Intelligent Waste Management Using WasteIQNet With Hierarchical Learning and Meta-Optimization

Effective waste management remains a critical pillar for sustainable urban development, particularly in rapidly growing regions like Delhi-NCR, where heterogeneous waste streams complicate classification. This study presents WasteIQNet, an intelligent, hierarchy-aware deep hybrid model for fine-grained waste classification across 18 categories structured under Wet and Dry waste types. The model integrates MobileNetV3 for semantic feature extraction with GraphSAGE to capture structural relationships among image representations. To address class imbalance and feature sparsity, the architecture incorporates Feature-wise Attention (FWA), Top-K Mixture of Experts (TopK-MoE), and advanced regularization techniques including Dynamic Sparse Training (DST) and Model-Agnostic Meta-Learning (MAML). We introduce a novel Hierarchical Tree Loss function to penalize semantically distant misclassifications by leveraging domain-specific waste hierarchy paths. The proposed framework is trained and evaluated on the WEDR dataset, a curated collection of 360,000 images reflecting real-world waste conditions across Delhi’s major landfill zones. WasteIQNet achieves a peak classification accuracy of 97.87%, demonstrating substantial improvements in both interpretability and generalization. This work contributes a scalable, robust, and deployment-ready solution tailored for real-time smart city waste segregation initiatives.