See attached file

The reviewed paper addresses the practically important topic of operational quantitative radar rainfall estimation, benefiting from the authors’ extensive experience in this domain. While the algorithms presented are grounded in well-established statistical and physical concepts and are clearly documented, they are not novel and introduce no fundamentally new elements. More recent literature includes machine learning approaches for radar QPE and geostatistics for radar-raingauge merging, which are not discussed in this work. Nevertheless, the paper’s key contribution lies in the accessibility of its data set and the description of the algorithms, which are made publicly available online. This facilitates benchmarking and further development, making the work valuable as a resource. This is a commendable and of great value to the community.

I have some comments regarding the description of the algorithms listed below. From my point of view, the manuscript can be published after consideration of these comments.  

• The evaluation of the old versus the new QPE algorithm suite is based on data of two different years. It compares the performance of the old suite in 2022 with that of the new suite in 2023. I have two concerns with this approach: a) The comparison is not using the same data. This makes it difficult to compare the algorithms due to interannual variability. I am a bit puzzled by this. Would it be possible to run the two different algorithm suites on the same input data? b) A period of one year is maybe a bit short if one wants to assess the performance of a QPE algorithm suite for all types of precipitation in a given region. Would it be possible to extend the study to 5 years or more?
• Machine learning opens new ways to process radar data and generate estimates of rainfall rates and amounts on the ground (QPE). The manuscript makes no mention of this new avenue for QPE. I propose to add a paragraph in the introduction to mention that this new avenue exists and briefly explain why the authors have opted not to use machine learning in the new QPE suite.
• The authors mention a latency time of about 2 minutes. a) How is latency defined / measured? Is it limited to the processing time of the QPE algorithms? Or does it include all the steps between the measurement at the radar site and the time when the multi-radar QPE product is ready for dissemination to customers? b) Is the latency of 2 minutes the latency in an operational setting averaged over a large period?
• As far as I understand the adjustment with gauges is updated every hour at clock hour. There is a risk of discontinuities (“jumps”) when one switches from a one-hour time interval with fixed adjustment factors to the next hour. Did the authors observe such jumps? What is the order of magnitude of the jumps (mean and maximum)? One could avoid the jumps by determining the adjustment factors with hourly aggregations but updated every 5 minutes. Please comment.
• The gauge network used for bias adjustment has a separation distance in the order of 30 to 40 kilometer. This may be sufficient for bias adjustment in stratiform rainfall with weak spatial gradients and small representativeness errors of the gauges. For convective rainfall with high spatial variability of hourly rainfall amounts the density of the gauges may not be sufficient. Did the authors look into this doing numerical studies? In stratiform rain I am confident that the bias adjustment improves QPE. In convective rainfall with a separation distance between 30 and 40 kilometer things are less obvious. Did the authors evaluate the performance of the bias adjustment specifically for convective rainfall?
• Some components of the QPE processing suite make use of dual-polarisation capability (for instance, the clutter removal algorithm). This is the way to go as dual-polarisation offers several advantages over single-polarisation QPE. There is however the question of operational robustness. Do the algorithms that use dual-polarisation moments seamlessly fall back to single-polarisation mode in case of issues or failures in the dual-polarisation measurements?
• Path attenuation is particularly large in the melting layer where snowflakes become coated with liquid water. How does the proposed attenuation correction perform in the melting layer?
• The new QPE algorithm suite includes a module for VPR correction. In the Netherlands with little beam shielding I expect VPR to become relevant at long ranges because of earth curvature and in regions where the lower beams are contaminated by strong clutter, for instance because of wind energy plants and ships. Can the authors provide some quantitative information about the role of VPR correction? What are typical VPR correction factors? What is the percentage of the ground pixels where VPR correction is relevant?
• I understand that the merging of different scans and the merging of different radars are done separately in two separate steps. What is the reasoning why it is done in two separate steps? From a statistical point of view it would be natural to do it in one single step in a manner that minimizes the final residual uncertainty. Please comment.
• As far as I understand the advection correction uses 14 images in 5 minutes. That is, there is an image every 21.4 seconds. If using 15 images, the interval is 20 seconds, a somewhat more natural number. Probably I do not correctly understand this technical detail.
• Uncertainty estimation and the generation of ensembles, two important aspects, are not mentioned in the article. Does the new QPE suite allow to estimate uncertainties and generate ensembles? If not, are there any plans to go in this direction?