AI-ENABLED URBAN WASTE SEGREGATION: A FIELD EVALUATION OF A LOW-COST SMART MONITORING SYSTEM

Abstract

The ability to effectively segregate urban waste is still limited due to the sub-optimal practices of disposal and the inability to afford monitoring systems that are intelligent. This paper designed and tested LiteWasteNet, a small convolutional neural network to classify organic and recyclable waste in a low-cost smart monitoring system. The empirical study was performed with 698 labeled RGB images, of which 398 are organic and 300 are recyclable, grouped in 74 source-image families to avoid data leakage due to augmentation. The performance of the model was evaluated using 5-fold stratified group cross validation and was compared with a baseline model, a histogram of oriented gradients support vector machine (HOG SVM). The results of LiteWasteNet is 82.95% for the accuracy, 82.69% for the macro f1 and 0.8958 for the receiver operating characteristic area under the curve which are all better than the baseline accuracy of 71.20%. The model had 15,114 trainable parameters, memory footprint of ~84KB and a median CPU inference latency of 1.54ms. The results show the limitations of computational resources in achieving reliable binary waste classification with limited validation controls for source dependence. The study offers an effective evaluation strategy for small, augmentation driven waste sets that can be replicated, and an efficient classification component for edge based monitoring. The model was found to perform well in the field, even though it is a variable model, with reduced values of accuracy when there is contamination or partial occlusion in the items.