A Deep Convolutional Neural Networks-Based Method for Inversion of Rough Surface Parameters

Deep convolution networks (CNN) is applied to inverse the rough surface parameters, including the root-me an-square height and the correlation length, from microwave images. We employ computational electromagnetic method to simulate the training data for deep CNN. The simulated backward scattering data is converted into microwave images as the inputs to the CNN. An inversion network of deep convolution neural networks with five cascaded convolutional-maxpooling layers and two fully connected layers is designed, including feature extraction and data regression by using convolution layers and fully connected layers. The simulated results demonstrate the feasibility to inverse the sough surface parameters from electromagnetic scattering fields by using deep CNN.