Inpainting Gaps and Blocked Areas in Weather Radar Networks with Deep Learning

One of the assumptions of correct weather radar measurements is the unblocked line of sight between radar and the observed precipitation. However, the majority of weather radar sites are imperfect in a way that physical obstacles obscure the line of sight at some of the azimuths and elevations. The first two radar sites with Meteopress solid-state C-band radars installed are not exceptions. There are buildings at the Brisbane Airport in Australia that reflect radar beams resulting in observing different azimuths as expected. The hill Loučná in the Czech Republic has two towers on top of it, one blocking the views to the north.

The first step in accounting for these imperfections is identifying and masking the corrupted data; the second is filling in the correct or estimated data. Here we present a novel solution utilizing a generative adversarial network (GAN) for inpainting missing data in the radar reflectivity measurements, which has been trained using a synthetic dataset with simulated blockages of correct measurements.

The model is trained to extrapolate from the previous radar images utilizing the information from the uncorrupted neighboring azimuths and elevations to predict accurate data that looks the same to the human eye as the truly measured data. The model has been tested quantitatively on the simulated dataset and qualitatively on Meteopress radar sites. Furthermore, the model transfers well to inpainting data missing for different reasons. Weather radar networks have gaps caused by neighboring radars too far away. We demonstrate the capability of the model to transfer to this setting on a case study simulating gaps in the networks and providing satellite imagery on the input.