For agriculture, it is extremely important to know how much it rained on a particular field. However, it is impossible to have local rain gauges everywhere. Therefore, remote sensing instruments such as radar are used to provide wide spatial coverage. It is a real problem to estimate precipitation as good as possible based on radar information.
In this paper I explore how to address this problem in a probabilistic manner using data mining techniques and the CRISP-DM (CRoss-Industry Standard Process of Data Mining). I compare a classic model optimising the least absolute deviation with a probabilistic based model using quantile regression based on multiple regressors. I use WSR-88D polarimetric radar meteorological data and I show the important attributions of the dual-polarization radars. My goal is to create new valuable variables and show their relevances. As a part of a framework to measure the performances I use Continuous Rank Probability Score, which makes possible to compare probabilistic and not probabilistic models.