Evaluating the impact of data pre-processing methods on classification of ATR-FTIR spectra of bituminous binders

Abstract

• Suitable normalization methods like NCV, NMO, RS, and NTS are for peak area/indices; NTS, NCV, AS, PS, and SNV for spectra. • Adaptive smoothness penalized least squares (aspls) is ideal for gradual material changes, not just additive detection. • Recommended DP methods ensure precise bituminous binder classification, aiding fingerprinting in various applications. • DP method effectiveness and choice of area or indices calculation depend on classification goals and data characteristics. • Using entire spectra or their first derivatives improves classification accuracy over indices or peak areas. Attenuated total reflectance Fourier-transform infrared spectroscopy (ATR-FTIR) is an essential tool for the analysis of bituminous binders due to its cost-effectiveness, user-friendliness, and non-destructive nature. However, its effectiveness is often hampered by challenges such as non-informative regions, lack of standardized analysis methods, and inconsistent baselines in spectral data. Addressing these challenges, this study aims to comprehensively evaluate the impact of various data pre-processing (DP) methods on ATR-FTIR spectra from diverse bituminous binder types, sources, and aging conditions. Using partial least squares-discriminant analysis (PLSDA) classification, the study assesses the effectiveness of baseline correction, normalization, and their combinations. The methodology involves analyzing peak areas, indices, entire spectra, and first derivative spectra to determine the most effective pre-processing strategies. Key findings reveal that the effectiveness of DP methods is influenced by the classification goals, characteristics of the spectral dataset, and the specific methods employed for input data preparation. The study demonstrates that using entire spectra or their first derivatives leads to higher classification accuracy compared to indices or specific spectral peak areas. The choice between peak area and indices calculation methods should align with the study’s objectives. For efficient and rapid selection of DP methods, tools like PLSDA are recommended. Among the normalization methods, normalization to constant vector length (NCV), normalization to change the maximum to 1 (NMO), robust scaling (RS), and normalization to sum (NTS) are suitable for peak area or indices-based classification. For entire spectra and their first derivatives, NTS, NCV, autoscaling (AS), pareto scaling (PS), and standard normal variate (SNV) methods are recommended. Regarding baseline correction, Adaptive Smoothness Penalized Least Squares (aspls) is suitable for studies focusing on gradual material changes, such as multi-level aging studies, but not for additive detection studies. The findings of this study provide valuable insights and practical recommendations for selecting appropriate DP methods, thereby enhancing the classification accuracy and reliability of ATR-FTIR spectral analysis of bituminous binders. This contributes significantly to the design of experiments, reduces operational risks, and optimizes resource utilization in the field.

Topics & Concepts