A comparison of classification methods to perform a typology of precipitation events for soil erosion modelling

Soil erosion is a major threat to soil resources. Our ACRP-supported project EROS-A aims to improve erosion modelling by investigating the role of extreme precipitation and associated erosivity on soil erosion in the Main Agricultural Production Zones (MAPZ) of Austria. For this purpose, it is important to separate precipitation events into different process types (e.g. convective and stratiform events), as these are expected to follow different distributions and can be modelled more accurately using a mixture model approach.

In this contribution, we assess the performance of different clustering methods to establish a process typology of precipitation events. The study uses high-resolution rainfall data with a time resolution of 5 minutes from 27 stations in the agricultural area of Austria. Hourly lightning data (ALDIS) is used as a conditional variable, as thunderstorms are a good indicator of convective events. In our approach, a precipitation event is defined as time spell when precipitation exceeds 0.1 mm per 5 minutes. Similar to a drought analysis, this can result in short, interdependent events. These are pooled using a minimum precipitation of 1.27 mm in 6 hours as an interevent time and volume criterium. The temporal characteristics of rainfall events are characterized by five indices: the amount (aggregated event precipitation), duration (the time between the start and end of the event), intensity (amount divided by duration), peak intensity (the maximum 5-min intensity), and the time-to-peak (relative to the duration of the event). These characteristics are typically dominated by small (positive) values and are thus assumed to follow a Gamma distribution. In addition, the binary lightning index was considered as this is expected to have discriminative power as well.

Based on the rainfall events so obtained, cluster analysis is performed using partitioning around medoids (PAM) with Gower metric transformed lightning index. For comparison, model-based cluster analysis for mixtures of multivariate Gamma distributions is conducted. The results are compared using the discriminative power of principal component analysis (PCA) and measures of cluster homogeneity and discriminability.  In a final evaluation, the discriminative power of the event classification is assessed in terms of event type distributions. Initial results indicate that the event indices contain a wealth of information that can be profitably used to establish a typology of precipitation events. The results will feed into future studies to perform rainfall simulations that can serve as an input for erosion scenario modeling for agricultural decision support.