Changes in daily precipitation frequency and distribution in Italy over the last 120 years

Authors


Abstract

[1] A new data set of 45 daily precipitation series, covering quite uniformly Italian territory for the period 1880–2002, was recovered. The series have been homogenized on daily basis, completed by means of statistical methods and grouped into five regions by a Principal Component Analysis. Seasonal and yearly total precipitation, number of wet days, and precipitation intensity were analyzed for each station record and averaged into five regional series for a synthetic description of the results. Proportion and frequency of daily rainfall amounts, belonging to six precipitation class-intervals, defined on the basis of some percentiles of the precipitation distribution, were also analyzed. The results show a negative significant trend in the number of wet days all over Italy, and a positive trend in precipitation intensity, which is significant only in the northern regions. The negative trend in wet days has persisted since the end of 19th century and is due to the marked decrease in the number of low intensity precipitation events. An increase in the number of events belonging to the highest intensity class interval was observed too, but only in northern regions.

1. Introduction

[2] One of the most important focuses of recent climatological studies is to characterize possible changes in climatic extremes, since they have the strongest impact on society. Modelling results have indicated increases in extreme precipitation in a warmer climate [Zwiers and Kharin, 1998; Kharin and Zwiers, 2000; Palmer and Räisänen, 2002]. Zwiers and Kharin [1998] studied changes due to CO2 doubling in the extremes of the surface climate, as simulated by the second-generation circulation model of the Canadian Centre for Climate Modelling and Analysis in 20-yr equilibrium simulations. They observed that precipitation extremes increase almost everywhere over the globe and return periods of extreme precipitation are shortened by a factor 2 and more in the 2XCO2 climate in many parts of the world. Moreover, the number of rainy days per year generally increases in polar regions and decreases in midlatitudes. Also Kharin and Zwiers [2000], by examining the extremes of precipitation in an ensemble of three climate change simulations for the period 1900–2100, found an increase almost everywhere and the observed increase was larger for larger return periods. Palmer and Räisänen [2002], by using a multimodel ensemble of 19 global coupled ocean-atmosphere climate models, estimated winter seasonal precipitation exceeding two standard deviations above normal will increase over much of central and northern Europe (with the probability of occurrence increasing to over 12%) over the next 100 years. They found a similar increase in the probability for the Asian monsoon region.

[3] Observational studies also suggest evidence of changes in climatic extremes. Karl et al. [1995] and Karl and Knight [1998] provide evidence of a statistically significant increase in precipitation greater than 50.8 mm per day in the United States from 1910s to 1990s. In a recent paper, Kunkel et al. [2003] analyzed a longer newly available data set of daily precipitation series to study the temporal variability of the frequency of short-duration extreme precipitation events for 1895–2000. They confirmed the results of Karl et al. [1995] and Karl and Knight [1998] for the recent period, but observed that heavy precipitation frequencies at the beginning of 20th century were nearly as high as during the late 20th century. Suppiah and Hennessy [1998] observed, for the period 1910–1990, an increasing trend in heavy daily rainfall (greater than the 90th and 95th percentiles) in Australia, both in summer and winter half years. Osborn et al. [2000] studied the intensity distribution of daily precipitation amounts in the UK and showed that it changed over the period 1961–1995, becoming on average more intense in winter and less intense in summer. Also Fowler and Kilsby [2003a, 2003b] observed that the frequency of extreme rainfall changed over parts of the UK in the period 1961–2000, in particular prolonged heavy rainfall events were increasing. In the Valencia region of Spain, González Hidalgo et al. [2003] observed a positive trend in most intense events as a percentage of annual rainfall, in spite of a negative trend in their volume.

[4] Results for Italy are presented in Brunetti et al. [2001a, 2002] for the period 1951–2000: the number of wet days has a clear and highly significant negative trend all over Italy and there is a tendency toward an increase in precipitation intensity which, in Northern Italy, is mainly owing to a strong increase in precipitation, which falls into the highest class-intervals, whereas in Southern Italy, it is spread over a wider part of the distribution of amounts of wet days.

[5] On a global scale, significant increases have been seen by Frich et al. [2002] in the extreme amount derived from wet spells and number of heavy rainfall events for the second half of the twentieth century. Easterling et al. [2000], in a brief review of climate extreme, observed that the tendency in most countries that have experienced a significant increase or decrease in total precipitation has been directly related to the change in the amount of precipitation during the heavy and extreme events.

[6] The relationship between the increase in total precipitation, and the frequency of heavy rain events was studied by Groisman et al. [1999], by applying a simple statistical model based on the gamma distribution to summer data of eight countries: Canada, the United States, Mexico, the former Soviet Union, China, Australia, Poland and Norway. The results showed that the shape parameter of the precipitation distributions remained rather stable, independent of total precipitation, while the scale parameter was more variable. If these results can be generalized and used as a model for the future, as total precipitation increases, a disproportionate increase in heavy precipitation has to be expected [Groisman et al., 1999].

[7] The relationship between precipitation intensity and total precipitation is, however, not general. In some places, such as Siberia in the summer season in the period 1936–1994, an increase in heavy precipitation was observed together with a tendency toward a decrease in total precipitation [Groisman et al., 1999]. A similar behavior was observed in Italy [Brunetti et al., 2000, 2001a, 2001b].

[8] One of the biggest problems in examining the climate record for changes in the occurrence of extremes is a lack of high-quality and high temporal and spatial resolution long-term data [Jones et al., 1999; Folland et al., 2000]. Furthermore, daily data are often not in digital form. In fact, as evident from the above-described state of the art, for most countries the analysis period starts from the end of World War II. Only in few countries (Costal regions of New South Wales and Victoria in Australia, United States, Norway, Natal in South Africa, and Northern Italy) analyses began near the start of the twentieth century [Easterling et al., 2000; Brunetti et al., 2000; Kunkel et al., 2003]. In recent years, many researchers have understood the necessity to set up new daily series data sets to extend further into the past the study of changes in extreme events. At the present the most relevant results come from the United States, where 920 station series were collected for the period 1895–2000, albeit with a nonuniform coverage [Kunkel et al., 2003]. As far as Europe is concerned, the most important attempt to realise a data set of daily resolution climatic time series was the European Climate Assessment (ECA) project [Klein Tank et al., 2002]. It focuses on the 20th century, from 1901 to 1999, and consists of about 200 series of minimum, maximum and/or mean temperature and daily precipitation amount. Unfortunately, only 50% of the series extends back to at least 1925. An attempt to test the series statistically with respect to homogeneity was made. As a result, a high percentage of series were labeled “doubtful” or “suspect” [Wijngaard et al., 2003]. This highlights the necessity for more effort from individual countries in setting up national homogeneous daily resolution databases to converge into wider international projects.

[9] Our goal is to do this for Italy. In this paper we will present the results of the analysis of a new daily precipitation database covering the last 120 years. The originality of this work lies, not only in the new data here presented, but also in the length of the series, in the wider metadata availability, and in the homogenization procedure performed on a daily basis.

2. Data

[10] Italy has some of the longest meteorological series in the world. After the observations recorded from 1654 to 1667 by the Accademia del Cimento, in Italy regular meteorological observations began in the eighteenth century in Bologna (1714), Padua (1725), Turin (1756), Milan (1763), Rome (1782) and Palermo (1791). In a number of other cities, e.g., Aosta, Florence, Genoa, Mantova, Modena, Naples, Parma, Pavia, Perugia, Udine, Trieste and Venice, observations started in the first half of the nineteenth century.

[11] This vast heritage of observations is still almost entirely available, even if it has been only partially transferred to a digitized form. Concerning precipitation, in the frame of some national and international projects we set up a new data set of daily series, the oldest ones starting from the second half of the 18th century. At the moment the database consists of 45 series quite uniformly distributed all over the Italian territory (Table 1 and Figure 1), but the data collection is still in progress to improve the area coverage. Within the same projects we also recovered monthly data for a larger number of stations (more than 100). This wider data set will be used in the near future to study total precipitation trends for the past two centuries with a higher and more uniform spatial coverage, in particular for Southern Italy where the density of daily series is rather low.

Figure 1.

Map of the stations (filled circles indicate stations that passed the homogenization procedure; open circles indicate stations rejected from the homogenization procedure). The five regions obtained from the PCA analysis (Figure 2) are indicated too.

Table 1. List of the Daily Precipitation Series, Location and Covered Period
Station NameStation CodeLat., degLon., degH, m aslSeries Length
AlessandriaALE44° 41′08° 45′981857–1986
L'AquilaAQU42° 21′13° 24′7351879–2002
ArezzoARE43° 27′11° 52′2741879–2002
BellunoBEL46° 07′12° 13′4041879–2002
BolognaBOL44° 29′11° 20′601813–2002
BolzanoBOZ46° 31′11° 21′2721921–2002
BraBRA44° 42′07° 52′3081862–2002
BrixenBRI46° 43′11° 39′5691921–2002
CagliariCAG39° 12′09° 19′551879–2002
CataniaCAT37° 30′15° 07′751901–2002
CuneoCUN44° 24′07° 31′5361879–1998
DomodossolaDOM46° 07′08° 17′3001872–1998
FerraraFER44° 49′11° 36′151879–2002
FirenzeFIR43° 47′11° 15′511860–2002
FoggiaFOG41° 28′15° 32′801901–2002
GenovaGEN44° 24′09° 08′531833–2002
LivornoLIV43° 33′10° 18′31876–2002
LocarnoLOC46° 10′08° 47′3791901–2000
LuganoLUG46° 00′08° 54′2761901–2000
MantovaMAN45° 09′10° 45′201840–1997
MessinaMES38° 12′15° 30′541881–2002
MilanoMIL45° 28′09° 11′1221858–2002
MontemariaMMA46° 44′10° 29′13231923–2002
MontevergineMOV40° 55′04° 47′2701884–2002
NapoliNAP40° 50′14° 15′1491866–2002
PadovaPAD45° 24′11° 52′141877–2002
PalermoPAL38° 06′13° 21′711787–2002
ParmaPAR44° 48′10° 18′571878–2002
PaviaPAV45° 10′09° 09′751883–1979
PerugiaPER43° 06′12° 23′5201874–2002
PesaroPES43° 52′12° 52′111871–2002
PiacenzaPIA45° 01′09° 43′501875–2002
Reggio CalabriaRCA38° 06′15° 39′151878–2002
Reggio EmiliaREM44° 42′10° 37′621879–2002
RomaROM45° 52′10° 45′731866–2002
RoveretoROE41° 54′12° 28′561921–2002
RovigoROV45° 52′11° 03′2061879–2002
SassariSAS45° 03′11° 46′91876–2002
SiracusaSIR40° 43′08° 36′2241874–2002
TarantoTAR37° 03′15° 17′231901–2001
Riva-TorboleRIV40° 27′17° 18′221921–2002
TorinoTOR45° 03′07° 40′2751802–2000
TrentoTRE46° 04′11° 07′1991921–2002
VallombrosaVAL43° 43′11° 00′9551872–2002
VeneziaVEN45° 26′12° 15′11900–2002

[12] Besides daily and monthly data we also set up a wide collection of metadata. Research activity concerning metadata was organized in order (1) to understand the evolution of the Italian meteorological network and (2) to reconstruct the history of all the stations with secular series.

[13] The network management displayed a very complex evolution, with short periods of rather high standardization and longer ones with the presence of different independent subjects (Central Office, Hydrological Service, Air Force). However, also in the period with higher standardization, there was not a complete homogeneity between instruments and methods used in the different observatories, as the most important ones were more autonomous because they had their own instruments and people with adequate knowledge to manage them and do experiments to check their reliability and calibrate different instruments. For these observatories the contribution of the Central Office in the station maintenance was rather marginal. In other cases, especially for the smaller observatories, the support of the Central Office was indispensable for both instruments and methods. The weak dependence of many observatories on the Central Office and the presence of different networks (Central Office, Hydrological Service, Air Force) for long periods are sometimes negative, because they lead to information losses or great difficulties in recovering station history information. On the other hand, the availability of data from different sources avoids the undesired problems produced by the simultaneous changes in the instruments or measurement methods of the whole network (not identifiable without information from different sources). An interesting example concerns rain gauge height above ground level. Even if all networks suggested the use of stations at ground level, a significant part of Italian observatories (e.g., Bologna Milan, Genoa, Rome, Palermo) used rain gauges on the top of buildings, more than 20 meters above ground level, for all observation periods.

[14] All the information on the evolution of the Italian network was collected in the frame of two national projects. One, “Reconstruction of the historical climate in the Mediterranean Area”, funded for the period 1997–1999 by the National Research Council, and the second, “CLIMAGRI” (www.climagri.it), financed by the Ministero delle politiche agricole e forestali, which is still in progress. At the conclusion of the first year (over a total of three) of the latter project, the authors produced a document containing a card for each data series where all the information about the station was summarized. Each card was divided into three parts. In the first part all the information obtained from articles and monographies was reported. In the second part there are abstracts from the epistolary correspondence between the stations and the Central Office. In the third part the sources of the data used to construct the record were summarized. An example of these cards is reported in the Web site of the project (www.climagri.it/FinaleMaugeriCap4.htm).

[15] Both the metadata on the evolution of the National network and that concerning the station cards contain a lot of information about instruments, station relocation and maintenance, changes of instruments and calibration problems, changes in the number of observations per day, in the observation time and in the people working at the station. All this information was very useful for a critical check of the quality of the data. At the end of the current projects involving the setting up of this data set, all the series will be available to the scientific community.

3. Data Homogenization

[16] Many studies over the past two decades have demonstrated that climate variability research is not possible without clear knowledge about the state of the data in terms of homogeneity. The real climate signal in original series is hidden behind nonclimatic noise caused by station relocation, changes in instruments and instrument screens, changes in observation times, observers, and observing regulations, algorithms for the calculation of means, and so on. Overviews focusing on the homogeneity problem and different ways of solving it are given in Peterson et al. [1998], and in the two volumes of the proceedings of the homogeneity seminars in Budapest [World Meteorological Organization (WMO), 1997, 1999]. However, we did not find in the literature any attempt at homogenizing daily series of precipitation.

[17] One of the biggest problems concerning daily precipitation data is that the series could be affected by two kinds of inhomogeneities: (1) in the precipitation amount and (2) in the number of rainy days. The latter can obviously generate an inhomogeneity also in the precipitation amount series. In fact, a period with some nonindicated missing data could be interpreted as a period with an underestimation of total precipitation and it could be badly corrected by increasing each single rainy day and originating some erroneous extreme events. So, we decided to check separately both total precipitation and the number of rainy days.

[18] Concerning the inhomogeneities due to the number of rainy days, their signal often disappeared from the homogeneity test curve if a different threshold to define a day as rainy was used (1 mm rather than 0.1 mm). This is due to instrument resolution: sometimes improvement in precipitation gauges leads to a higher number of rainy days with low precipitation amount [Nicholls and Murray, 1999]. Therefore the choice of a threshold to define the rainy days is not a negligible problem in this kind of analysis. Also the precipitation intensity, defined as precipitation amount per rainy day, is influenced by this choice. In the following statistical analysis all the precipitation events lower than 1 mm were neglected and considered as 0 mm events. Some other inhomogeneities in the number of rainy days were adjusted by the precipitation homogenization procedure, because, by correcting total precipitation, some days with below threshold amount were increased above threshold and vice versa, and the break signal in the test curve of the number of rainy days disappeared.

[19] We considered checks of both rainy days, and total precipitation, and we decided to correct only inhomogeneities in precipitation amounts if they corresponded to periods without problems in rainy days or with problems that could be solved by total precipitation homogenization. The remaining inhomogeneities in the rainy days were eliminated by invalidating the erroneous periods.

[20] Six of the 45 series displayed so many homogeneity problems that they were classified “not homogenisable”. At the end, a set of 39 single series constituted the final homogenized database.

[21] It was possible to perform a comparison of the identified inhomogeneities with the history of the stations and a high percentage of the breaks were explained through metadata. A total of 36 breaks were homogenized: for 15 of them there was a relocation or some changes in the instruments, 4 correspond to the World War periods and 5 to a re-start of the data collection after an interruption. So, a total of 24 breaks out of 36 were supported by information about the history of the stations.

4. Method for Checking and Adjusting Series on a Daily Basis

[22] Homogeneity testing and adjusting was performed in regional subgroups of 10 series using a version of the HOCLIS procedure [Auer et al., 1999] modified and adapted to daily data. HOCLIS rejects the a priori existence of homogeneous reference series. It consists of testing each series against other series, by means of a multiple application of the Craddock test [Craddock, 1979], in subgroups of 10 series. The break signals of one series against all others are then collected in a decision matrix and the breaks are assigned to the single series according to probability. This system also avoids trend imports and an inadmissible adjustment of all series to one or a few “homogeneous reference series.”

[23] The adjusting coefficients were calculated on a monthly basis, then they were fitted with a trigonometric function in order to obtain daily corrections without discontinuities from one month to another. These corrections were used to homogenize the daily records. Moreover, while in the HOCLIS procedure only one neighborhood series was used as reference to adjust one single break, here we considered more reference series and we calculated the correcting coefficients against each of them, and the average among the most coherent subgroup of coefficients (with the same yearly cycle) was used to homogenize the break.

[24] At the end, since the homogenization of daily data is a delicate topic, once all the series were homogenized, the homogeneity test was applied also to the precipitation intensity series (precipitation amount per rainy day) to verify if the procedure had introduced any errors, which was not the case. Moreover, a comparison between the original and the homogenized data set was performed in order to evaluate the degree of bias of the original data - are the data systematically biased or are all adjustments random? This issue was investigated by analyzing the adjustment series. As these series contain the factors that were used to increase or decrease original records in order to produce homogeneous data, this analysis will reveal any systematic errors in the original records. The results (not shown) highlighted that homogenization did not introduce any systematic change.

[25] After the homogenization, to prevent missing data from introducing any bias we used a procedure described in Karl et al. [1995] and widely used in many works [Karl and Knight, 1998; Brunetti et al., 2000, 2001a, 2001b] to estimate them. Basically, a gamma function is fit to each station's daily data for each month of the year. To determine if precipitation occurs on any missing day, a random number generator is used such that the probability of precipitation is set equal to the empirical one on that day. If precipitation occurs, then the gamma distribution is used to determine the amount that falls for that day, again using a random generator.

5. Regionalization

[26] The first step in analyzing the data was to cluster our data set into homogeneous precipitation areas by means of Principal Component Analysis (PCA) applied to monthly total precipitation series. PCA allows the identification of a small number of variables known as principal components (PCs), which are linear functions of the original data, that maximize the amount of their explained variance. The technique can be applied both to correlation and covariance matrixes. We used the correlation matrix R (in order to avoid a domination of the series with stronger variance) based on the monthly anomalies (in order to avoid the dominance of the annual cycle). R is defined by ZZt, Z being a matrix containing the standardised precipitation anomalies series (for a review on PCA see Jolliffe [1990, 1993]).

[27] The eigenvalues of R reveal that only 5 PCs account for more variance than the original variables (having eigenvalues greater than 1) and that these 5 PCs account globally for 65% of the variance of the standardised data. The weight of each station in each PC is given by the eigenvector components (the factor loadings) corresponding to the eigenvalues. The components of each eigenvector (the factor loadings) represent the correlation between the station series and the corresponding PC.

[28] The identified PCs (and consequently the corresponding loadings) are however, not unique. By rotating the PCs it is possible to obtain other sets of 5 PCs that account for the same fraction of variance of the data. The benefit of rotation is that it can be carried out to allow a more simple physical interpretation of the loadings. In our analysis we used VARIMAX rotation [see, e.g., Richman, 1986].

[29] The resulting loadings are 5 vectors of the 39 components. We represent the loadings on geographic maps, drawing contours through the points with the same loadings (Figure 2).

Figure 2.

Factor loading patterns of the first five rotated Principal Components obtained from the PCA.

[30] The loadings patterns allow the following regions to be identified: the north-western Italy (NW) with eight stations; the northern part of northeastern Italy (NEN) with nine stations; the southern part of northeastern Italy (NES) with seven stations; central Italy (CE), comprising also Sardinia, with eight stations; and the southern Italy (SO), comprising also Sicily, with seven stations.

[31] As indicated in Table 1, the longest series starts in 1787, but most of them have data only since 1880. Moreover, there are also some stations that start in only in 1920s. Sometimes the starting year is linked to the basin to which the station belongs, so we have different series lengths for the different regions: NW, NES and CE have data starting from 1880, but for NEN and SO most of the series have data only since 1920 and 1900 respectively.

6. Methods

[32] After setting up the final data-set in terms of homogenisation and gap filling, we calculated, for each station, some simple statistics of the precipitation series: the total seasonal precipitation (TP), the number of wet days per season (WD) and the mean amount of precipitation per wet day (precipitation intensity, hereinafter PI). Then we calculated the proportion of daily rainfall amounts, belonging to six precipitation class-intervals for each season and each year (C1, …, C6), compared with the corresponding total precipitation, and the number of events falling into these six classes (fC1, …, fC6).

[33] These categories were defined on the basis of some percentiles of the precipitation distribution of each series as follows: C1, precipitation lower than the 50th percentile; C2, precipitation between the 50th and the 75th percentiles; C3, precipitation between the 75th and the 90th percentiles; C4, precipitation between the 90th and the 95th percentiles; C5, precipitation between the 95th and the 99th percentiles; C6, precipitation greater than the 99th percentile. As suggested by Nicholls and Murray [1999] we chose percentiles, rather than some fixed thresholds to be used for all the stations (as in Karl et al. [1995]), to classify precipitation because of the strong differences between northern and southern Italy precipitation regimes: a precipitation event of 50 mm could be a normal event in some stations of northern Italy but an extreme event in southern Italy. Moreover, considering a single station, there are many differences also among the different seasons, and using the same threshold for each day of the year could lead to an overestimation of the number of extreme events in some seasons and an underestimation in others. Then, we decided to fit a precipitation distribution to each station and each day of the year, from which we estimated the thresholds (i.e., the millimeters corresponding to the different percentiles above defined to classify precipitation into the six classes). To calculate each daily gamma distribution we considered the data within a 31-day window around that particular day for every year of the series. At the end, for each station, we calculated 366 (by also considering the 29th of Febrary) threshold values for each precipitation category. After calculating TP, WD, PI, C1, …, C6, fC1, …, fC6 seasonal and yearly series for each station, they were converted into anomalies series and the regional average series were calculated by simply averaging the yearly and seasonal TP, WD, PI, C1, …, C6 anomalies series over all the stations of the area.

[34] The regional series were analyzed for trend with the Mann-Kendall nonparametric test. Use and computation of this test has been well described by Sneyers [1990]. The slopes of the trends were calculated by least squares linear fitting.

[35] In the Mann-Kendall test, for each element xi (i = 1 ,., n) of the series, the number ni of lower elements xj (xj < xi) preceding it (j < i) is calculated [Sneyers, 1990] and the test statistic t is given by t = Σi ni.

[36] In the absence of any trend (null hypothesis), t is asymptotically normal, independently of the distribution function of the data and u(t) = (t − 〈t〉)/equation image has standard normal distribution, with 〈t〉 and var(t) given by:

equation image
equation image

The null hypothesis can therefore be rejected for high values of ∣u(t)∣ being the probability α1 of rejecting the null hypothesis when it is true derived by a standard normal distribution table:

equation image

7. Results

[37] Figure 3 shows the yearly filtered (with a 31-year running window and 5-year σ gaussian filter) series for TP, WDs and PI.

Figure 3.

Yearly TP, WDs and PI series for the five regions. The data were smoothed with a 31-year width window and 5-year σ gaussian filter.

7.1. Total Precipitation

[38] TP shows no trend in northern regions (NW, NEN and NES) and a significantly negative trend in southern regions (CE and SO) with a decrease of about −10% per century in total yearly precipitation (Table 2). CE trend is mainly due to the spring season, that displays a −23% decrease per century, corresponding to −46 mm/100 y which is greater than half of the yearly trend amount (−82 mm/100 y). SO yearly trend (−9%/100 y, corresponding to −56 mm/100 y) is mainly due to the winter season, with a decrease of −17%/100 y (corresponding to −35 mm/100 y).

Table 2. Trends of TP, WDs, and PIa
 WSpSAY
  • a

    The values are expressed in trend amount per century. When the trend has a significance level lower than 90% only the sign is reported, otherwise it is expressed with the regression coefficient. The trends expressed as percentage variation per century (relative to the yearly and seasonal means of the 1961–1990 standard period) are indicated too. (Boldface values indicate significance greater than 99%; Italic values indicate significance greater than 95%).

NW
TP
   mm++
   %++
WD
   Rainy days−(3.4 ± 1.3)−(2.5 ± 1.4)−(7.5 ± 2.7)
   %−(18 ± 7)−(13 ± 7)−(9 ± 3)
PI
   mm/rainy day+++(0.8 ± 0.4)+(2.2 ± 0.8)+(1.1 ± 0.3)
   %+++(7 ± 4)+(17 ± 6)+(10 ± 3)
 
NEN
TP
   mm+++
   %+++
WD
   Rainy days+−(6.3 ± 4.9)
   %+−(7 ± 5)
PI
   mm/rainy day++++
   %++++
 
NES
TP
   mm++++
   %++++
WD
   Rainy days−(3.2 ± 1.1)−(2.5 ± 1.2)−(7.4 ± 2.4)
   %−(14 ± 5)−(12 ± 5)−(9 ± 3)
PI
   mm/rainy day+(0.8 ± 0.4)++(1.7 ± 0.4)+(1.3 ± 0.4)+(1.0 ± 0.2)
   %+(10 ± 5)++(17 ± 4)+(13 ± 4)+(11 ± 2)
 
CE
TP
   mm−(46 ± 12)−(82 ± 31)
   %−(23 ± 6)−(10 ± 4)
WD
   Rainy days−(5.5 ± 1.3)−(1.5 ± 0.9)−(4.0 ± 1.4)−(13.6 ± 2.6)
   %−(22 ± 5)−(11 ± 7)−(17 ± 6)−(15 ± 3)
PI
   mm/rainy day+(0.5 ± 0.3)++(1.1 ± 0.4) 
   %+(6 ± 4)++(9 ± 3) 
 
SO
TP
   mm−(35 ± 19)−(56 ± 32)
   %−(17 ± 9)−(9 ± 5)
WD
Rainy days−(4.0 ± 1.7)−(7.5 ± 2.5)
   %−(16 ± 7)−(11 ± 4)
PI
   mm/rainy day++++
   %++++

7.2. Number of Rainy Days

[39] WDs show similar behavior toward a decreasing trend in all regions, with the greatest decreases from 1930s to 1940s and from 1960s to 1980s.

[40] On a yearly basis there is a highly significant (greater than 90% in NEN and greater than 99% in all the other regions) negative trend in the number of WDs in all the regions (Table 2), ranging from −7% per century in NEN to −15% per century in CE (corresponding respectively to −6 and −14 days per century). On a seasonal basis, the trend is always negative and the temporal variability is similar in all regions (not shown). The strongest contribution to yearly trend comes from Spring (s.l. > 95% in NW and s.l. > 99% in NES and CE, with trend ranging from −14%/100 y in NES to −22%/100 y in CE, corresponding respectively to −3.2 and to −5.5 days) and Autumn (s.l. > 90% in NW and NES and s.l. > 99% in CE, with trend ranging from −12%/100 y in NES to −17%/100 y in CE, corresponding to −2.5 and −4 days respectively) in all regions but SO, where the greatest part of the trend is due to Winter.

7.3. Precipitation Intensity

[41] As a consequence of the decreasing trend in WDs all over Italy and of the nonuniform behavior of TP, PI shows a trend, with different coefficients and significance level, which is positive almost everywhere and in all seasons (apart from Summer in NEN, Spring in CE and Winter in SO).

[42] On a yearly basis, significant (s.l. greater than 99%) positive trends (Table 2) were registered in NW and NES, because of a strong decrease in the number of WDs. This trend is mainly due to summer and autumn (with a little contribution coming from winter, but only for NES). In CE there are significant positive trends only in autumn (s.l. greater than 99%) and winter (s.l. greater than 95%), but not in the year. In SO no significant values were found. This is because in southern Italy (CE and SO), both WDs and TP show a significant decrease.

7.4. Progressive Trend Analysis

[43] Trend values and significance vary with the length of the analyzed period. No trends in NEN and lower significance of the SO trends could be due to the later starting year of series (1921 and 1901 respectively). Therefore to realize a more comparable analysis, the trends were calculated also in a progressive way. This provided us also with a more detailed view of the time behavior of trends.

[44] The previously described trend analysis was then applied to the series starting from the i-th year, with i running from the beginning of each series to 1962 (we chose 1962 in order to have series not shorter than 40 years to calculate the trends), and ending with the last one. The results are shown in Figure 4 (only for the year) for TP, WDs and PI.

Figure 4.

Progressive trend analysis for (a) TP, (b) WDs, and (c) PI. Each point of a curve represents the regression coefficient (expressed as variation per century) calculated for the series beginning from the year in which it is located and ending with the last one (2002). Thick portions of the curves indicate trends with s.l. > 95%.

[45] Yearly TP have a significant negative trend in CE and SO only if the series starts before the 1910s. On a seasonal basis (not shown) the long period trend is always not significant, apart from the spring in CE, where it is significantly negative if the series starts before 1920s. Winter (not shown) is particularly interesting: in all regions but SO the trend becomes more and more negative the later the starting year is, even if it reaches significant values only in CE.

[46] Yearly WDs have a constant negative trend in all regions that is nearly always significant considering the subseries beginning before 1930s. Subseries starting after 1940 have a temporary attenuation of the trend strength, followed by a faster decreasing rate, reaching their lowest amount around 1960, with significant values in NES, CE and SO (the strongest trends are in NES and CE with a regression coefficient of about −45%/century for the period 1960–2002). This behavior seems to be due to the strong minimum in the number of rainy days located around 1940s (see Figure 3). Also for WDs, winter (not shown) is particularly interesting: it displays the same characteristics observed for TP.

[47] PI yearly trend is always positive with higher values the more recent the starting year of the subseries is. It has significant values for nearly all the subseries in NES, for those starting before 1930s in NW and only for those starting around 1920 in CE. This behavior is similar in all seasons with an increasing noise in the last years due to the decrease in the portion of series where the trend is calculated.

7.5. Analysis of Precipitation Categories

[48] The tendency in Italy toward an increase in the mean amount of precipitation per wet day can be statistically studied in more detail using the class-interval contributions (C1, …, C6). This allows a better evaluation than PI of the evolution of the shapes of the wet day amount distributions.

[49] The series of the relative contribution of the six class categories to total yearly precipitation are shown in Figure 5 for the five regions. Figure 6 shows the results of the application of trend analysis to seasonal and yearly regional C1, …, C6 series. There is a general tendency in northern regions (NW, NEN and NES) toward a decrease in the relative contribution of the lower classes (C1, C2 and C3) and an increase in that of higher classes (C4, C5 and C6). This is particularly evident in NW and NES on a yearly basis, where nearly all category trends reach significant values (negative for the lower classes and positive for the higher ones), mainly due to summer and autumn for NW and to all seasons for NES. CE has a similar well-defined behavior, even if the significance is lower (here the strongest contribution comes from Autumn). SO has a nondefined tendency in the precipitation distribution and rarely the trends are significant. NEN has few significant trends (only a positive trend in C6, for the year, and a negative trend for C2 in winter), probably because the series is shorter.

Figure 5.

Yearly series of the relative contribution of the six class-categories to TP for the five regions. The data were smoothed with a 31-year width window and 5-year σ gaussian filter.

Figure 6.

Trends (expressed as percentage variation per century relative to the mean value of the 1961–1990 standard period) of the relative contribution to TP of the six precipitation class categories. (Black = s.l. > 99%; grey = s.l. > 95%; light grey = s.l. > 90%).

[50] To understand if the above described behavior (the increase in the relative contribution of higher events to total precipitation amount and the decrease of the lower ones) is linked to a tendency of Italian precipitation to have more frequent extreme events, we also analyzed the number of events falling into each class category (fC1, …, fC6). Figure 7 shows yearly series of fC1, …, fC6 for each region.

Figure 7.

Yearly series of the number of events falling into the six class-categories for the five regions. The data were smoothed with a 31-year width window and 5-year σ gaussian filter.

[51] The results of trend analysis are shown in Figure 8 following the same representation procedure of Figure 6 for C1, …, C6. The signal is very clear: there is a strong and highly significant negative trend in the number of events falling into the lower classes. This decrease is evident in all regions, and also in many seasons. Positive trends in the number of events falling into the higher classes is observed too, but only in northern regions (as for NEN and NES in the year), while in CE and SO all categories display negative trends, even if the higher ones do not reach significant values.

Figure 8.

Like Figure 6, but for the number of events belonging to each class category.

7.6. Progressive Analysis of Precipitation Categories

[52] A progressive trend analysis was applied to class categories too (Figures 9 and 10). The clearest signal in the C1, …, C6 and fC1, …, fC6 trends are in the northern regions, where the highest class shows positive trends, often significant, for all the subseries. This tendency is evident also in CE and SO, but only in the subseries starting after 1900 in CE and after 1930 in SO, even if no significant values are reached. Also in the lowest class, the spatial pattern is similar: there are negative, often significant, trends in all NW and NES subseries, while in CE and SO they are negative, but not significant, only for subseries starting after 1900 and 1930 respectively.

Figure 9.

Like Figure 4, but for the relative contribution of each class category to TP.

Figure 10.

Like Figure 4, but for the number of events belonging to each class category.

7.7. Remarks

[53] On average the last decade presents the lowest number of rainy days per year and the highest amount of precipitation per rainy day (Figure 3). This tendency toward a decrease in precipitation events and an increase in their intensity persists for more than one century. Figure 4b (4c) shows that the WDs (PI) trend is always negative (positive) whatever the starting year of the series is.

[54] A minimum in the number of WDs similar to that of the 1990s happened also in the 1940s, but it was mainly related to a minimum in total precipitation (in that period the minimum precipitation amount of the series was reached) then no particular increase in the intensity was observed.

[55] By considering the time series of the class categories (Figures 5 and 7) it is worth noticing that there are high values between the end of the 19th and the beginning of the 20th Century, both in the relative contribution to total precipitation and in the frequency of the C4, C5 and C6 class categories. Their values are comparable with the recent ones for C4 and C5 but not for C6, for which the recent period shows the highest values of the series.

[56] A comparison with similar studies, performed by other researchers for other regions of the globe, is difficult to make because of the lack of long daily precipitation series. Only for the United States, daily data starting from the end of the 19th century have been studied [Kunkel et al., 2003]. Even if the analysis performed is different from ours, some similar results can be observed. They studied event durations of 1, 5, 10 and 30 days, and return period of 1, 5, and 20 years. For all combinations of duration and return period, heavy precipitation frequencies were relatively high during the late 19th/early 20th Centuries, decreasing to a minimum in the 1920s and 30s, followed by a general increase into the 1990s. For 1-day duration events, frequencies at the beginning of the 20th Century are comparable in magnitude to frequencies in recent decades, and for 5 and 10-day duration events, frequencies during 1895–1905 are only slightly smaller than those of the late 20th Century, while for the 30-day return period events frequencies during the late 19th/early 20th Centuries are smaller than in the recent period.

8. Conclusions

[57] An analysis of 45 daily precipitation records over 120 years for Italy was undertaken to identify any changes in the characteristics of precipitation distribution that may have occurred. The series were homogenized on a daily basis, completed by means of statistical methods and grouped into five regions by a Principal Component Analysis. The research concerned seasonal and yearly total precipitation, number of wet days, precipitation intensity and proportion and frequency of daily precipitation falling into six precipitation class-intervals defined on the basis of some percentiles of the precipitation distribution.

[58] The principal results are as follows:

[59] 1. The number of wet days in the year has a clear and highly significant negative trend all over Italy, ranging from −7% per century to −15% per century. The strongest contributions to yearly trend come from spring and autumn.

[60] 2. Besides the reduction in the number of wet days, there is a tendency toward an increase in precipitation intensity. This increase is globally less marked than the decrease in the number of wet days and reaches significant values in the year only in northern regions.

[61] 3. The increase in precipitation intensity is due to a decrease in the contribution of low precipitation categories to total precipitation and to an increase in that of higher categories corresponding to heavy precipitation events. This is more evident in northern regions (NW, NEN and NES) with highly significant trends and in CE even if the significance is low, while no clear signal is present in SO.

[62] 4. Considering the frequency of events falling into each category, rather than their relative contribution to total precipitation, a more uniform behavior all over Italy was observed: there is a highly significant decrease in the number of events falling into lower categories in all regions.

[63] 5. In northern regions there is also evidence of a significant increase in the number of events falling into the highest class-interval (comprising events above the 99th percentile).

[64] Italy, due to its orography, is particularly exposed to the negative effects of high intensity precipitation events. It is one of the European countries with the greatest human loss due to flood events. At the same time droughts are a frequent problem for agriculture, in particular during summer, when the lack of water resources often compromise the crop in large regions, especially in Southern Italy. The results of this paper highlight both these two aspects, indicating the last decade as the one with the lowest number of rainy days per year and the highest amount of precipitation per rainy day.

[65] However, the tendency toward a decrease in the number of rainy days, and an increase in intensity, that was already observed in a previous study for the last 50 years [Brunetti et al., 2001a], is not a peculiarity of recent decades, but it has persisted since the end of the 19th century. From this study, it is evident that most of this negative trend is explained by the marked decrease in the frequency of low intensity precipitation events (below the 75th percentile).

Acknowledgments

[66] We would like to thank the Ufficio Centrale di Ecologia Agraria (UCEA) and the Servizi Idrografici for their cooperation in giving access to databases. Special thanks go to the Consorzio Interuniversitario per il Calcolo Automatico (CINECA) for computing support. The research was partially developed in the framework of the FIRB and the CLIMAGRI projects.

Ancillary

Advertisement