A weather radar image prediction method in local parallel computation

Proposes a new method for the short term weather radar image prediction which depends on local information of images. A conventional method, called a kinetic method, estimates a moving vector of two radar echo images and predicts future radar echo images. This method cannot predict local changes of radar echoes in the image because they calculate the average radar echo velocity of the whole image. The proposed method treats local changes of radar echoes as local mapping between the images and predicts future radar echo images using a neural network as a nonlinear mapping. The performance of the proposed method is 12% better than a conventional method in the best case of critical success index (CSI) evaluation. The authors propose a linear approximation including terms of diffusion, translation and a sink and source to decrease the calculating costs. This approximation method provides almost the same prediction performance as the nonlinear method, and the calculating time is about 25% less than a nonlinear one