From ice phase to liquid precipitation at the ground level: weather radar and machine learning

The relationship between the dominant ice phase hydrometeors and the intensity of precipitation at the ground level has not been extensively studied. This thesis explore and make use of the links between the respective proportions of solid phase hydrometeors and the quantitative precipitation estimation at the ground level. The solid phase hydrometeors are obtained by applying a semi-supervised classification methods based on polarimetric radar measurements on a datasets acquired by the MeteoSwiss Monte Lema radar during seven events of intense precipitation. Precipitation intensity measurements are obtained from the CombiPrecip product that combines rain-gauge and radar precipitation estimates. In the first part of this thesis, the relationship between the percentage of aggregates (AG), crystals (CR) and rimed particles (RP) solid hydrometeors and the precipitation intensity in the vertical columns above the corresponding 1x1 km ground level pixels is studied. Several regression approaches indicate a relevant non-linear relationship between solid phase hydrometeor proportions and the precipitation intensity at the ground level. In the last part, machine learning algorithms are used to estimate liquid precipitation at the ground level from the different hydrometeor proportions of CR, AG and RP. Namely, a neural network with one hidden layer and a multiple linear regression are trained and tested by mean of cross-validation with solid phase hydrometeors and other non-independent parameters. The results of the machine learning algorithms show relatively good performances in estimating precipitation at the ground level. This approach, which make use of solid phase radar measurements, could be very relevant in case of complex topography.