SCANDY - A handheld Imaging System for Real Time Spoil Categorisation

This work builds on project C29048, which demonstrated the feasibility of using drone images and machine learning for spoil characterisation. While effective for Paddock Dumping, the drone-based approach faced operational constraints for Edge Dumping. This limitation prompted the development of the SCANDY project, which introduces a close-range, image-based system designed for practical deployment and automated spoil classification.

To ensure the effectiveness of SCANDY's classification framework, two distinct strategies were implemented and compared during model development:

• Image-level classification that treats each image captured of a pile independently, providing a baseline for feature learning and spoil characterisation. However, this method is limited by partial coverage of the pile, as each image represents only a portion of the overall pile material.
• Pile-level classification that aggregates predictions from multiple images of the same spoil pile. This method captures spatial heterogeneity and improves robustness by using multiple viewpoints. It forms the core of the SCANDY system. The pile-level approach demonstrated superior performance across datasets. On the QLD1 mine dataset, the model achieved an average accuracy of 95.1% and a weighted F1- score of 95.6%, with stable per-class performance despite class imbalance. On the QLD2 mine dataset, it attained an accuracy of 87.7% and a weighted F1-score of 86.1%, confirming its reliability across geologically distinct sites. However, minority categories such as Cat 1 and Cat 4 were either underrepresented or absent, limiting prediction coverage and highlighting the need for more balanced datasets.

For on-site deployment, the system is integrated into MineScanMateTM, a prototype mobile application and web interface designed for field operations. MineScanMateTM enables GPS-linked image capture, real-time predictions, and collaborative annotation, allowing field personnel to collect and classify spoil data on-site. While offline functionality ensures continued operation without network access, periodic internet connectivity is required for cloud synchronisation and model updates.

Together, these developments position SCANDY as a practical solution for spoil classification in operational mining environments. Its ability to deliver high-accuracy pile-level predictions, combined with mobile deployment and offline functionality, supports integration into routine field workflows.