Optimizing Waste Classification Using Transfer Learning and One Cycle Policy with Grad-CAM-Based Explainability

As the global population continues to grow, so does the volume of waste generated. Efficient waste management requires the separation of recyclable and organic materials from non-biodegradable waste. This paper presents a machine learning approach for classifying garbage images into two categories: Organic and Recyclable. Transfer learning was applied to finetune the pre-trained models-ResNet-50, DenseNet-121 and VGG16 using a dataset of 22,564 training images and 2,513 test images. A One Cycle Policy was adopted to optimize the learning rate, which accelerated convergence and improved generalization by reducing overfitting. The contribution of this work lies in integrating Grad-CAM-based visual explainability with One Cycle Policy optimization in the context of waste classification, providing a transparent and efficient pipeline. The models were evaluated using accuracy, precision, recall, F1 score and AUC-ROC. Among the models, ResNet-50 achieved the highest performance with 99.56 % accuracy. Grad-CAM was further employed to visualize the models' decision-making processes, thereby addressing the black-box nature of deep learning and enhancing interpretability by highlighting the most relevant regions influencing predictions.