Mapping of the Canopy Openings in Mixed Beech–Fir Forest at Sentinel-2 Subpixel Level Using UAV and Machine Learning Approach
Round 1
Reviewer 1 Report
This paper is aboud upscaling information by mixing reflectance data from Sentinel 2 and from UAV to see if canopy openness percent can be inferred from indices.
Some parts are not clear and must be improved.
The foremost critical part, as also expressed by the authors in section "2.4. Integration of Sentinel 2 and UAV data", is the integration of 10 m imagery from S2 with UAV imagery which is commonly has centimetric resolution. It is quite trivial to undersand that even considering 10 cm resolution of UAV imagery, this means about 100x100 = 10 000 UAV image cells inside each single S2 cell. Authors report more than twice the number of cells due to UAV resolution. The use of R for very large rasters requires careful strategies because many raster packages convert to R data frame format, increasing memory requirements drastically. From line 270 forward you talk about indices, but some of them are not reported in table 2 (NGRDI, VARI...) - why?
Line 205 - "cost-free software " ... SNAP is actually open-source software, so you can say "open-source" which is much more correct than "cost-free".
Line 281 - what do you mean a "fishnet grid"? do you mean a regular grid?
Line 294 - it is not clear which are the 155 predictors... I understand indices are used, but how do you reach so many? It should be clearly explained in this part how you reach 155 predictors.
Line 307 - citation of Kuhn and Johnson 2013 not in correct format
Line 409 - what is the use of visually examine 155 scatterplots ? The negative correlation can be clearly assessed by a coefficient of correlation R.
It seems that the COP is only weakly correlated with tested indices, likely due to the sensitivity of the indices to vegetatation that could also be present or absent in the background. This adds to mixing the signal that is tested for correlation. This is actually clear in figure 6c.
Figure 7c needs changes. White background (theme_bw() in case you use R) and also a 1:1 ratio with also a line with slope 1 would improve reading and comparing (see observed vs. predicted plots in other literature).
What do you think is the impact of forest density? For example could another factor like Hansens' forest cover map (www.sciencemag.org/content/342/6160/850) improve your result by acting as covariate? Or maybe you can use Hansen map to compare with your COP results (Figure 5). Another important test could be to verify if there is correlation between residuals and forest cover from this independent dataset.
Author Response
This paper is aboud upscaling information by mixing reflectance data from Sentinel 2 and from UAV to see if canopy openness percent can be inferred from indices.
Some parts are not clear and must be improved.
The foremost critical part, as also expressed by the authors in section "2.4. Integration of Sentinel 2 and UAV data", is the integration of 10 m imagery from S2 with UAV imagery which is commonly has centimetric resolution. It is quite trivial to undersand that even considering 10 cm resolution of UAV imagery, this means about 100x100 = 10 000 UAV image cells inside each single S2 cell. Authors report more than twice the number of cells due to UAV resolution. The use of R for very large rasters requires careful strategies because many raster packages convert to R data frame format, increasing memory requirements drastically. From line 270 forward you talk about indices, but some of them are not reported in table 2 (NGRDI, VARI...) - why?
Response 1: The spatial resolution of the UAV images was approximately 7 cm, i.e. 143 times finer than the S2_L2A; 1000 cm (S-2) / 7 cm (UAV) = 142.86 (143) ; 143x143 = 20499 times. The square dimension can be deceiving, so it is correct that with a resolution of 7 cm we obtain twice as many pixels on a 10m (Sentinel-2) grid as with a resolution of 10 cm. Probably it is the most important to perceive the very large disparity in resolutions (approx. 20000 cells), of two used sensors, and to harmonize these two scales were likely not so trivial task.
Only satellite indices offered by default in ESA SNAP Version 7.0 software for optical (Sentinel-2) image processing are considered. An algorithm (so-called processor) is given for each index used in SNAP, which enables their simple calculation. We are aware that in practice there are many also usable indices, such as the ones you mentioned, but we think that the indices included in the SNAP already sufficiently represent the variability of the forest cover that was analyzed in this research.
Line 205 - "cost-free software " ... SNAP is actually open-source software, so you can say "open-source" which is much more correct than "cost-free".
Response 2: We provided this change in text.
Line 281 - what do you mean a "fishnet grid"? do you mean a regular grid?
Response 3: These are quite complementary terms; the fishnet grid is a regular grid. Probably the Fishnet grid is a little more specific, in software like ArcGIS Fishnet presents a feature class containing a net of rectangular cells. We think that both terms are correct, but Fishnet is slightly more distinct. We wanted to say that it was a regular polygon grid without attributes. We added this small explanation in the text.
Line 294 - it is not clear which are the 155 predictors... I understand indices are used, but how do you reach so many? It should be clearly explained in this part how you reach 155 predictors.
Response 3: Explanation regarding the number of predictors is included in Line 263:
For each of the considered satellite images, (S2_20180725, S2_20180801, S2_20180829, S2_20180928) a total of 155 predictors were calculated; 4 Soil Radiometric Indices, 21 Vegetation Radiometric Indices, 5 Water Radiometric Indices, 5 Biophysical Indices, 10 Grey Level Co-occurrence Matrix (GLCM) parameter for each of S-2 band (B1-B12), that means 120 GLCM layers per S-2 image.
Line 307 - citation of Kuhn and Johnson 2013 not in correct format
Response 4: The reference is corrected in text.
Line 409 - what is the use of visually examine 155 scatterplots ? The negative correlation can be clearly assessed by a coefficient of correlation R.
Response 5: We agree that we were able to determine these relationships much easier using the correlation coefficients, which we preliminary also calculated, but we did not find a suitable way to present such a large number of variables and not to clog the manuscript. Instead, we decided to place more emphasis on the representation of R2 as to some point similar measure for the representation of the relationship between the variables.
In this case, the inclusion of R2 from the very beginning made it possible to monitor the improvement of the model building process from the very beginning i.e., from the process of transformation of input variables or S-2 bands, what is explained in the Discussion:
L663-L668 The model building process brought about the improvement of predictive performance on the training set in successive steps; (I) the highest R2 of the single feature (derived satellite index) using simple linear regression was up to 0.57, (II) the highest R2 using multiple features obtained from the single date, S-2 image and the best ML algorithm was 0.624, and (III) the highest R2 on the multi-temporal set of four consecutive S-2 images, using the best ML algorithm reached 0.697.
Another reason why we analyzed the relationships between COP and Satellite Indices graphically using local regression (Loess) is that we wanted to determine the character of the relationship between variables i.e. whether it is linear or not. This ultimately helped us to explain why Elastic Net, which belongs to the group of simpler linear models, showed such good performance. Namely, from the scatterplots shown, we can see that a linear relationship prevails between COP and the main predictors (ARVI, IPVI, NDVI, CI), which explains the efficiency of the selected model.
It seems that the COP is only weakly correlated with tested indices, likely due to the sensitivity of the indices to vegetatation that could also be present or absent in the background. This adds to mixing the signal that is tested for correlation. This is actually clear in figure 6c.
Response 6: The influence of the understory layer on surface reflectance measured by optical satellites is a common problem for forest stand discrimination. This has been observed and partly discussed in this paper, especially when extracting canopy masks from UAV RGB orthophoto images. At the investigated location, we have an additional problem with a large number of fallen trees, but with a still green canopy, which to some extent has an impact on the slightly lower overall accuracy of the best-obtained model (R2 = 0.7). On the other hand, the type of forest considered in this study usually does not have a dense layer of ground vegetation, while the soil is a very contrasting brown-red color, so this facilitates easier discrimination. Also, we think that by using a multi-temporal S-2 set that can detect differences in phenology between different types of vegetation cover, it was possible to distinguish canopy surface from ground vegetation because of slightly different phenological trajectories which is perhaps the main reason for observed significant improvement of the predictive performance of the multi-temporal set compared to single S-2 image.
Figure 7c needs changes. White background (theme_bw() in case you use R) and also a 1:1 ratio with also a line with slope 1 would improve reading and comparing (see observed vs. predicted plots in other literature).
Response 7: We agree and we adjusted all graphics in the paper accordingly.
What do you think is the impact of forest density? For example could another factor like Hansens' forest cover map (www.sciencemag.org/content/342/6160/850) improve your result by acting as covariate? Or maybe you can use Hansen map to compare with your COP results (Figure 5). Another important test could be to verify if there is correlation between residuals and forest cover from this independent dataset.
Response 8: Thank you for your question and comment. Here we have to emphasize the issue of the scale and decision-making level. We slightly improve the discussion and emphasize … that this work is related to the improvement of forestry planning and operational activities that could benefit from the availability of timely and accurate information on the state of forest cover. Please pay attention also to the rest of the included explanation in the discussion chapter for better clarification. So here we propose a solution based on the multi-sensor approach suitable for the operational forest management for the scale of the forest management unit (3000-6000 hectares) that is most common in Croatian forestry.
Larger scale forest cover maps, one that you mentioned and also pan-European tree density products from Copernicus are more suitable for the policy level on European or Global scale i.e. to obtain information about deforestation and related carbon emissions over the large areas (like Amazon or Kongo) and to commence global political processes that will mitigate this issue. In Europe, we can use tree density maps to asses regional problems such as the dieback of some main tree species caused by biotic factors (like ash dieback or spruce dieback). These scales are also suitable for more retrospective analysis back to last decade or two. In operational forestry, what is the most important, the end-users require actual, up to date overview of the forest cover, they also require very rapid information to see where to apply mechanization and human resources.
Regarding your last remark, based on our multiple computing attempts, it is not very likely that we could obtain some significant improvement from the residuals. However, what is important in this study is that the overall concept and approach that we implemented are feasible and satisfactory for the initial purpose. It also comprises trade-offs like relatively simple camera … Using a simple UAV system with only an RGB camera, which has some limitations in canopy segmentation, on the other hand, has advantages in wide operational applicability. Nowadays, systems such as DJI drones with RGB cameras are by far the most widespread of all UAV systems, while the processing of such images is far simpler than more complex multispectral, hyperspectral, thermal, and LiDAR images, which greatly facilitates the acquisition of information. We are aware that much better UAV systems exist which could obtain a significant improvement of canopy delineation (shadows, understory layer, fallen trees, etc.). Also, there are more advanced deep learning approaches that could make use of better and more accurate field data. However, in this study, we try to make a possible harmonization of the used methods what we also emphasize as a conclusion: the best performance shows a more general, simpler Elastic net model that is the most suitable for the simple data obtained from RGB imagery.
Reviewer 2 Report
Review comments: Manuscript ID:- remotesensing-1002424
General Comment
This is an interesting manuscript about a bi-sensor approach for mapping of forest canopy gaps. However, there are aspects that require improvement and clear presentation before being considering for publication. Specific comments and suggestions are included below.
Specific comments and suggestions
L58-59: “…ground measurments………These methods are labour intensive and cannot quickly provide accurate canopy cover estimates…..”, this is a strong claim. Usually ground measurements have better accuracy than other methods. Authors need to justify this claim.
L59-61: “…Recently, remote sensing methods, in particular Light Detection and Ranging (LiDAR), have been proven as the most advantageous approaches in terms of accuracy and cost efficiency …….”, there are other reports about the expensiveness of LiDAR for forest application as compared to others. Inclusive arguments also needed here.
L61-63: UAV based LiDAR is also criticized for its instability, which affects the accuracy.
L181-182: I suggest including explanation about the choice of survey month for this study.
L152-184: Authors need to explain why training and testing areas selected in two independent locations with different sizes. How to control other factors under such condition? Why not having the training and testing areas in one experimental locations? Which sampling strategy used for the training and testing sample selection?
L305-307: “….The 10-fold cross validation technique that was used concerning other resampling methods gives the best estimate of the actual RMSE and the most favorable trade-off between the bias and variance of the model…”, a 10-fold cross-validation approach is often criticized for validation approach. Authors need to justify their choice of this approach.
L409-413: those is description of methods. I suggest moving these to appropriate place under method section and directly present the findings here.
L599-727: the presented discussion is more of extensive explanation of the findings. As this is a methodological study, the results need to be discussed thoroughly. I suggest authors to reflect the following information during discussion: wow this study will increase our knowledge base and inspire others to conduct further research. how the results (not) support findings of earlier studies, whether your findings agree with current knowledge and expectations, any weaknesses in the results/approaches and suggest room for further research concerning that aspect of your analysis.
Author Response
General Comment
This is an interesting manuscript about a bi-sensor approach for mapping of forest canopy gaps. However, there are aspects that require improvement and clear presentation before being considering for publication. Specific comments and suggestions are included below.
Specific comments and suggestions
L58-59: “…ground measurments………These methods are labour intensive and cannot quickly provide accurate canopy cover estimates…..”, this is a strong claim. Usually ground measurements have better accuracy than other methods. Authors need to justify this claim.
Response 1: Here we refer directly to the work of Karhonen, L .; Karhonen, K.T .; Rautiainen, M .; Stenberg, P. Estimation of forest canopy cover: a Comparison 821 of Field Measurement Techniques. Silva Fenn., 2006, 40 (4), 577-588. DOI: 10.14214 / sf.315. At the beginning of their discussion, they, based on the results of the comparison of different terrestrial measurement methods they conducted, conclude that "The results of the comparison of measurement techniques confirm that the conventional methods cannot quickly provide accurate canopy cover estimates". This conclusion is also in line with previous experience in the field inventory of the authors of this article. We think that the accuracy of the field methods, which we agree are the most accurate concerning other measurement methods, is not questioned here. This statement primarily aims to say that these methods are labor-intensive and cannot quickly provide canopy cover estimates. We suggest the following modification of the above sentence:
Ground measurement methods, as the most accurate techniques, are labor-intensive and cannot quickly provide canopy cover estimates.
L59-61: “…Recently, remote sensing methods, in particular Light Detection and Ranging (LiDAR), have been proven as the most advantageous approaches in terms of accuracy and cost efficiency …….”, there are other reports about the expensiveness of LiDAR for forest application as compared to others. Inclusive arguments also needed here.
Response 2: We agree that this sentence is misguiding. The cost of airborne LiDAR is probably the biggest constrain why it is not used more frequently in the forest inventories. We emphasize here the UAV LiDAR cost efficiency but this technology is still in the experimental phase. So we rephrased the sentence:
Recently, remote sensing methods, in particular Light Detection and Ranging (LiDAR), have been proven as the most advantageous approaches in terms of accuracy and ease of data acquisition [18-21].
L61-63: UAV based LiDAR is also criticized for its instability, which affects the accuracy.
Response 3: We agree with this statement, we slightly modify the sentence:
UAV (Unmanned Aerial Vehicle) based LiDAR, despite observed shortcomings such as aircraft instability, has a lower cost, more convenient operation, and more flexible flight route design, as well as unique advantages in the possibilities of better discernment of the stand canopy details than the airborne LiDAR.
L181-182: I suggest including explanation about the choice of survey month for this study.
Response 4: We agree, we have included additional clarification:
The exact time of the survey, the end of July 2018, was selected due to the requirements of the forestry service. Since forest restoration activities at that time were carried out over the entire FMU (logging of remaining fallen trees after windthrows during the winter period and afforestation of the clearings) the interest of forestry service was focused on obtaining information on the spatial distribution of gaps throughout the FMU to perform the aforementioned silvicultural activities. For this purpose, a UAV survey was first performed at two independent locations, used in this study, where the occurrence of larger windthrows was detected. No particular sampling strategy or design was taken into account when selecting these two sites, only some preliminary knowledge about forest gap occurrence. After that, Sentinel 2 imagery was preliminarily used, which led to the idea of developing a new methodological approach that would combine both sensing methods, which is presented in this paper. For the survey, a forest service owned UAV was used with a simple RGB camera, which is otherwise more often used only for visual observation purposes.
L152-184: Authors need to explain why training and testing areas selected in two independent locations with different sizes. How to control other factors under such condition? Why not having the training and testing areas in one experimental locations? Which sampling strategy used for the training and testing sample selection?
Response 5: We partly explained these questions in the previous paragraph. We explained that sampling strategy was not preliminary planned and that the choice of the two locations was purely the judgement of the field personnel and UAV operator. We agree and provide more detailed discussion about this issue (701-722) that with preliminary planning of the UAV ground-truth locations in terms of the increasing number of (smaller) training/testing areas, distributed over the whole area, the most likely we could obtain more reliable prediction, what is important for the future similar studies.
We used training area only for the construction of the ML algorithms and evaluation of their performance using CV method. In addition, we used very remote testing area for additional model validation. However, for the purpose of the building of the final model and map creation, we joined the training and testing data, apply the chosen algorithm and constructed the final model on both areas. This is the most common approach in the similar cases of the prediction modelling.
L305-307: “….The 10-fold cross validation technique that was used concerning other resampling methods gives the best estimate of the actual RMSE and the most favorable trade-off between the bias and variance of the model…”, a 10-fold cross-validation approach is often criticized for validation approach. Authors need to justify their choice of this approach.
Response 6: We are referring here to the textbook from Kuhn and Johnson, Applied Predictive Modeling, Springer, 2013, p.615. (this citation is in the text). Their claim is based on a comparison between various resampling techniques; Leave One Out Cross Validation, Bootstrap, Repeated Training/Test splits, etc. provided by the authors. Our opinion is that there are no distinctive differences amongst resampling methods that would significantly improve the model building process in this study. k-fold cross-validation generally has high variance compared to other methods and, for this reason, might not be attractive. It should be said that for large training sets, the potential issues with variance and bias become negligible. However, it is one of the fastest resampling techniques what was very important for the training of complex algorithms used in this study.
We added additional explanation:
It is also one of the fastest resampling techniques what is important for the optimization of the computational processing time of the training of some more complex algorithms.
L409-413: those is description of methods. I suggest moving these to appropriate place under method section and directly present the findings here.
Response 5: We agree, we moved the paragraph in the method section.
L599-727: the presented discussion is more of extensive explanation of the findings. As this is a methodological study, the results need to be discussed thoroughly. I suggest authors to reflect the following information during discussion: wow this study will increase our knowledge base and inspire others to conduct further research. how the results (not) support findings of earlier studies, whether your findings agree with current knowledge and expectations, any weaknesses in the results/approaches and suggest room for further research concerning that aspect of your analysis.
Response 6: We improved the discussion accordingly:
The results of the research have shown that with the synergistic effect of two, in many respects very different EO systems, we can significantly improve the current practice of monitoring the state and disturbances of forest cover. This approach enables relatively fast and very precise insight of forests canopy over relatively larger areas. There are two particularly important aspects where the significant benefits of using the aforementioned approach could be identified. The first element is related to the improvement of forestry planning and operational activities that could take advantage of the availability of timely and accurate information on the locations and magnitude of forest gaps throughout the FMU. The second element is predominantly contribution in the remote sensing domain, that refers to the proxy way of improvement of the capabilities of Sentinel-2 for very fine gradation of canopy cover, at subpixel precision. The effectiveness of the combined use of these two sensors stems from the blending of their comparative advantages and the elimination of the weaknesses in each of them. Sentinel-2 is currently one of the most advanced satellite EO systems whose most important features are global coverage, high spectral resolution, and short revisit time, although its spatial resolution is too coarse for the discrimination of canopy openings smaller than 10 m. UAV systems, with ultra-high spatial resolutions and flexibility of manipulation, are nowadays unsurpassed in capabilities of precise detection of stand coverage, although they are limited in the operational area they can cover. By using the UAV system as ground-truth on pre-selected smaller training and validation areas and Sentinel-2 for wall-to-wall prediction with ML techniques, we extend in the surrogate fashion information from UAV over the whole FMU. Using a simple UAV system with only an RGB camera, which has some limitations in canopy segmentation, has advantages in wide operational applicability. Nowadays, systems such as DJI drones with RGB cameras are by far the most widespread of all UAV systems, while the processing of such images is far simpler than more complex multispectral, hyperspectral, thermal, and LiDAR images, which greatly facilitates the acquisition of information.
Round 2
Reviewer 1 Report
All comments were answered - paper is now fit to be published
Reviewer 2 Report
Review comments: Manuscript ID:- remotesensing-1002424
The manuscript (ID: remotesensing-1002424) entitled with “Mapping of the Canopy Openings in Mixed Beech-Fir Forest at Sentinel-2 Subpixel level Using UAVand Machine Learning Approach” has gone through a significant revision as compared to the earlier version. Major issues from my side were already taken into account.
The manuscript has merit as an attempt applying a bi-sensor approach for mapping of forest canopy gaps in the study region. The finding presents several advantages. Thus, I recommend considering this manuscript for publication.
