Robustness Spatiotemporal Clustering and Trend Detection of Rainfall Erosivity Density in Greece

Abstract

Soil erosion is affected by rainfall, among other factors, and it is likely to increase in the future due to climate change impacts, resulting in higher rainfall intensities. This paper evaluates the impact of the missing values ratio on the computation of the rainfall erosivity factor, R, and erosivity density, ED. The paper also investigates the temporal trends and defines regions of Greece with a similar monthly distribution of ED using an unsupervised method. Preprocessed and free from noise and errors rainfall data from 108 stations across Greece were extracted from the Greek National Bank of Hydrological and Meteorological Information. The rainfall data were analyzed and erosive rainfalls were identified, their return period was determined using intensity–duration–frequency curves and R and ED values were computed. The impact of missing data in the computation of annual values of R and ED was investigated using a Monte Carlo simulation. The findings indicated that missing rainfall data resulted in a linear underestimation of R, while ED is more robust. The trends in ED timeseries were evaluated using the Kendall’s Tau test and their autocorrelation and partial autocorrelation were computed for a small subset of stations using criteria based on the quality of data. Furthermore, cluster analysis was applied to a larger subset of stations to define regions of Greece with similar monthly distribution of ED. The findings of this study indicate that: (a) ED should be preferred for the assessment of erosivity in Greece over the direct computation of R, (b) ED timeseries are found to be stationary for the majority of the selected stations, in contrast to reported precipitation trends for the same time period, (c) Greece is divided into three clusters/areas of stations with distinct monthly distributions of ED.

Estimating Current and Future Rainfall Erosivity in Greece Using Regional Climate Models and Spatial Quantile Regression Forests

“…The second had an almost bimodal shape with its two peaks at July and October. The values of ED were slightly different from the ones already reported in our previous work [30], due to the application of the seasonal monthly conversion factors c m to a constant one.…”

“…During summer the convective activity over norther Greece produces higher R values than southern Greece, with the largest values occurring in the area of Thrace. The monthly temporal patterns illustrated in Figure 5, are compatible to the three areas that were identified in our previous study [30].…”

“…The climatic models from COordinated Regional Climate Downscaling EXperiment over Europe (EURO-CORDEX) [27], using high-resolution RCMs for the high greenhouse gases concentration scenario, RCP8.5, project a decrease of precipitation from 1971-2000 to 2071-2100 and for the medium concentration scenario, RCP4.5, project the same trend with a smaller magnitude [27,28]. Regarding rainfall intensity, in the form of heavy precipitation that exceeds the intensity at the 95th percentile of daily precipitation, the same models project diverging trends that are not statistically significant in most areas of Greece [27,28].In our previous studies about ED in Greece [29,30] it was proven that, in general, ED values are robust to the presence of missing values in contrast to R, which is highly affected, and specifically in Greece: (a) the values of ED are not significantly correlated with the elevation, (b) ED annual timeseries are found to be stationary, in contrast to reported precipitation trends for the same time period and (c) ED can be considered as spatially autocorrelated, as three contiguous areas were identified using clustering analysis, that had distinct temporal patterns. A comparison of our studies with an earlier study about R in Greece by Panagos et al [31] revealed that the previously reported R values were underestimated due to the presence of a significant volume of missing data in the precipitation records used in the calculations.In the Mediterranean region, the annual R model MedREM was developed using annual precipitation depth, the longitude and annual daily maximum precipitation data [32].…”

“…The estimation of current and future monthly erosivity was made under the assumption of future temporal stationarity of ED values, due to the fact that ED is related to seasonal rainfall intensity [15] and EURO-CORDEX GCMs-RCPs models does not project statistically significant trends in most areas of Greece [27]. Temporal stationarity of ED values already has been documented in Greece for the historical period [30].…”

“…In our previous studies about ED in Greece [29,30] it was proven that, in general, ED values are robust to the presence of missing values in contrast to R, which is highly affected, and specifically in Greece: (a) the values of ED are not significantly correlated with the elevation, (b) ED annual timeseries are found to be stationary, in contrast to reported precipitation trends for the same time period and (c) ED can be considered as spatially autocorrelated, as three contiguous areas were identified using clustering analysis, that had distinct temporal patterns. A comparison of our studies with an earlier study about R in Greece by Panagos et al [31] revealed that the previously reported R values were underestimated due to the presence of a significant volume of missing data in the precipitation records used in the calculations.…”