Hybrid MobileNetV2-ResNet50 with Random Forest Aggregation for On-Device Municipal Waste Classification and Capacitated Vehicle Routing Optimization

Abstract

Rapid urbanization in Busia Municipality, Uganda, has overwhelmed traditional waste management systems, leading to low collection coverage of only forty to fifty-five percent and widespread improper waste disposal practices. This paper presents a hybrid deep learning system for on-device municipal waste classification integrated with capacitated vehicle routing optimization. The proposed architecture combines MobileNetV2 and ResNet50 as parallel feature extractors, aggregates their outputs through a Random Forest meta-classifier, and achieves 99.78 percent classification accuracy across four waste categories: organic, recyclable, hazardous, and general. The model is deployed using TensorFlow Lite for edge inference with an average latency of 1.2 seconds per image, eliminating dependency on continuous internet connectivity. For collection logistics, a two-stage optimization framework implements a Cluster-First Route-Second heuristic with Haversine distance-based geographic clustering and nearest-neighbor sequencing to approximate solutions to the Capacitated Vehicle Routing Problem. Experimental results demonstrate a 22.1 percent reduction in total travel distance and 19.8 percent reduction in collection time compared to manual routing baselines. The complete system is implemented as a Flutter-based mobile application with Supabase backend, Firebase Cloud Messaging for real-time notifications, and Google Maps Directions API for live traffic integration. A comprehensive evaluation of 8,326 labeled waste images, training convergence analysis, ROC-AUC metrics, calibration curves, and ablation studies provides rigorous validation of the proposed approach. This work establishes that offline-first, edge-AI architectures can achieve state-of-the-art classification accuracy while simultaneously optimizing municipal logistics in low-resource settings.