Assessing the Accuracy of Gridded Precipitation Products in the Campania Region, Italy

Abstract

Accurate precipitation data are essential for hydrological modeling, climate studies, and water resource management. Indeed, there is an increasing focus on understanding shifts in precipitation events to monitor the risks of floods and droughts, as well as to ensure sustainable water resource management. This study compares four reanalysis and satellite precipitation products (ERA5-Land, CHIRPS, PERSIANN, and TerraClimate) with ground data from 2003 to 2022. Among the datasets evaluated, ERA5-Land has the best performance (overall) in reproducing ground data, with a minimal mean bias error (MBE) of 1.91 mm, the highest correlation coefficient (R² = 0.93), and the most favorable Nash–Sutcliffe efficiency (NSE = 0.93). In contrast, CHIRPS, PERSIANN, and TerraClimate significantly underestimate precipitation as compared to ground data. The categorical metrics also highlight ERA5-Land’s superior performance in identifying wet months. Spatial analysis shows that ERA5-Land and other datasets generally exhibit agreement regarding precipitation patterns. However, PERSIANN displays notable variances, particularly in northern regions, where it overestimates precipitation. To investigate possible changes in precipitation patterns, a longer period (1983–2022) is selected for trend analysis based on gridded precipitation products. Sen’s slope analysis does not reveal any significant annual precipitation trend. In autumn, the PERSIANN dataset indicates a significant increasing trend of +1.81 mm/year, which is also confirmed by ERA5-Land (+2.68 mm/year) and CHIRPS (+1.34 mm/year), although without statistical significance. The findings emphasize the need for more sophisticated satellite algorithms and integration with ground observations to improve precipitation accuracy.

1. Introduction

Climate change is altering precipitation patterns, affecting water availability, agriculture, and drought frequencies [1]. These changes have significant consequences for water resource management, ecosystem sustainability, and risk assessment in vulnerable regions. Assessing the relationship between heavy precipitation patterns and the occurrence of floods and landslides is challenging, so it is crucial to understand the effect of precipitation change on hydrological processes [2]. Water constantly moves through the atmosphere, land, and the oceans, as a key constituent of the global hydrosphere, to maintain ecological and environmental balance [3]. The variations in precipitation patterns significantly affect the water cycle, resulting in severe hydrometeorological hazards across the globe. In recent decades, we have seen significant fluctuations in climatic variables due to changes in precipitation patterns [4,5,6]. This variability affects weather patterns, surface water availability, and extreme precipitation [7,8]. Understanding precipitation variability on a spatial scale is crucial to mitigating climate-related impacts across various regions.

The actual estimation of precipitation is very difficult due its varying magnitude, state, and irregular spatial and temporal occurrence. These limitations in measuring and modeling precipitation further complicate our understanding of changes to this critical aspect of the hydrological cycle [9]. Satellite and reanalysis precipitation products each have unique features and offer distinct advantages over in situ measurements [10]. Satellite-based products encompass global coverage, high spatial (less than 0.05°) and temporal (with intervals as short as 3 h or less) resolution, and the accessibility of data that are nearly up to date. However, satellite-based precipitation products have some limitations [11]. As these datasets incorporate data from several sources, such as microwave, infrared, and in situ measurements, there is a possibility of encountering unidentified mistakes that are caused by devices or algorithms that may not accurately consider unique site-specific characteristics [12,13]. Additionally, their accuracy can also be affected by local conditions, geographical location, and elevation. On the other hand, reanalysis products integrate satellite data with a broad range of observational inputs and complex model simulations [14]. These datasets also have limitations in data-scare regions due to complex computational effort, delayed data availability, coarse spatial resolution, and reliability due to a lack of observational inputs. Therefore, a comparison of these two types of precipitation products with ground measurements of rainfall is crucial to understand their respective strengths and limitations in different topographic conditions. Remote sensing techniques are also used for the estimation of precipitation in complex topographic regions but these products have limitations due to coarse resolutions [15]. Radar-based precipitation products are also an important source of rainfall data [16]. Radar systems have a high spatial and temporal resolution and are useful for monitoring short-duration and convective rainfall events. However, their application is often limited by data availability and coverage. Despite these challenges, radar-based data remain valuable for localized hydrometeorological studies [17]. Several studies compared gridded precipitation products and concluded that ERA5, CHIRPS, PERSIANN-CDR, and TRMM perform better on a monthly scale [18,19,20]. However, recent research investigates how gridded precipitation products, either from meteorological reanalysis or from satellite observations, are suitable for hydrological modeling [21,22,23,24].

Southern Europe is strongly affected by severe precipitation, leading to geohydrological hazards [25]; high-resolution gridded products can help model hydrological impacts [26]. To study heavy precipitation events, Mediterranean regions such as the western and central basin are ideal due to their diverse topography [27]. In Italy, owing to the complex orography and the proximity to the sea, gridded products struggle to identify spatial precipitation patterns [28,29]; intercomparison with rain gauge data is still mandatory for hydrological modeling and climate studies [30]. An investigation into the spatial variability of precipitation, together with orography, was carried out to understand large-scale circulation patterns. The Campania region remains comparatively understudied despite being densely populated (the second-most densely populated area in Italy) and highly vulnerable to climatic variations. Given the limited availability of region-specific validation studies, assessing the performance of different precipitation products is crucial in this region to understand local climate trends. A recent study conducted in the Campania region found a drying trend in spring and summer, along with an increase in heavy rainfall events in autumn, mainly in mountainous regions [31]. Given the complex topography of the Campania region, comparing these precipitation products is useful in order to better understand climatic variations and their impacts. Mesoscale convective systems (MCSs) extend less than 50 km and they can be missed due to the sparse resolution of gauge stations [32]. As a result, high-resolution precipitation products like ERA-5 Land are essential for precipitation measurements, particularly during short-duration MCSs.

The main objectives of this study are (a) to compare the performance and accuracy of selected datasets with reference to ground station data by evaluating continuous and categorical metrics; (b) to analyze the spatial and temporal variability of precipitation across the region; and (c) to conduct statistical trend analysis for each gridded dataset with the novel pixel-based Sen’s slope technique [33]. The findings of the present study will help identify the most reliable precipitation datasets for the region and investigate their potential use in future hydrological and climate-related studies.

2. Materials and Methods

2.1. Study Area

The Campania region, located in southern Italy, covers a total geographical area of 13,595 km². It is located between 13°50′ E–15°45′ E longitudes and 40°50′ N–41°30′ N latitudes (Figure 1). Campania has a diverse landscape, including fertile plains, the Apennine Mountains, and coastal regions along the Tyrrhenian Sea. The region has a Mediterranean climate with hot, dry summers and mild, wet winters. On average, the highest rainfall occurs in late autumn (November), while the lowest occurs in July. The spatial distribution of average annual precipitation is heavily influenced by the region’s mountainous features. The tallest Apennine reliefs, such as Matese, Picentini, Partenio, and Cilento, receive between 1700 and 2000 mm/year. In contrast, the plain and coastal areas receive between 800 and 1000 mm/year. The eastern sides of the region (Sannio and Irpinia) are the driest, receiving 500 to 800 mm of rain annually, with less seasonal variability. From October to May, precipitation events in Campania are primarily caused by western and southern flows driven over the area by cyclonic circulation, which is caused by polar front oscillations. The Apennine reliefs also significantly influence precipitation dynamics from June to September. However, during the summer, rainfall events are generally influenced by different large-scale atmospheric dynamics, involving weak upper-level cyclonic conditions and negligible low-level pressure gradients [34].

2.2. Data Collection and Description

Ground station data were collected from the Civil Protection Department of the Campania Region from 2003 to 2022, focusing on 45 selected stations. Although there are around 100 stations in the region, some are newly installed, so the selection was made to ensure an even distribution of data-rich stations across the area. Thiessen polygons were also generated to compute the weighted average for each ground station, which was essential for assessing the accuracy of the selected datasets (Figure 1c). Satellite and reanalysis gridded precipitation data were obtained from the Google Earth Engine (GEE), using sources such as ERA5, CHIRPS, TerraClimate, and PERSIANN. Comprehensive details of the data collected for the study are provided in

Table 1. Summary of datasets used in the study. All URLs were accessed on 15 July 2024.

Dataset	Spatial Resolution	Time	Source	URL
ERA 5	0.1°	1983–2022	GEE	https://developers.google.com/earth-engine/datasets/catalog/ECMWF_ERA5_LAND_MONTHLY_AGGR
PERSIANN	0.25°	1983–2022	GEE	https://developers.google.com/earth-engine/datasets/catalog/NOAA_PERSIANN-CDR
TerraClimate	0.04°	1983–2022	GEE	https://developers.google.com/earth-engine/datasets/catalog/IDAHO_EPSCOR_TERRACLIMATE
CHIRPS	0.05°	1983–2022	GEE	https://developers.google.com/earth-engine/datasets/catalog/UCSB-CHG_CHIRPS_DAILY
Ground data	NA	2003–2022	Civil Protection Agency	https://www.regione.campania.it/regione/en/topics/civil-protection/civil-protection-agency
Digital Elevation Model (DEM)	30 m	2000	NASA SRTM	https://developers.google.com/earth-engine/datasets/catalog/USGS_SRTMGL1_003

. Cumulative sum [31] and Pettitt’s test [35] were applied to ensure the quality and consistency of the precipitation data. The study area was divided into four seasons—winter (Dec–Feb), spring (Mar–May), summer (Jun–Aug), and autumn (Sep–Nov)—for comprehensive seasonal analysis. Residual analysis was performed using both regionally averaged precipitation data and individual station observations. It is also important to note that ground station data may contain some errors due to instrument limitations or environmental factors.

The study followed a systematic approach to collect and analyze precipitation data. Ground station data were used as the baseline for the study to evaluate the selected datasets. The gridded datasets were evaluated for their accuracy by comparing them with ground station data using statistical tests to ensure data consistency and quality. Several tools and methods, including bias correction, categorical and continuous metrics evaluation, and spatial mapping, were employed to achieve a robust understanding of precipitation trends and distributions across the study area (Figure 2). The complete details of each dataset and the applied techniques are discussed below.

2.3. CHIRPS

CHIRPS is a quasi-global rainfall dataset comprising data from 1981 to present. It combines 0.05°-resolution satellite imagery with in situ station data to generate a gridded rainfall time series [36]. CHIRPS is renowned for its high spatial and temporal resolution data, which are based on long-term infrared cold cloud duration. For this study, CHIRPS data were acquired from GEE for the period from 1983 to 2022.

2.4. PERSIANN-CDR

PERSIANN-CDR was developed by the Centre for Hydrometeorology and Remote Sensing at the University of California. It covers the period from 1983 to present and has a spatial resolution of 0.25 degrees. This is also a quasi-global dataset and provides daily and monthly precipitation data [37]. It is generated using the concept of GridSat infrared data and making monthly adjustments according to the Global precipitation Climatology Project (GPCP) in order to assess the consistency of the two datasets at a monthly scale. In this study, monthly data were collected from GEE from 1983 to 2022.

2.5. ERA5-Land

ERA5-Land, developed by the European Centre for Medium-Range Weather Forecasts (ECMWF), offers reanalysis data that are designed for land surface. The dataset comprises continuous atmospheric and land surface data from 1981 to present, with a horizontal resolution of around 9 km. Due to its high accuracy, ERA5-Land is widely used across the globe [38]. ERA5-Land combines multiple sources, such as satellite observations, ground stations, and other measures, using data assimilation techniques.

2.6. TerraClimate

TerraClimate is a high-resolution global dataset developed by the University of Idaho, providing data relating to various climate variables from 1958 to present. TerraClimate has a spatial resolution of approximately 4 km and provides data on a monthly scale [39]. This dataset integrates multiple data sources, including satellite observations and ground-based measurements, to ensure accuracy and consistency in representing climate patterns over time. For this study, precipitation data from TerraClimate were collected from GEE. The complete flow of methodology is shown in Figure 2.

2.7. Continuous Metrics Evaluation

Monthly precipitation data were evaluated using the five standard techniques presented in

Table 2. Continuous metrics used for the evaluation of selected datasets.

Parameter	Expression	Range	Perfect Value
RMSE	$\sqrt{{\displaystyle \frac{{\displaystyle {\sum }_{i=n}^i}{\left(P_0^i-P_s^i\right)}^2}{n}}}$	[0, ∞]	0
MAE	${\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=n}^i}|P_0^i-P_s^i|$	[0, ∞]	0
MBE	${\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}(P_s^i-P_0^i)$	[−∞, ∞]	0
NSE	$1-{\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{\left(P_o^i-P_S^i\right)}^2}{{\displaystyle {\sum }_{i=1}^n}{\left(P_0^i-\overline{P_0}\right)}^2}}$	[−∞, 1]	1
R2	${\left({\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}\left(P_0^i{P}_s^i\right)-({\displaystyle {\sum }_{i=1}^n}{P}_0^i)({\displaystyle {\sum }_{i=1}^n}{P}_s^i)}{\sqrt{n{{\displaystyle {\sum }_{i=1}^n}{\left(P_0^i\right)}^2-({\displaystyle {\sum }_{i=1}^n}{P}_0^i)}^2}\sqrt{n{{\displaystyle {\sum }_{i=1}^n}{\left(P_s^i\right)}^2-({\displaystyle {\sum }_{i=1}^n}{P}_s^i)}^2}}}\right)}^2$	[0, 1]	1

in order to estimate the deviation of satellite and reanalysis data from ground observations.

The Nash–Sutcliffe efficiency (NSE) is used to determine the difference between the residual variance and observed variance. If the NSE equals one, it reflects the perfect agreement of satellite data compared to observed data. If NSE = 0, it indicates that the model (satellite or reanalysis product) has the same predictive skill as the mean of the observed time series (i.e., the model produces an estimation error variance equal to the variance of the observed time series). However, NSE < 1 indicates that the mean value of the observed data is a better predictor than the satellite data. To measure average bias, mean bias error (MBE) is used. If MBE has a negative value, it shows that simulated data underestimate the observed data. Conversely, if MBE has a positive value, it means the simulated data overestimate the observed data. The mean absolute error (MAE) determines the difference between the measured and expected average values. The coefficient of determination (R²) measures the proportion of variance in the observed data that is predictable from the simulated data. An R² value of one indicates a perfect match, while a value closer to zero indicates that the simulated data do not explain much of the variance in the observed data. The root mean square error (RMSE) method is used to determine the difference between observed values and values taken by the concerned product. The cumulative distribution function is also calculated for all datasets. Table 2 represents the parameters, expression, range, and perfect values of the matrices evaluation. Here, $P_0^i$ shows the observed precipitation, $P_0^i$ is the satellite-estimated precipitation and $P_s^i$ shows mean precipitation of the observed values.

2.8. Categorical Metrics Evaluation

Categorical metrics are essential for evaluating how accurately satellite and reanalysis products can detect precipitation based on ground data. In this study, five categorical evaluation techniques were utilized (

Table 3. Continuous metrics used for the evaluation of satellite and reanalysis datasets.

Parameter	Expression	Range	Perfect Value
Accuracy	${\displaystyle \frac{a+d}{e}}$	0–1	1
POD	${\displaystyle \frac{a}{a+c}}$	0–1	1
FAR	${\displaystyle \frac{b}{a+b}}$	0–1	0
POFD	${\displaystyle \frac{b}{d+b}}$	0–1	0
PSS	${\displaystyle \frac{a}{a+c}}-{\displaystyle \frac{b}{b+d}}$	−1–1	1

). The overall fraction of satellite-estimated precipitation is verified by accuracy. The Probability of Detection (POD) technique justifies the successful identification of precipitation events by satellite-estimated precipitation, with a higher POD value indicating the successful identification of events by the gridded precipitation product. However, the False Alarm Ratio (FAR) is the reciprocal of POD. The Probability of False Detection (POFD) measures the proportion of ‘no’ precipitation events that are mistakenly classified as ‘yes.’ A POFD near zero signifies an accurate evaluation. In general, the closer the accuracy, Probability of Detection (POD), and Peirce Skill Score (PSS) values are to 1, and the closer the False Alarm Ratio (FAR) and Probability of False Detection (POFD) values are to 0, the more accurate the satellite products are in detecting precipitation events. Table 3 summarizes the adopted categorical evaluation metrics, reporting their formulation, ranges, and perfect values. Here, the variable “a” shows the number of events that happened as per estimation, “b” shows the number of events that were estimated to happen but did not happen, “c” shows the number of events that were estimated not to happen but happened, and “d” shows events forecasted not to happen and that did not happen.

2.9. Bias Decomposition

Bias decomposition analysis was also performed to evaluate the performance of selected datasets (

Table 4. Parameters used for bias decomposition.

Parameter	Expression	Range
Systematic Bias (SB)	${\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{(\widehat{P_s^i}-P_0^i)}^2,$ where $\widehat{P_s^i}=\alpha +\beta P_0^i$	[0, ∞)
Unsystematic Error (NU)	${\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{(P_s^i-\widehat{P_s^i})}^2$	[0, ∞)
Percentage Systematic Bias (SB%)	${\displaystyle \frac{SB}{MSE}}\times 100$	[0, 100]%
Percentage Unsystematic Error (NU%)	${\displaystyle \frac{NU}{MSE}}\times 100$	[0, 100]%
Slope (β)	$\widehat{P_s^i}=\alpha +\beta P_0^i$	(−∞, ∞)
Intercept (α)	$\widehat{P_s^i}=\alpha +\beta P_0^i$	(−∞, ∞)

). The mean squared error was used to measure the total deviation between ground and gridded data, where total error is divided into systematic and unsystematic biases. Systematic bias reflects the model’s inability to replicate observed precipitation magnitudes, while unsystematic error captures residual noise. A spatial analysis of bias on both annual and seasonal timescales was carried out to investigate how orography affected the performance of the tested gridded precipitation products.

2.10. Sen’s Slope

The Sen’s slope method was used to determine the positive and negative trend in precipitation data with statistical significance [33]. This method uses the linear model to estimate slope trend and constant residual variance using the following formula:

$$Q_i={\displaystyle \frac{X_{j }- X_k}{j-k}} for j, k=1\dots \dots \dots ,n$$

(1)

where X_j and X_k are the data values at times j and k, respectively, with j > k; n represents the total number of observations. The slope is estimated N times, using all the possible couples (X_j, X_k) within the series. The N values of $Q_i$ are then sorted from the smallest to the largest, and the median of these values, which is known as the Sen’s slope estimator, is calculated.

$$Q_{med}=\left\{\begin{array}{c}{Q}_{(N+1)/2} if N is odd\\ Q_{(n/2)}+Q_{(N+2)/2} if N is Even\end{array}\right.$$

(2)

$Q_{med}$ reflects the direction of the data trend (i.e., increasing or decreasing), while the magnitude indicates the steepness of the trend. To determine if the median slope is statistically different from zero, the confidence interval around the slope can be computed as follows [40,41]:

$$Z_{\propto }=Z_{1-\infty /2} \sqrt{Var (S})$$

(3)

Here, Var (S) is the variance of the Mann–Kendall statistic (S); $Z_{\propto }$ represents the critical value for the desired confidence level. The pixel-based Sen’s slope analysis was performed in GEE, while the correlogram for four satellite datasets and ground data were generated in RStudio (version 4.3.1)using the ggplot2, readxl, and Gally libraries [42].

3. Results and Discussion

Table 5. Values of continuous metrics used for the evaluation of the satellite and reanalysis products.

Data	MBE (mm)	R2	NSE	RMSE (mm)	MAE (mm)
ERA5-Land	1.91	0.93	0.93	18.32	13.43
CHIRPS	−28.0	0.81	0.57	45.54	31.09
PERSIANN	−16.84	0.72	0.63	42.20	27.62
TerraClimate	−26.37	0.67	0.51	49.00	33.34

presents the results of continuous metrics on monthly data from four satellite products from 2003 to 2022. ERA5-Land shows a slight overestimation of precipitation, with an MBE of 1.91 mm, while CHIRPS (−28.0 mm), TerraClimate (−26.37 mm), and PERSIANN (−16.84 mm) exhibit significant underestimations. ERA5-Land has the highest R² value of 0.93, indicating a strong correlation with ground data, followed by CHIRPS (0.81), PERSIANN (0.72), and TerraClimate (0.67). NSE values further support these findings with ERA 5 (0.93), CHIRPS (0.57), PERSIANN (0.63), and TerraClimate (0.51), demonstrating ERA5-Land’s higher accuracy. The values of RMSE and MAE also confirm that ERA5-Land gives the most accurate estimates, while TerraClimate is the least accurate product. ERA5-Land has a slight overestimation of precipitation by 1.91 mm, while CHIRPS at −28.0 mm, TerraClimate at −26.37 mm, and PERSIANN mm at −16.84 mm are the products with substantial underestimations. ERA5-Land comes closer to the observed values, while the other products tend to substantially underestimate the levels of precipitation. A slight overestimation in precipitation by ERA5-Land was also observed by a recent study conducted in the Manas river in China and in the Alpine region (both regions have strong seasonality and elevated terrain that can impact orographic rainfall) [43,44]. Thiessen polygons slightly impacted the average precipitation across the region. The simple monthly average was 97.4 mm, while the Thiessen polygon-based average was 95.0 mm per month during the study period.

Figure 3a presents a comparison of the mean precipitation from 2003 to 2022 using different datasets. The ERA5-Land dataset exhibits a slight overestimation of precipitation amounts, whereas satellite products tend to exhibit underestimations. On the other hand, the CHIRPS dataset consistently shows the lowest amounts of rainfall for the entire study period. Figure 3b provides a representation of the average monthly rainfall in the study area from 2003 to 2022, emphasizing a clear and recurring seasonal trend. The winter season marks the peak rainfall in the region, while the summer months witness significantly reduced precipitation. Despite the fact that it reproduces the total annual precipitation well, PERSIANN underestimates seasonal variability. PERSIANN uses infrared satellite data, which sometimes misinterpret warm, non-rainy clouds as rain-producing systems. This issue is more common in summer, when short and local convective storms occur [45]. These storms are often intense but small, and PERSIANN’s coarse resolution cannot capture them accurately. As a result, it tends to overestimate rainfall, especially in areas with complex topography like Campania. Due to coarser spatial resolution, it cannot accurately capture convective storms, which can lead to the underestimation of precipitation records. A similar behavior has also been reported in other studies [46,47,48].

With the highest PSS (0.87), ERA5-Land effectively differentiates between event and non-event outcomes (

Table 6. Results of categorical metrics used for the evaluation of satellite and reanalysis products.

Data	PSS	POD	POFD	FAR	Accuracy
ERA5-Land	0.87	0.95	0.07	0.02	0.95
CHIRPS	0.76	0.82	0.05	0.02	0.85
PERSIANN	0.74	0.91	0.16	0.07	0.89
TerraClimate	0.72	0.79	0.08	0.03	0.83

). Its POD is also the highest at 0.95, capturing 95% of actual events, while PERSIANN follows at 0.91, CHIRPS at 0.82, and TerraClimate at 0.79. The POFD and FAR results also support ERA5’s efficient forecasts compared to other datasets. ERA5-Land’s accuracy is the highest at 0.95, outperforming CHIRPS, PERSIANN, and TerraClimate.

Table 7. Results of bias decomposition analysis with respect to ground data.

Satellite	MSE	MBE	Slope	Intercept	SB_%	NU_%
ERA5-Land	347.66	1.90	0.93	8.40	7.71	92.29
PERSIANN	1780.91	−16.84	0.58	23.19	65.22	34.78
TerraClimate	2454.27	−26.37	0.56	15.25	66.98	33.02
CHIRPS	2074.30	−28.03	0.56	13.26	82.89	17.11

shows the bias decomposition results of selected datasets. ERA5-Land shows around 7% of systematic bias, while CHIRPS has the highest systematic bias at 82.8%.

Figure 4 illustrates the seasonal time series comparison of ground data and remotely sensed datasets. The monthly time series at the same latitudes and longitudes were extracted for each satellite and reanalysis product. During the comparison of CHIRPS and ground station records, a strong interannual variability is evident, where the precipitation was underestimated by CHIRPS as compared to PERSIANN, which shows a significant overestimation compared to ground data during summer 2017. A similar underestimation behavior can be noted in the case of TerraClimate. However, the underestimation was more pronounced in the rainiest periods, e.g., during 2010 (Figure 4a). In the spring season, the ERA 5-Land dataset consistently provided higher estimates of precipitation, whereas the PERSIANN and CHIRPS datasets tend to record lower estimates (Figure 4b). Despite these differences, the data display fluctuations without any discernible consistent pattern across satellite products and ground stations. The recorded precipitation in the spring season varies from 120 to 400 mm. During the summer season, there is typically very little rainfall across the region (Figure 4c). The PERSIANN satellite product tends to overestimate the amount of precipitation (in summer) compared to the observed data, while other satellite products mostly underestimate it (Figure 4c). The average precipitation throughout the summer ranges from 50 to 200 mm. The data reveal that the summer of 2017 experienced the lowest amount of rainfall, while the summers of 2005 and 2018 had the highest precipitation. During the autumn season, the amount of rainfall is comparable to that of spring, typically ranging from 120 to 400 mm. The ERA5-Land dataset consistently matched the observed data, while other datasets underestimated the amount of precipitation. Among these datasets, TerraClimate exhibits the most significant underestimation. According to ERA5-Land and observed data, the autumn of 2010 stands out as the autumn with the highest amount of rainfall throughout the research period, with precipitation reaching up to 700 mm. Other datasets also demonstrate the highest levels of precipitation for that year, although to a lesser degree.

During the winter season, the region experiences the highest levels of precipitation, with up to 550 mm of rainfall. All gridded datasets generally underestimate precipitation when compared to ground data (Figure 4d). Winter 2016 was the driest winter, with the smallest amount of rainfall throughout the season.

However, ERA5-Land is the dataset that is closest to the actual observed values, while CHIRPS shows a significant underestimation. Interestingly, the data show the winter of 2021 as the wettest, with precipitation levels reaching up to 600 mm. However, CHIRPS, in contrast, greatly underestimates the amount of precipitation in that year. Recent studies further confirm that CHIRPS tends to underestimate precipitation in mountainous regions [10,49]. One key reason for this is that CHIRPS relies on smoothed estimates between gauge stations, which can miss localized heavy rainfall caused by topographic lifting.

Figure 5 presents an overall comparison of precipitation estimates. Although all datasets exhibit a strong correlation with ground observations (Table 4), there are significant inconsistencies in their estimations.

Correlation analysis provides vital insights into the efficacy and reliability of the various gridded products (Figure 6). Here “Corr” shows the correlation using all data points. “High” and “Low” show the correlations calculated, respectively, for values above and below the median of the ‘ground’ variable; the diagonal plots show how each variable is distributed. The ERA5-Land dataset shows highly robust and enduring relationships with ground data for both “High” and “Low” values, with correlation values of 0.925 and 0.891, respectively. These values indicate significant and statistically meaningful positive correlations. This suggests notable yet somewhat variable correlations. All relationships in the correlation matrix are statistically significant. This result confirms that the observed connections are robust and that they are unlikely a consequence of a random occurrence. CHIRPS and PERSIANN exhibit varying degrees of dependability based on correlation with the ground data. Figure 7 shows the residual plots between ground data and selected datasets. Overall, all the satellite-based datasets show asymmetric distributions, with prevailing negative residuals. In contrast, ERA5-Land shows a symmetric distribution, thus confirming its greater reliability. Detailed residual plots for every station are given in the Supplementary Material (Figures S1–S4).

The maps of the spatial distribution of the residuals of annual precipitation are given in Figure 8 for all the tested gridded datasets. All satellite products show the highest underestimation of precipitation where the reliefs are exposed to the west-to-east winds, i.e., the mountains spanning from north to south in the central part of the region, as well as the coastal reliefs in the southern part (Figure 1b). Moist fronts from the Atlantic Ocean, propagating along this direction, are indeed responsible for most of the rainfall in the area, with the additional moisture contribution coming from the western Mediterranean [27]. Interestingly, ERA5-Land, though showing a general better agreement with ground data, also shows a similar spatial pattern of the residuals, underestimating the precipitation around the same reliefs. It is worth noting that the non-homogeneous density of the rain gauges throughout the region likely affects this result, as ground stations may also underestimate precipitation in areas with sparse rain gauges. The maps of the residuals of seasonal precipitation, which are given in Figure S5 of the Supplementary Material, show a similar picture, with smaller residuals only being observed in summer, when little precipitation occurs, owing to the Mediterranean climate.

Figure 9, Figure 10, Figure 11 and Figure 12 show the spatial distribution of annual precipitation in the region from 2003 to 2022. It is evident from all satellite data that the northeastern part of the region receives a lower amount of precipitation, except for the PERSIANN data, which deviates from the other three datasets by overestimating precipitation in the northern region throughout the study period. However, the dryer trend in the southern region is consistent in all the datasets from 2016 to 2020. According to CHIRPS data, 2019 was the driest year in most of the region, while PERSIANN also shows 2019 to be the driest year, with 2006 being the second-driest year (Figure 9 and Figure 11).

ERA5 experiences relatively less-noticeable fluctuations compared to the satellite products. In 2006, ERA5 and PERSIANN highlighted the dry spell in the northern region, while CHIRPS and TerraClimate showed lower precipitation in the eastern regions. According to ERA5 and TerraClimate, the western and northwestern regions receive higher levels of precipitation compared to other regions.

The spatial distribution shows some fluctuations in annual precipitation patterns, but there is a slight decreasing trend in the southern regions, according to CHIRPS. This trend is more prominent in CHIRPS data from 2018 to 2022 (Figure 11). The variation in aggregated values of mean precipitation from PERSIANN at a standard spatial resolution shows a strong interannual variability compared to other datasets. This indicates that the meteorological reanalysis product offers a high interannual variability compared to satellite-based products, i.e., CHIRPS. There is a homogeneity in the precipitation patterns in all four datasets, e.g., the precipitation increased between 2007 and 2012, whereas it decreased gradually during the following 10 years. However, the recent (2018–2022) precipitation amounts are close to those previously (2003–2007). Overall, the northeastern region of Campania shows decreasing precipitation, which is in contrast to the southern region, where slightly increasing or stationary patterns dominate during the selected baseline period (Figure 11 and Figure 12).

The spatial variability of the performance of the analyzed datasets is due to the difference in retrieval algorithms, data assimilation methods, and the region’s complex orography. Convective systems, which are common in the Mediterranean region, can produce high-intensity rainfall over small areas and they are most likely to be missed by CHIRPS and PERSIANN, which rely on infrared-based observations [13,46,47,50]. In contrast, ERA5-Land and TerraClimate use station data and model physics, which improves the ability to capture synoptic patterns. However, recent studies show that these two datasets may smooth local extremes in areas of steep gradient [51,52]. Previous studies in this region have shown the orographic enhancement on windward slopes, as well as rain-shadow effects on the leeward side [27,53].

Systematic biases also reflect the influence of the algorithms and adjustment methods used in each dataset. TerraClimate uses WorldClim but is sensitive to errors where reference climatology does not capture fine-scale microclimates [51].

The spatial distribution of overall annual precipitation trends from 1983 to 2022 confirms this general no-change pattern in all datasets, with CHIRPS, PERSIANN, and ERA5-Land reporting very-small positive Sen’s slope mean values of +0.0129 mm/yr (p = 0.2664), +0.0233 mm/yr (p = 0.0760), and +0.0283 mm/yr (p = 0.1246), respectively (

Table 8. Seasonal and overall trends in total monthly precipitation based on Sen’s slope estimator and the statistical significance test (p-value). Values represent Sen’s slope (mm/season/year) and corresponding p-values, respectively. Bold values indicate statistically significant trends.

Dataset	Overall Trend	Winter	Spring	Summer	Fall
PERSIANN	0.023/0.076	1.53/0.289	0.18/0.683	0.74/0.357	1.81/0.029
ERA5-Land	0.028/0.124	1.61/0.345	0.72/0.448	0.54/0.616	2.68/0.091
TerraClimate	−0.007/0.577	0.01/1.000	−0.50/0.435	0.29/0.666	0.06/0.972
CHIRPS	0.012/0.266	0.52/0.584	0.31/0.584	0.22/0.552	1.34/0.115

), as well as an even smaller negative mean value for TerraClimate, at of only −0.0073 mm/yr (p = 0.5773). None of the slopes detected are of statistical significance. This result confirms what was recently observed for the precipitation trend between 1979 and 2019 in southern Italy, where change points were detected, rather than a continuous trend [30].

A non-parametric linear trend was also applied to the monthly mean precipitation time series (1983–2022) (Figure 13). The trends detected were, in most cases, not statistically significant. It is worth noting that precipitation data often exhibit weak statistical trends due to high natural variability. Several studies with longer temporal data have also reported non-significant trends [54,55,56]. Among the datasets, strong positive trends during autumn were detected with ERA5-Land data (2.68 mm/yr, p = 0.0911) and PERSIANN data (1.81 mm/yr, p = 0.0294), with the latter being the only statistically significant trend, having a p-value below the 0.05 threshold (Table 5). TerraClimate, in contrast, showed no significant trends in any season. CHIRPS indicated weak, non-significant positive trends in all seasons, with the strongest being in autumn (1.34 mm/yr, p = 0.1157). A similar comparative study conducted in India also highlighted the limitation of CHIRPS in recording extreme weather events [57]. Overall, the increasing trend in autumn, suggested by three of the datasets analyzed and with the smallest p-values, seems the only significant one, while in the other seasons, there is no evidence of any variation in the total precipitation amount.

Limitations and Future Direction

Although all tested datasets show a significant correlation with ground data, the results highlight that limitations still exist. Gridded datasets may not accurately capture localized precipitation occurrences due to their coarse resolution. Moreover, the treatment of precipitation fluctuations related to elevation is not consistently accounted for in all datasets, which could affect the precision of trend analyses in areas with substantial changes in elevation. While this study focused on standard validation measures to evaluate precipitation datasets, future analyses should more deeply investigate spatial skill metrics, spatial uncertainty, and spatial density sensitivity in order to resolve dataset biases across complex terrains.

To overcome these constraints, it is crucial to combine satellite data with observations taken from the ground to improve the total precision. To enhance the understanding of climatic patterns, it is necessary to improve the algorithms used for precipitation retrieval. Given that precipitation trends are frequently not statistically significant, extending the study period to incorporate a longer time frame may enhance the reliability of the analysis [58]. It would also be helpful to enhance the spatial resolution of the datasets to accurately capture localized patterns of precipitation, so as to set up models that take into consideration the variations in precipitation that are connected to changes in elevation.

4. Conclusions

Among the gridded precipitation products investigated, ERA5-Land proves to be the most dependable option, as it much better resembles ground precipitation data compared to CHIRPS, PERSIANN, and TerraClimate. For accurate rainfall estimates, ERA5-Land is the best choice because it has only a small overestimation, a strong correlation with the observed data, and good values of performance metrics like R², NSE, and POD. In contrast, CHIRPS, PERSIANN, and TerraClimate have notable drawbacks, such as a considerable underestimation of rainfall and a reduced precision in capturing spatial and temporal fluctuations. Spatial analysis also reveals a consistent correlation between ERA5-Land and the observed precipitation patterns. Differently, satellite-based products tend to overstate the precipitation levels in the northern regions and display inconsistencies in other locations, with the largest overestimation in the areas where orographic effects are responsible for the highest levels of precipitation. This issue also affects the accuracy of ERA5-Land, which shows the worst performance in the same areas. The seasonal analysis reveals that ERA5-Land exhibits remarkable precision in the autumn and winter seasons, whereas other datasets tend to underestimate precipitation levels. Sen’s slope analysis (on a monthly scale) showed no significant trends in the total annual precipitation of the whole region. Only in autumn, increasing trends are detected by most datasets, which are statistically significant at the 5% level only in one case (PERSIANN), but with small p-values for ERA5-Land and CHIRPS. The obtained results indicate that in the study area, characterized by complex orography and proximity to the sea, and by a non-homogeneous density of operating rain gauges, gridded precipitation products can supplement ground data for climate analyses.