Pseudo-3D Radar Reflectivity By Applying U-Net Models to Global Passive Microwave Brightness Temperatures

Weather-radar coverage around the world is sparse. They are mostly located in the most developed countries, sometimes only around densely populated areas, with an almost non-existent coverage of the oceans and underdeveloped countries. The satellite passive-microwave Brightness Temperature (TB) observations, with a global coverage and a high revisiting rate, have been mainly used in surface precipitation retrievals. However, these observations can also be used to indicate the intensity of convective systems. This study attempts to relate passive microwave TBs to the radar observables so that the intensity of convection can be intercompared between radar and passive microwave observations. The Global Precipitation Measurement (GPM) core satellite, with both radiometer (GPM Microwave Imager, GMI) and radar (Dual-frequency Precipitation Radar, DPR) sensors, serves as the perfect medium to train a set of U-shaped Convolutional Neural Network (CNN, U-Nets) to estimate the maximum heights of selected radar reflectivity values from TB information alone at the pixel level. As such, a collocation step is done to match TB observations from 13 microwave channels (i.e., from 11 to 183 GHz) to their radar equivalents (i.e., top heights of 15/20/30/40 dBZ) into 48x48 pixels rasters. Considering the size of such dataset, the models are trained using only one year of data (i.e., 2020) and are tested on the following one (i.e., 2021). Four different U-Nets are consequently trained to regress the maximum height of their respective reflectivity at the pixel level. Current results present high false alarm rates (i.e., any estimation above 0 km when nothing was observed by the DPR; with 50.1%, 58.2%, 68.6%, and 74.2% for 15, 20, 30, and 40 dBZ respectively), resulting from the lack of insight provided by TB alone and the uncertainty linked with the use of the DPR, but with great Pearson’s correlation coefficients (i.e., 0.88, 0.87, 0.83, and 0.68 for 15, 20, 30, and 40 dBZ respectively), coefficients of determination (i.e., R²; with 0.73, 0.69, 0.59, and 0.02 for 15, 20, 30, and 40 dBZ respectively, and Structural Similarity Index Measure (i.e., SSIM; with 0.79, 0.78, 0.82, and 0.91 for 15, 20, 30, and 40 dBZ respectively). Results suggest that there is a great potential to retrieve pseudo-radar profiles from passive microwave TBs and fill the radar coverage gaps around the globe.