To quickly and non-destructively monitor the degree of cadmium contamination in lettuce, visible-near infrared spectroscopy is used to classify cadmium contamination in lettuce. Lettuce leaves’ visible-near infrared reflectance spectra were collected to analyze the variation in reflectance spectra under different cadmium pollution levels (0, 5, 10, 20 mg·kg^-1) in soil, with lettuce as the research subject. The spectral analysis reveals that within the wavelength range of 510 to 730 nm, the reflectance of lettuce leaves in the visible-near infrared spectrum decreases and then increases with the increase in cadmium content in the soil. Within the wavelength range of 730 to 799.53 nm, the reflectance of lettuce leaves under 5 and 20 mg·kg^-1 cadmium stress is higher than the CK, while under 10 mg·kg^-1 cadmium stress, the reflectance is lower than the control group. Additionally, an absorption valley was observed at 762.199 nm. In establishing a cadmium pollution monitoring model for lettuce, various preprocessing methods were applied to the raw spectra to improve the signal-to-noise ratio. These methods include smoothing (SG), multiplicative scatter correction (MSC), standard normal variate (SNV), mean normalization (MN), SG+MSC, SG+SNV, SG+MN, SG+first derivative (FD), and SG+second derivative (SD). Based on the principal component analysis (PCA), dimensionality reduction was performed on the original spectra and spectra subjected to various preprocessing methods. Subsequently, the reduced data was divided into training and testing sets in a 4∶1 ratio. These sets were then used to establish classification monitoring models for cadmium pollution in lettuce by combining particle swarm optimization-random forest (PSO-RF), genetic algorithm-optimized support vector machine (GA-SVM), backpropagation neural network (BP-NN), extreme learning machine (ELM), and Naive Bayes, followed by analysis and comparison. The results demonstrate that among the different models, the PSO-RF (SG) model achieves the best recognition performance, followed by the GA-SVM (SG+FD) model and the ELM (MSC) model. The training accuracy of the PSO-RF (SG), GA-SVM (SG+FD), and ELM (MSC) models is 100%, while their testing accuracies are 100%, 83.33%, and 79.17% respectively. On the other hand, the BP-NN model and the Naive Bayes model perform relatively poorly. The training accuracy of the BP-NN (SNV) model is 42.72% with a testing accuracy of 50%. The Naive Bayes (SG+FD) model achieves a training accuracy of 71.84% and a testing accuracy of 83.33%. It indicates that applying visible-near infrared spectroscopy combined with particle swarm optimization random forest modeling can provide a novel approach for studying the monitoring of heavy metal contamination in lettuce.