New insights into atmospheric predictability from global high-resolution non-hydrostatic simulations
Kinetic energy (KE) and tracer spectra derived from global atmospheric simulations with a mean grid spacing of 13 km from the Finite‐Volume Cubed‐Sphere (FV3) numerical forecast model are presented. The FV3 KE spectra compare well with previous observation-based spectra and possess a well-resolved structure at both synoptic and mesoscale. Also, the KE spectra reveal that the effective model resolution is around 6dx resulting from the FV3 configuration. Trace spectra show the dependence on the height, which implies the influence of the discontinuous flow near the surface on the predictability at different altitude. FV3 BT spectra and GOES-R BT spectra are qualitatively similar at both synoptic and, which is likely associated with convection.
Deep learning project: Lightning Detection from Weather Radar
In this study, we use Radar Images in deep learning algorithm to detect lightning. Radar reflectivity represents the quantity and size of water and ice particles in the atmosphere. This value does not directly relate to the presence of lightning, but similar processes that produce high reflectivity values also lead to a greater probability of lightning. We use radar data along with lightening labels from the Geostationary Lightning Mapper to train deep learning models for lightning detection. The radar image was captured once every 5 minutes.The lightning strikes were captured once every 20 seconds and combined into one lightning label every 5 minute. These data are available from Mar, 2018 to Oct, 2019, giving around 150,000 images in total. We use data augmentation and downsampling to overcome the unbalanced nature of the data and reduce the memory demands of the model. We test UNet, Google Inception v3 and ResNet architectures initially. In the initial testing, UNet performed the best. Training a new UNet model from scratch we find that it can reasonably predict lightning locations from radar data, with an F1 score of 0.29.
More involved details are included in our report: IST597_team4_final_report