|The overall idea that observed and modelled flow should have comparable Hausdorff dimensions (here, think about the RD) would remain solid to me if the method being introduced could be independently used for calibrating models. Unfortunately, the method being introduced requires other existing objective functions in reducing the mismatch between observed and modelled streamflows. In this line, it becomes difficult to quantify the significance of the contribution of the new method to the overall performance of the model. In other words, we can have Nash-Sutcliffe Efficiency (E) of, say, 0.95 (indicating a perfect or nearly perfect model) with or without the introduced method. Then, why the new method? However, the question is: Can it be possible to have E = 0.95 when the results of the model does not make sense with respect to RD? Intuitively, the answer is NO. Perhaps to avoid subjectivity of regarding the choice of E and RD to characterize the model's optimality, I suggest that the authors come up with a new metric (NM) in terms of the product of E and RD. In this way, an ideal model would have NM=1.|
Alternatively, the authors could explore the possibility of incorporating RD within E. Of course, the well-known metric E comprises comparison of observed and modelled series in terms of (a) correlation, (b) measures of dispersion, and (c) difference in mean. This can be found in a number of articles, for instance, see the formula for computing the Kling-Gupta Efficiency (Gupta et al., 2009). The idea would be to try to replace correlation in E with RD and you assess model performance. In this way, you can come up with a formula which considers RD, ratio of standard deviation of observed flow to that of simulated flow, and ratio of mean of observed flow to that of simulated flow. The findings would be in terms of whether combination of RD with measures of dispersion, and means could lead to comparable model performance like for the case when E is used alone. This is just some suggestion which the authors could try to explore.
Lastly, the paper is to lengthy and its contribution not straightforward. I strongly suggest that the authors make the paper concise. In this line, one may find it apparent that further analyses (of trying to link RD to (i) the variation of sub-flows separated by WETSPRO, and (ii) parameters of the hydrological models) were somewhat unnecessary and could have instead engendered more uncertainties in results. There are some hydrological models that purely depend on flow splitting for their calibration and the authors need to know that HBV is not such a model. In other words, flow splitting is not that relevant for the introduced method for simplicity. Conclusively, these further analyses (of flow splitting and linking RD with model parameters) should be discarded from the revised manuscript or placed in a supplementary material.
My conclusion is that the authors require time to do more work which can make the paper compact and/or straight forward
Gupta H V, Kling H, Yilmaz KK, Martinez GF (2009) Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. Journal of Hydrology 377:80–91. doi: 10.1016/j.jhydrol.2009.08.003
Figure 3: Also include the latitudes and longitudes around the DEM of Dong.