*5.2. Comparison of CNN Analysis Results*

We extracted the image after two 7 × 7 convolution layers to analyze whether the model can extract the haze characteristics [30–32], as shown in Figure 8.

**Figure 8.** Convolution feature extraction diagram (**a**) Convolution diagram during severe haze; (**b**) Convolution diagram with a lower haze level.

Figure 8a is a convolution result diagram when the haze is severe in spring, and the bright block is in the upper left corner of the image. We find a cloud layer in the area comparing it with the original image, which indicates that the cloud layer appears bright in the convolution result. The remaining areas with severe haze are darker, where the difference between emissivity and reflectance is more significant. Figure 8b is a convolution result graph containing minimum cloud information and a lower haze concentration level. The brighter feature in Figure 8 is the area where the AOD is low. The difference between the emissivity and the reflectance is slight. We found that the haze level prediction model can effectively distinguish the image characteristics of different haze concentrations by comparing the results. We labeled the cloud information of the image when marking the dataset, avoiding the mistakes where the cloud was identified as haze.

We used the MOD02-1 km data of the Beijing area in 2013 and 2014 as the training set and test set of the haze level prediction model. We extracted satellite images from the MOD02-1 km data so that the training and test sets contained 730 satellite images. We marked the haze level on the training set.

To verify whether the model can effectively establish the correlation between satellite image and PM2.5 concentration and to compare it with the traditional inversion method, we conducted the same linear regression between the output of the haze level prediction model and the PM2.5 daily average concentration. The results are shown in Figure 9.

**Figure 9.** Linear regression between haze level using CNN and PM2.5 (**a**) Linear regression for CNN and PM2.5 in Spring; (**b**) Linear regression in Summer; (**c**) Linear regression in Fall; (**d**) Linear regression in Winter.

In Figure 9a, the y-intercept *a*<sup>0</sup> is −13.88. The slope *a*<sup>1</sup> is 34.74. The correlation coefficient R is 0.90. In Figure 9b, the y-intercept *a*<sup>0</sup> is 25.07. The slope *a*<sup>1</sup> is 17.11, and the correlation coefficient R is 0.65. In Figure 9c, the y-intercept *a*<sup>0</sup> is −20.74. The slope *a*<sup>1</sup> is 40.61. The correlation coefficient R is 0.93. In Figure 9d, the y-intercept *a*<sup>0</sup> is 9.03, and the slope *a*<sup>1</sup> is 40.67, the correlation coefficient R is 0.65.

A comparison with the traditional inversion method is shown in Table 4.

**Table 4.** The correlation coefficient between traditional inversion and level prediction.


From Table 4, the correlation coefficient of the haze level prediction model based on the convolutional neural network is superior to the traditional inversion method in spring, summer, and fall. In particular, the summer and fall results are improved by 12% and 39%, respectively, which indicates that the haze level prediction model can provide a better PM2.5 concentration prediction than the traditional inversion method. Furthermore, all correlation coefficients in the haze prediction model are above 0.6, indicating a strong correlation between haze level and PM2.5 concentration.
