Development of a Quantile Regression Model for Replicating Extreme Precipitation in Different Climatic Zones of Morocco
Author(s):
Oumechtaq Ismail1*, Oulidi Abderrahim1 and Bahaj Tarik2
In Morocco, where floods represent the most frequent risk, the modeling of this phenomenon primarily relies on precipitation data. Given the dispersion
and scarcity of meteorological stations in the country, as well as gaps in the recorded series, many researchers turn to satellite data to obtain this crucial
information. However, the direct use of satellite data can introduce biases, necessitating research efforts to correct them. Several studies have proposed
solutions, such as the merging of satellite products or the use of arithmetic corrections like averaging, but these focus mainly on the central part of the data
(low to moderate precipitation), leaving extreme events, which cause floods, uncorrected.
Our contribution is divided into two parts: first, we evaluated the reliability of different satellite data sources, including GPM, ERA5, TRMM, and PERSIANN.
T
hen, we developed a quantile regression (QR) model tailored to each homogeneous region (five regions) in terms of climatic factors responsible for
generating extreme events. This model was designed to accurately reproduce extreme events, even in areas lacking measurement stations. We used the
Kullback-Leibler (KL) distance to determine the rate that best reproduces ground precipitation. Subsequently, a comparison of the selected model with the
linear regression model was made to demonstrate that the latter cannot accurately inform about extreme precipitation.
T
he results reveal a variation in the rate of the selected regression model from one area to another, as well as the coefficients defining the contribution
of each satellite product, depending on the specific rainfall regime of each homogeneous region. We validated our model through different approaches:
validation using extreme precipitation, validation using other stations outside the DGM network, and another validation using extreme events beyond the
period [2000-2018] used to construct the QR model.
