Please see the attached file.

This manuscript presents an improved method for automatically classifying all sky images in 7 different categories (6 cloud types plus cloudless conditions). The topic is relevant as enhanced observations of clouds are needed both for meteorological and climate research as well as for photovoltaic power plants management. However, the real advancements presented in this manuscript are relatively marginal (as at the end, the new method improves from 96-97% accuracy to 98% accuracy) and restricted to cloud classification in 7 categories from two pre-defined datasets (so it is not clear if they involve images with other cloud classes, or other conditions such as rain or fog). In addition, no reference is made to cloud cover (or cloud amount) nor to cloud movement, which may be key for PV plant management. Nevertheless, the paper is in general technically sound, and could be suited for publication in Atmospheric Measurement Techniques journal provided that it is previously improved by considering some suggestions and applying some technical corrections. 

As an atmospheric scientist, I will focus my review on the introduction, data, and results. Unfortunately, I don’t feel capable of commenting on the methodology itself, as my knowledge of machine learning techniques is very limited.

Introduction

L. 33. What do you mean by “its decimeter-level observation”?

L. 35. What do you mean by “its equipment”?

L. 36 and Fig. 1. I would say that this is not necessary. Everyone knows that the view from the above is different of the view from below. The differences in detail and area observed are also quite well known. Moreover, there are other satellites that give much more detail of clouds, despite the image is never as detailed as from the ground. 

L. 39. Johnson 1989  is indeed a pioneer work, but it's not about the TSI, but about an original prototype of the “whole sky camera” (WSI). In addition, I would say that Shields is not a coauthor of that report. If you want a reference for the TSI, you can use Long et al 2006: Retrieving Cloud Characteristics from Ground-Based Daytime Color All-Sky Images; C. N. Long, J. M. Sabburg, J. Calbó, D. Pagès; Journal of Atmospheric and Oceanic Technology vol. 23, 5(2006) pp: 633-652.

L. 46. Taravat et al 2015 is a too specific reference for a so general statement, which may be found in atmospheric radiaton textbooks of review papers. Moreover “by suppressing short-wave and long-wave solar radiation” is note quite precise. First, other wording (absorbing, scattering,…” could be used; second long-wave solar radiation sound strange, as I think you refer to long-wave (infrared) radiation which is emitted by the ground (and clouds) not to solar radiation as such. Please clarify.

L. 50. I would use “visual” instead of "manual” observations.

L. 51. What do you mean by “low efficiency”?

L. 54. “Home and abroad” sounds strange in a science paper. Use “worldwide” instead.

L. 55-56. This sentence is repetitive. 

L. 58. “stratus nimbus” is not a cloud genera. It could be “nimbostratus” 

L. 60. There are studies that also used feature extraction before Hu et al 2018.

Figure 2. I would say is the other way around (a/b). For sure, images in (b) are not from Cazorla et al 2008.

L. 93-94. Explain in few words what it is CloudNet, CloudA, AlexNet.

L. 112-114. Please rewrite and clarify. This is the result of the present study? Or is like a summary of the previous paragraph?

L. 127-128. What does “subsoil” cloud image classification mean? Are you anticipating a result of your study in the introduction section?

Methods

Figure 3 and Table 1. Why stages go from 1-5 in Table 1 and from 0-4 in fig. 3?

L. 164-168. These sentences are a repetition of introduction.

Dataset and experiment

L. 327-328. “each ground-based cloud image sample contains ground-based cloud images taken at the same time” Please rewrite or explain.

Section 3.1.2. Apparently, the text is exactly the same as in section 3.1.1. Please do not repeat and focus on the differences between both datasets. Explain for example if the two datasets contain subsets of images which are the same or if, contrarily, they are totally different. Explain how the “true” classification has been established (visual inspection of images?)

Please do not use “Cs” for clear sky. This is confusing as Cs means “cirrostratus”, which is another cloud genus. For clear sky you may use CS (uppercase) or Cl. 

L. 366. “accuracy rate” is repeated.

Is there an index for each genus? What does “n” mean? Should TP, TN, … carry an “i” subindex? Are accuracy, precision,… overall indexes or they correspond to each cloud genus?

L. 370-373. Explain better, and use correct wording (False Positive is repeated). One may think that a sample is either classified in the correct genus or not. So it’s not clear how do you have 4 options. 

L- 373. “precision” instead of “accuracy”?

Eq. (19). If Pr and Re are already totals (sums) I don’t understand what are you summing to obtain F1. 

Results and discussion

Table 4, 5. Why there is a single Accuracy value but values for each cloud genera for the other indexes? 

L. 403. The correct classification of the Cu is the largest because in the datasets the number of Cu images is the largest too. I mean that the absolute number is not particularly relevant. 

It should be noted that all indices derive from the numbers in the confusion matrices. Therefore, I would present first the matrices, and then the indices, which somewhat summarize what is given in the matrix.

L. 459-461. You don’t need to repeat all numbers that are given in the tables. Eventually, you can highlight some numbers in the discussion.

Conclusion

L. 480. You should highlight, at least here (also in the abstract), that the accuracy that you reach is in a classification in 7 classes. There are other papers that use more (and less) cloud categories, so it’s important to make sure that the occasional reader knows to what are you referring to.

English and technical corrections.

Acronyms should be defined the first time they appear both in the abstract and in the text.

L. 9. Clouds impact Earth radiation, not only its “prediction” 

L. 22. “Accuracy” respect to what? Which is the reference? In other words who or how was the “true” cloud classification established?

L. 27. “covering” instead of “accounting”, I would say.

L. 32. “of the outside to the inside”, it should be “from above”.

L. 45. “budget balance”, I think one of the two words is enough.

L. 70 (and many other places). Do not repeat “Singh et al. (Singh and Glennen, 2005)…”; You can simply write “Singh and Glennen (2005)…” 

L. 331. “diagram” is not the adequate word, in my opinion.

L. 414. Is “Ablation” the adequate wording?