The authors have attempted to solve an important problem of increasing resolution of available sediment D50 in contiguous USA. The authors should be lauded for attempting to tackle this problem using a unique approach of predicting D50 for regions that lacked measurements using a Gradient-Boosting based machine learning method. The paper is well written, though the presentation can be improved. The approach taken by the authors is unique, though I don’t think the resultant synthetic dataset generated meets the standards of the measured datasets usually published in this journal. I say this because, it is well known that a ML model’s predictive capability is constrained by the range of values present within the training data-set. Even if a trained model shows decent prediction capability for the testing data-set, there is no guarantee that the model will be able to predict correctly for cases that have input parameter values beyond the range of the training data. When the authors synthetically generated the data for the whole US using a model trained using (extremely sparse) 2577 spatial locations, they didn’t mention about the range of values for the input parameters in the 26000 flowlines, which were later used to generate the synthetic dataset. I would recommend the authors to submit this paper to a different journal, e.g. WRR, Advances in Water Resources, Geoscientific Model Development, JGRE, etc., and increase the discussion about why the 13 parameters that SHAP value indicate to be the ones most responsible for determining the D50 of sediment at a stream location.  

As the above statement about the dataset suitability for the journal could be deemed subjective,  I leave the final decision on suitability of the synthetically generated dataset for ESSD to the Editors. Going ahead, I will only comment on the technical and presentation issues of the paper. Please find detailed comments in the attached document. 

The authors have taken up an important and interesting problem of estimating D50 in streams across the United States, when limited data exists to make these estimates. I think the data set and analysis is publishable and will be of use to the scientific community, despite the inherent uncertainties in the estimates in stream systems where little or no training data is available. The data used for the analysis has been collected for many years, and if the community waits until sufficient data is present to train the estimation model more accurately, we will be waiting for a very long time. With this being said, I think the authors need to provide more clarification regarding where their model estimates can be considered stronger and where they should be considered weaker; also some discussion regarding the representativeness of bed sediment data at gauge stations is warranted.

My biggest concern is that smaller streams may have more limited training data than exists for larger streams, and the D50 estimates for these smaller streams will involve greater uncertainty. This is based on the following: (a) a greater percentage of large streams are gaged by the USGS than small streams; (b) most of the data that does exist has D50 in the sand range – this suggests that smaller, steeper first and second order streams that are more likely to be gravel/cobble bedded are poorly represented in the data set, even though these smaller streams may dominate the total length of streams in the database. To clarify this issue as to whether my concern is founded, it would be helpful for the authors to provide a figure with two histograms for comparison: (1) a histogram that bins the stream database data according to stream size (x-axis) and analyzes total length of stream in each bin (y-axis); (2) a histogram that utilizes the same stream size (x-axis) bins and analyzes the number of D50 data points in each bin (y-axis). The variable chosen to represent stream size will need to be a surrogate such as upstream catchment area of the reach (totdasqkm) or stream order (streamorde) attributes. If that analysis confirms that my concern is founded, my recommendation will probably be that the analysis should be stratified into a dataset with higher confidence (larger streams) and a separate dataset with lower confidence (smaller streams).

Secondly, I think some discussion is warranted regarding what the D50 data at the gauge stations represents. Gage stations are established at cross sections in the stream where flow measurements are convenient and with conditions conducive to high quality flow measurements – the issue of whether the bed sediment composition represents the reach is generally not taken into account when the gage station location is established. A large percentage of gage stations are at bridges. Due to flow constriction at flood stage, bridge sections are more likely to be subject to significant scour and fill – thus, with the bed sediment being more representative of the pool than the riffle or cross-over in a reach. Particularly in gravel-bedded streams, the difference in bed sediment composition between the pool and the riffle can be substantial. This does not invalidate the data; the analysis just needs to be clear that the bed sediment at the gage station may not always be representative of the reach, which will help the user of the data understand its limitations. I would recommend citing the Edwards and Glysson methods document (https://pubs.usgs.gov/twri/twri3-c2/), which probably best characterizes how most of the bed sediment samples were collected and composited at a cross section by the USGS over the years. A short conversation with Molly Wood or Tim Straub of the USGS would also be useful to inform your discussion of these issues.

The authors of this paper use machine learning techniques to calculate the median sediment particle size (D50) in U.S. streams. The scarcity of in situ measurements and continuous regional maps make this a worthwhile challenge. A total of 2577 D50 measurements and 76 predictive attributes were used to train a machine learning model and subsequently generate a D50 map for the contiguous U.S. The machine learning model used is a Gradient Boosting variant called XGBoost, its hyperparameters were optimised using the Optuna framework. The model is further improved by trimming the input features through iterative calculation of their feature importance scores. While the main contribution of this work seems to be the resulting National D50 map, I consider the clearly documented ML approach along with a couple of insightful comments on the use of said algorithms to be at least as valuable. 

The article is well written and organised and for the most part seems methodically sound, at least from my machine learning point fo view. 

My first mayor remark concerns the actual usefulness of the final data product when taking the model performance into consideration: although the KGE might be an established performance metric in the field of hydrology, the testing R2 metric does not point to great predictive accuracy. Ultimately, the model usefulness should be assessed by experts in the field of hydrology (which I am not) and an extended discussion on different performance metrics (including some that facilitate  physical interpretability like RMSE) would help with this assessment. 

My second mayor remark regards a possible sample bias and echoes that of the second reviewer albeit from a data focused point of view. The large disparity in counts shown in Fig. 1d) as well as the fact that over 10% of the samples were measured at the same source (USACE Mississippi River main stem) make the question of data representativeness an important one. Along the histogram suggested by the second reviewer, I would suggest a dedicated discussion on how XGBoost handles skewed datasets and their impact on prediction performance.

Beyond this two points, I think the articles makes a good use of the existing data and a well informed use of machine learning to produce a new data product, and I would like to see this research published. 

Further minor comments are included on the supplement PDF