the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 Dec 2025
A 225-Year (1799–2024) Homogenized Daily Water Level Series of the Vistula River in Warsaw
Abstract. We present a 225-year (1799–2024) homogenized daily water level series for the Vistula River in Warsaw, comprising 82,453 observations. The construction of this consistent dataset required adjustments for changes in gauge location, shifts in gauge zero, differences in historical measurement units, and calendar discrepancies between the Julian and Gregorian systems. A small number of missing observations were reconstructed using stage–stage relationships established between overlapping periods of observation at the Warsaw gauge and parallel measurements from downstream stations along the Vistula. The resulting dataset offers a robust foundation for long-term hydrological, climatic, and socio-environmental research. The dataset is openly available at Zenodo repository: https://doi.org/10.5281/zenodo.16919654 (Sobechowicz et al., 2025).
- Preprint
(1857 KB) - Metadata XML
-
Supplement
(159 KB) - BibTeX
- EndNote
Status: open (extended)
-
RC1: 'Comment on essd-2025-538', Anonymous Referee #1, 18 Jan 2026
reply
-
AC1: 'Reply on RC1', Łukasz Sobechowicz, 11 Mar 2026
reply
We would like to thank the Reviewer for their thorough reading of our manuscript and for providing such constructive feedback. These insightful comments have been immensely helpful in improving the quality of our work.
Specific comments:- A simple linear regression based on the neighbouring gauges has been used to fill the data. They have also selected specific highs and lows in the data for this purpose. The authors have observed a lag of up to 4 days for the waters to travel to the neighbouring gauges. Apart from this, the authors also use the maximum, minimum, and onset points of the flood data to fill in data that do not represent these conditions. In short, I find the gross assumption of an average time lag and the use of specific points overly simplistic, given the nonlinearity of the hydrodynamics of flows of varying intensity. More evidence of these kinds of assumptions would be useful for the readers to understand the reliability of the filled data. For instance, using cross-correlation between the time series of different gauges to show the time lag.
Authors' Response:
We thank the Reviewer for this insightful and methodologically significant comment. We fully agree that the hydrodynamics of the Vistula River are inherently non-linear and that wave propagation velocity varies depending on discharge intensity. To address the concern regarding the reliability of our lag assumptions, we have conducted an extensive empirical verification using the Cross-Correlation Function (CCF). This analysis was performed on three high-resolution historical datasets to ensure that the chosen lags are not merely "simplistic assumptions" but represent the robust physical characteristics of the river reaches in a daily temporal resolution.
- Methodological Approach: First-Order Differencing
To ensure statistical rigor, all CCF calculations were performed on first-order differenced data. We deliberately avoided using raw water level values because such series are highly non-stationary and exhibit strong autocorrelation. In our preliminary tests, raw data consistently yielded a maximum correlation at Lag 0, which is a known artifact of spurious correlation in hydrological time series (where shared seasonal trends mask the actual wave propagation). By applying differencing, we isolated the short-term hydrological signals (impulses). This methodological rigor allowed us to identify the true physical travel time of the flood wave with high statistical confidence.
- Empirical Evidence for Travel Time Lags
The analysis was conducted across extensive datasets to capture different hydrological regimes:
- Warsaw – Cypel Mątowski (Reach 1): Analyzed for 1806 and 1818–1828 (approx. 4,400 daily pairs). The CCF reached its maximum at Lag 4, confirming our initial assumption. Monthly heatmap analysis shows that in periods of high activity (e.g., August), the correlation coefficient reaches at exactly 4 days.
- Warsaw – Toruń (Reach 2): To verify this reach, we analyzed two distinct periods:
- Early Period (1817–1830): Approx. 4,900 daily pairs.
- Later Period (1908–1924): Over 6,200 daily pairs.
In both cases, despite the century-long gap and potential changes in the riverbed, the strongest statistical correlation was consistently observed at Lag 2.
- Addressing Hydrodynamic Non-linearity
The Reviewer correctly pointed out the non-linearity of flows. Our monthly heatmap analysis indeed captures this phenomenon:
- During spring freshets or high-flow months (e.g., March), the correlation is sometimes distributed between two or three days.
- However, Lag 2 (for Toruń) and Lag 4 (for Cypel Mątowski) remain the modal values, the most statistically frequent and strongest signals across the entire dataset.
In a model operating on a daily resolution, adopting these modal lags is the most reliable approach for data reconstruction. The stability of these signals across more than 15,000 analyzed days (total) demonstrates that these average values are robust empirical reflections of the river's behavior during the study period.
The suggested revisions will be included in the revised version of the manuscript.
- The authors could have used robust techniques like Long Short-Term Memory (LSTM) neural networks for data filling, as they can capture non-linear temporal patterns while selectively retaining information relevant to the current output in long time series. The authors need to present the existing variations in time lag across different gauge locations and discuss the associated uncertainties. Please refer to the work done by Ren et al. (2022) in this regard.
Authors' Response:
We would like to thank the Reviewer for suggesting the use of LSTM neural networks. In response, we have incorporated this robust technique into our workflow. Rather than simply replacing the initial linear models, we have chosen to provide both LM and LSTM reconstructions in the final dataset. This enables us to discuss the associated uncertainties more effectively and demonstrates the leap in accuracy achieved by using neural networks for capturing non-linear temporal patterns, as suggested by the Reviewer. We have incorporated the recommended reference (Ren et al., 2022) to support this methodological shift.
1. Model Architecture and Training Protocol
To ensure maximum reliability and avoid the pitfalls of stochastic initialization, we implemented a rigorous training framework:
- Three Specialized Models: Based on data availability and gauge locations, we developed three distinct LSTM models:
- Warsaw–Cypel Mątowski (1799–1828): Reconstructing gaps for 1800–1816.
- Warsaw–Toruń (1817–1830): Reconstructing gaps for 1817.
- Warsaw–Toruń (1908–1924): Reconstructing gaps for 1914–1918.
- Hyperparameter Optimization: We conducted a comprehensive Grid Search to identify optimal configurations for window size (7, 15, 31 days), LSTM units, dropout rates, and batch sizes. The final models were selected based on the lowest Mean Squared Error (MSE) on the validation set.
- Overfitting Prevention: Models were trained for 50 epochs with Early Stopping and Gradient Clipping (clipvalue=1.0) to ensure stability and generalizability.
2. Validation methodology
Standard k-fold cross-validation is unsuitable for autocorrelated hydrological time series as it violates the independence assumption. Random partitioning causes 'data leakage', where future observations inform past predictions, resulting in artificially inflated performance metrics. Instead, we employed a Chronological Split (Training/Validation/Testing sets). This ensures that the testing data remained entirely "unseen" by the model during training, providing a true assessment of its predictive performance on continuous historical gaps.
3. Feature Engineering and Seasonality
To enhance the models' understanding of river dynamics, we utilized a multi-feature input vector:
- Hydrological Inputs: Raw water levels and daily increments from the reference stations.
- Cyclical Encoding: To account for the strong seasonality of the Vistula (spring snowmelts vs. summer droughts), we applied trigonometric transformations (Sine/Cosine) to the temporal data. This allows the LSTM to process the time of year as a continuous, circular feature, which is crucial for hydrological accuracy.
The suggested revisions will be included in the final version of the manuscript.
- The validation of the proposed linear-regression-based data filling is incomplete. A robust k-fold cross-validation is needed to assess the accuracy of the proposed data-filling methods. The data needs to be split iteratively into training and validation sets. Importantly, the error needs to be reported using RMSE, NSE, KGE, etc.
Authors' Response:
We have evaluated both the original Linear Models (LM) presented in the article and our newly developed LSTM architectures. Detailed results are summarized in the table below. The analysis indicates that the LM models exhibit significant inconsistency, with performance varying substantially across different years. In contrast, the LSTM models demonstrated superior accuracy and higher stability, consistently yielding lower error margins. Consequently, the LSTM-based reconstruction results will be integrated into our database, and all corresponding updates will be incorporated into the revised manuscript.
- The authors themselves say that the bed level varied extensively due to gauge relocation and anthropogenic activities. They have taken these into consideration and modified the zero level. They need to comment on the reliability and accuracy of these adjustments.
Authors' Response:
The corrections were based on documented gauge relocations and archival leveling records. The reliability of these corrections is particularly high for data recorded after 1834, as that was when the first professional leveling and connection to the height reference system were established. However, data from before 1834 may contain errors due to the reconstruction of the gauge following winter damage, which lacked a reference to a known benchmark. The precision of subsequent zero-level adjustments to the gauge depended on the accuracy of the geodetic instruments available at the time, and it is not expected to exceed 1 cm.
- Even high-resolution remote-sensing-based digital elevation models may not accurately represent riverbed topography. In addition, extensive cross-sectional surveys are needed to simulate the correct water levels that reflect the corresponding conditions. Therefore, the discharge time series is often highly useful to the hydrological community for validating hydraulic/hydrological models or for climate research. Can the authors provide a discharge time series or comment along these lines?
Authors' Response:
Hydrological data on daily discharges of the Vistula River at the Warsaw gauging station are collected and made available by the IMGW - Institute of Meteorology and Water Management in digital form for the period from 1 November 1950 to 30 October 2024. They are available in Open Access format at: https://danepubliczne.imgw.pl/.
Attempting to reconstruct time series of discharges for the Vistula River in Warsaw for the period before 1950 is definitely justified, extremely difficult and a very interesting challenge that we would like to take on in the future. This work will require a lengthy search for information on historical hydrometric measurements (cross section of the channel, flow velocity) for the Vistula River collected in various archives in Poland, Germany and Russia. Due to the very complicated history of Poland, these measurements have been widely dispersed and, at this stage, their storage locations for the 19th century have not been identified. Russian archives in particular are currently inaccessible to us. To sum up, in the future we will attempt to reconstruct and make available a series of historical values of the Vistula River discharges in Warsaw.
The minor comments are as follows,
- Line 68, “km XXX of the Vistula River”, is difficult to understand. Please change similar lines in the manuscript.
- Line 127, what do the authors mean by “km 421 + 600”? Please modify similar lines to provide greater clarity to readers.
Authors' Response:
Location refers to the official river chainage (river kilometer system) used in Poland, measured along the river course upstream from the river mouth. we have changed the indicated sections to improve readability.
- In Table 4, what does the last column indicate?
Authors' Response:
The last column contains information about the number of observation pairs used to construct each LM.
- Line 290, please use a clearer term than “early measurements”.
Authors' Response:
We have changed the indicated sections to improve readability.
- Lines 389 -417: the summary need not include the uses of the dataset.
Authors' Response:
When writing the summary, we followed the guidelines for authors. One of the elements was to indicate the potential use of the dataset.
-
AC1: 'Reply on RC1', Łukasz Sobechowicz, 11 Mar 2026
reply
Data sets
Daily Water Levels of the Vistula River at Warsaw, 1799–2024: A Complete and Homogenized Long-Term Record Ł. Sobechowicz et al. https://doi.org/10.5281/zenodo.16919654
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 417 | 191 | 32 | 640 | 75 | 33 | 28 |
- HTML: 417
- PDF: 191
- XML: 32
- Total: 640
- Supplement: 75
- BibTeX: 33
- EndNote: 28
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors have done an impressive job of creating a 225-year dataset of daily water levels in Warsaw. Such a long data record is crucial for climate research, hydrological modelling, and flood risk management. They have exhaustively included data from publications and yearbooks (somewhere in different units and different locations) prior to 1981. Nevertheless, there are some serious technical concerns that need to be addressed before publication. They are as follows,
The minor comments are as follows,
Reference
Ren, H., Cromwell, E., Kravitz, B., and Chen, X.: Technical note: Using long short-term memory models to fill data gaps in hydrological monitoring networks, Hydrol. Earth Syst. Sci., 26, 1727–1743, https://doi.org/10.5194/hess-26-1727-2022, 2022.