<![CDATA[Waste management in Indonesia remains a major challenge due to increasing waste volumes and the low efficiency of manual sorting processes at Landfills (TPA). This study aims to improve the performance of an automated waste classification system for three categories: organic, inorganic, and hazardous and toxic waste (B3) using deep learning-based computer vision technology. The proposed method is the EfficientNetB0 architecture with a transfer learning approach, whose performance is compared with four other pre-trained architectures (VGG-16, InceptionV3, MobileNetV2, and ResNet50). The dataset used consists of 7,003 valid images collected from public sources and manual acquisition after a data cleaning process. The dataset is divided into 70% as training data, 20% as validation data, and 10% as test data. Data augmentation and class balancing strategies are used to increase variation and overcome data imbalance between classes. Training is conducted in two stages: Feature Extraction and Fine-Tuning, with consistent hyperparameters for a fair comparison. Performance evaluation is performed using accuracy, precision, recall, and f1-score metrics. The test results show that EfficientNetB0 managed to achieve the best performance with an accuracy rate of 96.87%. Modern architectures like EfficientNetB0 have proven capable of extracting complex features with good computational efficiency, thereby holding the potential for use in AI-based automatic waste sorting systems to support more effective and sustainable waste management.]]>
