This study introduces a new DL architecture that enables a better prediction of high precipitation fields and outperforms the current state-of-the-art in this area: none of the other models they considered were able to simulate heavy precipitation at the 3 hourly lead times. Overall the paper is clearly written and the description of the DL models are particularly clear. As well as improving the results for the problem they considered, I think the architecture could be used for broader applications in the sciences, where multiple scales are ever-present. Also: great figures!  I noticed a couple of grammatical errors, so a final proof read is in order.

In “Deep Learning Model based on Multi-scale Feature Fusion for Precipitation Nowcasting”, the authors present a methodology to produce near-term (3 hour) forecasts for rainfall in China.  They then compare the results from their MFF model to other nowcasting methodologies using several metrics.   The MFF consistently has better performance and shows marked improvement in capturing the maxima in rainfall as compared to the more traditional methods.  It is an interesting paper that is suitable for publication in GMD once the minor issues below are addressed.  And while the paper is mostly readable as is, I would encourage the authors to edit the paper for grammar, spelling, and word choice, as sometimes it is difficult to understand their intent.

Broad comments:

1. I’m not overly familiar with the method described here, but for other machine learning applications, when a model has to extrapolate beyond the training dataset, it often performs poorly. What have you done here to mitigate that?  Does your training dataset encompass the full range of precipitation rates you would ever see?  What happens if you have a 1 in 100 year rainfall event?  How would the forecast be able to handle that event if it is outside the bounds of the training dataset?  Some discussion of these limitations or why they are not limitations is warranted.
2. I agree that your method seems to be doing a better job of capturing the extremes in the precipitation than the other methods presented here, but, particularly for your second case study, the MFF model seems to be overestimating values as compared to the GT. While the smoothing effect you mention can explain some of this, more discussion about why this overestimation occurs is warranted.  From a forecasting perspective, I would think false positives, if they come frequently enough, could be as detrimental as false negatives, as people will start ignoring forecasts if they are always overestimating rainfall. Again, this is a relatively minor issue, but I think a little more nuanced discussion is required.

Specific comments:

Line 122: I assume you mean the spatial resolution of one radar image (all grid boxes combined) is 5 km?  It’s unclear from your phrasing.  Also, where does the 205848 number come from?  Is that the temporal component? 

Line 252: Define RAPS. 

Line 263: Do you mean 2018 – 2021?

Line 310: Should be 5 June 2022.