the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 11 Jul 2025
Towards impact-based early warning of drought: a generic framework for drought impact prediction in the UK
Abstract. Drought impact forecasting is essential for enhancing preparedness and mitigation strategies. However, identifying key predictors and achieving reliable predictions remains challenging. Previous studies have shown promise in developing indicator-impact relationships and yet these are often region- and impact type-specific. Here, utilized the European Drought Impact Inventory (EDII), and a wide range of meteorological and hydrological predictors, including the Standardized Precipitation Index (SPI), Standardized Precipitation-Evapotranspiration Index (SPEI), and soil moisture indices (SSMI), to develop a generalized forecasting framework for predicting drought impacts in the UK across different lead times. We firstly compared multiple machine learning models for drought impact prediction and identified Random Forest (RF) as the most effective model. Our results show that RF delivers the highest accuracy for short-term forecasts (0–3 months), with performance declining beyond six months, similar to trends observed in weather prediction models. At longer lead times, the model incorporates a broader set of predictors to maintain accuracy. Key findings highlight the importance of long-accumulation-period drought indicators, particularly SPEI24, and deep-layer soil moisture (SSMI L4), which were identified as the most influential predictors. A generalized model approach was employed, aggregating drought impacts from various regions, and the model was validated using unseen datasets from within the UK, using parts of the EDII UK dataset held back from the training, confirming its robustness. A pilot application to a completely different country (Germany) highlights the potential for extrapolation to new domains. Gridded impact predictions were also developed, and successfully captured the spatial distribution of observed impacts, and a spatially explicit evaluation showed reasonable agreement between predicted and observed drought impacts. Although uncertainties persist, particularly for long lead times, our findings suggest that a generalized approach based on hydrometeorological indices provides an effective framework for operational drought impact forecasting, supporting early warning systems and decision-making in drought risk management.
- Preprint
(1806 KB) - Metadata XML
-
Supplement
(260 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-3176', Anonymous Referee #1, 08 Sep 2025
-
AC1: 'Reply on RC1', Burak Bulut, 09 Nov 2025
>> We thank the reviewer for their positive assessment of the manuscript and appreciate the constructive feedback, which we will address to further improve the work.
>> We agree that a more explicit quantification of multicollinearity would strengthen the discussion and improve transparency. In response, we will follow the reviewer’s suggestion and include additional analyses, such as a correlation matrix or variance inflation factors, to demonstrate the extent of interdependence among the predictors used in the linear models. The manuscript will be revised accordingly, and the results of these additional analyses can will be provided in the supplementary materials.
>> We appreciate the reviewer’s concern regarding the dataset. The impact dataset used in this study is derived from the European Drought Impact Inventory (EDII), for which our team provided and curated the UK component. At present, this represents the most comprehensive and peer-reviewed source of drought impact data for the UK. Although preliminary data for more recent years exist, they have not yet been fully processed, validated, or published, and therefore were not used in the current study.
We will explicitly state the training period (1970–2012) in the abstract to improve clarity. While the current dataset limits the model’s direct application to more recent, unprecedented droughts (e.g., 2018–2022 events), the proposed modelling framework is transferable, and forthcoming updates to the UK EDID dataset will enable future studies to extend forecasts to these recent events.
Finally, following the suggestion of the second reviewer and to address concerns regarding potential confusion between EDII and EDID datasets, we propose to remove the EDID data citation and reference from the manuscript, as these data were not used in the present study.
Minor comments
>> We agree that referring to the upper tercile as “extremes” may overstate the nature of the threshold, since it includes the top 33% of values rather than very rare events. To address this, we note that in the manuscript we retained the original naming convention from Shorakaya et al. (2022) for consistency with the literature. However, for clarity in the discussion, we are planning to describe the terciles as minor, moderate, and significant, corresponding to the lower, middle, and upper terciles, respectively. The lower tercile represents minor impacts, the middle tercile moderate impacts, and the upper tercile significant impacts. This terminology more accurately reflects the statistical distribution of the data while preserving the impact-based focus of our study.
>> We will review the manuscript carefully and revise the identified passages (Lines 60, 142, and 449) to improve clarity. In addition, we will examine other sections with similar constructions to enhance readability throughout the manuscript.
Citation: https://doi.org/10.5194/egusphere-2025-3176-AC1
-
AC1: 'Reply on RC1', Burak Bulut, 09 Nov 2025
-
RC2: 'Comment on egusphere-2025-3176', Kerstin Stahl, 28 Sep 2025
Dear Burak and colleagues,
apologies for the delay of this review and thank you for the opportunity to comment on this manuscript. I enjoyed reading about your study and provide my comments below. They are minor.
Regards
Kerstin
(Kerstin Stahl)
Summary
The manuscript reports a study that applies a range of advanced statistical models to the task of predicting drought impact occurrence under certain meteorological and soil moisture conditions in the UK. To do that it used training data from a text-based category-coded impact database and indices from operational monitoring or hydrometeorological conditions. The contribution is a valuable application that tests and compares various statistical modelling options that have not previously compared to that extent. The study also evaluates the potential operational use of this application, specifically prediction/forecasting of impacts with various lead times. The paper is well generally well written and makes an important contribution. A few aspects need some improvement to provide a more focused and consistent message and therefore the impact the paper deserves. They should be fairly easy to implement.
Main comments
The title is quite long. In light of potentially misleading terminology, I suggest that 'impact-based' as well as "generic" might be removed from the title. But there may also be other solutions.
My main comment relates to those two terms and the focus and consistent message of the paper. The two terms come with ambiguity and are used very differently in the literature. I think their precise use could be improved throughout the manuscript.
(1) "impact-based forecasting" is used interchangeably with "impact forecasting/prediction" in the title and text. While cited sources have used that term, some of the literature on climate impacts uses "impact-based-forecasting" differently, e.g. selecting ensemble members of a physical model based on impact information; the impact, however, is then not directly forecasted. Strictly speaking, "impact forecasting/prediction" might therefore be more correct. However, at least consistency and introduction/discussion might be improved on that.
(2) "generic framework" suggests to me a standardized procedure of applying this in operation as indicated that it will be and/or a procedure that is transferrable to other data and regions. It speaks a bit against the rather detailed analytic comparison and analysis of several statistical models and the different forecasts that is the main aim and in fact in my opinion the main value and contribution of the study.
For consistency in the aims and main contribution made with the study, I strongly suggest to either tone down this 'generic framework aim' or explain in more detail what it is exactly in the end - perhaps including a flow chart or so. The methods generally have been applied previously, so what the general methodological 'developing' is, might also be clarified.
Figure 1 goes a bit into that direction but is not entitled "framework". So which part of it is the framework? And would that operational framework always use all model options? i.e. train all, but then apply/predict with the best? Or how is this transferred to the framework of application?
Data statement and line 102ff
The latest (and likely last) version of the EDII is available with doi and should be cited as:
Blauhut, V., Stephan, R., and Stahl, K.: The European Drought Impact report Inventory (EDII V2.0), [data set], Uni Freiburg, Freiburg, https://doi.org/10.6094/UNIFR/230922, 2022
This is our preferred reference, because the website that is given by the authors no longer functions correctly.
I don't see the reason to cite the new EDID database in the data statement as it was not used. Please consider that while EDID ingested a major part, but not all content of the former EDII, along with other databases, it uses different categories and different attributes than EDII. Therefore, naming it here and in lines 102ff as one and the same is misleading as using it for the same purpose might provide different time series of NI etc.
For your information: A paper on the new db is in progress and about to be submitted to NHESS. Futhermore, guidelines for interested contributors on how to transfer EDII-categories into the new EDID-systems are already available:
Szillat, K., Hlavsová, M., Rossi, L., Blauhut, V., Stahl, K.: Transformation of text-based drought impact data from EDII (European Drought Impact report Inventory) to EDID (European Drought Impact Database): Guidelines. Freiburg HydroNotes no. 8. https://doi.org/10.6094/UNIFR/271380, 2025
Minor comments
"Short term" = 0-3 months? For weather/floods anything more than a few days would be considered long term and not short term. Discussion and use of terminology might be improved.
line 13 'Here, "we" used ...?
Line 344: what is meant by 'regional information'? The impact occurrence? Why not use 'predictor' and 'predictand' or 'response' or so or better even, a variable name.
line 674. EDID is not global. Replace with something global or say 'Europe'.
Figures
Figure 5 what are dashed and solid lines? A legend with symbols and line types is strongly preferable over the difficult to read/miss caption text.
Also, in Figure 2 it is confusing that some components have a legend and others don't. It should at least be consistent.
The caption in Fig. 11 needs to name all panels (or a to d...etc..) but just singling out some is inconsistent.
Citation: https://doi.org/10.5194/egusphere-2025-3176-RC2 -
AC2: 'Reply on RC2', Burak Bulut, 09 Nov 2025
>> We thank the Kerstin Stahl for her positive assessment of the manuscript and appreciate the constructive feedback, which we will address to further improve the work.
>> We thank the reviewer for this suggestion. After clarifying the terminology used in this study regarding the generic framework and impact-based forecasting, we can either retain the current title or revise it to improve clarity and avoid potential misunderstandings. Two possible revised titles are:
(1) “A framework for drought impact prediction and early warning in the UK”; or
(2) “Developing a UK framework for drought impact forecasting and early warning.”
We will carefully consider these options and select the most appropriate version in the revised manuscript.>> Thank you for your comment regarding the terminology. We have clarified the distinction between impact prediction and impact forecasting in the revised manuscript. Specifically, impact prediction refers to modelling observed impacts using historical data, while impact forecasting refers to anticipating potential impacts using lagged indicators. This clarification (given below) will be added to the Introduction.
“In this study, impact prediction refers to modelling observed impacts using input data from when the impacts have already occurred (in other words, zero leading time). In contrast, impact forecasting uses lagged indicators to anticipate potential impacts in advance (i.e., 1-3-6 ... 24-month leading time). While some literature uses the term “impact-based forecasting” more broadly or differently (e.g., selecting model ensemble members based on impact information without directly forecasting impacts), we adopt this terminology consistently throughout the paper to maintain clarity.”
>>Thank you for these helpful comments. Based on these clarifications regarding the scope and purpose of the generic framework, including the selection of the best-performing model and its transferability to other regions, we will update the manuscript text to improve clarity for readers. Figure 1 will also be revised accordingly to explicitly illustrate the framework and its main steps.
In this study, we aimed to develop a generic framework that uses the lumped impacts across the entire United Kingdom for training, rather than training separate models for each NUTS1 region. The primary goal of this approach is to produce a robust model that can later be applied to other climatically similar countries (e.g., Germany) or regions where local impact data are insufficient to train separate models. The framework itself defines a structured modelling approach in which all available indices are used as input predictors and all available impact data are used as the predictand. Although the framework allows for multiple modelling approaches to be tested, in practice the best-performing model identified from the prediction step is used for forecasting.
As part of this study, we compare different modelling approaches—ranging from linear models to machine learning models, which are generally used in the literature—to identify the best-performing model for use within this framework. Using the best-performing model (Random Forest in our case) from the prediction step (zero lead time), we generated impact-based forecasts by incorporating lagged indicator information to anticipate potential future impacts, ranging from one month up to 24 months ahead.
The main contribution of this study is threefold: first, the comparison of different models trained on all available UK data to identify the best-performing approach for impact prediction; second, the validation of this generic model’s transferability using unseen spatial and temporal data across the UK as well as on unseen datasets from Germany, including evaluation of gridded impact predictions at the NUTS1 level for both regions to allow direct comparison with previous studies; and third, the application of this generic framework to generate impact-based forecasts using lagged indicators, which are subsequently validated.
>> We thank the reviewer and will correct the EDII citation as suggested.
>> We thank the reviewer for this comment. After consideration, we have decided to remove the EDID reference and citation from the text, as it was not directly used in our study. We agree that mentioning it alongside the former EDII database could cause misunderstanding or confusion, given that EDID covers only part of EDII and uses different categories and attributes. This change has been applied in the relevant sections to ensure clarity and accuracy.
Minor comments
>> All minor comments regarding edits, grammar, and sentence corrections will be carefully considered and addressed in the revised manuscript.
Figures
Figure 5 what are dashed and solid lines? A legend with symbols and line types is strongly preferable over the difficult to read/miss caption text.
>> Thank you for this comment. The dashed lines indicate the training results, while the solid lines show the validation AUC results. We agree that the current caption is difficult to read. In the revised manuscript, we will add a clear legend with symbols and line types to the figure for clarity.
Also, in Figure 2 it is confusing that some components have a legend and others don't. It should at least be consistent.
>> Thank you for pointing this out. We will ensure consistency across all components of Figure 2, so that all elements either include a legend or are clearly labeled, making the figure easier to interpret.
The caption in Fig. 11 needs to name all panels (or a to d...etc..) but just singling out some is inconsistent.
>> Thank you for this suggestion. We will revise the caption of Figure 11 to name all panels (a–d) consistently, rather than singling out some panels, to improve clarity and readability.
Citation: https://doi.org/10.5194/egusphere-2025-3176-AC2
-
AC2: 'Reply on RC2', Burak Bulut, 09 Nov 2025
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 1,151 | 70 | 25 | 1,246 | 41 | 28 | 33 |
- HTML: 1,151
- PDF: 70
- XML: 25
- Total: 1,246
- Supplement: 41
- BibTeX: 28
- EndNote: 33
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript "Towards impact-based early warning of drought: a generic framework for drought impact prediction in the UK" is well written and presents a relevant study. The authors evaluate multiple models for predicting drought impacts, carefully train and validate them with different lead times. Overall, the study is rigorous, the results are clearly presented, and the framework seems solid and well analyzed. I identified two shortcomings that should be revised.
First, the discussion of multicollinearity is unsatisfactory. Readers need to trust the authors’ assertion that multicollinearity is present but varies. Finally, this concern is brushed aside in line 605 with the sentence: “Although SPI and SPEI indices are highly correlated in the UK, the RF model is capable of managing this multicollinearity.” However, multicollinearity is indeed a problem for Linear Regression and LASSOCV, which are compared in the first step before being outperformed. A more explicit quantification is necessary (e.g. showing a correlation matrix or variance inflation factor).
Second, the dataset (1970–2012) is quite outdated, even though more recent data (up to 2024) seems to be available, as noted in the data availability statement. This weakens the study’s relevance and raises questions about whether it still represents state-of-the-art work. This is particularly disappointing since the introduction references more recent drought events (2018–2019: Turner et al., 2021; 2022: Barker et al., 2024) and sets expectations that are not met when entering the methods section. I suggest clearly stating the training period already in the abstract (e.g. “trained with data from 1970–2012”) and including a discussion on whether the model would be capable of forecasting more recent, unprecedented droughts. Ideally, if possible, predictions for these years could be shown.
Minor comments
• Defining the upper tercile as "extremes" seems overstated, particularly since the threshold includes the upper 33%.
• Several passages are overly long and difficult to follow due to heavy nominalization. Please streamline them for clarity or remove if they add no value. Examples: Line 60/ Line 142/Line 449