We rely on two datasets: (1) operational forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) and (2) the ERA5 reanalysis dataset , covering the period 1950–2024. ECMWF forecasts are used operationally for real-time event tracking, while ERA5 is employed to construct the historical catalogue of events. The variables considered are daily mean temperature at 2 m above ground level, accumulated daily precipitation, and daily maximum sustained wind speed. The analysis focuses on Europe, within the geographical domain [30° N, 75° N, 30° W, 40° E].

The operational single-level forecasts are produced with the ECMWF Integrated Forecasting System (IFS) and are distributed under the ECMWF open data license (https://apps.ecmwf.int/datasets/licences/general/, last access: 20 June 2025). We use the Atmospheric Model high-resolution 15 d forecast (HRES), provided at a spatial resolution of 0.25° × 0.25°, with four forecast cycles per day (00:00, 06:00, 12:00, and 18:00 UTC). For consistency and timeliness, we retain only the forecasts initialized at 00:00 UTC, which are disseminated between 05:45 and 06:12 UTC according to the ECMWF schedule.

For each forecast, we use temperature at 2 m and neutral wind components at 10 m (u and v) at 3-hourly lead times (3, 6, 9, 12, 15, 18, 21, and 24 h), as well as accumulated precipitation at 24 h lead time. Daily mean temperature at 2 m is computed by averaging the eight 3-hourly temperature values. Daily maximum sustained wind speed is calculated by first deriving wind speed at each lead time as u2+v2 and then retaining the maximum value over the day.

ERA5 reanalysis data are obtained from the Copernicus Climate Data Store (https://cds.climate.copernicus.eu/, last access: 2 September 2026) at a spatial resolution of 0.25° × 0.25°, with a typical latency of about five days. We retrieve daily variables using the CDSupdate Python package , which directly provides daily aggregates. The same variables as for the forecasts are used: daily mean temperature at 2 m, 24 h accumulated precipitation, and daily maximum wind speed.

Events detected in real time using forecast data are re-evaluated every six months using ERA5 reanalysis and subsequently integrated into the historical database. We expect only limited differences between the two products, as ERA5 is initialized from ECMWF operational analyses at zero lead time.

We use the International Disasters Database (EM-DAT, ; https://public.emdat.be, last access: 2 September 2026) as a reference dataset for optimizing the RHITA algorithm parameters. EM-DAT is the most comprehensive global database of disasters and catalogues high-impact events at the country level, including those driven by natural and technological hazards. For each recorded event, the database provides information such as casualties, economic losses, temporal extent, and event magnitude. Data are compiled from a wide range of sources, including UN agencies, non-governmental organizations, reinsurance companies, research institutes, and press agencies. Owing to this diversity, the dataset is not fully homogeneous and may contain reporting gaps and biases .

The objective of RHITA is to construct a historical catalogue of significant disasters that is consistent with EM-DAT while offering a more quantitative and dynamical representation of events from the perspective of the underlying hazards. Focusing on Europe, we align EM-DAT with our optimization framework by filtering events based on the disaster type and disaster subtype fields. This results in three event categories: heatwaves, cold spells, and storms. The detailed filtering criteria are summarized in Table .

Although EM-DAT data are available from 1900 onward, records prior to 2000 are particularly affected by reporting biases and inconsistencies . For this reason, we restrict the analysis to the period 2000–2023.

The RHITA algorithm is implemented as a Python package (https://github.com/gcazzaniga/rhita, last access: 2 September 2026) developed for the spatio-temporal detection of hazards associated with extreme weather events. The algorithm processes three-dimensional gridded datasets (longitude, latitude, time) in three main steps, described below and illustrated in Fig. .

Thresholding. Identification of grid cells exceeding a quantile-based or fixed threshold q, computed or selected over a historical reference period.

The algorithm depends on four key parameters: the quantile-based threshold q, the spatial distance threshold d, the minimum event area Amin⁡, and the centroid linking distance dc. These parameters must be selected to reflect the relevant spatial and temporal scales of each hazard type. For the threshold q, we use the 0.99 quantile for extreme high temperatures, extreme winds, and heavy precipitation, and the 0.01 quantile for extreme low temperatures. Quantiles are computed at each gridpoint using all days from the 1950–2023 reference period.

This choice represents a trade-off: the threshold must be sufficiently extreme to identify intense events, while remaining sensitive enough to capture less intense but potentially impactful events driven by long duration, large spatial extent, or high vulnerability of affected areas. Using quantile-based thresholds allows adaptation to local climatological conditions, although it also implies sensitivity to the chosen reference period. The remaining parameters (d, Amin⁡, dc) are optimized independently for each hazard type. The optimization procedure is described in Sect. .

Once events are detected, several metrics are computed to characterize their intensity, extent, and duration. These metrics are listed in Table , together with their mathematical definitions. For each hazard type and for metrics such as mean excess, maximum excess, total impacted area, and duration, we estimate the return period RP over the full historical time series and the entire European domain, as a measure of event rarity. The return period of a metric m is computed following as 1RPm=λ⋅11-FM(m), where λ = 1/μ, μ is the mean annual event occurrence rate, and FM(m) denotes the cumulative distribution function of metric m.

In this study, RHITA is applied to detect heatwaves, cold spells, and storms using ERA5 reanalysis and ECMWF forecast data. In principle, however, the algorithm can be applied to any gridded dataset to provide a Lagrangian detection of spatio-temporal extreme events.

As described in Sect. , the EM-DAT database is used as a reference dataset for optimizing the RHITA parameters d, Amin⁡, and dc. The optimization objective is to maximize the detection of extreme weather events associated with significant societal impacts.

For each hazard type, physically plausible parameter ranges are defined based on the characteristic spatial and temporal scales of the underlying atmospheric processes (see Sect. S1 of the Supplement for details). A regular parameter grid is constructed, and a grid search is performed to evaluate all parameter combinations.

First, we identify all parameter triplets (Amin⁡,d,dc) that maximize the detection sensitivity – the so-called recall in machine learning terminology , defined as 2sensitivity=true positivestrue positives+false negatives.

A positive detection is defined as an algorithm-detected event that satisfies two conditions: (1) spatial overlap with at least one country affected according to the EM-DAT record, and (2) temporal overlap within a ±10 d window around the EM-DAT event start or end dates. This temporal tolerance reflects the typical time scales of extreme weather phenomena, ranging from a few hours to one or two weeks .

Whenever multiple parameter triplets achieve the same maximum sensitivity, the optimal configuration is selected by minimizing the cost function S: 3S=nsNdet+ncNdet+Nobs-Ndetmax⁡Nobs,Ndet, where ns is the number of erroneously split events, nc the number of erroneously merged events, Ndet the total number of events detected by the algorithm, and Nobs the number of events in the reference dataset.

The cost function consists of three normalized terms, each bounded between 0 and 1:

a splitting penalty quantifying erroneous event fragmentation;

Because the reference dataset is affected by reporting gaps and biases inherent to human-compiled disaster databases, false positives (i.e., algorithm-detected events absent from EM-DAT) are intentionally not penalized in the optimization procedure. For the same reason, the sensitivity values reported here should be interpreted as a partial evaluation of the algorithm against documented, impact-relevant events, rather than as a comprehensive assessment of its meteorological detection skill: events that are meteorologically extreme but do not appear in the impact-oriented EM-DAT record are not necessarily missed detections in a physical sense, and a full quantification of detection skill would require an independent meteorological reference not available here.

Heatwaves are identified as temperatures exceeding the 0.99 quantile, while cold spells correspond to temperatures below the 0.01 quantile. For heavy precipitation and wind extremes, the 0.99 quantile of daily accumulated precipitation and daily maximum sustained wind speed is used, respectively. For all hazard types, a minimum event duration threshold of three consecutive days is imposed.

The optimization is performed using EM-DAT records from 2000 to 2023, which are randomly split into training and validation sets. Two-thirds of the years are used for training, while the remaining one-third is used for validation. The optimized parameters for each hazard type are reported in Table . The optimization procedure yields sensitivities exceeding 0.75 for all hazards except cold spells.

Heatwaves and cold spells are optimized jointly, as reported in Table . This choice is motivated by the fact that temperature extremes over Europe are strongly linked to recurrent large-scale meteorological patterns, in particular quasi-stationary anticyclonic circulation and atmospheric blocking . However, the algorithm exhibits lower detection skill for cold spells in both the training and validation phases. A plausible explanation is that cold spells are reported between 2 and 3 times more frequently than heatwaves in the EM-DAT database. This higher reporting rate likely includes many moderate events, whereas the algorithm applies a symmetric detection threshold designed to capture only the most extreme occurrences, leading to a larger fraction of missed cold spell events. However, using separate thresholds would implicitly calibrate the algorithm to EM-DAT reporting biases; our choice of a symmetric threshold preserves a physically based and comparable definition of temperature extremes.

Parameters for heavy precipitation are optimized separately using EM-DAT records filtered specifically for storm-related events (see Table ). For wind detection, we adopt the same parameter set as for heavy precipitation. This choice reflects the difficulty of isolating wind-specific disasters in EM-DAT, as no single, unambiguous criterion exists to identify events driven exclusively by strong winds. Consequently, sensitivity metrics are not reported for wind events.

To construct the European climatology, only detected events that, at least at one time step, impact a European country are retained. Events occurring entirely outside European countries within the analysis domain (e.g., exclusively over North Africa) are excluded from the climatology presented here.

Over Europe, RHITA detects a total of 760 heatwaves, 681 cold spells, 805 heavy precipitation events, and 686 strong wind events over the period 1950–2024. All detected events within the analysis domain are available for browsing and downloading through the RHITA web tool.

Figure  shows the spatial distribution of detected hazards across European countries. For each country, a pie chart represents the relative proportion of the four hazard types: heatwaves (red), cold spells (blue), extreme wind events (orange), and heavy precipitation events (light blue). In Southern Europe, RHITA detects a predominance of heatwaves and extreme wind events in Spain and Portugal. In Italy and Greece, as well as in most Western European countries, extreme precipitation and extreme wind events represent the largest fraction of detected hazards. In the United Kingdom and Ireland, heatwaves account for a substantial share of detected events, with Ireland also exhibiting frequent strong wind events, while cold spells contribute more strongly in the United Kingdom. Central and Eastern Europe are characterized by a dominant contribution from heavy precipitation and wind events, with heatwaves generally detected more often than cold spells. In Northern Europe, the detected events are more evenly distributed across the four hazard types.

It is important to note that this aggregation is based on the geographical impact footprint of the detected events. Consequently, a single event can contribute to multiple countries when its spatial extent affects more than one national territory.

To investigate large-scale spatial patterns in the temporal evolution of extreme event frequencies, detected events were aggregated by European subregions. The subregions considered are Central and Eastern Europe, Northern Europe, Southern Europe, and Western Europe, following the EuroVoc classification maintained by the Publications Office of the European Union . The list of countries included in each subregion is provided in Sect. S2.

For each subregion and hazard type, trends in yearly event frequency were assessed using the Mann–Kendall test . When statistically significant, the magnitude of change was quantified using Theil-Sen’s slope estimator . Figure  summarizes the results, showing the change in yearly event frequency per decade for statistically significant trends only.

Heatwaves exhibit a robust and significant increase in frequency across all four European subregions. The strongest increase is detected in Southern Europe, with an increase of 0.96 events per year per decade, followed by Western Europe (+0.73), Northern Europe (+0.56), and Central and Eastern Europe (+0.45). Cold spells display an opposite behavior, with decreasing frequencies, particularly in Northern Europe, where the yearly frequency declines by approximately 0.87 events per decade.

Trends in heavy precipitation and strong wind events are more heterogeneous across subregions. Nevertheless, a significant increase in strong wind events is detected in Western Europe (+0.33 events per year per decade), along with a significant increase in heavy precipitation events (+0.32 events per year per decade).

In addition to event frequency, trends in event characteristics were analyzed at the European scale using the metrics defined in Table : mean excess, maximum excess, total impacted area, and duration. Trend detection was performed over the 1950–2024 period using the Mann–Kendall test at a 5 % significance level, with trend magnitude estimated using Theil–Sen's slope. This analysis is performed on the aggregated European climatology. As a result, regional-scale signals may be partially masked by spatial averaging. All figures supporting this analysis are provided in Sect. S3.

Heatwaves exhibit statistically significant increasing trends in several metrics. The maximum excess temperature increases by 0.07 °C per decade, while the mean excess increases by 0.02 °C per decade. The total impacted area per event also shows a significant increase of approximately 80 000 km² per decade, corresponding to roughly twice the surface area of Switzerland. No statistically significant trend is detected for event duration.

Cold spells display negative slopes for maximum excess temperature, mean excess temperature, and total impacted area. However, none of these trends are statistically significant at the 5 % level. A statistically significant trend is detected for event duration, but the associated slope is close to zero, indicating a negligible change over the analyzed period.

For heavy precipitation events, no statistically significant trends are detected for any of the analyzed metrics. The trends in maximum excess (p = 0.094) and mean excess (p = 0.072) are close to the significance threshold, while total impacted area (p = 0.493) and duration (p = 0.851) show no evidence of systematic change.

Extreme wind events show no statistically significant trends in mean or maximum excess, nor in total impacted area. A statistically significant trend is detected for event duration; however, the associated slope is close to zero, indicating no substantial change in duration over time.