From Topological Signatures to Textural Patterns: Enhancing the Precision of Sand Classification

Traceable sourcing of sand is a significant challenge in the global construction industry. Fast, accurate and inexpensive methods to classify sand by its original mining source is a critical component of more sustainable and equitable supply networks. This study explored various machine learning and deep learning methods to enhance classification accuracy. While promising, initial attempts with Convolutional Neural Networks (CNNs) and Siamese Networks did not yield satisfactory results, mainly due to overfitting, as evidenced by high training accuracies but lower validation accuracies. A notable improvement was achieved with transfer learning using the VGG-19 model, reaching a validation accuracy of 81.6%. However, the most effective results were obtained through traditional machine learning models combined with manual feature extraction. Applying XG Boost with manual feature extraction significantly enhanced classification accuracy, achieving 86.18%. Following manual feature extraction, a simple Multi-Layer Perceptron (MLP) model with feature standardization resulted in a high validation accuracy of 90.45%. A modified approach, excluding three poorly performing classes, further improved the validation accuracy to 98.18% in a 5-class configuration. These findings highlight the effectiveness of blending advanced machine learning techniques with targeted manual feature extraction for sand sample classification, offering substantial potential for practical applications in various sectors.