ResFusionNet-TSMT: A residual network for pesticide detection using surface-enhanced Raman spectroscopy

Surface-Enhanced Raman Scattering (SERS) has emerged as a promising tool for rapid pesticide detection. However, its accuracy is often hindered by spectral interference and signal variability. To address this, we propose ResFusionNet-TSMT, a novel deep learning framework for simultaneous pesticide classification and concentration quantification. The model uniquely integrates the local feature extraction capability of Residual Networks (ResNet) with the global dependency modeling of Transformers. It incorporates a dual-stream architecture for processing raw and multi-scale spectral inputs, an attention pooling mechanism to focus on discriminative peaks, and a Transformer encoder for robust feature fusion. Furthermore, a novel class attention mechanism optimizes joint learning between classification and regression tasks, significantly improving accuracy especially for low-concentration samples. The proposed model achieved a classification accuracy of 99.21 % (F1 = 99.08 %) and a mean absolute error of 0.4640 ppm (R² = 0.8764) for quantification, outperforming both traditional machine learning and state-of-the-art deep learning approaches. Ablation studies and visualizations validate the model's effectiveness and its ability to capture characteristic spectral features, demonstrating its strong potential for application in agricultural safety monitoring.