3D-sersnet based pesticide residue detection in okra using hyperspectral imaging technology

Modern agricultural methods are leading to a noticeable increase in pesticide residues in vegetables, which can be harmful to health. To address this issue, this work investigates a hyperspectral imaging (HSI)-based deep learning approach for classifying pesticide residue levels in okra. A dedicated HSI dataset can be created using the Resonon Pika-L camera, covering the spectral range of 400–1000 nm with 300 bands, for okra samples treated with pesticide residues at varying concentrations of 0.5 ml, 3 ml, and 6 ml per liter of water, representing low, medium, and high levels, respectively, along with organic okra samples. To effectively process the rich spectral-spatial information, a novel deep learning framework, a 3D Squeeze Excitation-Residual Spectral Network (3D-SERSNET), is proposed. This model accepts raw 3D hyperspectral cubes as input and integrates 3D convolutional layers with Squeeze-and-Excitation (SE) blocks to capture inter-channel dependencies and enhance spectral feature learning. This study performs both binary classification (presence or absence of pesticide) and multi-class classification (pure, low, medium, and high). For comparison, the machine learning (ML) algorithms are evaluated, and the results show that the K-nearest neighbor achieves the highest accuracy of 79.3% for the four-class classification, while the Decision Tree classifier reaches 84.6% for the two-class classification. Finally, the proposed 3D deep model records an accuracy of 98.07%, whereas the model without the SE layer achieves 77.83%. These findings show that the proposed model outperforms ML algorithms in detecting pesticide residues in okra and is adaptable for future integration with portable HSI systems.