*3.4. Identification of Various PAHs on Fruit and Vegetable Surfaces*

Because there is more than one PAH residue on the surface of fruits and vegetables in the real environment, the main purpose of this work is to realise the identification of various PAHs, without considering a single class. And the competitive adsorption of SERS can also lead to different contributions of PAHs to the SERS signal at the same concentration and can even cover the signals of other PAHs, resulting in low efficiency and accuracy in identifying various PAHs through the manual analysis of spectra. Consequently, in this study, a lightweight network combined with the SERS spectra of various PAHs was used to construct a classification model for intelligent, accurate identification.

The SERS spectra of various PAHs were obtained using the flexible β-CD@AuNP/PTFE substrate, as shown in Figure 5. Figure 5A shows the SERS spectra of Pyr, Nap, and BaP. It is evident that the characteristic peak of Nap at 505 cm−<sup>1</sup> is different from those of

Pyr and BaP, and the characteristic peaks of Pyr at 587 and 1399 cm−<sup>1</sup> are unique without overlapping. Similarly, the characteristic peak of BaP at 607 cm−<sup>1</sup> is unique. These three targets have unique characteristic peaks that provide a basis for the identification of subsequent detection.

**Figure 5.** SERS spectra of BaP, Nap, and Pyr at a concentration of 10 μg/mL (**A**); SERS spectra of various PAHs (**B**), from top to bottom: (a) BaP + Nap, (b) Pyr + Nap, (c) BaP + Pyr, (d) BaP + Nap + Pyr.

Figure 5B shows the SERS spectra obtained after mixing Pyr, Nap, and BaP. From spectra d of BaP + Pyr + Nap, it is evident that the number of characteristic peaks is more than those of BaP + Nap, Pyr + Nap, and BaP + Pyr (Figure 5B a, b, c). Therefore, the identification performance for BaP + Pyr + Nap may be higher than that for the other types of spectra in subsequent identifications. Additionally, the spectra of a, b, and c in Figure 5B all contain the characteristic peaks of the two PAHs; moreover, there are many overlapping characteristic peaks that make quick and intuitive manual identification difficult to achieve. Consequently, the combination of the SERS spectra of PAHs with lightweight networks is an effective and robust method for constructing recognition models.

The identification results of the SERS spectra of various PAH residues on fruit and vegetable surfaces using the model constructed with three lightweight networks (SqueezeNet, MobileNet, and ShuffleNet) are shown in Table 1. The results obtained from SqueezeNet, with *ACCT*, *ACCV*, and *ACCP* values of 99.57%, 93.22%, and 94.48%, respectively, were unsatisfactory. Based on the *Precision*, *Recall*, and *F*1-*score* of the prediction dataset, SqueezeNet is the best at identifying the mixed spectra of BaP + Pyr + Nap, primarily because of the distinct features of the mixed spectra. But the identification of the BaP + Pyr and Pyr + Nap spectra by this network is poor, indicating that the extracted features are not sufficiently rich, which is consistent with the confusion matrix predicted by the SqueezeNet model (Figure 6A). MobileNet performs better than SqueezeNet, with *ACCT*, *ACCV*, and *ACCP* values of 100%, 94.92%, and 96.06%. The *Precision*, *Recall*, and *F*1-*score* of BaP + Pyr are 100%, 93.94%, and 96.88%, respectively, which are considerably higher than those of SqueezeNet, indicating that MobileNet can effectively capture BaP + Pyr features. Detailed prediction results were obtained from the confusion matrix of the MobileNet model (Figure 6B). Unfortunately, the ability of the network to recognise BaP + Pyr + Nap is low. ShuffleNet achieves the best identification results, with *ACCT* = 100%, *ACCV* = 96.61%, and *ACCP* = 97.63%. These conclusions are also evident from the confusion matrix of the ShuffleNet model shown in Figure 6C, with the identification results of BaP + Pyr + Nap and BaP + Nap by the ShuffleNet model all being correct, and only three Pyr + Nap samples being misclassified as BaP + Pyr. The reason for this result can be found in spectra b and c in Figure 5B, the main characteristic peaks of the spectra of Pyr + Nap and BaP + Pyr being provided by Pyr, and the other weak characteristic peaks from BaP and Nap showing little difference. In general, the results indicate that a lightweight network combined with SERS

provides a fast, accurate, and intelligent method for identifying various PAH residues on fruit and vegetable surfaces.

**Table 1.** Identification results of SERS combined with a lightweight network for various PAH residues on fruit and vegetable surfaces.


Abbreviations: *ACC*, accuracy of correct classification; *ACCT*, *ACC* of the training dataset; *ACCV*, *ACC* of the validation dataset; *ACCP*, *ACC* of the prediction dataset.

**Figure 6.** Confusion matrix of (**A**) SqueezeNet, (**B**) MobileNet\_V1, and (**C**) ShuffleNet\_V1.
