This paper proposed a machine-learning based framework to examine the processes and controls of regional floods over eastern China. Authors utilized the stream station network including observations of annual maximum flood peak during 1980-2017, to analyse flood clusters in spatial extents and intensities. The structure of the paper is clear, however, there are some concerns.

Specific comments:

1) For extreme floods cluster in space and time, it is quite urgent to get the water depth spatial distribution and variation in the short time, like serval days. Instead, authors used a time window-15 days, to analyse the flood frequencies. Could authors demonstrate the extreme flood’s distribution in a period and define the water depth for extreme floods?

2) In addition, the AMF denotes annual maximum flood peaks. Authors mentioned “The 15-day time window moves from the first to the last date of AMF occurrences for each year. We thus obtain all qualified clusters” in lines 149-150. Does it mean that only one polygon is selected in every year?

3) In the methodology part, it is difficult to understand that why authors choose use the inversed ranks in Equation (1) for AMF to represent the severity of RegFl.

4) Authors used three machine learning algorithms, DBSCAN, K-means and conditional random forest for identification, characterization and statistics respectively.  For each algorithm, it requires training and test. Could authors show the model performance in each algorithm and discuss the influence of model uncertainty in each step impacting on the following model’s training and test?

5) The predictors are in different spatial resolutions and time scales. Could authors provide more details about data preprocess?

This study analyzed the floods in eastern China from a regional perspective. The authors developed a novel approach for identifying the regional floods and attributing these floods to their driving factors. Compared to isolated floods, like flash flooding in a small watershed scale, regional floods often cause more catastrophic disasters. Such studies, like this one, are always encouraged to improve our understanding of characteristics and generating mechanisms of regional (i.e., substantial) floods.

However, a minor revision is needed before the manuscript can be recommended for acceptance.

Comments:

1. For the Identification method, the transition from step 3 (DBSCA clusters; Figure 2c) to step 4 (largest convex-hull polygon within a 15-day window) is very confusing. If multiple clusters have been identified in step 3, will you merge them to a bigger one in step 4? Please revise this part to make it clearer.
2. The statistical modeling and its associated results are not central to the main content of this manuscript. I suggest removing these elements and considering expanding them into a separate paper.
3. In line 181, please briefly explain “flood ratio” as the way how you explain the “unit peak discharge”.
4. Consider changing the subtitle 2.3 from “Empirical analyses” to “Regional floods attribution”.
5. In section 3.2, please further explain “the role of orographic lifting in enhancing rainfall intensity.” Is there any strong relationship between elevation and rainfall intensity? Also, consider citing the following paper in Section 3.2.

Houze Jr, Robert A. "Orographic effects on precipitating clouds." Reviews of Geophysics 50, no. 1 (2012).

1. Consider changing “Mild RegFls” to “Moderate RegFls”.
2. In figure 9, it shows that rainfall peaks in about 2days in advance of soil moisture and peak flow. My interpretation is that extreme rainfall caused the increase in soil moisture. In other words, the watershed antecedent soil moisture does not matter in extreme floods. Please defend it.

Guo Yu, Ph.D.,

Desert Research Institute,

Reno, NV, United States