A hierarchical class-based machine learning approach for on-site detection of explosives and precursors using NIR...

Accurate and reliable on-site detection of explosives and their precursors remains critical for first responders and EOD teams. While portable spectroscopic techniques such as Raman and FTIR are widely used for this purpose, their limited sensitivity and robustness when analysing heterogeneous real-world specimens such as Homemade Explosives (HMEs) lead to high false-negative rates in the field. This study presents a cloud-based hierarchical machine learning framework integrated with handheld NIR spectroscopy to improve detection robustness. The two-level approach provides high-sensitivity screening at Level 1 (explosive/precursor family detection) and refined subclass characterization at Level 2 (e.g., moisture state, composition). The model was trained on approximately 17,000 NIR scans (~5000 averaged samples from ~550 substances). Performance was assessed using bootstrapping and an independent EOD operational dataset (195 explosives, 10 precursors, 65 true negatives). The best ensemble model achieved 96.3% accuracy on the independent EOD data set, with a 3.1% false-negative rate (96.9% true positive rate) and 0% false-positive rate among the chemical families represented in training. This corresponds to a substantial reduction in false negatives compared to previously reported portable NIR and Raman studies. Importantly, the hierarchical structure ensures that missed subclass identification does not compromise initial threat detection, providing an inherent safety margin for field screening. These results demonstrate that combining hierarchical classification with large-scale, variability-driven training significantly enhances the reliability of portable NIR-based detection. The cloud-based framework enables continuous model updates through the integration of field data. Remaining limitations include intrinsic NIR insensitivity for certain inorganic salts and dependence on representative training variability. Future integration with complementary techniques such as Raman spectroscopy offers a pathway toward comprehensive, field-deployable multi-sensor detection systems.