Akademik Veri Yönetim Sistemi
EXPERT SYSTEMS WITH APPLICATIONS, cilt.283, 2025 (SCI-Expanded)
This study investigated the classification of banana samples from 33 distinct environments using hyperspectral reflectance data and various machine-learning algorithms. Feature selection was performed using Recursive Feature Elimination, reducing the dataset from 750 to 50 wavelengths. The data was partitioned into training, testing, and validation sets using stratified sampling. Traditional machine learning algorithms, including KNN, CART, SVM, Naive Bayes, LDA, QDA, MLP, Ridge, Bagging, Random Forest, AdaBoost, LightGBM, and XGBoost, were evaluated and optimized using Optuna. Deep learning architectures, such as CNN, CNN-LSTM, ConvLSTM, GRU, LSTM, and BiLSTM, were also assessed and tuned using Bayesian optimization and early stopping. Performance metrics included accuracy, training and testing times, and confusion matrices. Among the traditional models, SVM (C = 1000, polynomial kernel) and MLP (hidden_layer_sizes=(100, 50), identity activation, lbfgs solver) achieved the highest test accuracies of 96.07 % and 92.13 %, respectively. XGBoost and Bagging also performed well, with test accuracies of around 82 %. For the deep learning models, GRU (Adamax optimizer, learning rate = 0.01, 100 epochs) outperformed the others with 85.32 % test accuracy, followed by CNN at 81.39 %. LSTM and BiLSTM achieved test accuracies of 76.70 % and 67.78 %, respectively, whereas CNN-LSTM and ConvLSTM struggled at 56.28 % and 43.87 %. This study highlights the trade-off between model complexity and computational efficiency, with GRU and CNN offering a good balance for real-time applications. These findings provide valuable insights for selecting suitable models for hyperspectral data classification in agricultural and environmental monitoring.