Combination of gray level features with deep transfer learning for copra classification using machine learning and neural networks

Copra (dried coconut) is used for oil production and raw materials for its by-products. Traditionally, Coconuts are halved and sun-dried in the field. Fumigation using sulphur is employed in the industry to maintain its colour and prevent microbial growth from inhibiting it. The proposed study aims to classify the sulphur-fumigated copra and normally dried copra to benefit the buyers. Images of copra were collected from various drying industries and segmented to exclude irrelevant parts. A novel approach is introduced by combining GLCM (Gray-Level Co-Occurrence Matrix) features with features extracted from four transfer learning models. These concatenated features were evaluated using various machine learning classifiers and neural networks. Among the classifiers tested, Neural Network-based Pattern Recognition (NNPR) achieved the highest accuracy of 99.6%, sensitivity of 99.64%, specificity of 99.64%, F1-Score of 99.6, and a Kappa score of 0.99, demonstrating its superior performance. Other classifiers, such as Logistic Regression (98.3% accuracy, 0.96 Kappa), Kk-Nearest Neighbour (KNN) (98.3% accuracy, 0.96 Kappa), and Random Forest (98.9% accuracy, 0.97 Kappa), also performed well but slightly lower than the neural network. This methodology outperforms existing literature and provides a robust solution for accurately classifying sulphur-fumigated copra, ensuring its practical utility for farmers and buyers in the copra industry.