|Authors have included some of my comments in the revised version of the manuscript. However, it is clear that Authors are resisting to look at some of the comments which actually raise serious objections on the efficacy of the proposed two-steps blending approach and consequently the validation of results. In the previous version also, I had commented that details were not provided on many critical aspects. Authors have admitted the limitations of the previous version of the manuscript. |
Now in the revised version, they have somehow tried to satisfy my comments by adding more analysis but again the presented details in the methodology are hardly sufficient. In summary, I feel resistant from Authors to perform any further investigation about their approach, analysis with more years of data and provide more details on the parameter estimation using a dataset. In my view, the revised manuscript tries to demonstrate a potential approach with lack of in-depth analysis. I would recommend major revision. Authors may want to work on my following comments:
 Section 3.2 says “The goodness-of-fit of the Student’s t distribution for the bias between GR and SPE is examined graphically by using a quantile-quantile plot at the training sets (Fig. 3). It is found that they are close to the diagonal red line.”
Please look at Figure 3, the distribution is completely skewed. Blue dots are hardly lying on the diagonal red line. If the distribution is not properly set, then the parameter estimation and results will have errors.
Regarding the estimation of parameters, please look at Robertson et al. 2013, ‘Post-processing rainfall forecasts from numerical weather prediction models for short-term streamflow forecasting’ where MCMC approach of estimating parameters were changed to MAP approach based on the available length of data. Also please look at the parameter estimation section in Shrestha et al. 2015, 'Improving Precipitation Forecasts by Generating Ensembles through Postprocessing.'
This is one of the reasons why I had requested Authors to explain the steps to estimate parameters using some example dataset. It seems that Authors are hesitating to present this in the manuscript. Instead, Authors have added few more equations in the methodology section to make it appear more descriptive.
 In response to my comments, now in Section 2, it is mentioned that “The rain gauge data are spatially interpolated with a 0.25° x 0.25° resolution in the study region for each rainy day using a bilinear interpolation approach. The 34 grid cells with the gauge sites are assumed as ground references (GR) in the blending process.”
This poses a serious limitation on the analysis. Given the complexity of the region, a simple bilinear interpolation approach is hard to justify. Look at figure 1, the elevation changes from 785 to 6252. The rain gauges stations are also far from each other, they are not dense.
 In response to my comments, Authors wrote that “This study aims to develop a newly TSB algorithm on the multi-satellite precipitation data fusion in a certain time in regions of interest. Given that the larger challenge in the TP is to provide more accurate rainfall in a spatial domain, we are trying to overcome the shortage of limited rain gauge network based on the available SPE with spatial advantage using the TSB method in the NETP as a demonstration purpose. We agree that the satellite data are available for several years, but the exploration of long-term periods for the TSB method is another critical issue, e.g., the consideration of time impact on the fusion result.”
This response is hardly justified because Authors can repeat the validation for other years, especially when datasets are online available. Without this, I would have low confidence in most of the discussion in the result section.
 Authors have tried to satisfy my comments by adding a small section on comparing the proposed approach with existing BMA and OOR approaches. There are no details on BMA and OOR. It is entirely left up to the readers to figure out all these from previous literature.
I am repeatedly asking for details because the research topic which Authors are trying to address in this manuscript is extremely challenging. If the selection of distribution, parameter estimation etc. have known drawbacks then demonstrating better values of RMSE, MAE and CC does not prove the efficacy of the TSB approach.