Deep Learning-Based Emulation of Radiative Transfer Models for Top-of-Atmosphere BRDF Modelling Using Sentinel-3 OLCI
Round 1
Reviewer 1 Report (New Reviewer)
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Thanks
Author Response File: Author Response.pdf
Reviewer 2 Report (New Reviewer)
I acknowledged the work and the work can be published. However, I have a few major comments which should be addressed.
1, I strongly disagree that this is a physics-informed method. Physics-informed models involved modelling physics into the machine learning or deep learning models in the neural network structure or loss function. Neither of them is seen in the proposed method. Simply using physics models to generate training data is a not a physics informed method. Actually, simply using physics model to generate training data is less applicable than using really observed samples for training and has many limitations which is never mentioned by the authors (please see my below comments and add them in discussion). Simply using physics model to generate training data is something called ‘emulator’ that has been published in the literatures for years before deep learning. I cannot support publication if the title is not changed (as the title is misleading).
2, This is for TOA reflectance BRDF modelling rather than surface BRDF modelling which is more widely applied in the literature. The authors should make this clear. So suggested title is “A deep learning method for Top of the Atmosphere (TOA) BRDF modelling using Sentinel3 OLCI data”
3, The input parameters for the simulated data are not clear at all. They are critical and very important for the success of the trained model and for the readers to understand the application domain of your trained models. The authors failed to discuss this and failed to make the texts clear. The authors should also discuss that the model can only applicable to the cases within the range of data in Table 1.
In the texts the VAZ is 0 to 50 but in the Table it is 0 to 60.
What is the step in each variable in Table 1.
Other parameters (Cm and water vapor) are fixed to what values ?
100,000 samples exactly – I don’t think so due to the combinations. Otherwise how the 100,000 is derived?
It is a bit strange to have OLCI viewing angle <68.5 but simulated viewing angle is max at 60 or 50 (not sure).
Any particular reason why solar angle starts from 10 not from 0?
RAA from -185 to 185 not from -180 to 180?
More importantly, what is the input parameters (e.g., AOD and N) in section 4.5 for application in real images? This is extremely important.
4, Introduction
“However, semi-empirical BRDF models cannot derive the full benefit of satellite observations and geospatial data hosted on CCPs due to the limited number of parameters for BRDF modeling.”
I disagree. This is a bit over-statement in order to highlight the importance of your work. Most of the semi-physical based method only used angle parameters but they allow the model parameters (like Ross-Li) to change with season and location so that other parameters (like Cha or LAI) despite not modelled but implicated in the model parameter dynamics. I don’t think the authors can criticize the semi-physical BRDF that much. Please introduce your work in another way, e.g., highlight the importance of TOA BRDF.
5. Some reference is not correct ! Please check all of them. “For instance, 109 in [22, 23], and [25], Support Vector Machines (SVMs), Random Forests (RFs), 110 and Artificial Neural Networks (ANNs) were trained using simulated data 111 generated by physical RTMs to retrieve surface or atmospheric parameters.” There references are not about surface or atmospheric parameters.
6 Some tables are mis-referenced. I believed because you added new tables/figures but forgot to change their reference numbers.
7, what are the fine-tuning parameters and how the fin-tuning is implemented?
More importantly, what is the input parameters (e.g., AOD and N) in section 4.5 for application in real images? This is extremely important.
8, “lines 359-361” how RMSE is used for sample selection? Does the authors on purpose select those samples with low RMSE?
9, it is very weird to reshape the input 9 parameters into 3*3 in order to use 1D CNN. Why not simply using deep NN without convolution? How to arrange the 3*3 of the 9 variables in Table1?
10, Fig. 14 should use true color display – current color setting is not readable and cannot differenate different land cover types.
11, Please discuss more limitation of the method, e.g., link with above the emulator method.
Line 289, “Since PROSAIL and 6S simulate directional reflectance between 400 nm 289 and 2,500 nm,” at how many nm interval?
Line 419, pooling not polling
Author Response
Please see the attachment.
Thanks
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report (New Reviewer)
Thanks for the authors to addressing my concerns. I believed the authors addressed them well.
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
Please see Attachment.
Comments for author File: Comments.pdf
Reviewer 2 Report
In general, this paper is quite well organized and written. There are only minor errors in sentence construction and spelling that should be easily cleaned up in the final edit.
A few comments:
1. Lines 361-365 make some qualitative statements about results that really should be quantitative. Maybe the authors could quantify and indicate what “significantly” means and why that supports their conclusion?
2. RSME, MAE, and R^2 metrics are introduced, but I don’t see MAE results, only RSME and R^2. It’s likely that MAE alongside RSME is not very interesting. Maybe the authors could address this, or show MAE results? The text referencing Table 4 even says Table 4 contains MAE. It does not.
Overall, this paper is recommended for publication after addressing these minor points.