Surface‐Enhanced Raman Scattering Imaging Assisted by Machine Learning Analysis: Unveiling Pesticide Molecule Permeation in Crop Tissues

Abstract Surface‐enhanced Raman scattering (SERS) imaging technology faces significant technical bottlenecks in ensuring balanced spatial resolution, preventing image bias induced by substrate heterogeneity, accurate quantitative analysis, and substrate preparation that enhances Raman signal strength on a global scale. To systematically solve these problems, artificial intelligence techniques are applied to analyze the signals of pesticides based on 3D and dynamic SERS imaging. Utilizing perovskite/silver nanoparticles composites (CaTiO3/Ag@BONPs) as enhanced substrates, enabling it not only to cleanse pesticide residues from the surface to pulp of fruits and vegetables, but also to investigate the penetration dynamics of an array of pesticides (chlorpyrifos, thiabendazole, thiram, and acetamiprid). The findings challenge existing paradigms, unveiling a previously unnoticed weakening process during pesticide invasion and revealing the surprising permeability of non‐systemic pesticides. Of particular note is easy to overlook that the combined application of pesticides can inadvertently intensify their invasive capacity due to pesticide interactions. The innovative study delves into the realm of pesticide penetration, propelling a paradigm shift in the understanding of food safety. Meanwhile, this strategy provides strong support for the cutting‐edge application of SERS imaging technology and also brings valuable reference and enlightenment for researchers in related fields.


Text S1: Calculation of the Enhancement Factor (EF) for thiram
The enhancement factor (EF) of thiram can be calculated as [1]: ISERS was the peak intensity of the 1380 cm −1 peak in the SERS spectrum and IRaman was the peak intensity of the same vibrational mode.
Thiram with a mass of 0.0005 g were taken for Raman test.2.5 μL pesticide were added to the 5 μL Ag@BONPs, and were absorbed by a capillary with an inner diameter of 0.5 mm.SERS detection was performed under a 10-fold mirror, and the laser diameter was 1545 nm.
Both of these spectra were recorded using a 633 nm laser under the same conditions and the exposure time was set to 3 s with a one-time accumulation.

Text S2: FDTD principle and simulation process
According to the previous DLS result, the diameter of Ag@BONPs was set to be 58 nm, the diameter of CaTiO3/Ag@BONPs was set to be 110 nm and the nanogap between adjacent nanoparticles was set to be 2 nm.The surrounding medium was set to be 1.0 (nmedium = 1.0).Simulated result was shown in Figure 2B, it could be seen that there are hot spots between silver nanoparticles, which was shown with red arrow, and the maximum electric field intensity of the sample was about 36.7 V/m.
According to the EFEM calculation formula, Simulated result was shown in Figure 2B, it could be seen that there are hot spots, which were shown with red arrow.For DAg= 110 nm, EFEM = 127 4 ≈ 2.6 × 10 8 .From the calculated results of EFEM, it could be found that the electromagnetic field enhancement effects of Ag@BONPs and CaTiO3/Ag@BONPs are extremely obvious, which is consistent with the experimental trend.

Figure S11. Comparison of the current method with conventional methods in imaging. Raman imaging of thiram in pulp (A) and SERS imaging using Ag@BONPs (B) and
CaTiO3/Ag@BONPs (C) substrates.

S6. Machine learning Text S3: Detailed process of machine learning
Firstly, the Raman spectrum data of pesticides were read, and the dimensionality of the data was reduced according to the distance of 10 sites before and after the characteristic peak interval of the pesticide.In addition, in order to extract the features of pesticides more deeply, the "procomp" function in R language was used for principal component analysis (PCA), and the new features were combined to describe the changes of feature peak position and peak intensity.The two principal components with the largest cumulative variance contribution rate were PC1 and PC2.The R language "ggplot" function was used to visualize and compare the Raman spectral data of multiple groups of samples using the eigenvector of the covariance matrix.The Raman spectra are projected to a fraction plot in proportion to the load, the PCA 2D fraction plot, juxtaposed with a 95% confidence ellipse.In order to effectively distinguish the component information in the spectral image, the "fviz_nbclust" function of R language was used to view the best k value of the best K-means clustering, and then the "kmeans" function was called to complete the clustering.From the degree of boundary between pesticides in the visualization results, the clustering module had good intra-group similarity and inter-group difference.In order to measure which module is most likely to be the pesticide part, the "pure pesticide" data is used to evaluate the similarity of all channel results in each K-means clustering module in the image.The evaluation method is to call the similarity measure based on Euclidean distance of the R language "matchFactor" function.Thus, the data as pesticide group and non-pesticide control group were selected.The obtained data were divided into training set and test set, and then the R language "svm" function was used to fit the SVM model based on the training data set, and the performance of the model was verified on the test data set, that is, the confusion matrix and ROC curve were output.

Figure
Figure S6. A. Scanning electron microscope (SEM) and Energy Dispersive Spectrometer (EDS) mapping of the CaTiO3/Ag@BONPs in the sprayed state.The respective distributions of silver (B), calcium (C), and titanium (D) elements.

Figure S7 .
Figure S7.The chlorophyll content in cucumber after spraying with CaTiO3/Ag@BONPs. A. Standard curves for UV absorption of chlorophyll.B. Comparison of chlorophyll concentrations.The data are mean ± SD from three individual experiments.
where E0 is the incident electric field intensity, E0 = 1 V/m, and Eout is the electric field intensity of the position of hot spots caused by the incident light, E'out refers to the field evaluated at the scattering frequency.The local electric field intensity of SERS (|Eout| 2 × |E'out| 2 ) is approximately equal to the surface localized electric field intensity of nanoparticles (|Eout| 4 ).Therefore, the above formula can be changed into   = |  | 4 | 0 | 4 .For DAg= 58 nm, EFEM = 36.7 4 ≈ 1.8 × 10 6 .

Figure S12 .
Figure S12.Specific steps of elution. A. The process of water and CaTiO3/Ag@BONPs elution from the pericarp.B. The process of water and CaTiO3/Ag@BONPs elution from the pulp.C. Processes before and after CaTiO3/Ag@BONPs elution from the pericarp.D. Processes before and after CaTiO3/Ag@BONPs elution from the pulp.

Figure S14 .
Figure S14.Representative SERS spectra of acetamiprid and chlorpyrifos in fruits and vegetables.

Figure S16 .
Figure S16.ROC curves between different fruits and vegetables.

Figure S18 .
Figure S18.Baseline correction for SERS spectra in plants.A. SERS spectra with plant fluorescence.B. Comparison of SERS spectra before and after correction.