An automated waste classification system using deep learning techniques: toward efficient waste recycling and environmental sustainability

Nahiduzzaman, M ORCID: https://orcid.org/0000-0003-4126-0389, Ahamed, MF ORCID: https://orcid.org/0000-0002-7014-3205, Naznine, M ORCID: https://orcid.org/0009-0007-0296-9981, Karim, MJ ORCID: https://orcid.org/0009-0006-4226-3652, Kibria, HB, Ayari, MA, Khandakar, A, Ashraf, A, Ahsan, M ORCID: https://orcid.org/0000-0002-7300-506X and Haider, J ORCID: https://orcid.org/0000-0001-7010-8285 (2025) An automated waste classification system using deep learning techniques: toward efficient waste recycling and environmental sustainability. Knowledge-Based Systems, 310. 113028. ISSN 0950-7051

Preview

Published Version Available under License Creative Commons Attribution. Download (31MB) | Preview

Abstract

The growing negative effects caused by inadequate waste processing have led to the widespread implementation of waste classification systems. One effective approach is to develop an automated classification system that uses advanced waste recognition technology. This method can decrease the amount of manual labor required for waste separation and recycling activities. In the present study, a novel three-stage waste classification system was proposed. It incorporates the parallel lightweight depth-wise separable convolutional neural network (DP-CNN) in conjunction with the ensemble extreme learning machine (En-ELM) classifier. Waste items are first classified into two main categories: biodegradable and non-biodegradable. The dataset is then split into nine distinct categories in the second stage based on the overall waste characteristics. The final stage of the classification process involves a more detailed granularity, as all images are assigned to one of thirty-six specific classes. With an average accuracy, precision, recall, f1, and ROC-AUC values of 96 %, 95.0 ± 0.02 %, 95.0 ± 0.02 %, 95.0 ± 0.02 %, and 98.77 %, respectively, the proposed model demonstrates promising results in the first stage of two-class classification. Advancing to the second stage, the framework showed excellent results in nine-class classification, with performance rates of 91.00 %, 90.0 ± 0.04 %, 89.44 ± 0.06 %, 89.66 ± 0.05 %, and 98.57 %, respectively. Similar to the previous stages, the model continued to perform effectively in the third stage, achieving 85.25 % accuracy, 85.02 % precision, 85.25 % recall, 84.54 % f1-score, and 98.68 % AUC in the thirty-six-class classification. The En-ELM classifier, a fusion of pseudoinverse ELM (PI-ELM) and L1 regularized ELM (L1-RELM), achieved impressive results with an average testing time of 0.00001 s. A novel comprehensive dataset titled TriCascade WasteImage, which combines four smaller preexisting datasets, was used to measure the performance of the DP-CNN-En-ELM model. With only nine layers and 1.09 million parameters, the proposed approach precisely extracts pertinent information from images to classify diverse waste materials. The effectiveness of the model is confirmed by comparing it to advanced transfer learning methods. Various explainable AI (XAI) methods are used to explore the interpretability and decision-making capability of the proposed model. Additionally, this study presents a comprehensive prototype hardware architecture for rapid waste categorization in an augmented environment, enabling autonomous waste sorting in industrial applications.