Surface Ice Detection Using Hyperspectral Imaging and Machine Learning

Abstract

Highlights: What are the main findings? Hyperspectral imaging combined with machine learning enables accurate surface ice detection. Support Vector Machine (SVM) and Random Forest (RF) classifiers were evaluated on real hyperspectral datasets. The models generalize well across coated and uncoated surfaces, including challenging dark coatings. Spectral band reduction was analyzed, revealing trade-offs between classification accuracy and computational efficiency. A multiclass classification approach was introduced to differentiate between rime and glaze ice. What is the implication of the main finding? Demonstrates the feasibility of non-contact, hyperspectral-based methods for automated ice detection. Validates robustness of machine learning models across diverse material surfaces. Supports the development of efficient, real-time ice detection systems through reduced spectral dimensionality. Enhances safety applications by enabling discrimination between different ice types. Provides a foundation for integrating hyperspectral AI systems in energy, transportation and infrastructure monitoring. Ice formation on critical infrastructure such as wind turbine blades can lead to severe performance degradation and safety hazards. This study investigates the use of hyperspectral imaging (HSI) combined with machine learning to detect and classify ice on various coated and uncoated surfaces. Hyperspectral reflectance data were acquired using a push-broom HSI system under controlled laboratory conditions, with ice and rime ice generated using a thermoelectric cooling setup. Support Vector Machine (SVM) and Random Forest (RF) classifiers were trained on uncoated aluminum samples and evaluated on surfaces with different coatings to assess model generalization. Both models achieved high classification accuracy, though performance declined on black-coated surfaces due to increased absorbance by the coating. The study further examined the impact of spectral band reduction to simulate different sensor types (e.g., NIR vs. SWIR), revealing that model performance is sensitive to wavelength range, with SVM performing optimally in a reduced band set and RF benefiting from the full spectral range. A multiclass classification approach using RF successfully distinguished between glaze and rime ice, offering insights into more targeted mitigation strategies. The results confirm the potential of HSI and machine learning as robust tools for surface ice monitoring in safety-critical environments.