3.4.1. Solution Based on DNN

The atmospheric duct parameter inversion model based on DNN consists of four hidden layers, and the nodes of each hidden layer were 1024, 512, 256 and 32. The input data was the AIS signal power data, and the output data was the surface or elevated duct parameters. Surface duct parameters consisted of duct height and strength, and elevated duct parameters consisted of atmospheric duct top height, the slope of the base layer, and duct layer thickness and strength. The hidden layer activation function was Rectified Linear Unit (ReLU), and the expression is:

$$f(\mathbf{x}) = \max(x, 0) \tag{11}$$

The adaptive moment estimation method is used for optimization, and can dynamically adjust the learning rate in the process of training to adapt to different weight parameters and achieve better optimization results.
