Hyperspectral imaging for the elimination of visual ambiguity in corrosion detection and identification of corrosion sources

Dayakar N Lavadiya,Hizb Ullah Sajid,Ravi K Yellavajjala,Xin Sun

DOI: https://doi.org/10.1177/14759217211041690

2021-08-28

Structural Health Monitoring

Abstract:The similarity in the hue of corroded surfaces and coated surfaces, dust, vegetation, etc. leads to visual ambiguity which is challenging to eliminate using existing image classification/segmentation techniques. Furthermore, existing methods lack the ability to identify the source of corrosion, which plays a vital role in framing the corrosion mitigation strategies. The goal of this study to employ hyperspectral imaging (1) to detect corroded surfaces under visually ambiguous scenarios and (2) identify the source of corrosion in such scenarios. To this end, three different corrosive media, namely, (1) 1M hydrochloric acid (HCl), 2) 3.5 wt.% sodium chloride solution (NaCl), and (3) 3 wt.% sodium sulfate solution (Na 2 SO 4 ), are employed to generate chemically distinctive corroded surfaces. The hyperspectral imaging sensor is employed to obtain the visible and near infrared (VNIR) spectra (397 nm–1004 nm) reflected by the corroded/coated surfaces. The intensity of the reflectance in various spectral bands are considered as the descriptive features in this study, and the training and test datasets were generated consisting of 35,000 and 15,000 data points, respectively. SVM classifier is trained and then its efficacy on the test data is assessed. Furthermore, validation datasets are employed and the generalization ability of the trained SVM classifier is verified. The results from this study revealed that the SVM classifier achieved an overall accuracy of 94% with the misclassifications of 18% and 13% in the case of NaCl and Na 2 SO 4 corrosion, respectively. Reflectance spectra obtained in the VNIR region was found to eliminate the visual ambiguity between the corroded and coated surfaces and, identify the source of corrosion accurately. Further, the range of key wavelengths of the spectra that play an important role in the distinguishability of coating and chemically distinctive corroded surface were identified to be 500–520 nm, 660–680 nm, 760–770 nm, and 830–850 nm.