Summary

Reanalysis products are globally gridded climate observations produced using numerical weather prediction models within a data assimilation framework (e.g., ERA5; Hersbach et al., 2020). These products often misrepresent climate variability and patterns in regions where localized conditions significantly influence climatology and where observational networks are sparse. However, reanalysis products effectively capture large-scale atmospheric dynamics (e.g., circulation patterns). This information can be combined with local observations to develop a transfer function, which can be used to downscale their data through statistical methods such as bias correction (e.g., ibicus; Spuler et al., 2024) and Perfect Prognosis (e.g., pyESD; Boateng & Mutz, 2023). Additionally, analogue models can be used to identify historical atmospheric patterns or weather states, which helps fill in missing observations or extend observations back in time. 

The manuscript describes a Python package (RASCAL v1.0.0) that incorporates all the necessary steps for using analogue models and Principal Component Analysis (PCA) to combine reanalysis (ERA20C and ERA20M) and station observations (precipitation and temperature) to reconstruct and extend climatological time series (as illustrated for four stations in Spain). The manuscript is well-written and the authors have clearly presented their work. However, I have major comments that should be addressed before publication, along with some minor comments for further clarification or additional detail.

General Comments

Introduction and motivation seciton: The authors provide detailed information about the decline of climate observation networks and its consequences. However, the emphasis on these issues overshadows the main focus of the manuscript, which is gap-filling or reconstructing missing records. I encourage the authors to shorten the introduction and focus more on the methods used, as the paper does not address the global increase of observational networks.

Evaluation Metrics: The manuscript lacks clarity on how the reconstruction was compared to actual observations. It would be beneficial if the authors elaborated on this. For instance, which period was used for training or identifying the analogue patterns, and which period was retained for independent performance assessment? Even if cross-validation was employed, a detailed explanation is necessary. Additionally, using reanalysis observations as a reference for evaluating reconstruction performance is questionable. Station-based reconstructions, being trained specifically for a station, will naturally outperform reanalysis. The critical aspect is the prediction error on the daily scale between observations and reconstructions for the period not included in model training. The authors should consider splitting the data into training and testing sets or randomly excluding parts of the time series for evaluation.

Model Development: The package documentation needs improvement. More detailed instructions on using the package and its functionalities should be provided. Although the documentation will evolve as user numbers grow, the current state is insufficient for users to get started easily. If this is done properly, the code snippets in the manuscript can be removed and referred to the documentation website.

Package Installation: I encountered issues while installing the package from both GitHub and PyPi. The PyPI upload lacked the necessary modules, causing import errors post-installation. The GitHub repository also lacked a setup configuration file, complicating installation after cloning. I urge the authors to test the installation in an independent environment to ensure functionality and highlight this in the installation section of the documentation.

Testing the Package: The authors should consider uploading synthetic datasets or the actual datasets used in the illustrative study. This will enable users to test the package before adapting it to their needs. Current software design decisions seem to rely heavily on the specific data used, which may hinder easy testing by new users. I believe reducing the size of the predictors (reanalysis data) to only the area used in the study or even the PCA time series will not be that huge to upload on Zenodo as supporting material for the manuscript. 

By addressing these comments, the manuscript can be significantly improved, providing a more robust and user-friendly tool for climate observation gap filling and extension

Specific comments

L3  “Reanalyses face challenges in examining..” change examining to representing

L4-5: add a sentence about how empirical downscaling resolves the issue of reanalysis mentioned

L6: “designed to address gaps in observational climate data..” modify the sentence to highlight how it addresses gaps (designed to fill gaps?)

L9: “outperforming reanalysis in conveying climatic characteristics.” more details is missing before this statement. Maybe only mention this when you’ve introduced the basis of the evaluation

L18-19: “However, there are challenges to consider, such as the requirement for long-term data series and susceptibility to disruptions caused by changes in land use or urbanization processes..” the author mentioned that the package is useful for addressing gaps in observational data with limited long-term data. But is mentioned here that a long-term data series is required. Please clarify. Also, rephrase susceptibility to disruption to something specific.

L52: “implicitly account for all the involved physics through complex mathematical..” delete all here since such statement is overstretching 

L66: “satellite measurements are crucial for assessing the Earth’s atmospheric conditions and perform the numerical weather prediction..” you meant “..for monitoring Earth's atmospheric conditions and evaluating the performance of numerical weather predictions” 

L67: “thanks to satellites, they may be behind the gradual decrease in the number of operational surface meteorological stations around the world..” I couldn’t understand the point raised here. Maybe modify the sentence and rephrase “thanks to satellites” 

L70-71: I don’t understand the point raised here too since reanalysis data also relied on weather stations. Please kindly clarify. 

L86-87: I don’t understand what the authors meant by may be behind the steady and heterogeneous decrease observed… please clarify. 

L92: “are the result..” change to as a result of 

L98-100: The author should be specific about the categories of downscaling. This should be Dynamical downscaling (e.g., RCMs) and Empirical Statistical Downscaling(ESD). ESDs are grouped into Model Output Statistics, Perfect Prognosis and Weather generators (in which analogue models are used). 

L109: “climate change scenarios..” you mean climate change information

L112: The authors should clarify what is meant by the regionalization method

L139: The resulting PCs from the EOFs are predictors or predictand?

L281: So if the objective is to evaluate low temporal resolutions like monthly, why not train the model directly with monthly data? The authors should present more evidence as to why it is accepted to train the model on high resolution but evaluated on low resolution

L316: The authors should clarify what they meant by the main hydrological resource

L331-332: The authors should provide references to support the statement about the longest records

L341-342: The authors should provide more details on why these predictors were selected and also test the sensitivity of the reconstruction to other predictors. Even though the authors mention that details are presented in other studies, a summary of the reason here would be useful to readers 

L369: You mean Figure 4f

L372: Where is it shown that the reconstruction is sensitive to the pool size selection?

L374: What is scientific inquires?

References 

Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049. https://doi.org/10.1002/qj.3803

Spuler, F. R., Wessel, J. B., Comyn-Platt, E., Varndell, J., & Cagnazzo, C. (2024). ibicus: a new open-source Python package and comprehensive interface for statistical bias adjustment and evaluation in climate modelling (v1.0.1). Geoscientific Model Development, 17(3), 1249–1269. https://doi.org/10.5194/gmd-17-1249-2024

Boateng, D., & Mutz, S. G. (2023). pyESDv1.0.1: An open-source Python framework for empirical-statistical downscaling of climate information. Geoscientific Model Development Discussions, 1–58. https://doi.org/10.5194/gmd-2023-67

This paper describes the structure and performance of an open tool developed by the authors to complement and extend climatological time series based on statistical downscaling of reanalysis/GCM data. It could contribute substantially to climatological studies especially on statistical properties of climatic variables at stations subject to localized effects. My comments are as follows.

1. Since selection of appropriate predictors is crucial to acquire the best estimation, it is desirable to state more specific and systematic methodology to select predictors from numerous meteorological variables in the reanalysis/GCM dataset at lines 186-188 and lines 341-343.

2. Since RASCAL is a downscaling tool of reanalysis/GCM data, I think its strength comparing to existing downscaling methods should be clarified rather than describing improvement from the reanalysis data in detail (in conclusions for example).

3. To evaluate the real performance of the estimation model, observed data used for model validation should be excluded from model calibration. (For example, by splitting the observed period into calibration and validation periods)

4. There are many options (closest, average, quantile) to adjust estimation to observation data. In such case, I think robustness of the selected option should be tested to make sure if it is valid only for the calibration data. (For example, again, by splitting the observed period into two or several periods and comparing the selected option)

5. For temperature reanalysis data for the four stations in chapter 5, are they corrected according to difference of elevation? It seems simple elevation correction using typical adiabatic lapse rate gives quite good estimation. The reconstructed data should be compared to the corrected reanalysis data.

The followings are minor corrections for specific words or phrases.

6. L16-17: “thus…biases” is repeated.

7. L56-57: “From a few hundreds of…to several thousand” According to Fig.1a, several thousand (around 7,000) at the end of the nineteenth century and several tens of thousand (around 30,000) at the end of the twentieth century.

8. L129: Title of Section 3.1 appears here, but Section 3.2 doesn’t exist.

9. L218: “In this method is possible” > “In this method it is possible”?

10. L219: “The the number of” > “The number of”?

11. L321: “sumarized” > “summarized”

12. L322-323 and L333: “observations of precipitation and temperature since 1893” and “Observations…available since 1948” are confusing as explanations of the data period of Retino station.

13. L389: “ahow” > “show”

14. Table 1: It is better to change the order to S, N, C, R according to the order in Figs. 4-10.

15. Fig.4, Fig.5, Fig.6 captions: “a, d, g, j, m” > “a, d, g, j”, “b, e, h, k, n” > “b, e, h, k”, “c, f, i, l, o” > “c, f, i, l”

16. Fig.7, Fig.8, Fig.9, Fig.10 captions: “a, b, c, d, e” > “a, b, c, d”, “f, g, h, I, j” > “e, f, g, h”, “k, l, m, n, o” > “I, j, k, l”