Tackling Drizzle Bias in the Euro-Mediterranean region: The Multivariate Adjustment Solution

Precipitation plays a pivotal role as a climatic indicator across various domains, particularly in the context of climate change studies. However, its accurate simulation and projection remain challenging due to its inherent stochastic nature. Climate models frequently exhibit a tendency to overestimate the frequency of light precipitation events and underestimate the magnitude of extreme precipitation totals, a phenomenon commonly referred to as the 'drizzle bias'. Consequently, while the total precipitation amounts may be adequately represented, discrepancies often emerge in the frequency distribution of rainy days. This discrepancy challenges the model’s ability to capture the precipitation patterns, impacting climate related assessments and predictions.

This study seeks to mitigate the 'drizzle bias' in simulated precipitation, by introducing and implementing two distinct statistical methodologies aimed at improving the precision of simulated and projected rainy-day counts within the broader Euro-Mediterranean region. The first approach, which mimics the convention, involves adjusting the number of rainy days based on the assumption that the relationship between observed and simulated rainy days remains constant over time (thresholding). In contrast, the second approach employs a machine learning model specifically Random Forests (RF), to statistically minimize the drizzle bias based on a function of several simulated climate variables.

The findings suggest that utilizing a multivariate approach yields outcomes comparable to traditional thresholding techniques when adjusting for sub-periods characterized by similar climatic patterns. However, the efficacy of the RF method becomes apparent when addressing periods marked by extreme bias, distinguished by substantially different frequencies of rainy days. These notable deviations predominantly manifest within the Mediterranean domain, particularly in the extremely arid regions. Specifically, within the Mediterranean area, simulated rainy-day counts exceed 100% (relative proportion), with percentages in the African segment of Mediterranean reaching up to 200%.

Furthermore, the study reveals that while the prevalence of the thresholding method is prominent in Eastern Europe and select Mediterranean locations, the RF method demonstrates superior performance for stations exhibiting significant disparities between the two methodologies, notably in the Balkan Peninsula.

Concluding, by employing innovative statistical techniques, this study enhances our understanding of precipitation modelling and highlights the importance of tailored methodologies, particularly in regions characterized by distinct climatic characteristics, such as the Mediterranean.