*3.4. The Results of ANNs*

As shown in Figure 2, we adopted the three-layer neural network structure; the number distribution of three-layer neurons was N-2N-N. In this section, the influence of the value of N and activation function is analyzed. N is set at 5, 10, 20, 30, 40, 50 and 60, respectively. Tables 5–7 show the RMSEs, Rs and MDs of the wind speed retrieval using ANN models with different activation functions and N values, respectively. As in Section 3.3, the bold font represents the best result, and the performance of the high wind speed models was analyzed in three data intervals. Figure 8 shows RMSEs and Rs of the wind speed retrieval models in a more intuitive form, i.e., in the form of line chart.

**Table 5.** RMSEs of the ANN models.


**Table 6.** Rs of the ANN models.


**Table 7.** MDs of the ANN models.


**Figure 8.** RMSEs and Rs of the wind speed retrieval models using ANN models with different activation functions and different numbers of neurons.

In the low wind speed interval, it is obvious that the choice of activation function hardly affected the ANN models, as shown in Figure 8. However, the increase of the number of neurons significantly reduced the accuracy of the models, although the accuracy variation was not very significant. In the high wind speed interval, the increase in the number of neurons had little effect on Sigmoid and Tanh, but it had an obvious effect on ReLu. As shown in Table 6, for a low wind speed interval, when the activation function was Sigmoid and N was 10, the performance of ANN was the best. For a high wind speed interval, when the activation function was Tanh and N was 30, the performance of ANN was the best. Overall, although the underestimation of ANNs at high winds was smaller than that of regression trees, the retrieval performance of ANNs was slightly worse than that of the regression tree modeling methods. In order to facilitate a comparison with other methods, scatter plots of low wind speeds retrieved by the Sigmoid function and of high wind speeds retrieved by the Tanh function are presented as examples, as shown in Figure 9.
