Assessing PCA and band selection approaches in hyperspectral classification of construction waste

This paper presents a machine learning pipeline for sorting construction waste using hyperspectral imaging. It compares principal component analysis (PCA) and several feature and band selection methods, including mutual information (MI), Fisher's Score (FS), sequential forward selection (SFS), sequential backward selection (SBS), and categorical maximum spectral difference (CMSD). The study investigates the impact of dimension reduction methods on the classification process of construction materials using a convolutional neural network (CNN) based on VGG-19. Evaluation metrics include accuracy, precision, recall, and F1 score. The results show that targeted band selection approaches outperform transformation-based techniques and achieve accuracies of over 90%. While the imaging format enables spatially resolved acquisition across heterogeneous particle surfaces, this study intentionally focuses on spectral signatures to establish a robust spectral baseline for construction waste classification.