A High Spatiotemporal Enhancement Method of Forest Vegetation Leaf Area Index Based on Landsat8 OLI and GF-1 WFV Data
Round 1
Reviewer 1 Report (New Reviewer)
The article has generated high resolution LAI products by using the LAI inversion model (SA-BPNN) in the northwestern part of the Inner Mongolia Autonomous Region. It has primarily employed Landsat8 OLI and GF-1 WFV reflectance products as well as field-based LAI observations. The study has novelty aspects and is supported by the data used. However, the current form of the manuscript needs improvement. The specific comments are given below.
Major Comments:
1) Abstract: It is not written concisely. Many large sentences need to convert into smaller sentences.
2) Introduction: it is well written.
3) Study area (L125): How many plots were collected by using LAINet observation?
4) Table 1: Under Acquisition time, 151, 186, and so on are related to what? It is not clear. Is it DOY. Mention in the Table caption
5) Fig 3: Why only 2013 and 2016 is selected here and based on what logic ?
6) Figure 4: GLASS LAI and the inversion LAI have a large difference in most of the years (espl in 2015 and 2016 ), Justify why
7) Whereas in Figure 5, they had R2 of 0.99. I think Figs 4 and 5 are inconsistent
8) Add limitations of the study under discussion.
9) In Many places abbreviations are not introduced in their first use
10) In Many places, sentences are long and hard to follow. Authors should polish writing throughout the ms
Minor comments
· Rewrite L56-58 as the LAI of the Lei bamboo is repeating
· L90, L93: Reference style is inconsistent
Author Response
Dear reviewer,
We sincerely acknowledged for your kind critical comments and suggestions in improving the quality of the manuscript. We have carefully revised the manuscript.
The following is response to your comments and suggestions.
Sincerely,
Xin Tian,
Institute of Forest Resource Information and Techniques, Chinese Academy of Forestry
Email: [email protected]
Author Response File: 
Author Response.docx
Reviewer 2 Report (New Reviewer)
The manuscript proposes the creation of 30 m spatial resolution, 8 day temporal resolution LAI product by fusing an existing, low spatial/high temporal resolution LAI product (GLASS) with a local high spatial/low temporal resolution LAI map produced using Landsat-8/GF-1 data by a neural network inversion model. While the basic idea is valid, due to issues outlined below, I regrettably have to recommend rejection.
1. The BPNN model is trained using local training data that comprises of 20 plots from a single area of approximately 1 km^2 (Figure 1). As an empirical model, the performance of BPNN will crucially depend on how well the training data capture the variation of LAI and the reflectance variables in the target population. Given that training data is available from only a single small area, I have serious doubts that it is a sufficiently representative sample of the whole forest population in the larger study area. The effect of this is potentially visible in Figure 4, where the BPNN LAI predictions seem to saturate at around LAI=3, while GLASS is able to capture fuller range of LAI values. Evaluation of the training data is further hindered by lack of details in the manuscript; it would be good to have a table that shows statistics on the distribution of LAI and other variables, such as species composition, in the training data.
2. Both in the results and in the conclusions, the authors claim that the proposed approach is effective and produces good results. However, I am not convinced that the presented evidence is sufficient to make that claim. Figure 6 shows the most crucial validation result, where the proposed fused LAI product and GLASS are compared to LAINet field-measured time series. The authors claim that the fused product has a slightly better RMSE compared to GLASS. From Eq. (7), RMSE depends on the mean of squared residuals. Now, looking at Figure 6a, except for DOY 225, the GLASS LAI values are consistently much closer to the field measured values than the fused LAI values. Thus, the RMSE of GLASS LAI should be smaller. I suspect that the RMSE values presented in Figures 6b&c were calculated against the fitted regression line, instead of the field-measured values as should have been. Furthermore, the scatterplots 6b&c would be better when presented in the conventional form, where x- and y-axis have the same scale. Unit line (y=x) should also be included.
In conclusion, the evidence in Figure 6a does not support the claim that the fused LAI product is good. Based on the result, it seems to actually perform worse than the original GLASS product. Of course, it is only a single validation point, which brings up the next issue.
3. Optimally, there would be multiple validation locations across the study area, instead of a single location that is located very near to the site used for training data measurements. But due to the usual poor availability of LAI field measurements, this limitation is somewhat understandable. In the manuscript, the validation over the whole study area is done primarily by comparing the average values of fusion LAI and GLASS, which measures the systematic average difference between the two. However, as the purpose of the study is to produce a LAI product with high spatial and temporal resolution, it could be more meaningful to examine the pixel-level difference at multiple locations similar to Figure 6a. Due to different resolution (30 m vs. 1 km) this is of course complicated to do quantitatively.
4. While the structure of the article is good, there are some parts that are difficult to follow due to lack of clarity.
I do think that the manuscript is salvageable, but it requires substantial reworking to ensure that the conclusions and the evidence to support those conclusions are sufficiently in harmony.
Minor comments:
Line 75: Remove "However," as the paragraph is not related to the previous paragraph.
Line 100: It is unclear who "The researcher" is.
Line 125: What was the size of the field plots?
Lines 144-154: Unclear
Lines 154-158: What was the FROM-GLC-seg used for?
Figure 2 & Lines 236-241: There is a discrepancy between the text and the figure. In the figure, BP is done first, then the SA. While according to the text, SA is done first, then the BP.
Table 3: Please also include the number of trainable parameters in the BPNN.
What response and loss functions were used in the BPNN?
Lines 274-276: "...at time t_0 and the predicted LAI at time t0,..." Is this correct?
Lines 279-287: Unclear
Line 289: 750x750 pixel window would correspond to a 22.5x22.5 km window, which seems a bit large.
Line 290: This needs to be explained in more detail. If I understand correctly, from the pixels within the window a total of 40 pixels were selected for the data fusion using Eq. (2). Were these the 40 pixels with the largest weights W_ijk?
Figure 3: These scatter plots would also benefit from having the same scale on x- and y-axis, see Comment 2. Please also check that the RMSEs were calculated with respect to the LAI2000 values and not the line fit.
Lines 346-352: Unclear
Line 444: Remove "surface reflectance"
Lines 459-464: Unclear
Lines 474-480: Please elaborate what you mean by gap filling and how it is related to what is done in this study.
Author Response
Dear reviewer,
We sincerely acknowledged for your kind critical comments and suggestions in improving the quality of the manuscript. We have carefully revised the manuscript.
The following is response to your comments and suggestions.
Sincerely,
Xin Tian,
Institute of Forest Resource Information and Techniques, Chinese Academy of Forestry
Email: [email protected]
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report (New Reviewer)
The authors have improved the overall quality of the manuscript and addressed all queries.
Reviewer 2 Report (New Reviewer)
The authors have done an excellent work in revising the manuscript. The additional validation data and upgraded results are sufficient to put my previous concerns on the performance of the proposed method to rest.
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
The study used a combined STARFM and SA-BPNN method to estimate the high spatiotemporal (30m/8day) LAI at a forest study area based on Landsat8, GF-1 data and 1km GLASS LAI product. I read it with interest however I found several critical issues in the current version:
1) In the method part, to retrieve 30m LAI from Landsat 8 and GF-1 data, a machine learning model (SA-BPNN) was trained from ground truth data. As is well known that the success of any machine learning model depends heavily on its training samples, in section 2.2.2 the details of this key field data have not been given, and even the total number of the measurements is unknown, from Figure 1, we can learn that the LAI-2000 field plots clustered at about 1km*1km region, which cannot represent the whole study area of 45000km^2. Figure 3 shows the SA-BPNN inversed LAI against field data, is this the independent field data, or the data has already been used for training? What is the training accuracy of the model? What is the applicability of this model or method? The authors failed to answer these important questions throughout the whole manuscript.
2) In Figure 4, the inversed LAI from the original image displays the “block” effect, I guess this is caused by the mosaic of the input images which have different sun or observations angles, which has not been considered in the model training step.
Also, I have a few suggestions:
3) The language needs substantial improvement, lots of grammar errors, please smooth the language and seek more guidance from the senior authors.
4) The 250m GLASS LAI has been available to the public, its higher resolution may assist the STARFM results.
Author Response
Dear reviewer,
We sincerely acknowledged for your kind critical comments and suggestions in improving the quality of the manuscript. We have carefully revised the manuscript.
The following is response to your comments and suggestions.
Sincerely,
Xin Tian,
Institute of Forest Resource Information Techniques, Chinese Academy of Forestry
Email: [email protected]
Author Response File: Author Response.docx
Reviewer 2 Report
This work is focused on the determination of forest stand vegetation LAI using Landsat8 OLI and GF-1 WFV reflectance products and then the authors combined a Spatio-temporal adaptive reflectance fusion model (STARFM) with a LAI inversion model based on SA-BPNN. Although the obtained accuracies are high, there are many major issues with the manuscript as the authors can find some of them in the following.
1. Introduction
Many scholars have introduced artificial neural networks into regression problems and some other DLs have been used if not particular about LAI. (e.g. 1-D CNN has been applied and you can see a lot of paper which focused on it). The authors must clarify why they only chose BPNN.
Sentinel-2 data has also been used for Spatio-Temporal Fusion Algorithm (The author cited a related work). However, there are no description about Sentinel-2 data in the Introduction section. I'm not sure why the authors ignored.
2.2.2. Field measurements LAI dataset
The authors should have specified how many rings were used to calculated LAI from LAI-2000 data.
2.2.3. Remote Sensing Image Dataset Preprocessing
Reference is missing after the FLASSH atmospheric correction module.
2.3.1. LAI Estimation model
It is unclear why this configuration (shown in Table 3) was selected.
2.4. Accuracy assessment
How was R2 calculated?
3.2. Time series of LAI assimilation
The white regions were confirmed in Figure 7. What are they?
4. Discussion
The advantages of the proposed methods should have been appealed.
Did you apply the PROSAIL radiation transfer model in some points?
Author Response
Dear reviewer,
We sincerely acknowledged for your kind critical comments and suggestions in improving the quality of the manuscript. We have carefully revised the manuscript.
The following is response to your comments and suggestions.
Sincerely,
Xin Tian,
Institute of Forest Resource Information and Techniques, Chinese Academy of Forestry
Email: [email protected]
Author Response File: Author Response.docx
Reviewer 3 Report
The article "A high spatial-temporal enhancement method of forest vegetation leaf area index based on Landsat8 OLI and GF-1 WFV data" uses a data fusion method to demonstrate the fusion of high spatial resolution Landsat and high temporal resolution MODIS surface reflection data methods. And then obtained 8-day and 30m 14 forest stand vegetation LAI time series from 2013 to 2017( 121st to the 305th day of each year). This method is of great importance for obtaining LAl products of forest stand vegetation with large area and high temporal and spatial resolution.
1. In Figure 1(a), the translation of Inner Mongolia is incorrect.
2. Lines 162-163, the author mentions "... and refers to the RPC file and DEM data that comes with the image download, or-162 thorectify the GF-1 WFV image, eliminating the geometric distortion caused by the influ-163 ence of the mountain, ...". It is necessary to clearly explain the source and spatial resolution of the DEM.
3. Has the author fully considered that the collected Landsat8 OLI and GF-1 WFV data are affected by clouds? Part of the image data in Figure 4 is affected by cloud cover and it seems that this cannot be ignored.
4.In Section 3.1, it is suggested that specific validation data sources should be added. If the data used for accuracy verification (LAI2000 ground measurements observed in situ at the Genhe Ecological Station) and the data used for model construction are the same set of data, the accuracy evaluation will not be convincing enough.
5. According to the experimental data in the main text, the analysis object time is from the 121st to the 305th day of each year from 2013 to 2017. That is, it does not cover the whole year. Therefore, I recommend changing the presentation in the abstract (line 15) and in parts of the text to clarify the time frame of the study.
Author Response
Dear reviewer,
We sincerely acknowledged for your kind critical comments and suggestions in improving the quality of the manuscript. We have carefully revised the manuscript.
The following is response to your comments and suggestions.
Sincerely,
Xin Tian,
Institute of Forest Resource Information and Techniques, Chinese Academy of Forestry
Email: [email protected]
Author Response File: Author Response.docx
Reviewer 4 Report
This paper aims to generate forest stand vegetation LAI products with high spatiotemporal resolution. The LAI retrieval model was established based on the back propagation neural network with simulated annealing algorithm (SA-BPNN) and ground LAI measurements. Then, to overcome the limitations of long revisit cycle and cloud contamination, the spatio-temporal adaptive reflectance fusion model (STARFM) was employed to generate high spatiotemporal LAI products by the combination of Landsat-8/GF-1 LAI estimates and GLASS LAI products. This study would be of interest to the community and the readership of Remote sensing. However, several major issues should be addressed to improve this paper. Please find my detailed comments below.
It is necessary to clarify the novelty of this paper carefully. The LAI retrieval method (BPNN) and the STARFM model have been widely used in many previous studies. So, what is the novelty of this study?
Page 4, Lines 123:124: What is the spatial and temporal resolution of GLASS LAI products? Please introduce this information comprehensively.
Page 4, Section 2.2.2: How many the field LAI measurements for LAI-2000 and LAINet? What is the proportion of training and validation dataset?
Page 4, Lines 133-137: What is the overall accuracy of the FROM-GLC-seg product?
Page 4, Lines 168-170: This sentence is not clear. Please revise the grammar error.
Page 4, Lines 170-172: Is this reasonable that authors reprojected and resampled GLASS LAI? The spatial resolution of GLASS is much coarser than Landsat-8/GF-1, indicating additional errors would be introduced if authors processed GLASS data.
Page 5, Lines 179: “respectively fused with MODIS product GLASS LAI”? I do not understand why GLASS LAI is MODIS product.
Page 6, Figure 2: Since the GF-1 and Landsat-8 have the different band numbers and spectral response functions, how can authors consider these issues when training BPNN?
Page 9, Equation (6): What is the meaning of Z or x in this equation?
Page 9, Line 282: Where is the MAE result? Please check the entire manuscript carefully.
Page 10, Section 3.2: Why did authors compare SA-BPNN derived LAI with GLASS LAI? Any implications?
Page 12, Lines 347-349: First, the grammar of this sentence is not correct. Second, this sentence is wrong because the RMSE of fusion LAI was higher than that of GLASS LAI. Third, I do not think the fusion LAI performed better than GLASS LAI (their R2 were comparable but the RMSE of fusion LAI was much higher than that of GLASS LAI). This result cannot be acceptable because the uncertainty of fusion LAI was quite large. How can authors explain this issue?
Page 15, Lines 370-373: This sentence is too long and its grammar is not correct. I strongly suggest that authors should read the entire paper carefully to avoid such mistakes.
The citation format of many reference papers in the manuscript was not correct (For instance, Page 2, Line 51; Page 2, Line 53; Page 2, Line 84; Page 2, Line 90…). Please revise it carefully.
Author Response
Dear reviewer,
We sincerely acknowledged for your kind critical comments and suggestions in improving the quality of the manuscript. We have carefully revised the manuscript.
The following is response to your comments and suggestions.
Sincerely,
Xin Tian,
Institute of Forest Resource Information and Techniques, Chinese Academy of Forestry
Email: [email protected]
Author Response File: Author Response.docx
