A Comparison of MRMS and GREMLIN Reflectivity Using the Method for Object-Based Diagnostic Evaluation (MODE)

Radar reflectivity is critical in the verification of convection-allowing forecasts, where small-scale features such as thunderstorms can be resolved on the model grid. Additionally, reflectivity is a key source of information for data assimilation used to generate initial conditions for these forecasts. The Multi-Radar Multi-Sensor system (MRMS) has been relied upon for both of these applications. MRMS is a mosaic product which integrates data from a variety of sources, most notably reflectivity from individual radar sites in the United States and Canada. Its standardized, fine spatial and temporal resolution makes it appealing for use in data assimilation and verification. However, it is not without limitation; radar coverage is still lacking in areas of complex terrain or over open water. Such gaps propagate into MRMS products, which causes a deficit in data quality that could hinder the performance of the models that ingest it or influence the verification of convective-scale forecasts in certain regions. One proposed solution to fill these gaps is satellite-derived reflectivity products, such as GOES Radar Estimation via Machine Learning to Inform NWP (GREMLIN). GREMLIN is a convolutional neural network that utilizes infrared satellite observations to estimate the composite reflectivity field. While its resolution is coarser, it is available in areas of complex terrain and over open water. This work aims to evaluate GREMLIN’s performance using the Method for Object-Based Diagnostic Evaluation (MODE) to provide perspective on particular aspects of GREMLIN’s output and serve as an evaluation of GREMLIN’s potential to become a supplement or even a replacement to MRMS. Preliminary findings suggest that GREMLIN’s known limitations make perfect matches rare. Other metrics such as increased area and decreased complexity suggest that it sometimes merges several storm objects into one. Overall, though, areas where GREMLIN’s performance is lacking seem to be artifacts of previously identified limitations. Additional work is ongoing to continue investigating this comparison with respect to several case studies, seasonality, and storm mode.