*3.3. Atmospheric Duct Classification Model*

The atmospheric duct classification model adopted the DNN of three hidden layers, and the nodes of each hidden layer were 512, 256 and 32. The input layer data involved the AIS signal power data, and the output layer data involved the atmospheric duct type data. The type of data consisted of three numbers, and the format and value are shown in Table 3.

**Table 3.** Type data format.


The hidden layer activation function is tanh, and the expression is:

$$\tanh(x) = \frac{1 + e^{-2x}}{1 - e^{-2x}} \tag{9}$$

The output layer activation function is softmax, which is one of the most common activation functions. The expression is:

$$S(x) = \frac{1}{1 + \varepsilon^{-x}} \tag{10}$$

The Stochastic Gradient Descent Method was selected for optimization; it splits the dataset into batches and randomly selects a batch to calculate and update the parameters.
