The paper proposes a new CHUP-BMA ensemble forecasting method by incorporating the CHUP-derived posterior distribution of the observed flow into the BMA framework. It has the advantage that the initial state constraints can be considered in the BMA while avoiding the normal quantile transformation of the HUP-BMA method. Based on deep learning, an ensemble forecasting scheme considering input, model structure, and parameter uncertainty is constructed in Three Gorges Reservoir, China, and the effectiveness of the CHUP-BMA method in reducing forecast uncertainty is verified. The study is innovative and theoretically rigorous and has promising results with solid application potential. Some questions need further discussion.

1. The sources of Figure 3 and Table 2 need to be explained to improve the reasonableness of the paper.

2. Various model inputs (e.g., rainfall, tributary flows, etc.) exist in the interval basins. The article only considers the input uncertainty of rainfall, and it is suggested to add a reason for this in subsection 3.2.1.

3. Line 255. You should briefly introduce the LSTM in subsection 3.2.2 to improve the paper's readability. In addition, it is recommended to cite references more relevant to the LSTM.

4. Deep learning parameters significantly impact forecast accuracy, so it is recommended to show the values of deep learning parameters. The study should concentrate on ensemble forecasting methods rather than deep learning models. Therefore, the model parameter values can be shown in the appendix.

5. Line 369, add a description of the member type with better forecast accuracy, i.e., the input composition, the model structure, and the objective function of the selected parameters.

6. There are numerous evaluation metrics in deterministic and probabilistic forecasting. Briefly explain the reasons for the metrics chosen in the paper.

7. Line 465, replacing 'concentration' with 'sharpness' as 'reliability (α_index), concentration (IGS),' should correspond to the name of Figure 13. 

8. To improve modeling rationality, explain why observations are used as model tributary inputs in training and validation periods. 

9. In the outlook, adding the construction of the CHUP-BMA method using a more flexible vine copula will make the CHUP-BMA method more competitive.

The authors combine the copula-based hydrological uncertainty processor (CHUP) and Bayesian model averaging (BMA) to obtain a novel approach to statistical post-processing of hydrological ensemble forecasts. The proposed approach is promising and the presented results are fair, but the paper needs some improvement and it also raises some questions.

Major comments:

1. L28: The cited paper Sloughter et al. (2010) deals with post-processing wind speed forecasts. The BMA model for precipitation is introduced in Sloughter et al. (2007).

2. I am also missing references to BMA models for hydrological forecasts, e.g. Hemri et al. (2013) or Baran et al. (2019).

3. Eq.4: In the original description of the HUP, different CDFs are considered for the forecasts and the observations, moreover, in the former case it is considered as an initial estimate. Does such a relaxation make sense here as well?

4. Section 3.1.2. introducing the HUP follows the structure of Sections 2.1.2 – 2.1.4 of Darbandsari and Coulibaly (2021); however, one should mention that the Markov process of Eq.5 is stationary and define exactly how θ_t in L169 is related to Eq.7 (see Darbandsari and Coulibaly, 2021, Eq.10).

5. L310: “The IGS metric indicates the sharpness of the probabilistic forecast”. The IGS, similar to the CRPS addresses simultaneously both calibration and sharpness, as indicated in the cited work of Gneiting et al. (2005). Hence, I think referring to IGS as a measure of concentration is slightly misleading.

6. In Section 4, I would definitely consider the corresponding scores (or at least some of them) for the ensemble forecasts as well.

7. What can be said about the statistical significance of the score differences between HUP-BMA and CHUP-BMA?

Minor remarks, typos:

1. L205-206: “It has been studied that the BMA method with sliding windows can obtain better probabilistic forecast performance”. Better compared to what?

2. L307: “indicative function” → “indicator function”

References

1. Baran, S., Hemri, S. and El Ayari, M. (2019) Statistical post-processing of water level forecasts using Bayesian model averaging with doubly-truncated normal components. Water Resour. Res. 55, 3997–4013.
2. Hemri, S., Fundel, M. and Zappa, M. (2013) Simultaneous calibration of ensemble river flow predictions over an entire range of lead times. Water Resour. Res. 49, 6744–6755.
3. Sloughter, J. M., Raftery, A. E., Gneiting, T. and Fraley, C. (2007) Probabilistic quantitative precipitation forecasting using Bayesian model averaging. Mon. Weather Rev. 135, 3209–3220.

In this round of revision, the authors have satisfactorily addressed all my concerns and significantly improved the manuscript and the language. The authors have proposed a novel approach, combining the copula-based hydrological uncertainty processor and Bayesian model averaging, to obtain statistical post-processing of hydrological ensemble forecasts, significantly reducing flood forecasts. I recommend accepting the revised manuscript.