Machine-Learning-Based Precipitation Reconstructions: A Study on Slovenia’s Sava River Basin

Round 1

Reviewer 1 Report

I thank the authors for the opportunity to reflect on the topics discussed. Although there are interesting insights into the capabilities of the analysis methods and the approach is quite convincing, the proposed research downplays all the climatic and environmental aspects that may underlie data comparisons that may be uneven and/or even incomplete. The structure of the dataset should be examined, highlighting any gaps, criticalities and inhomogeneities in the sample. The weather data recorded in 1870 does not have the same sensitivity as today's, the reading cadence and aggregations have changed. This should also be examined. Does the data refer to the same site, or has the monitoring point changed over the decades? Certainly, the critical review of the analysis methods is interesting, but no substantive comparison with other methods is provided. Beyond duration, the validity of hydrological variables based, for example, on tree rings could be compared by giving them a fair amount of space. To be convincing, the authors should find arguments and applications on verifiable datasets, commenting on the results.
A discussion of statistical parameters and indicators with their respective pros and cons might help a reader less familiar with the details of the proposed subject matter.
Thank you

Author Response

Thank you for your insightful comments and suggestions. We have taken them into careful consideration and provided the following detailed responses:

The dataset we utilized, known as the Long-term Alpine Precipitation Reconstruction (LAPrec), offers gridded fields of monthly precipitation for the Alpine region, spanning eight countries. Derived from station observations, LAPrec meets high climatological standards, ensuring temporal consistency and accurate spatial pattern reproduction in complex terrains. The dataset's temporal extent, covering over a century, makes it a robust foundation for historical climate analysis in a region significantly impacted by climate change.

LAPrec combines two primary data sources:

1. HISTALP (Historical Instrumental Climatological Surface Time Series of the Greater Alpine Region): Provides a homogenized station series of monthly precipitation from the 19th century. The version we used, starting in 1871, employs 85 almost continuous series that are uniformly distributed across the Alpine region.
2. APGD (Alpine Precipitation Grid Dataset): A high-resolution grid dataset of daily precipitation from 1971–2008, constructed from over 8500 rain gauges. It represents one of the densest in-situ observation networks globally, covering the entire Alpine region.

The LAPrec dataset was synthesized using the Reduced Space Optimal Interpolation (RSOI) method, a linear model between station and grid data. This method was developed collaboratively by the national meteorological services of Switzerland (MeteoSwiss) and Austria (ZAMG). For a more in-depth understanding of the methodologies and further insights, we direct the reviewer to the detailed documentation available here: https://cds.climate.copernicus.eu/cdsapp#!/dataset/insitu-gridded-observations-alpine-precipitation?tab=overview as well as in the following scientific papers: https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD029910,

 https://rmets.onlinelibrary.wiley.com/doi/10.1002/joc.4343,

 https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2009JD013047.

While LAPrec is optimal for long-term temporal coverage and high temporal consistency, it's essential to note its limitations. For instance, the dataset might not be ideal for applications focusing on small-scale spatial patterns or requiring high absolute precision. Users should also be aware of potential errors, underrepresentation of high altitudes, and other issues that might lead to precipitation underestimation.

We agree with the reviewer's observation regarding the changing climate conditions over the decades and the potential impact of dataset selection on results. However, our choice of LAPrec was grounded in its construction using state-of-the-art climatological approaches, ensuring a homogeneous dataset aligned with European meteorological office standards.

To address the reviewer's concerns comprehensively, we have enriched the manuscript with a detailed description of the dataset and an elucidation of the statistical metrics employed for model performance evaluation.

Reviewer 2 Report

Line 171: The authors mentioned that RMSE were smaller than 120 mm. Does the value have specific meaning, or is it to point out those two RMSE were the smallest values?
And, the R² are too.

Line 131~: From the tables 1 to 4, the model performances look getting better, based on the R². However, the final R² was 0.327, it is very hard to say the value is acceptable. Typically, R² needs to be greater than 0.6 or 0.7 for acceptance.
And, at the discussion section, NSE came up, they look acceptable.
How about using consistent criteria from result to discussion sections?

Line 217: There is no conclusion section. There is a need to add the section.

Figures 3 and 4: The texts are somewhat small to recognize.

Author Response

Response to comment on Line 171: The values less than 120 mm are highlighted as the lowest RMSE scores.

Response to comment on Line 131: We acknowledge the importance of consistent criteria and have chosen to use RMSE, NSE, and KGE as the primary metrics throughout the paper. This decision is in line with common machine learning practices for assessing model performance on unseen data. However, recent research has successfully developed seasonal (April-May-June-July) precipitation reconstructions using tree rings in Europe, specifically in the Lower Kızılırmak River Basin, Turkey, with an R^2 value of 0.36 (https://www.mdpi.com/1999-4907/13/4/501). Thus, seasonal precipitation reconstructions using tree rings with R^2 values exceeding 0.3 are considered acceptable. It's worth noting that when using tree rings, the skill in precipitation reconstruction is traditionally lower than in streamflow reconstruction. This was confirmed by [5], where the skill for seasonal (AMJJAS) streamflow reconstruction ranged from 0.50 to 0.74. Streamflow, which integrates the entire hydrologic cycle (including precipitation, evapotranspiration, infiltration, and runoff), is typically better represented in tree-ring growth.

Response to comment on Line 217: Thank you for highlighting the absence of a conclusion section. We will include a dedicated 'Conclusion' section in the revised manuscript to summarize our findings and their implications.

Response to comment on Figures 3 and 4: We have increased the font size on all figures to ensure readability.

Round 2

Reviewer 1 Report