Operational implementation of a random forest approach to derive quantitative precipitation estimation with measurements from the Swiss polarimetric radar network

Quantitative precipitation estimation (QPE) is crucial for hydrological, climatological and meteorological studies. Weather radars provide remote observations of hydrometeors with a wide-range coverage. However, an accurate retrieval of precipitation intensity from weather radar observations remains a challenge.
In this study, we present an approach using random forest regression to derive QPE from operational polarimetric radar observations. Random forest regression is a machine learning approach based on an ensemble of decision trees. The model “Rainforest” (RF) is trained with a database containing six years (January 2016 to December 2021) with observations from 288 rain gauges and polarimetric radar observations from five dual polarization Doppler C-band radars (Swiss weather radar network). Each radar performs plan indicator (PPI) scans at 20 elevations in 5 minutes.
The complex topography in Switzerland poses many challenges to derive an accurate QPE. Such challenges include beam shielding, ground clutter and the complex spatial structure of precipitation. Hence, it is important to use information from the entire vertical column of radar. Here, the vertical column of radar observations that results from 20 radar sweeps are aggregated to a single value at the ground level onto a Cartesian grid by applying an exponential weighting function in the vertical.
Based on the evaluation of the feature importance, we include 13 predictors in RF: horizontal and vertical reflectivity, fraction of the contribution of each radar to the grid cell, height of each radar volume above the grid cell, static visibility of the radar volume, specific differential phase shift, co-polar correlation coefficient and spectral width. The legacy model presented by Wolfensberger et al. (2021) includes temperature estimates given by the numerical weather prediction model COSMO-1 at each model elevation. Because of constraints in the operational implementation, this multi-layer predictor is replaced by the single-layer 0°C isothermal altitude. Further evaluations show that this change does not affect the overall performance of the model.
The model is then trained using the extensive database of station observations and the corresponding grid cells of aggregated input variables. A five-fold cross-validation shows a promising performance in the accuracy of the model at an hourly time resolution. RF outperformed the single-polarized operational QPE model at MeteoSwiss.
As the gauge observations of the Swiss network are logged at 10 min time steps, the temporal resolution of the radar measurements have to be aggregated from 5 to 10 min. This is done through an average over two consecutive time steps of radar volume scans. For the operational implementation, QPE maps are produced at a 5min temporal resolution by applying a temporal disaggregation scheme based on the Swiss Z-R relationship.
In summary, we present the implementation of a machine learning model for operational QPE maps at a 5 min resolution on a 1x1 km2 grid in Switzerland. The new model, which runs in real-time with operational dual-polarization data, is able to outperform the currently operationally implemented radar-only QPE at MeteoSwiss. The novel approach shows great potential for meteorological and hydrological studies in complex terrain regions.