Detecting shallow precipitation from conical-scanning radiometer observations using a Random Forest model over the Netherlands

Over the past decades, spaceborne radiometers have proven to be valuable input to realize a global coverage of precipitation estimates. However, retrieving accurate shallow precipitation estimates from radiometers remains challenging. The signal related to precipitation formed close to the Earth’s surface is difficult to distinguish from dry weather, especially over land.  Despite the relatively low precipitation rates that are often associated with shallow precipitation, its persistent presence results in a significant contribution to the total amount of rainfall over the mid- and high latitudinal regions. Hence, correct identification is important.

This study aimed to improve our understanding of the radiometric signatures of shallow precipitation from passive microwave observations by implementing a Random Forest (RF) model. RF is chosen because of its limited risk of overfitting and the ability to physically interpret the resulting model structure and parameters. The RF model is applied to brightness temperature observations from all channels onboard the Global Precipitation Measurement (GPM) Microwave Imager (GMI) during 2017-2020 over The Netherlands (52°N). A high-quality gauge-adjusted radar product is used as reference. The echo top height retrieved from the two radars in The Netherlands (Herwijnen and Den Helder) are used to classify the GMI footprints to either dry, shallow (<3km) or non-shallow (>3km) regime.

Hyperparameter settings, such as the depth of the model, and choices such as the number of years the model is trained on or the threshold to classify footprint as dry, shallow, or non-shallow regime have a limited effect on the performance of the RF. In general, the model tends to wrongly classify dry footprints as wet (both shallow and non-shallow). The model showed a seasonal dependency, with the best performance in summer. Preliminary results also showed a strong seasonal effect when excluding all footprints within 40km distance of the coast. These results indicate that four different parameter sets representing each season are required. Furthermore, the specific years the model is trained or tested on are found to strongly affect its performance. Currently, additional variables (such as ERA5 freezing level, two-meter air temperature) and simultaneous observations from the GPM dual-frequency precipitation radar (DPR), are included to further improve and understand the performance of the RF model.