Hybrid Explainable Deep Learning Framework for IoT-Enabled Smart Waste Classification Using Transfer Learning and Frequency Analysis

Abstract

Rapid urbanization and increase in the generation of municipal waste have necessitated the need for intelligent waste classification for effective recycling and sustainable resource management. Conventional waste classification schemes encounter difficulties due to environmental variations, background interference, illumination variations, similarity in appearances of different types of wastes, and manual separation. In this work, a Hybrid Explainable Artificial Intelligence Framework for IoT-Enabled Smart Waste Classification is presented through transfer learning via Spatial and Frequency-Domain Feature Learning. A custom Smart Waste Classification Dataset (SWCD-2026) of 20,000 images in balanced numbers for plastic, paper, glass, metal, and organic wastes was created in diverse environmental conditions. Normalization, augmentation, and frequency domain transformation of images improved the features representation. Hybrid feature fusion used the complementary spatial and frequency domain features, while explainable visualization helped to discover discriminative image regions for classification decision-making. An experimental evaluation was carried out to validate the effectiveness of the approach and achieved 98.767% accuracy, 98.541% precision, 98.924% recall, 98.732% F1-score, and 99.184% ROC-AUC values.