Universitas Dian Nuswantoro conducted a study applying machine learning algorithms to classify waste images gathered from Temporary Disposal Sites in Bogor City. The activity involved Ritzkal, R.; Muljono, M.; Andono, Pulung Nurtantio; and Santoso, Heru Agus as participants. The study focused on identifying the status of waste containers through image preprocessing, feature extraction, and classification methods.
The preprocessing phase included resizing, noise reduction using Non-Local Means, histogram equalization, and color normalization in the YUV color space. These procedures contributed to improved image clarity and contrast. Feature extraction integrated pixel, statistical, histogram, HSV, and edge-based descriptors. The resulting high-dimensional feature set captured color and structural characteristics of waste objects.
The study used 750 balanced images classified into empty, filled, and full categories to train and evaluate six conventional classification algorithms: K-Nearest Neighbors, Logistic Regression, Random Forest, Naive Bayes, Stochastic Gradient Descent, and Decision Tree. Performance evaluation used accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix analysis. Grid-search hyperparameter tuning with five-fold cross-validation was applied to support model robustness.
The experimental findings showed that K-Nearest Neighbors achieved the highest performance, with an average accuracy of 97.6%, an AUC of 0.991, and an F1-score of 0.976. Random Forest and Logistic Regression also produced commendable results. Naive Bayes, Stochastic Gradient Descent, and Decision Tree demonstrated reduced accuracy, associated in the source with assumptions regarding feature independence, non-linearity, and tendencies toward overfitting. The findings describe a methodical approach to waste classification and report the model's potential for real-time application in smart waste management systems, supporting sustainable urban waste handling in the context of Bogor City.