Model-agnostics for real waste management using image classification

Automated waste segregation has emerged as a critical requirement for sustainable and safe urban waste management. While deep learning–based image classification models have demonstrated promising accuracy in waste identification, their black-box nature limits trust and reliability in real-world deployments. This paper presents a systematic evaluation study that combines the EfficientNet-B0 architecture with Local Interpretable Model-Agnostic Explanations (LIME) to investigate not only whether a waste classifier achieves high accuracy, but also whether its correct predictions are driven by semantically meaningful visual features rather than spurious background correlations. The proposed approach was initially benchmarked on the standard CIFAR-10 and CIFAR-100 datasets to establish baseline robustness, before being applied to the RealWaste dataset. The EfficientNet-based model achieves classification accuracies up to 93.08% on RealWaste, and across five random seeds, it achieves the highest mean test accuracy (91.30% ± 1.11%, vs. 90.04% ± 1.22% for ResNet50, 89.33% ± 1.13% for InceptionV3, and 86.02% ± 3.12% for MobileNetV2) while using fewer parameters than all models, except MobileNetV2. Beyond accuracy, LIME-based explanations analyze the validity of model decisions, revealing both meaningful feature utilization and instances of spurious visual correlations, supported by deletion/insertion AOPC scores and stability analysis. The study highlights the importance of combining predictive performance with explanation-driven evaluation and demonstrates the potential for explainable deep learning for future research on automated waste management. All results are based on controlled experimental conditions; scalability and real-world generalization require further validation.