Please find the review attached. 

Marine low-clouds cover the majority of the ocean, and play an important role on the Earth’s radiation budget. Due to a lack of local or ground-based observations, satellites become powerful tools for MLC measurement, while satellite observations over nighttime are still relatively limited. Thus, this study by Wu et al. introduced a deep-learning based method for the classification of MLC and their mesoscale morphology using MODIS observations, and a global dataset is developed as well. Both all-day model and day-time model were developed and evaluated. It is interesting to find some differences on the daytime and nighttime MLC, and distinct seasonal variations were also noticed for different MLCs. The new method as well as the resulting dataset is an important addition for the community, and the paper is well organized and presented. The paper could be considered for publication after considering following suggestions.

1. The quality of the training and testing dataset has been essential for DL-based models, so the datasets for the training should be carefully constructed. The 2.2 Data session gave some information on the dataset, while missed some as well. For example, Figure 1 gave some examples of MLCs of different kinds, and how was the original training dataset classified? The independency of training and testing dataset is also important, so I would suggest to introduce the testing and evaluation dataset at the Data session as well.

2. Cloudy and atmospheric properties show clear seasonal variations. For example, surface and atmospheric temperatures may significantly different from season to season, and this is also true for clouds. It is mentioned that only the results over the first half of 2014 were used for data training. Would such choice of results from half a year influence the DL performance?

3. Would it be possible to include the exact variables of input for different models in the flowchart of figure 2? This would be very helpful to better understood the details of the model efficiently.

4. The example of solid stratus show relatively regular linear structure, and are such structures natural? Please double check.

5. The training model based on daytime results is extend to nighttime observations. This is essential for the work, and could be tricky. The validation of the model for nighttime observations is very important, while only some examples were shown in Figure 4. Would it be possible to improve the validation to ensure the reliability of the results for nighttime?

6. Figures 7 indicates clear day and time differences between RFO of different MLCs. Could the authors give some discussions on the reasons for the differences?

Please see my comment as attached.