of manuscript should be modified, firstly.

1 Table 1 requires supplementary references about these Species diversity indices.

2 Line 137, how to understand the sentence “Tree species, stem diameter, and height were collected within a 9-ha plot on 225”?

3 line 147, All trees were marked with metal tags for relocation and measurement

4 figure 2, figure 3 and figure 4 don’t showed in the paper.

5 For the part of 2.4 Introduction of sensors, remote sensing images and extracted features need to be further refined.

6 Using total station, it is very difficult to measure tree heights. How do you obtain the tree height?

7 in table 5, how do you calculate the results of RMSE, R² and MEA? ensemble models? What is RMSE SD?  Does it mean the SD of RMSEs extracted from base models?  These results should be more clearly described for readers.

8 Figure 5 and Figure 6, based on the title of figure, six diversity indices should be plotted for each remote sensing data. However, it is rather puzzled for readers to judge these plots extracted from Sentienl-2, RapidEye, or simulated LiDAR.

Author Response

Comments and Suggestions for Authors

 

Reviewer 2

In this manuscript, various remote sensing images are used for estimating tropical forest diversity with stacking ensemble learning. The experiment approach and results are extremely interested. However, it is hard to see any scheme in this piece of writing; it is very confused for readers, specially for the parts of introduction and results. I suggest that the structure of manuscript should be modified, firstly.

 

We have greatly modified the structure of the manuscript in the revision submitted to better order and outline each aspect of the study.

 

1 Table 1 requires supplementary references about these Species diversity indices.

Agreed, we have added reference for each of these.

 

2 Line 137, how to understand the sentence “Tree species, stem diameter, and height were collected within a 9-ha plot on 225”?

Yes, it is a little unclear as to how the 9-ha study site is set up for tree measurements. We revised this sentence to indicate that the 9-ha area was sub-divided into 225, 20m x 20m (400 m²) plots for making systematic tree measurements and orientation while inside the study area. We hope it is clearer with these changes. 

 

3 line 147, All trees were marked with metal tags for relocation and measurement

We have made this chance to the sentence.

 

4 figure 2, figure 3 and figure 4 don’t showed in the paper.

We are not sure what happened to the figures that were embedded in the manuscript originally submitted. Something may have happened to them while submitting the paper or while transferring them to reviewers. In any case, we also found them missing from the manuscript copy we downloaded with reviewer comments. We have added these back to the revised manuscript submitted.

 

5 For the part of 2.4 Introduction of sensors, remote sensing images and extracted features need to be further refined.

Agreed. We have made several refinements to this section. The Materials and Methods section has been reorganized and revised according to this and other reviewer comments. To better outline data development and analytical section we have ordered separate sections that are 1) Study area, 2) Tree data, 3) Remotely sensed data, 4) Diversity indices, 5) Sample design, 6) Ensemble models, and 7) Data analysis and model validation. We added a new paragraph to the bottom of the Sample design section to better explain how all 0.10-ha plots and remotely sensed data were combined for use with modeling, validation, and spatially explicit mapping. We have also added a work flow figure showing data development, modeling and analysis steps.

 

6 Using total station, it is very difficult to measure tree heights. How do you obtain the tree height?

Agreed. Tree height was measured separately using a TruPulse 360° compact electronic laser hypsometer, from the base to the highest point of the tree. We have added this information to the Materials and Methods Section.

 

7 in table 5, how do you calculate the results of RMSE, R² and MEA? ensemble models? What is RMSE SD?  Does it mean the SD of RMSEs extracted from base models?  These results should be more clearly described for readers.

These training model performance statistics are only reported in Table 4, which is probably what is meant by this comment. We explain in the caption that error statistics are generated from 10-fold cross validation (CV) resamples as part of ensemble model training. Technically, model control functions such as cross validation resampling are called independently from the ensemble model syntax. So, ensemble model performance statistics are independent of the base learners at this point. To make this clearer, we have re-written the Table 4 caption and revised methods creating a separate and improve 2.6. Ensemble models section. A new workflow Figure 5 has also been added to show the steps taken.

 

8 Figure 5 and Figure 6, based on the title of figure, six diversity indices should be plotted for each remote sensing data. However, it is rather puzzled for readers to judge these plots extracted from Sentienl-2, RapidEye, or simulated LiDAR.

 

Agreed and thank you for this comment. Since we give all the model details in Table 4 and Table 5, we present scatterplots from only the combined sensor models which had better performance in nearly all cases. We have reworded Figure 5 and Figure 6 captions to clarify that only combined ensemble model predictions are compared with observed validation data for each α-diversity index in these figures.

 

 

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Authors,

Thank you very much for submitting your work "Model Ensembles for Mapping Tree Species Diversity with 2 Multi-source Satellite Data in an Ecuadorian Dry Tropical Forest" to the Remote Sensing journal. The manuscript is timely and falls within the scope of the journal. I perused the manuscript, and in general, it reads well. The authors used in-situ collected tree-level geospatial data, other variables, and multiple satellite data to map tree species diversity in Ecuadorian tropical dry forest (Is it dry tropical forest or tropical dry forest? The latter seems natural to me). The author's approach is valient; however, at this stage marred by several methodological inconsistencies in terms of logical flow and coherency. In general, the introduction needs to be trimmed by 10- 15% and should build the story as to why the author's study deserves to be explored/investigated to meet their objectives. Methods and results need major shuffling and fine-tuning of the story. Too many tables and figures may hinder readers from getting the most out of this manuscript. Therefore, I am suggesting a major revision. The attached pdf (remotesensing-2057203-peer-review1) has my comments/suggestion for authors.

Regards,

 

Reviewer

Comments for author File: Comments.pdf

Author Response

Comments and Suggestions for Authors

 

Reviewer 3

Thank you very much for submitting your work "Model Ensembles for Mapping Tree Species Diversity with 2 Multi-source Satellite Data in an Ecuadorian Dry Tropical Forest" to the Remote Sensing journal. The manuscript is timely and falls within the scope of the journal. I perused the manuscript, and in general, it reads well. The authors used in-situ collected tree-level geospatial data, other variables, and multiple satellite data to map tree species diversity in Ecuadorian tropical dry forest (Is it dry tropical forest or tropical dry forest? The latter seems natural to me). The author's approach is valient; however, at this stage marred by several methodological inconsistencies in terms of logical flow and coherency. In general, the introduction needs to be trimmed by 10- 15% and should build the story as to why the author's study deserves to be explored/investigated to meet their objectives. Methods and results need major shuffling and fine-tuning of the story. Too many tables and figures may hinder readers from getting the most out of this manuscript. Therefore, I am suggesting a major revision. The attached pdf (remotesensing-2057203-peer-review1) has my comments/suggestion for authors.

 

We have addressed the general critiques in the above comments and made specific changes according to comments in the original article submitted. These we address line by line below. 

 

Lines 2-3: Change title to better describe your results. It currently highlights satellite data. Also, usually people use alpha (α) rather than symbol while searching either change here in title or change in keywords.

My suggestions.

 

Ensembles machine learning models for mapping tree species diversity with multi-source ground and satellite data in Ecuadorian Tropical Dry Forests

 

Ensemble machine learning models for mapping alpha diversity of tree species using multi-source satellite data in an Ecuadorian tropical dry forest

We appreciate recommendations to revise the title as follows:

 

 “Ensemble machine learning for mapping tree species alpha-diversity using multi-source satellite data in an Ecuadorian Seasonally Dry Forest”.

 

Our approach applies a stacked ensemble of separate and moderately tuned ML model types that are combined into a single meta-model for prediction. This aspect of the modeling is hopefully better reflected in the title. We also address this with further changes in the main text, in response to other comments. Hopefully adding “Seasonally Dry” to the title also helps readers better understand the type of topical forest studied. We have also added a few sentences to the first paragraph of the Introduction to better describe the tropical forest type studied.    

 

Line 27: It would mislead readers in thinking that LiDAR data were used.  However, it wasn't the case. Either delete this or say simulated LiDAR

Good suggestion. We changed the keyword to read “simulated LiDAR”.

 

Line 45: Why GEDI? As stated, all resolutions have been improved (spatial, temporal, radiometric, and spectral), can they be used for similar task? basically differentiation of 2D vs 3D data or complementarity.

To better outline the rational for bringing together these sensor types, we changed the next sentence on starting on line 45 to read:

 

 “The Global Ecosystem Dynamic Investigation (GEDI) satellite mission and waveform light detection and ranging (LiDAR) provide complementary measurements of vertical forest canopy structure linked to tree diversity for all tropical environments [15, 16]”

 

Hopefully this change and the citations given are sufficient rational for the multi-sensor approach taken and multiple data sources used are meant to be complementary.

 

Line 84: It appears, authors are trying to describe methods. Can you make your introduction short and put this table in method? Why is this (S), and, numerator in Species richness (D) equation is s? (lower-case s vs upper case)

Agreed, we have moved Table 1 to section 2.4. of Methods to describe each of these. Also, we have changed the species richness equation to show that S is the number of species occurrences in a sample that is referenced in the Piélous J evenness equation.

 

Line 93: Somewhere, I saw the space-borne (GEDI), and air-borne lidar data were not available for your study area? isn't it.

To clarify our need to simulate LiDAR variables, we revised this sentence to read:

 

“Because of inadequate data quality and the small extent of our study area, we could not use GEDI waveform LiDAR and airborne data were not available”.

 

Hopefully this is sufficient explanation that there were no airborne LiDAR for the study area. There were also very few GEDI LiDAR shots fully within the 9-ha plot (n = 10, and cover 25m radius). We show the GEDI shot locations in Supplementary 5, Figure 6. Most of these in the study area had a low data quality flag, likely because of dense cloud cover when data were collected. 

 

Line 147: I would simply say field survey, to sound simple. Or say reconnaissance survey

Agreed. We changed the wording to “field survey”.

 

Lines 159-161: Why is this special treatment for your testing data and not for training data?

Good question. We have changed terms to reference the different simulated plots used throughout the manuscript and show their treatment in the new workflow Figure 5. Training and testing data are the 55 non-overlapping samples used for model development. These are resampled using 10-fold cross validation during model development. The other simulated plots are left out of model development for validation purposes. We have refined methods adding the section 2.5. Sample design to better describe how and why these were developed differently. Random validation plots were better shuffled with the 3 sampling iterations (20 validation plots each time) because there is more space in the study area to randomly allocate samples at a 30m distance to reduce direct overlap with Testing/Training plots. This is not better explained in the Sample design section.

 

Lines 163 – 164: I would rather use a informed distance from variogram (range) as a threshold to calculate training and testing plots.

A semivariance approach would also work well here, potentially using a similarity index as we have with the correlogram approach. We chose to use permutational and multivariate Mantel correlograms that have shown to be an equally powerful tool for evaluating spatial autocorrelation with ecological data (Borcard and Legendre 2012).

 

Borcard, D.; Legendre, P. Is the Mantel correlogram powerful enough to be useful in ecological analysis? A simulation study. Ecology, 2012, 93, 1473-1481.

 

Lines 167 – 172 and Lines 194 - 196: Where are they?

Good question. We do not know how these figures disappeared, but they were original version of the manuscript submitted. We have added them back to the revised version submitted.

 

Line 198: I would place this part before as 2.3.

We have made this change in the revised manuscript submitted as well as re-organized section following the outline below. 

 

Up until there authors outlined how they have collected trees along with sampling design. I suggest authors to first describe

 

1) how they collected tree-level data,

 

We have added further details about tree measurements, such as instrumentation used to measure tree heights and diameters in the Tree data section. Sampling and plot design methods have also been published from previous studies that we referenced here.  

 

2) how they used Leica system - how does it work in terms of collecting 3-D clouds?

The Leica total station does not collect 3-D point cloud data. It was only used as a surveying instrument to record the UTM coordinate and elevation of each tree. Tree height data were taken separately using a laser range finder. We have added more details on how these measures were taken in the Tree data section to help clarify.

 

3) remote sensing data and only data

 

Done.

 

4) diversity indices and vegetation indices.

 

Done.

 

5) sampling design (comparing plot and satellite-based indices)

 

Done.

 

6) Data analysis

 

Done. We have reorganized Methods according to the outline above, as suggested by the reviewer.

 

6.1 In the data analysis part, mention what features/variables were used and how they were related.

All features developed were used in the feature selection and ML modeling process. However, we did not specifically explain how remotely sensed data and the training and validations samples were brought together, which was commented on by another reviewer. We have added a paragraph about this to the bottom of the Study design section. In addition, we have revised the Data analysis section to better describe model development, validation, and other analyses. We’ve added workflow Figure 5 also to a separate section 2.6. Ensemble models as suggested by the reviewer.

 

6.2 I would only put what equations/models were used.

We have made this change to in the data analysis section, describing only the model types used.

 

Use of 10-fold cross-validation is not required or not fully explained why. If separate datasets were created for testing samples why 10-fold cross-validation? Not clear.

We have revised the Data analysis section to help clarify the approach taken so that it is clear what level of testing and validation being developed to evaluate model performance and why. 

Admittedly, the study area is very small with potential for overlap and spatial autocorrelation between data used and held out of modeling. We believe it was important to devise and report model performance with the different methods for sampling the tree data. It is also customary to report model training performance from 10-fold CV outputs that use all sample data with replacement. An advantage of resampling training and test sample data is that there was no spatial overlap between these samples, but some potential for spatial autocorrelation among samples. We considered that model validation from data left out of training was also important to our analyses but would partially overlap with training/testing inputs. The two different treatments of validation data that we distinguish as random and regularly spaced (60m x 60m) 0.10 ha plots was used to determine which of the three sensor models may consistently perform better than others. We’ve added a few further details to describe each stage of data and model development in the Sample design and Data analysis sections to better clarify their purpose.

 

I haven't personally used the "caret" package for ensemble machine learning. At least for categorical variables ensemble approach is only valid for binary response variable. However, for continuous data (regression case) I don't know.

We are using all ML models in regression mode that is automatically assumed with ‘caret’ model implementation using a continuous response variable. It is common to get a warning message in R and associated ML packages when a categorical variable (e.g., a factor or integer with 2 levels) is treated as a continuous variable.

 

6.3 ) I suggest naming equations x,y,z without diving too much into the bagging boosting now. (We used heterogeneous models (this is the key to ensemble machine learning).

We’ve reorganized these section and paragraphs to separate modeling procedures from validation and other analyses. We moved a paragraph describing each ML method to the very bottom of the Ensemble model section. 

 

6.4) Model validation (validation statistics).

Done.

 

I highly doubt that any variables were removed by RFE function.

That is not often the case with features selection, but sometimes all predictors are retained depending on the data inputs. We have added a column to Table 4 showing the number of variables used and retained from RFE optimization. From this and previous work, we have found RFE to be an effective method for variable selection as compared to other approaches (e.g., Boruta). Some recent comparisons seem to point to Boruta as the best method for feature selection, but we have not found that to be the case, as many variables are often retained with that would be dropped using RFE.  

 

Lines 214-216: Write short and direct sentence. ndvi vs rndvi is already a different terminology with equation in table.

We have shortened the sentence to read as follows:

 

“We used further band combinations for leaf-on and leaf-off images interchanging red with red-edge bands that may enhance differences in tree canopy reflectance (Table 1)”.

 

Lines 271-218:  Data source? all equation/description based on Sentinel?

We have changed to the caption to clarify that bands and VI equations are based on Sentinel-2 imagery only:

 

“Table 2. Spectral bands and VI according to Sentinel-2 satellite imagery from leaf-on and leaf-off periods to be used as predictor variables in ensemble models of α-diversity measures.”

 

Line 221-230: Can be shortened, providing info on variety of sentinel bands at spatial resolution of 10 - 60 m and referring to table should be sufficient. Lengthy description of original data seems obscuring original point. Table 3 does not correspond with what authors are referring in this text.

We agree and have shortened this paragraph and corrected the references to Table 2 and Table 3. 

 

Line 232: Merge this paragraph with previous one, only highlighting key points.

We have made this change and moved the reference to Supplementary 2 to the bottom of a separate paragraph describing RapidEye satellite imagery.

 

Line 241-245: First sentence is hard to read. I believe separate paragraphs for each sensor would be better to communicate with readers. I would also suggest, I haven't looked at the latter part, different vegetation indices /spectral indices from each sensor would provide more clarity as to what data set the information came from. Such as for RapidEye based NDVI (R sub ndvi or Rndvi and Sentinel based NDVI as S_ndvi...so on.

We have revised this sentence to read as follows:

 

“As with RapidEye, we further developed Sentinel-2 VI (Table 3) interchanging red-edge for red bands in addition to developing indices with a wider dynamic range important under dense forest canopy conditions [46].”

 

Agreed, it is a good idea to add a prefix to distinguish between Sentinel-2 and RapidEye bands and VI. However, in the original version of the manuscript we distinguish RapidEye bands and VI by adding an “r.” to each variable as shown in Figure 8 (now Figure 9). We believe that using an additional prefix for Sentinel-2 bands and VI would add confusion to an already complicated set of variable names.

 

Lines 250 – 253: Second last paragraph of Introduction, authors invoke the issue of quality and study area.

We have clarified our use of tree data to serve as LiDAR analogues by shortening and combining the second and thirds sentences of the paragraph to read as follows:

 

“Because of insufficient LiDAR for the study area, we used tree height and elevation measurements as information analogous to LiDAR data from a total of 5129 tree UTM coordinates that were well distributed across the study area.”   

 

Line 332: reconnaissance survey?

We changed this to “field survey” to be consistent with a change in Methods.

 

Line 334-336: Why? If goal was to compare these two; field survey could have done to take cluster of trees. Authors earlier mention tree coordinates were precise (<5 cm). If so. Just based on the coordinates raster values could have extracted. Isn't it?

 

We agree with the reviewer that it would have been useful to have LAI measures clustered in areas across the study site to compare with individual Sentinel-2 pixels. We were not able to do this because of time and logistical constraints. We have revised our description of these measures and comparisons in Methods and Results to clarify their purpose. These measurements were designed to systematically assess general LAI conditions across the 9-ha study area traveling between georeferenced sub-plot corners to evaluate disturbance at the same time. The subplot corners were relatively easy to locate, most of the time, and helped to stay oriented inside the study area. We anticipated having LAI for all sub-plot corners in the study area. Upon implementation, the field LAI measures took much longer than expected and we were able to complete only a reduced but representative sample across the middle portion of the 9-ha area (six transects, 20m apart, at 75 subplot corners). We believe that these data help to establish that Sentinel-2 LAI measures developed from the biophysical processing tool in ESA’s SNAP software generally compare well with forest conditions on the ground. 

 

Lines 371-373: How the ensemble models was run? did you use same equation as a meta-learner for all models? did you include all base-models in ensemble models? It is also not clear how the ensemble machine learning models was created.

 

Thank you for pointing this out and we agree that it is necessary to clarify exactly what was done. We have revised our description of the approach in the second paragraph of the Data analysis section to specify more clearly how models were developed and evaluated. We recognize that it is not always easy to understand precisely what is going on with the varied ensemble methods implemented in open-source software packages. In the Ensemble model section, we have revised text to make specific reference to the R packages used and model functions at each step that is now shown more clearly in the Figure 5 workflow diagram. 

 

Line 337: Can you put info this and next table into one using important information (adj.R2, error rate, and p value only. These figures and tables (run-ons) are difficult to summarize information.

 

Thank you for this suggestion. As recommended, we combined the two tables into as single table but thought that it was important to retain equal reporting for both validation data sets. Therefore, we changed column headings and dropped the Error Rate calculations that were not as essential to our analysis. This will help condense methods and results reporting also.

 

Line 382-387: I suggest authors use scatter plot for sensor model that was better in terms of predicting such variables. Others, if needed, can be placed as supplementary materials.

 

We are keeping scatter plots from combined sensor models as is since they performed consistently better, but not always at each test and validation step. We also reviewed error and goodness of fit statistics in Tables 4 and 5 and did not find a case where a single sensor model was shown to be consistently better than the combined models. For example, the combined model for Piélou’s J showed better performance with training and randomly distributed validation plots, but slightly lower R² values than single sensor models with the evenly spaced validation plots. All the model outcomes and comparisons are in Tables 4 and 5. We mainly used scatter plots from the combined models in Figure 6 and Figure 7 to show how the two validation data sets look when comparing observed and predicted values.

 

Lines 403-407: I highly doubt it. If it was so strongly correlation, why did RFE not flag this variable in the first place? How accurate the simulated Lidar canopy heights ware? If multiple points for a tree were evaluated then, average slope of tree or crown radius etc would be a proxy. In same trees spatial variation is expected, which is not a big deal in satellite image if entire tree falls within a pixel.

 

Understood, but we do not state that elevation and tree height are strongly correlated, nor do we state that there are multiple points from any one tree. We indicated that average height and elevation are positively correlated only (r = 0.54). We have added further information in Methods so that it is clearer how LiDAR analogs were simulated. We do point out that species richness showed a strong negative correlation with tree density (r = -0.74) that is positively related to tree height (r = 0.74), which is probably why average tree height does show up as important in this model as does minimum height in the Shannon’s H′ model (but pretty low on the list). RFE is unlikely to drop variables such as these, that are moderately correlated but add explanatory power. In our study area, there are many more trees per hectare and species on the low elevation sites that are in a transitional area between Tumbesian shrub and deciduous tree dominated formations. Even though all or most of the ecological reserve is on relatively level terrain, small changes in local relief and elevation can have a large impact on tree density and diversity. These conditions are described in the second paragraph of the Discussion. In our models and study area, elevation just happens to be a better variable than LiDAR height. However, tree heights and other canopy profile metrics from GEDI or airborne LiDAR would likely add to remote sensing models of tree diversity, as has been shown elsewhere. We’ve added analyses with and without the use of elevation data in Supplementary materials (Supplementary 5) to help indicate the importance of tree height data for estimating species richness.

 

Lines 438 – 441. Isn't this expected? isn't alpha-diversity measures of species diversity?

Yes, this is true and should be expected, which we point out in the Discussion. We could probably delete rows two and three of Table 6 (Mantel tests) but keeping them for now. We are just being consistent in using multivariate Mantel and partial Mantel tests. The first row examines if distance apart is a factor or if there is a strong or consistent gradient with distance, which there was not. In the second and third rows, if α-diversity were not significantly related to species composition (even when controlling geographic distance) we would want to check the data.

 

Line 460 - Try avoiding table/figure reference in the discussion

Agreed, we have deleted the reference to Table 5.

 

Line 512-513. is this true from Figure 8, especially in case of lidar data?

Yes, good point. Not so true from Figure 8 (now Figure 9). We do think that this is still true from running the species richness model without elevation variables and finding that LiDAR minimum height becomes the top predictor with little or no loss of explanatory power. We are leaving these comparisons in Supplementary material (Supplementary 5) for now to avoiding lengthening the article but believe this part of the discussion is sufficiently supported by these assessments. 

 

Line 594: In ensemble model knowledge of domain is important, without extensive parameter tuning may hint that authors used bunch of variables without knowing what variables could potentially explain variance. As I explain, parameter tuning is can still be important as ensemble model does  not always guarantee best model prediction. 

 

True. We made a small clarification that it is the hyperparameters we mean here. GBM, XGB linear and XGB tree models have several tuning parameters that require several iterations to optimize (e.g., grid search for number of boosting iterations, shrinkage, max depth etc.), which can sometimes lead to over fitting (e.g., excellent goodness of fit, but very high model error) or greater computation time. Random forest only has a few parameters (number of trees and variables tried at nodes) and are not prone to overfitting as with boosting, which can sometimes outcompete the ensemble approach as you mention. In our view, it is more practical to develop moderately tuned base models that are then weighted for prediction. It is also clear from the references cited that there is good precedent for developing predictors from the multiple sensor types, VI, biophysical variables and simulated data we have chosen. ML methods can seem like a sausage grinder with so many variables, however efforts to use hyperspectral remotely sensed data with minimum noise fraction/principal components, ‘spectral species’ and similarity indices for data reduction can look very similar. 

 

Author Response File: Author Response.docx

Reviewer 4 Report

Dear Sirs,

I much appreciated you paper for its methodological rigour and the high detail applied in description of methods and results. Anyway, just this plethora of analyses and figures makes partly the presentation of data slightly confused, and the designation of tables and figures (same numbers of the main paper) in supplementary material enhances this impression. Besides, I think it is a misprint, but Figures 2 and 3 in Methods and Figure 4 in Results are lacking, although their captions are present. I think that this considerable work still needs a work that still need to be focused on some fundamental features, al least at level of data selection and presentation. Some minor criticisms are reported in attached files.

I suggest a reconsideration after major revision.

Best regards

Comments for author File: Comments.pdf

Author Response

Comments and Suggestions for Authors

 

Reviewer 4

 

Dear Sirs,

I much appreciated you paper for its methodological rigour and the high detail applied in description of methods and results. Anyway, just this plethora of analyses and figures makes partly the presentation of data slightly confused, and the designation of tables and figures (same numbers of the main paper) in supplementary material enhances this impression. Besides, I think it is a misprint, but Figures 2 and 3 in Methods and Figure 4 in Results are lacking, although their captions are present. I think that this considerable work still needs a work that still need to be focused on some fundamental features, al least at level of data selection and presentation. Some minor criticisms are reported in attached files.

 

We are not sure what happened to the figures that were embedded in the manuscript originally submitted. Something may have happened to them while submitting the paper or while transferring them to reviewers. In any case, we also found them missing from the manuscript copy we downloaded with reviewer comments. We have added these back to the revised manuscript submitted.

 

Lines 31-41: Here you explain that SDTF need protection and remote sensing analysis, but you do not describe the main features and floristic composition of such forests, at least in summary. You must consider that it is a very peculiar type, maybe not known by a certain part of readers. Even if you give more details in the following Materials and Methods

Good suggestion. We have added the following sentences to the first paragraph:

 

“Banda et al. [5] defines five distinct floristic groups of SDTF for northern South America. Groups differ by woody plant composition, but all contain deciduous vegetation that shed leaves for 3 to 5 months a year during periods with <100 mm of rainfall per month. Ecuador itself contains three distinct SDTFs identified as Piedmont Tarapo-to-Quillabamba, Central Andes Coast and Central inter-Andean Valleys floristic groups [5], that can be further subdivided according to sub-regional and endemic flora [9].”

 

Line 84 (Table 1) - The absence of line spaces renders confusing table reading. Please distribute adequately.

 

Agreed. Hopefully this will be taken care up with formatting in the final publication. However, we expanded the Description column and added line divisions between each row to improve readability. 

 

Lines 217 and 268 - According to text, maybe you have mistaken Sentinel-2 and  RapidEye in Tables 2 and 3?

Thank you, that is right. We have corrected the Figures referenced in the text so that it applies to the right table.

 

Lines 582-611 - Such Conclusions, so extensive, seem enough to another part of Discussion. This session needs more synthesis and assertiveness

Thank you. We made changes in Conclusions to assert key findings and outcomes. However, the conclusions seem reasonable in length and state our main findings.

 

 

 

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The revised manuscript has been greatly improved. Before publication, two suggestions should be considered:

(1) In the part of 2.3, common descriptions of remote sensing data, such as widely used RapidEye and Sentinel-2, can be reduced.

(2) In the Conclusions, the main conclusions need to be further summarized and clearly expressed

Author Response

Reviewer 2

The revised manuscript has been greatly improved. Before publication, two suggestions should be considered:

• In the part of 2.3, common descriptions of remote sensing data, such as widely used RapidEye and Sentinel-2, can be reduced.

Good advice. We have reduced our description of the two sensor types to the greatest degree possible without losing information on our treatment of these data sources.

(2) In the Conclusions, the main conclusions need to be further summarized and clearly expressed

Agreed. We have added a small amount of detail and further refined Conclusions to represent main findings from the study more clearly.

Reviewer 3 Report

Dear Authors,

The revised version of the manuscript significantly improved informed by comments by reviewers. Almost all comments/suggestions were followed. I still am not sure about the choice of words on lidar proxy using only one set of observations on each tree. Ensemble models developing with moderately tuned hyperparameters and potential autocorrelation issues that seem to be evident should be discussed/presented differently. My comments and concerns are in the attached document.

Comments for author File: Comments.docx

Author Response

attached please find the response.

Author Response File: Author Response.pdf

Reviewer 4 Report

Dear Sirs,

I read the second version of your paper and I argued that you added lacking figures, giving a complete outlook of your work, as well as suggestions by the other commendable reviewers. I think your paper can now be published in the present form.

Best regards

Author Response

Reviewer 4

I read the second version of your paper and I argued that you added lacking figures, giving a complete outlook of your work, as well as suggestions by the other commendable reviewers. I think your paper can now be published in the present form.

Best regards

Thank you. We have made a small number of additional changes to reduce the Methods section and clarify principal findings in the Conclusions.

Round 3

Reviewer 2 Report

The revised manuscript has been greatly improved, and I suggest to publication now.

Author Response

We have further revised Conclusions to better state principal findings and outcomes from the study. 

Reviewer 3 Report

Dear Authors,

The revised version of the manuscript, judging from highlighted text, seems improved and technically sound. I can understand the behavior of RFE on feature selection.

If spatial autocorrelation was not present then validation statistics on regular-spaced and randomly selected data should have shown competitive statistics. Acknowledging such issues informs the reader about potential issues; this is fine. 

The manuscript at this stage is almost ready for acceptance. I would suggest removing for example part from the conclusion. It looks more like results and discussion. Try generalizing the conclusion, but do not overgeneralize it as was in the first draft. 

 

Author Response

We greatly appreciate critiques provided by this reviewer, which have improved the articles contents, organization, and reporting.  

If spatial autocorrelation was not present then validation statistics on regular-spaced and randomly selected data should have shown competitive statistics. Acknowledging such issues informs the reader about potential issues; this is fine. 

At the bottom of the third paragraph in results, we have added a sentence to further acknowledge spatial correlation between random validation plots that can impact comparisons in such a small study area as ours. 

"Some spatial correlation among randomly placed validation plots in the 9-ha study area likely contributed to greater overall model performance (i.e., adj. R²) observed from these comparisons."

In the previous revision, we've also provided an explanation at the bottom of the second paragraph of Result about why this was most important to models including Sentinel-2 bands, which have a 10m to 20m pixel resolution, which did affected closely spaced plots regardless of resampling. 

The manuscript at this stage is almost ready for acceptance. I would suggest removing for example part from the conclusion. It looks more like results and discussion. Try generalizing the conclusion, but do not overgeneralize it as was in the first draft. 

We have further revised Conclusions to better highlight principal findings and outcomes from the study. We have retained some detail, but eliminated most others that are more thoroughly covered in the Results and Discussion