From hyperspectral to multispectral: Robust corrosion detection in iron and steel using minimal bands

Abstract

Abstract Corrosion in iron and steel structures is a major challenge in coastal environments, where exposure to saline conditions accelerates material degradation and demands continuous monitoring. Multispectral cameras are a promising tool for detecting corrosion in a noninvasive, noncontact, accurate, and fast manner. The wavelengths used by the multispectral camera significantly impact detection performance, which is also influenced by the type of coating on the metal. This study presents a hyperspectral analysis to identify the three most informative bands for detecting corrosion across different coating types and to transfer the selected bands to a multispectral system. For this purpose, a dataset is created consisting of five sample sets with uncoated surfaces and various coatings (differing in paint and color). The optimal spectral bands are selected for each sample set using a Sequential Forward Selector (SFS) and a Random Forest (RF) classifier. Band and pretrained transferability are evaluated across datasets. The generalizable model, incorporating all samples, identifies 518, 655, and 838 nm as the most relevant wavelengths. Incorporating texture features from the Gray-Level Co-occurrence Matrix (GLCM) improves accuracy by up to 6%. Robustness is validated using an independent dataset with unknown coatings. These bands are subsequently mapped to a Silios commercial multispectral camera (547, 665, and 826 nm), and the generalizable hyperspectral model is applied to the resulting multispectral images. This model achieves high accuracy, confirming its cross-sensor applicability. This approach enables future deployment on UAV platforms for large-scale, automated corrosion monitoring in coastal environments.