Machine-learning-assisted classification of construction and demolition waste fragments using computer vision: Convolution versus extraction of selected features

Improper sorting of construction and demolition waste (CDW) leads to significant environmental and economic implications, including inefficient resource use and missed recycling opportunities. To address this, we developed a machine-learning-assisted procedure for recognizing CDW fragments using an RGB camera. Our approach uniquely leverages selected feature extraction, enhancing classification speed and accuracy. We employed three classifiers: convolutional neural network (CNN), gradient boosting (GB) decision trees, and multi-layer perception (MLP). Notably, our method’s extraction of selected features for GB and MLP outperformed the traditional CNN in terms of speed and accuracy, especially for challenging samples with similar textures. Specifically, while convolution resulted in an overall accuracy of 85.9%, our innovative feature extraction approach yielded accuracies up to 92.3%. This study’s findings have significant implications for the future of CDW management, offering a pathway for efficient and accurate waste sorting, fostering sustainable resource use, and reducing the environmental impact of CDW disposal. Supplementary materials, including datasets, codes, and models, are provided, promoting transparency and reproducibility. •Classifiers were trained to recognize construction and demolition waste (CDW).•CDW fragments were recognized from RGB images.•Features were extracted for GB and MLP models; CNN employed convolution.•GB and MLP outperformed CNN in terms of speed and accuracy.•GB: 92.3% overall accuracy, MLP: 91.3%, and CNN: 85.9%.