Advanced Hyperspectral Signature Processing for Chemical Stress Detection in Vegetable Leaves Using Hierarchical Feature Extraction and Enhanced Ensemble Model

The increasing concern about the presence of pesticides in vegetable leaves has underscored an urgent need for real-time, nondestructive, and accurate detection methods. Traditional methods are reliable but laboratory-based, costly, and unsuitable for field monitoring. In this study, we propose an efficient learning model pipeline that uses hyperspectral reflectance signatures to detect pesticide residue in plant leaves. We extract a comprehensive set of 39 domain-specific features based on vegetation indices, red-edge metrics, spectral statistics, and derivative profiles. To enhance the performance, use a multilayer perceptron to extract more features. A feature fusion module is used to combine both domain-specific features and features extracted by a multilayer perceptron. Further refinement is achieved through a feed-forward attention scoring module that dynamically weights important features. The efficiency of the system is evaluated using an enhanced extra trees classifier, which shows superior classification performance and stability across different feature formats. With cross-validation, our model achieves an accuracy of 94.69%, significantly outperforming conventional classifiers such as convolutional neural networks, support vector machines, and ensemble models such as random forest and extra trees. This framework not only improves interpretability and performance but also provides a foundation for a real-time, on-site pesticide monitoring solution.