Enhancing Insecticide Classification Accuracy with Modified ChemNet Architecture

Proper detection of insecticides is also vital since it is the determinant of agricultural safety, efficiency, environmental protection and environmental compliance. In this paper, a modified ChemNet architecture is introduced, which is modified to the classification of insecticide in fruits. Compared to the traditional ChemNet, which leaves out chemical priors and residual learning, the proposed model has certain improvements: Squeeze-and-Excitation (SE) blocks for adaptive channel recalibration, residual blocks coupled with SE units to improve feature extraction, Self-Attention mechanism to learn long-range dependencies, and the swish activation function to support gradient flow and non-linear representation. Also, the network applies progressive dropout, early stopping, and class-weighted loss to address overfitting and unbalanced samples. The validation and training of the proposed model were trained using the Kaggle Banana Insecticide Dataset comprising a total of 6,103 images. The data is grouped into six distinct categories: monohigh (high level of banana treated with mono-type insecticide), monolow (low level of mono-type insecticide), novahigh (high level of nova-type insecticide), novalow (low level of nova-type insecticide), natural (banana samples that were never treated) and rotten (samples of banana that were biologically spoiled). This composition ensures that the dataset encompasses a wide range of levels of insecticide contamination and natural spoilage, enabling a strong foundation for training and evaluation. This will ensure the quality of the model in the classification of the data. The experimental findings show that the modified ChemNet achieves an overall classification accuracy of 81.02% and demonstrates good generalization across the chemically heterogeneous classes. These outcomes indicate the effectiveness of the proposed modifications to transform ChemNet for image-based insecticide detection in agricultural images and suggest its potential application as a tool for food safety monitoring that can be expanded.