Graph neural network-based hyperspectral analysis for rapid coal ash content estimation

Abstract Ash content is a critical coal quality indicator that determines coal grade and commercial applications, making its rapid and accurate measurement essential. Traditional slow ashing methods are operationally cumbersome, time-consuming, and pose environmental and safety concerns. Hyperspectral imaging technology offers a promising alternative with its non-contact detection and superior spatial-spectral information fusion capabilities. However, existing machine learning and deep learning approaches, while capable of modeling complex spectral-analyte relationships, often overlook spatial information and struggle to effectively capture the heterogeneous distribution characteristics of ash minerals in coal samples. Therefore, this study proposes a fast estimation method of hyperspectral coal ash based on graph neural network (GNN), using a hybrid graph convolutional network-graph attention networks encoder and an adaptive spatial-spectral edge construction strategy. This method models the mineral distribution pattern by constructing a spatial graph, constructs a spectrogram to capture the mineral spectral characteristics, and optimizes it for different ash content ranges. In the experiments of concentrated coal (ash range from 9% to 18%) and tailings coal samples (ash range from 49% to 77 %), the average absolute errors of this method are 0.862% and 1.18 %, respectively, which are significantly better than traditional machine learning methods and deep learning methods. The interpretability analysis shows that the absorption characteristics of hydroxyl-containing minerals corresponding to the 1200–1400 nm band contribute the most to the ash estimation, which verifies the effectiveness of the GNN in coal quality analysis and provides practical guidance for coal automatic characterization and industrial quality control.