Gust front detection algorithm based on deep convolutional neural network

As a cold outflow feature of strong convection, gust front is an important boundary layer convergence system. But its automatic monitoring and identification have always been a difficult point in daily meteorological services. In this paper, an automatic recognition algorithm for gust front is designed based on deep convolutional neural network(d-CNNGFDA. Based on Faster RCNN algorithm and Inception V2 network model, end-to-end automatic recognition of narrow-band gust front echoes from radar echo data is realized by redesigning the input and output terminals. According to the rotation invariance of the radar data around the radar center, the data samples are added and the complexity of feature extraction is reduced. The model is trained by 200, 000 steps through the Nanjing radar data from 2007 to 2011, and the total loss function value converged to 0.003. Analysis of the recognition results shows that the recognition rate is 100%, the missing rate is 0%, and accuracy rate is 87% in the training samples. By evaluating the generalization ability of 32 voluminal scan data from Hefei Doppler weather radar on June 5, 2009, the recognition rate is 91.7%, the missing rate is 8.3% and the correct rate is 73.3%.