Cadmium contamination significantly impacts the physiological growth, yield, and quality of rice, posing a considerable food safety risk. Therefore, rapid and accurate detection of cadmium stress in rice is crucial for agricultural management and food safety. However, traditional methods for acquiring plant stress-related physiological information are often cumbersome, time-consuming, and unsuitable for timely and effective crop stress mitigation, remediation, and precision management. This study aims to establish a rapid, efficient, and accurate method for detecting cadmium stress levels in rice, enabling simple and fast discrimination of the stress severity. We prepared arginine-modified flower-like silver nanoparticles (Ag NPs-Arg), which effectively enhanced the Raman signals of cadmium stress response molecules in rice, namely salicylic acid (SA), malondialdehyde (MDA), and peroxidase (POD). Subsequently, quantitative prediction models (PLSR, SVM, CNN-Transformer) for these three molecules and a discrimination model for cadmium stress levels were developed. The results indicated that all three quantitative models could effectively predict the contents of the target molecules. The CNN-Transformer model demonstrated superior prediction accuracy, with determination coefficients (Rp2) of 0.889 for SA, 0.832 for MDA, and 0.802 for POD. An objective weighting method was employed to integrate the SA, MDA, and POD indicators for a comprehensive assessment of cadmium stress levels. The resulting stress level discrimination model achieved an accuracy of 95.83%. This approach more comprehensively and robustly captures the characteristic physiological and biochemical changes induced by cadmium stress and is suitable for the real-time monitoring of cadmium stress levels in rice.