Optimizing Construction and Demolition Waste Sorting for Sustainable Environmental Management: A Reproducible Deep Learning Framework with Transparent Data Augmentation

Abstract

Considering the increasing amount of global waste, improvements in environmental management systems are becoming progressively more important, particularly in reducing pollution and optimizing resource recovery. Construction and Demolition Waste (CDW) constitutes a substantial portion of global waste, indicating the potential need for innovative environmental management strategies to improve existing practices. Deep learning addresses this environmental management need by enabling algorithms to perform object detection tasks using image inputs. Many studies use deep learning algorithms to analyse CDW datasets using data augmentation policies, which are mostly abstract and hardcoded. Therefore, comprehensive testing of augmentation policies for the reproducibility and progress in CDW detection is necessary. This study bridges the gap by systematically evaluating 80 augmentation settings across five state-of-the-art (SOTA) deep learning models for CDW detection. Through 1600 experiments, we observed that geometric, noise, and composition-based augmentation methods improve mean Average Precision (mAP) by almost 10%, while colour-based augmentation policies show even adverse impact. Optimized augmentation policies applied on a benchmark CDW dataset and SOTA detectors could achieve a plausible mAP of almost 90% with inference time of only 5.7miliseconds per image. Our work demonstrates transparent augmentation settings to guide future deep learning research in CDW identification, enabling direct comparisons and reducing redundant experimentation. These findings contribute to scalable environmental management solutions by providing a reproducible deep learning with transparent data augmentation policies.

Keywords: Environmental Management, Deep Learning Models, You Only Look Once (YOLO), Real-Time Detection Transformer, Construction and Demolition Waste (CDW), Data Augmentation