Image-Based Waste Classification Using CNN Method: A Comparative Study with Decision Tree, Random Forest, and SVM for Sustainable Waste Management

Abstract

The increasing volume of municipal waste has intensified the need for accurate and efficient automated waste-sorting technologies to support sustainable waste management. Although Convolutional Neural Networks (CNNs) have become the dominant approach in image classification due to their ability to learn feature representations automatically, their effectiveness under limited-data conditions remains insufficiently explored. This study investigates the performance of CNNs in comparison with conventional machine learning algorithms, namely Support Vector Machine (SVM), Random Forest (RF), and Decision Tree (DT), for image-based waste classification. The experiments were conducted using the Garbage Classification Dataset consisting of 4,133 images distributed across seven waste categories. The proposed framework involved image preprocessing, Histogram of Oriented Gradients (HOG) feature extraction for machine learning models, stratified data partitioning with a 70:20:10 ratio, model training, and evaluation using accuracy, precision, recall, and F1-score metrics. The results demonstrate that SVM achieved the highest accuracy of 67.15%, followed by Random Forest (65.70%), Decision Tree (39.61%), and CNN (23.67%). A notable finding of this study is that the CNN model, despite its superior theoretical capacity for automatic feature learning, produced the lowest classification performance among the evaluated approaches. This outcome suggests that training a CNN from scratch on a relatively limited dataset with considerable inter-class visual similarity is insufficient to learn highly discriminative feature representations. In contrast, HOG-based feature engineering provided more structured and stable visual descriptors, enabling conventional machine learning algorithms to achieve better generalization performance. These findings indicate that deep learning models do not necessarily outperform traditional machine learning approaches in all scenarios and that dataset characteristics play a critical role in determining model effectiveness. This study contributes empirical evidence that, in resource-constrained environments and limited-data settings, the combination of HOG and SVM can serve as a more accurate and computationally efficient alternative to CNN-based approaches for automated waste classification. The findings provide valuable insights for the development of practical intelligent waste-sorting systems that support sustainable waste management initiatives.