Classification of translucent flesh defects in phulae pineapples using stacking ensemble classifiers and deep neural networks

• Developed a dataset specifically for Phulae pineapples, including samples with and without translucent flesh defects • Developed a robust framework for classifying Phulae pineapples into "normal" and "translucent flesh defect" categories • Evaluated and compared conventional classification, stacking machine learning, and deep learning methodologies • Stacking machine learning achieved superior performance with a test accuracy of 93.06%, sensitivity of 96.30%, and specificity of 97.78% Phulae pineapples, a valuable agricultural product from Thailand’s Chiang Rai Province, have significant economic importance but face quality challenges due to “translucent flesh defects” which reduce their marketability and shelf life. This study classified Phulae pineapples into “normal” and “translucent flesh defect” categories using three methodologies: conventional classification, stacking machine learning, and deep learning architectures. Conventional classification paired manually engineered features with classifiers, providing moderate baseline performance. Stacking machine learning outperformed the other methods, achieving a test accuracy of 93.06%, sensitivity of 96.30%, and specificity of 97.78% by integrating outputs from 144 baseline models into a meta-classifier. Deep learning architectures demonstrated potential with automated feature extraction but encountered overfitting and computational challenges due to the limited size of the dataset. These findings highlighted the superior balance of accuracy and generalization of stacking machine learning as the most suitable solution for addressing translucent flesh defects, improving quality assessment, and reducing postharvest losses in Phulae pineapples.