Sustainable Smart Waste Sorting Through Reject-Aware IoT and Calibrated Vision: A Measurement-Calibrated Selective Routing Prototype

Smart waste systems can support sustainable recycling only when sensing, classification, and routing decisions are linked to transparent uncertainty management. This study aims to develop and evaluate a measurement-calibrated, risk-aware selective routing prototype for sustainable smart waste sorting under bench-scale conditions. The proposed prototype integrates an HC-SR04 ultrasonic sensing layer, Arduino Nano control, ESP8266 wireless synchronization, calibration-based fill level conversion, median filtering, exponential smoothing, hysteresis, and persistence-driven alert logic with a conveyor-based single-item vision-routing layer. The sorting layer routes items into glass, metal, plastic, trash, or a reject stream using a reject-aware VGG19 classifier with true logit temperature scaling and calibrated confidence gating. A quantitative prototype-level experimental evaluation was conducted using a five class benchmark of 2527 TrashNet images under a stratified 75/10/15 hold-out design, with cardboard and paper serving as outlier-exposure samples for the reject class. The final model achieved 88.13% test accuracy, 88.15% weighted F1, and 96.60% reject class F1. With a final policy tau = 0.70, the system routed 76.09% of routing class test items with 92.66% routed precision, while reject recall reached 98.66% and reject false acceptance was limited to 1.34%. Structured perturbation testing showed that severe visual degradation reduced automatic routing coverage but mostly shifted uncertain cases toward rejection. The contribution of this work is a transparent prototype-level mechanism that links calibrated sensing, calibrated visual confidence, selective rejection, and actuation-aware routing to support recycling stream purity under bounded uncertainty. Future work should validate the approach under multi-bin, multi-object, and field deployment conditions.