Monitoring Mega-Crown Leaf Turnover from Space

Round 1

Reviewer 1 Report

This manuscript proposes a novel approach for leaf turnover assessment using satellite data. The manuscript presents a significant contribution to future efforts in the field and could improve the leaf-level evaluation by the means of increasing the accessibility of data, i.e. using satellite data, and accuracy of the assessment. Following are the comments and suggestions for authors:

Comments:

1- The title should be capitalized; "Monitoring Mega-Crown Leaf Turnover From Space"

2- line 17: "have never been used". I think "before" is redundant.

3- line 27: were significantly "correlatED" ...

4- The introduction is well-written and presents a solid background of the work, but I feel the need of more quantitative results added to this section from related and similar studies. This can highlight the gaps and problems the authors' approach is addressing.

5- The materials and methods section is well-written, however, it could benefit from better subsections. the authors can maybe number different sections, and also assign a specific one to data (materials).

6- I may have missed something, but I could not figure out how the probability mentioned in line 273 is evaluated, i.e. what assumptions were used for pdf evaluation and what model?

7-  Authors need to include more previous related works in the discussion section and compare their findings with the ones found in such studies. Although as mentioned earlier in the manuscript, there are no studies focused on using satellite data in this field, but other approaches used to address similar problems should be included and discussed in the discussion section.

8- You can add more clarifications on why you got poor results and how future work can address such issues. Is there anything that can be changed in this study itself to improve the results?

Author Response

This manuscript proposes a novel approach for leaf turnover assessment using satellite data. The manuscript presents a significant contribution to future efforts in the field and could improve the leaf-level evaluation by the means of increasing the accessibility of data, i.e. using satellite data, and accuracy of the assessment. Following are the comments and suggestions for authors:

We thank the reviewer for their comments and suggestions and we answer each of the points raised below.

Comments:

The title should be capitalized; "Monitoring Mega-Crown Leaf Turnover From Space"

Response: Done.

2- line 17: "have never been used". I think "before" is redundant.

Response: Removed.

3- line 27: were significantly "correlatED" ...

Response: We are unsure of the reviewer’s request in this instance as the sentence reads as the reviewer has suggested.

4- The introduction is well-written and presents a solid background of the work, but I feel the need of more quantitative results added to this section from related and similar studies. This can highlight the gaps and problems the authors' approach is addressing.

Response: We have included a more thorough summary of related studies in the Introduction at lines 119 to 157, parts of which are shown below:

“…  Spectral indices based on the visible colour space have also traditionally been used by close-range remote sensing of individual phenology (i.e. “phenocams” and Unmanned Aerial Vehicles, UAVs, or “drones”) [10,11,25–27], which often lack the ability to record outside the visible spectrum.  Digital cameras have been attached to towers and used to take time-lapse photographs from above the canopy at a number of sites. In the Brazilian cerrado, a normalised chromatic coordinate index extracted from digital camera imagery showed changes in greenness that mirrored ground observations of leafing strategy at the individual and species-levels [11]. An analysis using a tower-mounted hyperspectral camera to monitor individual crowns in a Brazilian tropical forest during the dry season showed increases in reflectance, especially in the green and NIR regions, and significant changes in vegetation indices (such as NDVI) concurrent with leaf flushing [28]. In both cerrado and rainforest landscapes, tower mounted camera studies have demonstrated that community-wide analyses of spectral indices fail to detect seasonal changes apparent when individual canopies and species are differentiated [11,12,28].

Among several proposed colour indices, the Green Leaf Index (GLI) [28] – also known as the Green Leaf Algorithm, GLA [30]and similar to the Visible-Band Difference Vegetation Index, VDVI [31], and the Excess Greenness Index, ExG [10] - can be formulated as (2G-(B-R))/(2G+B+R), where G, R and B are spectral reflectance values at the red, green and blue spectral regions. GLI is sensitive to leaf chlorophyll content and can be derived from Landsat 8, MODIS, Sentinel-2 and Sentinel-3 orbital sensors [32]. The GLI has been used previously to detect changes in chlorophyll content in crops [33] and foliage cover in rangelands [30]. Visible-based indices have been successful in detecting canopy phenology from daily close-range imagery [10,25], but have seldom been assessed for satellite imagery, probably because the atmosphere is less transparent in these wavelengths, and the difference in reflectance from green leaves is much stronger between red and infra-red bands than between any of the visible spectrum bands. Still, as NIR bands are more sensitive to leaf mesophyll structure and water content, a more pigmented-oriented index in the visible range may still be of use when tracking single individuals.

A recent study using images from UAVs to predict individual-tree leaf phenology in a tropical forest in Panama found the best performing models (r²=0.84) to derive from a combination of colour and texture metrics [10]…”

5- The materials and methods section is well-written, however, it could benefit from better subsections. the authors can maybe number different sections, and also assign a specific one to data (materials).

Response: Done

6- I may have missed something, but I could not figure out how the probability mentioned in line 273 is evaluated, i.e. what assumptions were used for pdf evaluation and what model?

Response: The probability referred to here is simply whether the data point from the satellite time series fell within (1) or without (0) one of the leaf loss or renewal events. The models are binomial generalised linear mixed models to account for the presence/absence structure of the response variable.

7-  Authors need to include more previous related works in the discussion section and compare their findings with the ones found in such studies. Although as mentioned earlier in the manuscript, there are no studies focused on using satellite data in this field, but other approaches used to address similar problems should be included and discussed in the discussion section.

Response: We have included additional discussion and references to previous work in the Discussion section, for example at lines 404 to 417 as follows:

“The greenness indices from Sentinel-2 show the most promise for detecting individual crown leaf turnover from space. As expected from previous work using UAVs and tower-mounted digital cameras [10,11,24,28], greenness declines during leaf senescence and loss and increases during leaf flush for individual crowns at our site. NDVI outperformed GLI when detecting leaf senescence and loss at our site (standardized estimates of normalized NDVI and GLI are -0.76 and -0.56 respectively). As NDVI correlates almost linearly with LAI [24] this result indicates that it is leaf loss, rather than the change in leaf chlorophyll content, that dominates the signal of this event in satellite data. Despite significant correlations in our results, uncertainty is still very high with just over 50% leaf senescence or loss events correctly identified at the lowest observed NDVI values relative to surrounding forest and just 15% leaf renewal events correctly identified at the highest observed GLI values. A recent study using UAVs to monitor individual crown phenology at a tropical forest in Panama found that models based on the Green Chromatic Coordinate alone (a visual-based index similar to the GLI used in this study) performed worse than models combining colour and texture metrices (r² = 0.52 versus 0.84) [10]. ”

8- You can add more clarifications on why you got poor results and how future work can address such issues. Is there anything that can be changes in this study itself to improve results?

Response: We have expanded the section in the Discussion that addresses the uncertainties in our method and the improvements that we could make in the future at lines 443 to 468 as below:

“Despite our hope that the commonly used monthly-interval for focal crown observations from the ground might be sufficient to confirm the satellite signal for leaf turnover, it appears that there is too much uncertainty in both ground and satellite data sources to build a reliable model to scale-up detection of Moabi leaf turnover outside our study area at the current time. The poor temporal resolution of both the ground observations (due to the monthly field work schedule) and satellite observations (due to flight schedules and cloudiness) are major contributors to this uncertainty. To improve detection rates in future analyses under the current suite of available satellite products we could collect higher resolution ground data, for example using daily or weekly measurements provided by field personnel, phenocams or UAVs (see Figure 8). UAVs provide an ideal intermediary between ground and satellite observations, often providing for example optical data with 5-10cm spatial resolution, with the advantages of flying beneath the clouds and covering a much larger area than possible by ground observations and should yield more precise canopy delineations [10].

More frequent data collection should also help improve the assumptions we made when allocating the “periods of likely leaf turnover”. We designed the current leaf turnover protocols based on the monthly frequency of ground observations (Figure 1). A brief investigation of the remote sensing data suggests that elevated greenness associated with leaf flush may persist beyond the window allocated in the present study and that finding a more optimal window of leaf turnover could improve the strength of the predictions from our models (Supplementary Table 1). We will be able to carry out a formal window selection in the future when additional higher frequency ground data are available.  We should also consider that human error may have lent some uncertainty to the ground observations. However, the observers at Lope are highly trained and have been consistent throughout the time period [21]. Human observation more generally brings a number of benefits, such as intelligent looking – searching for multiple cues and moving position to gain a better vantage point – and is fairly insensitive to variable lighting, compared to an automated system. A combination of human and automated observation methods and an increase in frequency would be advantageous to reduce uncertainty in phenophase detection in future analyses. “

Reviewer 2 Report

The paper presents a novel approach for the detection of leaf senescence/loss and leaf renewal events of a tropical tree (Moabi) from Sentinel-1 and Sentinel-2 data. For that, ground observations on eight Moabi trees for a period of 5 years that were collected in monthly time intervals were utilized. Although the sample data set is quite small (containing 18 ? leaf-turn over events) and the temporal resolution of the ground truth data is limiting the analysis, the research is interesting and the article is well written. Some improvements are suggested in the method part. These include a visual representation of the mapped area and information on the stability of the surrounding buffer area that is utilized for time series normalization. Furthermore, effects of the lower forest vegetation layers on the satellite signal could be included in the discussion.

 

Detailed comments

Page 4, line 173: Do you mean ”increments from 0 (no coverage) to 8” instead of “increments from 0 (no coverage) to 4”?

Page 4, line 188: “may not be recorded” instead of “may not recorded”?

Page 5, Figure 1. : Why is a different brown shade utilized for the leave senescence and loss event in figures (A) and (B)?

Page 6, Figure 2.: Closing brackets for observation period (2015-2019) is missing.

Page 6, line 213: It would be useful to have a small map of the study area showing the delineated tree crowns (with their numbers) and buffer area.

Page 7, line 236: Could you please include the wavelength of the Sentinel-2 bands that are used in the equations.

Page 7, line 243: The authors normalize the satellite time series of the selected trees against the buffer areas with forest cover around the tree crown. How stable are the index time series for these buffer areas and what is the phenology there? What is the overall level of vegetation indices observed for the crown and buffer area?

Page 7, analysis: Could you please specify the number of observations utilized in modeling? Has there been validation of the models with a test set?

Page 9, Figure 3: The time series of the index-deviations from the surrounding forest look still quite noisy and with similar or even larger deviations outside the leaf turnover periods. How much variation is within the crown areas? Could you add the standard deviations to the figure?

Page 10, Figure 4: Please correct figure caption “…loss from Sentinel 1 and Sentinel data…” to “…loss from  Sentinel 1 and 2…”.

Page 14, Discussion:

The authors report the low detection rate of the leaf loss and leaf renewal events with the proposed method. It would be useful to discuss also the effect of the lower canopy layers on the observed vegetation index change during the leaf loss events and the potential masking of the targeted phenological events and potential saturation issues of vegetation indices.

Furthermore, are there any other variables (possibly non remote sensing) that could be included into the prediction models?

Author Response

The paper presents a novel approach for the detection of leaf senescence/loss and leaf renewal events of a tropical tree (Moabi) from Sentinel-1 and Sentinel-2 data. For that, ground observations on eight Moabi trees for a period of 5 years that were collected in monthly time intervals were utilized. Although the sample data set is quite small (containing 18 ? leaf-turn over events) and the temporal resolution of the ground truth data is limiting the analysis, the research is interesting and the article is well written. Some improvements are suggested in the method part. These include a visual representation of the mapped area and information on the stability of the surrounding buffer area that is utilized for time series normalization. Furthermore, effects of the lower forest vegetation layers on the satellite signal could be included in the discussion.

Response: Many thanks to the reviewer for highlighting changes needed to improve our manuscript. We have addressed each of the points raised in this summary in the detailed comments below.

Detailed comments

Page 4, line 173: Do you mean ”increments from 0 (no coverage) to 8” instead of “increments from 0 (no coverage) to 4”?

Response: The field protocol is to record from 0 to 4 with ½ point increments. It is in essence the same as recording 0 to 8 but we tried to accurately described the method used in the field, to avoid any confusion if a reader was to compare with other papers or the raw field data. We have updated the text to read “recorded as a 9-point scale in half integer units from 0 (no coverage) to 4 (full coverage)”, which we hope removes any confusion.

Page 4, line 188: “may not be recorded” instead of “may not recorded”?

Response: Done.

Page 5, Figure 1. : Why is a different brown shade utilized for the leave senescence and loss event in figures (A) and (B)?

Response: We have altered this figure so that the brown shades are now the same. The reason for the difference before was simply because the focus of part B is on the leaf renewal stage. We agree with the reviewer that it is simpler to use the same shade.

Page 6, Figure 2.: Closing brackets for observation period (2015-2019) is missing.

Response: Done.

Page 6, line 213: It would be useful to have a small map of the study area showing the delineated tree crowns (with their numbers) and buffer area.

Response: We have now included an aerial view of the study area and the eight focal Moabi trees in the new Figure 3. The caption is included below:

Figure 3. Aerial view of eight focal Moabi crowns at Lopé NP, Gabon (small yellow circles) and their 100m buffers (surrounding yellow circles). A. Focal Moabi crowns were identified and drawn by hand using high resolution (<2 m pixels) imagery available in the Google Earth™ and Microsoft Bing™ platforms and GPS coordinates from the ground. We delineated forest in a 100m buffer around the crown boundary of each focal tree erasing the focal crown and other Moabi crowns contained within the buffer. B. Tree 3231 is adjacent to two Moabi crowns which are not part of this study but were excluded from the 3231-forest buffer to avoid signal contamination. C. Tree 3029 sits at the forest-savanna edge and thus c.a. 58% of the buffer was excluded due to being within the savanna. Background image is provided by ESRI World Imagery.

Page 7, line 236: Could you please include the wavelength of the Sentinel-2 bands that are used in the equations.

Response: We have included information on the wavelengths of the bands at line 247 as follows:

“The central wavelengths for the Sentinel 2 bands are as follows: Band 2 (blue): 0.490nm, Band 3 (green): 0.560nm, Band 4 (red): 0.705, Band 8 (near infrared): 0.865. “

Page 7, line 243: The authors normalize the satellite time series of the selected trees against the buffer areas with forest cover around the tree crown. How stable are the index time series for these buffer areas and what is the phenology there? What is the overall level of vegetation indices observed for the crown and buffer area?

Response: We have now included five additional supplementary figures showing raw data timelines per tree for each satellite data source to allow the reader to scrutinise the variability of the canopy, and buffer time series and the impact of normalising the data relative to surrounding forest. We have also included overall summary statistics per data source to show the overall level of vegetation indices observed. These are detailed in the text at lines 283 to 285 and in the new Table 1 as follows:

“The canopy, buffer and normalised canopy time series per Moabi tree for all data sources are shown in in Supplementary Figures 1 to 4 and the overall mean values (and standard deviations) for all observations per satellite data source are shown in Table 1.”

Table 1. Summary satellite data for eight focal Moabi crowns at Lopé NP. This table shows the overall mean (and standard deviation) of all crown-level means per satellite data source. Satellite data were sourced from the VV and VH Synthetic Aperture Radar (SAR) bands of Sentinel-1 and from the Normalized Difference Vegetation Index (NDVI) and the Green Leaf Index (GLI) derived from Sentinel-2. ‘Canopy timeseries’ refers to the mean values extracted for each focal crown at each time step, ‘Buffer time series’ refers to the mean values in the forest buffer surrounding each focal crown at each time step and ‘Normalised canopy time series’ refers to the difference between the canopy and buffer values.

Data	Obs.	Canopy time series	Buffer time series	Normalised canopy time series
Sentinel-1 VV	837	0.18 (0.07)	0.22 (0.04)	-0.04 (0.07)
Sentinel-1 VH	837	0.04 (0.02)	0.05 (0.01)	-0.01 (0.02)
Sentinel-2 NDVI	477	0.44 (0.19)	0.43 (0.18)	0 (0.04)
Sentinel-2 GLI	477	0.02 (0.03)	0.01 (0.02)	0 (0.01)

Note: Obs. = number of observations

Page 7, analysis: Could you please specify the number of observations utilized in modeling? Has there been validation of the models with a test set?

Response: We now include the number of observations per data source in Table 1 (837 observations for all models based on Sentinel 1 data and 477 observations for all models based on Sentinel 2 data). As regards a test set, we consider this analysis to be the test set and don’t have additional unutilised data to compare our models results with.

Page 9, Figure 3: The time series of the index-deviations from the surrounding forest look still quite noisy and with similar or even larger deviations outside the leaf turnover periods. How much variation is within the crown areas? Could you add the standard deviations to the figure?

Response: We have now included five supplementary figures showing time lines and boxplots per tree for each data source to illustrate this data better. As the reviewer has noticed, there remain large deviations in the time series unexplained by our ground observations for the leaf phenology.

Page 10, Figure 4: Please correct figure caption “…loss from Sentinel 1 and Sentinel data…” to “…loss from Sentinel 1 and 2…”.

Response: Done.

Page 14, Discussion:The authors report the low detection rate of the leaf loss and leaf renewal events with the proposed method. It would be useful to discuss also the effect of the lower canopy layers on the observed vegetation index change during the leaf loss events and the potential masking of the targeted phenological events and potential saturation issues of vegetation indices.

Response: We agree that lower canopy layers are likely to influence the observed vegetation indices but unfortunately we don’t have quantification of these. We have included a comment on this issue in the Discussion at lines 418 to 427 as follows:

“Normalising the NDVI canopy time series relative to surrounding forest at our study site very much improved the utility of this data source in detecting leaf loss and senescence (the standardized estimate of the normalised canopy data was 2.8 times the strength of the estimate from the original canopy data; Table 2 and Figure 4). This improvement serves to emphasise the importance of compensating for the likely influence of variable lighting and atmospheric conditions on the estimation and interpretation of NDVI [28]. Furthermore, although we did not quantify the amount of subcanopy cover under our focal Moabi trees, we expect that any amount of green leaves under the main crown would further reduce signal differences from leaf-on phases or from the surrounding green canopies. However, the impact of the normalisation on the other data sources and for detection of leaf renewal was minimal. “

Furthermore, are there any other variables (possibly non remote sensing) that could be included into the prediction models?

Response: In future analyses we hope to include additional explanatory factors in our analyses to better understand the pattern of leaf phenology at this site. However, the objective of this particular study was to assess the utility of remote sensing products to monitor leaf turnover and thus only remote sensing variables are considered.

Reviewer 3 Report

"Monitoring mega-crown leaf turnover from space" is a study adopted the well know techniques of RS and used the imagery data of Synthetic Aperture Radar (SAR) and Sentinel to monitor the leaf turn over. Although, this study is a good effort and used a well adopted methodology but can not reach to a significant conclusion or outcome. As still there are too much uncertainty in result and findings, this study may be continue till to find a good result and soil findings. Although MM section and result and discussion sections are written good. While there are specific comments are mentioned in attached reviewed report.  

Comments for author File: Comments.pdf

Author Response

"Monitoring mega-crown leaf turnover from space" is a study adopted the well know techniques of RS and used the imagery data of Synthetic Aperture Radar (SAR) and Sentinel to monitor the leaf turn over. Although, this study is a good effort and used a well adopted methodology but can not reach to a significant conclusion or outcome. As still there are too much uncertainty in result and findings, this study may be continue till to find a good result and soil findings. Although MM section and result and discussion sections are written good. While there are specific comments are mentioned in attached reviewed report

We thank the reviewer for taking the time to read and comment on our manuscript. We list below specific responses to comments and suggestions made by the reviewer in the pdf.

Response to comment at lines 30-31 and 397: While uncertainty is high in our predictions, we are reporting the first attempt at orbital monitoring of individual crowns in an under-studied area of the world and as such, believe our contribution to be an important one and worth sharing at this point to encourage other innovations and improvements.

Response to comment at Line 67: We have reviewed the Introduction and feel that all detail there is necessary to describe the background and context of our analysis.

Response to comment at Line 100: We have linked these two paragraphs as requested.

Response to comment at Line 183-185: We acknowledge that the method for determining the periods of leaf senescence / loss and renewal is complicated but feel that the written description accompanied by the figure and the R code are sufficient to give a detailed description.

Response to comment at Line318: We used generalised linear mixed models as they enabled us to model the hierarchical structure of our dataset and the binomial nature of the residuals.

 

Round 2

Reviewer 3 Report

Article looks much improved as previous version. I have one suggestion suggestion about English of the article, many times authors used non academic languages and many time mentioned we show and so on, here i suggest please try to use academic language like in articles and research papers. 

Author Response

We thank the reviewer for taking time to read our new revision and glad that they find it improved.

We are content with the style of language used in this paper and prefer an active style of voice in academic publications.