Bayesian optimization of super-resolution for crop disease detection

Abstract

This work proposes a Bayesian hyperparameter optimization framework for enhanced deep super-resolution networks (EDSR) to improve lesion-aware disease classification. A 4× degradation model was applied to PlantVillage images, and Optuna was used to optimize residual blocks, learning rate, feature channels, and patch size. Lesion-aware color gray-level co-occurrence matrix (Color GLCM) features were extracted and classified using extreme gradient boosting (XGBoost) under 5-fold cross-validation. Results show that degradation reduced classification accuracy to 58.15%, while baseline super resolution (SR) improved it to 74.26%. Bayesian-optimized SR further increased accuracy to 75.56% and reduced standard deviation from 0.74 to 0.30, indicating improved generalization stability. The proposed framework demonstrates that hyperparameter-optimized SR enhances lesion texture restoration and provides an efficient preprocessing strategy for precision agriculture applications.