A statistical model to relate landslide activity with rainfall in the Emilia-Romagna region (Modello statistico per la previsione operativa dei fenomeni franosi nella regione Emilia-Romagna)

This paper deals with a statistical model to define warning thresholds for shallow and deep-seated landslides in Emilia-Romagna, Italy. The study was performed in the framework of a joint research project between the Earth Sciences Department of Florence University and Emilia-Romagna Regional Meteorological Survey (ARPA-Meteo), with the aim of implementing a warning system for prediction of landslide risk in real time. The implemented model compares historical rainfall series with landslide mobilization, and has been tested in the Reno valley, Northern Apennines. The landslide typology in the study area includes earth slides caused by prolonged rainfall, usually consisting of reactivations of pre-existing dormant landslides (BERTOLINI & PELLEGRINI, 2001), as well as less frequent shallow landslides and earth flows. A careful analysis of the relationship between lithology and landslides has shown that the mobilizations take place prevalently in clayey soils (BENEDETTI, 2003). A broad and accurate database on landslide occurrence and daily rainfall was available for the investigated area. The rain gauge of Porretta Terme (mean annual rainfall slightly higher than 1200 mm per year) was selected as the representative rain gauge for the analysis because of both its central location and long series of daily rainfall data; furthermore, that station is able to transmit data in real time. Using landslide archives and rainfall data dating back to 1950, it is shown that landslide behaviour more or less corresponds to the periodic pattern of precipitation with a lag of approximately six months, which is considered to be related to the time required for the ground water level to reach a critical point for failure. The analysis was then continued with a statistical model concentrating on the pluviometric thresholds for landslides. The rainfall series were statistically analysed in order to find the limit between ordinary and anomalous events. The implemented model was based on cumulative daily rainfall with windows of variable length (from 1 to 365 days) sliding with daily steps from 1950 to 1996. For each cumulative length, the original cumulated values were transformed into the standard normal distribution, checking the fit with the original data (GOOVAERTS, 1997). Then, a standard deviation figure was associated with a corresponding value of the actual cumulative rainfall. Values were then selected to represent the probable thresholds between ordinary and anomalous conditions upon which alert curves were built. An alert system scheme was built by trial and error technique, using for calibration the 1950-1996 datasets of rainfall and landslides. The approach used for the definition of the warning procedure accounted for the typology and different behaviour of landslides. Regarding shallow landslides, cumulated rainfall from 1 to 10 days are analysed and compared with two alert curves, indicating a situation of medium and high risk in cases of exceeding the 1,5σ and 2σ curves, respectively. For deep-seated landslides, since they are affected by prolonged rainfall and eventually by a trigger event occurring usually at the end of rainy periods, the model takes into account both long and short period cumulated rainfall. Cumulated rainfall from 10 to 360 days is first examined. In cases of no overcoming of the 1,5σ alert curve, the risk is considered at the ordinary level. Otherwise, the cumulated rainfall from 1 to 5 days before the date of analysis was considered. In case of overcoming of the 1σ alert curve, the model indicates a high-risk situation; otherwise, a condition of medium risk is issued. Preliminary results obtained for the calibration period (19501996) indicated that in 81% of days with landslides the antecedent rainfall curve exceeded one of the warning threshold curves: for most of them (43%) the model prompted a high level of risk, whereas in 38% of cases a condition of medium risk was pointed out. The model was only unable to predict less than 19% of the mobilizations. The model was successively tested on an independent dataset, constituted by rainfall data and landslide mobilizations recorded in a 3-month period (September-December 2002) within the study area. Results of this validation phase indicated a similar model performance, in terms of relationship between warning level and landslide occurrence, but highlighted also a high number of false alarms, thus limiting its applicability for civil protection purposes.