To improve the results obtained by Tiranti and Ronchi (2023), we considered an updated and extended analysis that encompasses 123 widespread landslide events from 1960 to 2023 (collected from historical archive of national newspaper “La Stampa” - http://www.archiviolastampa.it - and from the regional technical reports archive - https://www.arpa.piemonte.it/ricerca/pubblicazioni?field_tema_target_id=18), distinguishing shallow landslide events through a classification into seven severity classes related to the number of landslides that occurred during a rainfall event (Table 1).

A weak upward trend is highlighted from 1960 to 2023, indicating that shallow landslide events are progressively becoming more intense in terms of the number of landslides occurring during a single rainfall event, considering landslide events having a severity class ≥4 (moderate or higher) (Figure 1A).

On the contrary, the number of shallow landslide events shows a decreasing trend in the same time interval divided into five-year periods, as illustrated in Figure 1B.

We used ERA5-Land hourly precipitation data from 1 January 1960, to 31 December 2023, for Piemonte and neighboring areas (longitude: 6.5°E to 9.5°E; latitude: 44°N to 46.5°N), sourced from the Copernicus Climate Data Store. For a preliminary validation of applied bias correction technique, the hourly data were aggregated into daily cumulative precipitation to match the temporal resolution of the regional dataset developed by the Regional Agency for Environmental Protection of Piemonte (ARPA Piemonte), which uses OI techniques. The OI dataset contains daily cumulative precipitation (along with maximum and minimum temperatures, not used in this study) from 1959 onwards, on a 0.125° grid (∼15 km resolution) covering the Piemonte region. The ERA5-Land data were interpolated onto the OI grid points using the “remapcon” function provided by the Climate Data Operators toolset (Schulzweida, 2023).

To minimize biases in the ERA5-Land data, robust Quantile Mapping (QM) was applied daily, based on the 1991–2020 climatological reference period, using the fitQmapRQUANT function in the “qmap” package in R software (Gudmundsson et al., 2012). To evaluate the performance of QM we computed and plotted (see Figure 2A of Section 3) the average monthly precipitation over the entire study period. For further validation at a finer scale, we applied a geographical mask for the Piemonte region to both the ERA5-Land dataset and the adjusted dataset. For both, we calculated the mean daily regional precipitation for each day from 1 January 1960, to 31 December 2023. In order to test the performance of each dataset after QM bias correction in recognizing wet (≥1 mm rain) vs. dry (<1 mm rain) days, contingency tables were calculated, performance metrics were computed, and a performance diagram was plotted (see Figure 2B of Section 3) using the “verification” package in R software, with the OI dataset serving as the reference.

After checking the validity of daily QM for correcting biases in ERA5-Land data, 24-h rolling cumulative precipitation totals were calculated for all possible 24-h windows (e.g., 00:00–23:59; 01:00–00:59 D+1; 02:00–01:59 D+1, etc.), creating 24 distinct datasets. This step was decisive for testing the R-SLEWS of Piemonte, particularly for rainfall events that last for several days, where the total precipitation over a 24-h window might exceed the threshold, even if individual days do not. The rolling windows were essential to capture events where the cumulative precipitation exceeds the threshold outside of the standard 00:00–23:59 timeframe, ensuring that no significant events were overlooked due to the fixed daily boundary. These events, which might exceed landslide-triggering thresholds but go undetected by standard daily 00:00–23:59 UTC precipitation data, are now accounted for by this rolling approach. The QM-corrected coefficients were applied to each of 24-h datasets to align the ERA5-Land data with the observed OI precipitation, ensuring consistency across all datasets used for validation and modeling. The treatment of the climate datasets described in this paragraph was performed using the Climate Data Operators (Schulzweida, 2023).

Each of the 24 dataset was then analyzed to assess their ability to predict shallow landslide events triggered by rainfall. For each dataset, we identified the dates when the average 24-h cumulative precipitation exceeded the trigger thresholds defined by the R-SLEWS developed by Tiranti et al. (2019): 70 mm for hilly areas and 120 mm for mountainous areas. These dates were subsequently compared against a database of 120 shallow landslide events, collected from technical reports and newspaper archives. It is important to note that some of these landslide events were very widespread and affected both mountainous and hilly areas. For the purposes of this study, these widespread events were also divided into 179 sub-events (fragmenting a single event into sub-events based on the geographical environment involved) and analyzed individually.

In the updated study here presented, the dataset was composed of 185 sub-events derived from a total of 123 widespread events, adding to the previous 120 events those that occurred between 2020 and 2023:

• 2–3 October 2020: the landslide event involved both Alps and TPB environments with a range of shallow landslides between 500 and 1,000 (high severity class);

For each 24-h dataset, we applied an elevation mask to extract precipitation data for areas at elevations between 400 and 700 m asl (designated as “hills”) and higher than 700 m asl (designated as “mountains”). For each dataset, we generated a list of dates on which precipitation exceeded the triggering threshold (either 70 mm or 120 mm, depending on elevation and related geological features) within moving 24-h windows. These lists were then automatically cross-referenced with the dataset of shallow landslide events compiled for Piemonte to identify all matches within each 24-h window. We then refined the results by removing any duplicated dates.

To further explore the detection capabilities of our approach, we applied running sums to each of the 24 datasets. This allowed us to obtain cumulative sums over two consecutive timesteps, for each dataset. Such data were then used to compile 48 datasets, each covering a time window of 48 h (00:00 D – 23:59 D+1; 01:00 D + 00:59 D+2, and so on). From each dataset we then extracted list of dates on which precipitation exceeded the triggering threshold (either 80 mm or 170 mm, depending on elevation and geological features) within moving 48-h windows. After cross-referencing such lists with the shallow landslide dataset to identify all matches, the results were then refined by removing any duplicated dates.

The final output of our analyses, for both mountainous and hilly regions, was a list of dates on which bias-corrected precipitation surpassed the triggering thresholds, coinciding with an observed shallow landslide event. In this study, if a shallow landslide event was widespread and extended over multiple consecutive days, we considered it as successfully identified if the triggering threshold was exceeded on at least 1 day during the event duration.

Mean monthly precipitation of the OI dataset, the ERA5-Land dataset and the QM-corrected dataset are represented in Figure 2A. Raw ERA5-Land data generally capture the observed precipitation patterns from the OI dataset, correctly identifying May, October, and November as the wettest months. However, ERA5-Land systematically overestimates average monthly precipitation (Pearson correlation coefficient: 0.88; RMSE: 33.82), with this overestimation being less pronounced in spring (MAM) and fall (SON) but markedly higher in summer (JJA) and winter (DJF). The application of QM to the ERA5-Land data noticeably reduces these biases and aligns with the trends and values observed in the OI dataset (Pearson correlation coefficient: 0.96, RMSE: 9.03).

The effectiveness of daily QM in reducing biases is confirmed when the analyses are repeated for the mean daily regional precipitation, in order to evaluate the ability to distinguish wet and dry days. The QM-corrected dataset, when compared to the observed data, exhibits a high correlation (Pearson correlation coefficient: 80%), with a RMSE of 4.43 mm and a SD of 6.98 mm. The performance diagram of the bias-corrected dataset in identifying wet vs. dry days is represented in Figure 2B. The statistical metrics (POD: 0.76; SR: 0.73; CSI 0.60; bias: 1.04) confirm that the bias-corrected dataset is reasonably accurate in detecting and forecasting precipitation; notably, the bias value indicates that the QM-corrected dataset does not significantly over-estimates or under-estimates the number of wet days (Roebber, 2009).

Based on these results, daily QM emerges as an acceptable approach for bias correction of ERA5-Land precipitation data over Piemonte, showing good skills in capturing the natural variability of rainfall events. Consequently, the QM indices calculated using daily QM were applied to obtain the 24-h rolling cumulative precipitation.

Following the approach of Tiranti and Ronchi (2023), we categorized the landslide events based on their geographical setting into two groups: Alpine areas (mainly characterized by crystalline bedrocks) and Apennines/hilly areas (mainly characterized by sedimentary bedrocks). The latter group includes all shallow landslide events that occurred in the TPB, the Turin hills, and the Apennines foothill. Widespread shallow landslide events affecting multiple environments were divided into sub-events, which were then grouped based on the specific environment where each sub-event took place and treated accordingly. Final shallow landslide dataset included 123 events, that could further be subdivided in 87 sub-events in Alpine environments, and 84 sub-events in the hilly/Apennines environments (of which: 8 in the Apennines, 13 in the Turin hill, 77 in the TPB).

New approach allowed to correctly identify 61 shallow landslide events based on 24-h cumulated rainfall, out of the 123 collected in the landslide database of Piemonte. We correctly identified 33 shallow landslide events taking place in Alpine areas, and 49 taking place in the hills. Interestingly, the detection capability was only marginally increased by the application of sliding windows: daily cumulated rainfall between 00:00 and 23:59 UTC allowed to detect 59 events, while the analysis on different 24-h sliding windows provided only two additional events. Remarkably, such additional events were observed in the data regarding Alpine areas, triggered by heavy precipitations (exceeding 120 mm/24 h). The sliding window approach, on the other hand, had no appreciable effect on the detection of events taking place in hilly areas, which can be triggered by less intense rainfall (>70 mm/24 h).

When 48-h periods were considered, we successfully identified 68 events (of which 34 in the mountains and 54 in hilly areas). Once again, the application of sliding 48-h windows did not impact the detection of events in the hilly areas, while it sensibly improved the performance in Alpine areas (6 additional events). Only when 24-h periods and 48-h periods were pooled were we able to correctly identify 73 events, representing nearly 60% of the shallow land slide events recorded in our dataset. Table 2 shows in details the number of hits and misses alarms we obtained with our procedure.