Supporting data for "Using AI-based Deep Learning Algorithms for Nowcasting Cloud Evolution"

Under climate change, it seems that extreme rainfalls occur more frequently around the world, posing significant threats to the lives of people and the safety of properties. To this end, this study, ''Using AI-based Deep Learning Algorithms for Nowcasting Cloud Evolution'', focuses on using AI-based deep learning algorithms for accurate nowcasting of cloud evolution. A high-resolution radar echo map mosaic dataset collected by the Meteorological Bureau of Shenzhen Municipality, China, is adopted. This new image dataset includes all the rainfall events in Guangdong Province of China between 2010 and 2020 with a spatiotemporal resolution of 1 km and 6 minutes, providing abundant details for the training of deep learning algorithms. The raw outputs from the weather radars are the logarithmic radar reflectivity factors (expressed as the echo intensity hereafter in this study) measured with dBZ between 0 and 70. Such factors are linearly transformed to pixel values between 0 and 255 to generate grey scale radar echo maps for the extrapolation methods and deep learning algorithms. The pixel values are stored in the square matrix for the convenience of data manipulation. To capture the complete evolution of rainstorm clouds in rainfall events, mosaic images of different radars are generated with the Constant Altitude Plan Position Indicator(CAPPI) images. Notably, the radar echo intensity can be converted to rainfall intensity, and then we can nowcast rainfalls. Each radar echo map covers an area of 256 km×256 km with a spatial resolution of 0.01°×0.01° (each pixel represents about 1 km×1 km area), and each radar echo map sequence lasts at least 120 minutes with a 6-minute time interval. Therefore, each sampled sequence includes 20 (= 120/6) frames, 10-frame radar echo maps as input data to those nowcasting models and 10 frames as the ground truth for evaluating 1-hour (= 10×6) model nowcasting results.