Thursday, 31 August 2023
Boundary Waters (Hyatt Regency Minneapolis)
Super resolution can be used to enhance the spatial resolution of gridded data. Traditionally, the super resolution has been performed using interpolation methods such as nearest neighbor, bicubic interpolation algorithms [1]. However, super resolution based on machine learning methods has been studied recently and shown its superiority in computer vision and image processing fields compared to the traditional interpolation schemes. The research on super resolution in radar meteorology is also important since the weather radar has limited range resolution and ray spacing for the scan. This can be enhanced by the super resolution based on machine learning method using the low resolution of images as an input data and producing high resolution of images. In this paper, we propose Attention U-net machine learning model for single image super resolution of radar Plan Position Indicator (PPI) images. We used reflectivity from National Weather Service, dual-polarized Doppler radar (WSR-88D) data in Dallas Fort Worth (DFW) to train and test the model. We gridded the radar data with 256 km for latitude and longitude and saved it as 256x256 size Portable Network Graphics (PNG) image files so that it can have 1 km spatial resolution. The image files were downsampled to 64x64, 32x32 and used as training data while the original 256x256 size images were used as test data. Therefore, the machine learning model performs 4x, 8x super resolution of PPI images. We focused on the hailstorm occurred on April 3, 2012 in DFW. We compared the super resolution performance of machine learning model to nearest neighbor, bicubic interpolation methods and we found that the presented Attention U-net model outperforms traditional interpolation methods in both PPI images and metrics used for evaluation and has similar performance to U-net model. Furthermore, we cropped with the size of 64x64 and 32x32 where there was high reflectivity from the original 256x256 image and performed 4x, 8x super resolution respectively.
[1] Geiss, A., and J. C. Hardin, 2020: Radar Super Resolution Using a Deep Convolutional Neural Network. J. Atmos. Oceanic Technol., 37, 2197–2207, https://doi.org/10.1175/JTECH-D-20-0074.1
[1] Geiss, A., and J. C. Hardin, 2020: Radar Super Resolution Using a Deep Convolutional Neural Network. J. Atmos. Oceanic Technol., 37, 2197–2207, https://doi.org/10.1175/JTECH-D-20-0074.1

