APPLICABILITY OF DETECTING SURFACE WEATHER FRONTS USING CONVOLUTIONAL NEURAL NETWORK BASED DEEP LERANING

Abstract

 A deep learning model using a convolutional neural network (CNN) and U-net is built to study the applicability in detecting surface weather fronts in Japan and surrounding sea. First, a CNN model is used to predict whether there are fronts (1) or not (0) in the region. If CNN outputs 1, frontal locations can be predicted by assembling binary classification of front/no front at each grid point within the region using a U-net. The predictability of CNN/U-net stems from their ability to match locations of outstanding horizontal gradients with observed frontal locations. Our study shows that CNN can achieve higher accuracy in filtering out no front cases by using fewer predictors than the U-net. Overall, the predictability of frontal locations varies with season and differs when stationary fronts are present or not, which suggests that dynamics related to seasonal frontogenesis can influence the spatial distribution of predictability of fronts.