the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 06 Oct 2025
Seasonal forecasting of water table elevation in shallow unconfined aquifers with a case study in the Umbria Region, Italy
Abstract. Accurate seasonal forecasting of water table elevation is critical for effective water resource management in unconfined aquifers, particularly under climate variability and anthropogenic pressures. This study presents a novel methodology for predicting water table elevation on seasonal timescales by coupling reanalysis and seasonal forecast data of soil moisture with a calibrated nonlinear transfer model. The approach leverages ERA5 reanalysis and SEAS5 seasonal forecasts to estimate flux toward the aquifer and forecast water table elevation. A case study in the Umbria Region of central Italy demonstrates the model's ability to simulate and predict monthly water table fluctuations. Two modeling strategies are compared: a static calibration approach (OPT1) and a dynamic calibration approach (OPT2), where model parameters are updated by considering different time periods. Both options yielded skillful forecasts across lead times of 1 to 6 months, with OPT2 showing slightly improved stability in forecast performance metrics. Results confirm the feasibility of incorporating seasonal climate forecasts into operational groundwater prediction frameworks. As expected, forecast accuracy is limited by the skill of precipitation predictions, especially during autumn and winter. The proposed framework lays the groundwork for anticipatory aquifer management and early warning systems under evolving hydroclimatic conditions.
- Preprint
(3494 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- AC1: 'Comment on egusphere-2025-3677- Minor clarification on Figure 1 by Authors', Miriam Saraceni, 11 Nov 2025
-
RC1: 'Comment on egusphere-2025-3677', Anonymous Referee #1, 21 Nov 2025
The manuscript is fairly well written, and tries to tackle a relevant research question – establishing simple models for groundwater level predictions allowing to perform seasonal forecasts of this important hydrological variable.
However, I do have major concerns about the validity of the reported work in the current form. I hope these can be addressed by the authors, to ensure the relevance of their work.
In the conventional understanding, “numerical” groundwater models are physically-based, i.e. try to encode physical processes e.g. via solving governing equations of Darcy’s flow or similar. Your model should be considered an empirial or statistical model, similar to the widely used nonlinear transfer function noise models - and it should be discussed in this context. Another question is: Why do you not simply use those very well-established TFN models? See them e.g. implemented in the python package pastas https://ngwa.onlinelibrary.wiley.com/doi/full/10.1111/gwat.12925
The empirical formula used does not allow for memory effects in groundwater levels that go beyond the monthly scale – they assume direct scaling from each month’s recharge to the same month’s groundwater level – again, this limits the applicability of the model to very shallow aquifers without impact from slower-reacting groundwater flow or recharge. Again, the question arises why you do not use established TFN models, which, in many implementation handle this issue, i.e. include past stresses in their response function.
Having said that, I struggle seeing that sound conclusions can be reached based on results from a single wells. Groundwater level hydrographs and responses to climatic input (or, related soil moisture and derived recharge) are highly diverse. I strongly suggest to apply this method for various wells, which should be a doable task, potentially spanning different aquifer settings etc., assuming that this will yield interesting insights both for the authors and any reader of the manuscript.
You deselect wells that show (i) human influence, and (ii) limit yourself to a quite narrow (and shallow) range of groundwater depths (lines 65 ff). I guess I understand (i) for reasons of simplicity/model development. However, I would postulate that in the end you are in particular interested in forecasts in exactly such disturbed aquifers – you yourself set the scene in the introduction with the context of water management strategies. There is little or no management in undisturbed wells. Furthermore, limiting yourself to shallow wells you once more exclude relevant aquifers – which, in many cases for drinking water or other water supply – originate from deeper aquifers/wells due to water quality concerns etc. Your choices severely limit the applicability of your method within the context you yourself set out – at the very least this needs to be discussed. Also, please note that there are enough relatively simple modelling approaches – such as the mentioned TFN - that also allow easy integration of interference such as pumping – e.g. https://doi.org/10.1016/j.jhydrol.2016.01.042
Finally, why do you not use precipitation (or net precipitation, i.e. P – ET) as input directly? Or at least test this as another alternative. After all, soil moisture in reanalysis datasets is a modelled product itself, with inherent uncertainties, based on the climate data of the reanalysis. Especially as global-scale reanalysis must contain soil parameterization with very limited validity at point scale.
Some more detailed comments:
Line 38: “Recently, as an alternative option, […]” There are many alternative options. Please elaborate a bit more.
Line 40: ”reanalysis covers all the world and is fully open access” There are different reanalysis datasets, and this certainly does not apply universally
Performance evaluation: I am missing a benchmark performance of the model, without the effect of SEAS5 forecast uncertainty. Best would probably be the performance of the model forced with ERA5 input directly. Furthermore, how skilfull is the model really? A reasonable KGE at monthly scale for these kind of time series can probably be achieved very easily, as, in principle, we see the same/a very similar seasonal patterns with winter (spring) highs and summer (autumn) lows repeated every year. Those originate from the seasonal patterns of (net) precipitation, which the seasonal forecast easily reproduces. As a benchmark, the performance of a simple climatology should be considered – or similar. I.e. calculate the average water table across all January, February etc values to obtain the climatology – and then compare this against the actual monthly timeseries. Any forecast model has to be better than this “forecast” that can be obtained “for free”.
Table 4, 5, 6: Performance values should probably provided with to valid digits
Citation: https://doi.org/10.5194/egusphere-2025-3677-RC1 -
AC2: 'Reply on RC1', Miriam Saraceni, 26 Nov 2025
Dear Anonymous Referee #1,
Thank you for the attention you paid to our paper. The issues you pointed out are of great interest, and we will address them in the revised version of the paper, within the fixed deadline.
We do hope that in such a way the quality of our manuscript will improve.
Citation: https://doi.org/10.5194/egusphere-2025-3677-AC2
-
AC2: 'Reply on RC1', Miriam Saraceni, 26 Nov 2025
-
RC2: 'Comment on egusphere-2025-3677', Anonymous Referee #2, 19 Jan 2026
In this study a system is developed to provide seasonal outlooks of groundwater levels up to 7 months into the future. The method uses soil moisture estimates from the SEAS5/ERA5 systems to predict the groundwater levels with natural fluctuations. The method is tested on a single monitoring well in Italy using monthly mean groundwater levels. The results show that reasonably accurate forecasts can be provided, but large deviations for higher groundwater levels in springtime. The use of soil moisture rather than meteorological input directly is an interesting approach to forecast groundwater levels, which I did not encounter before. Below I outline some major concerns, with minor line comments further down.
Major comments:
- The introduction section does mention any review of studies attempting to generate seasonal groundwater levels outlooks/forecasts, while literature is available. I recommend including a one or two paragraph review of existing work and clearly stating what the proposed methodology might add to that work (i.e., using different input data – soil moisture).
- The methodology is tested on a single well, which makes it difficult to generalize the results, that also shows clear seasonal fluctuations and probably a good predictability. It would be good if the Authors could extend the work to include a few more wells, to see if/how the method generalizes.
- The approach to how the groundwater level model is developed is only briefly described (section 2.2) and I found it challenging to follow without any details about what was done here. Things become a bit clearer when the case study is described in Section 3 / Formulas 9 and 10. I would prefer Section 2.2 to be more in-depth, and generalizable to other monitoring wells with different model structures.
- There is no estimation or even discussion of the model parameter uncertainty, which probably is substantial. I think this should at least be discussed but preferably added to the simulation results.
- The most important part that I am missing is a comparison with other ‘naïve’ forecasts such as persistence or climatology. This is a very common practice in the forecasting community and helps identify the source of the forecast skill. I think adding such forecasts from simpler systems is essential to truly understand if the newly developed system brings additional forecast skill. I recommend adding such naïve forecasts and adding skill scores (i.e., CRPSS) to investigate this. I suspect simpler systems without the soil moisture input data, particularly when taking annual seasonality into account, might perform similarly.
Minor comments:
L14: “due to groundwater”?
L19-20: Please specify or rephrase; the potential is determined by many more factors in terms of quality and quantity than just the water table.
L20: Perhaps remove words such as “very” here, and similar instance hereafter.
L31: remove “numerical”, there are other model types that can do this.
L33: remove “much more challenging indeed”
Figure 1: What is OPT1? Please clarify this in the figure caption.
L56: What is done in the other cases?
L68-69: This is a limitation of the approach that could be discussed. Also, please specify how it was determined that no other (substantial) influences are present?
L124: What are the reference values? Please clarify
L144: instead of “improve calibration”, perhaps “improving verification” is meant here?
L179: No other piezometers fulfill these criteria?
Table 4: Please add numbers with 2 decimals
L250: I think the raw value says little, and 0.5 meters seems quite a large MAE to me. Perhaps add the variation of the GWL fluctuations for comparison?
Figure 4: Please consider providing Excel tables with the raw values of all simulations and measurements to reproduce this figures and the table.
L286-287: This statement is a bit vague, please clarify what is exactly meant here.
L330: Add: “for this data and case study area” or something similar, because we do not know how well the results generalize based on one well.
Citation: https://doi.org/10.5194/egusphere-2025-3677-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 251 | 77 | 30 | 358 | 19 | 18 |
- HTML: 251
- PDF: 77
- XML: 30
- Total: 358
- BibTeX: 19
- EndNote: 18
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors wish to revise Figure 1 of the published preprint of the manuscript. This revision does not affect the presented methodology, analysis, or results. On the contrary, it addresses two minor typographical errors in the original flowchart. The first issue is that the soil moisture forecast, provided by SEAS5 and used in the forecasting procedure, should have been written as θf , as the superscript f indicates the forecasted value. The second one is that, for the same reason, the forecasted relative flux toward the aquifer should have been written as Fg ts,f instead of Fg ts. In fact, the superscript f highlights the distinction between the datasets used for model calibration and those for the seasonal forecasting of the water table elevation. In addition, the figure layout has been refined to improve clarity and readability. The color scheme of several blocks has been adjusted, and the directional arrows between the blocks have been slightly modified to make the procedural flow more intuitive. The authors hope that these minor graphical and labeling revisions enhance the clarity of the proposed methodology.