High-level data fusion using majority voting for the classification of spray paint spectroscopic data

A high-level data fusion approach for classifying spray paint samples from five major U.S. manufacturers, each represented by five color groups (black, blue, red, silver/gray, and white), was investigated. Spectral data were collected using four analytical techniques: Fourier transform infrared (FTIR) spectroscopy, Raman spectroscopy with laser excitation at 532 nm and 785 nm, scanning electron microscopy coupled with energy-dispersive spectroscopy (SEM-EDS), and UV-Vis microspectrophotometry (MSP). Their combined use and discriminating ability were evaluated. Each dataset was independently modeled using five supervised machine learning classifiers: Naïve Bayes, k-nearest neighbors (KNN), support vector machine (SVM), random forests, and extreme gradient boosting (XGBoost). The intermediate predictions from each classifier were integrated using the majority voting mechanism to yield a final class assignment, forming a high-level data fusion scheme. The proposed approach consistently outperformed individual instruments, achieving near-perfect classification accuracy across several color blocks, particularly for red and blue paints. Among classifiers, generally, Random Forest and Naïve Bayes provided the most stable performance, while SVM with a linear kernel and XGBoost showed lower accuracy. The findings confirm that fusing complementary spectral information improves discriminative ability, reduces redundancy, and creates a computationally efficient, reproducible framework for objective evaluation of source-level questions arising from forensic paint examinations. Overall, the developed model mirrored the process followed by forensic paint examiners in recognizing relevant spectral features from the various techniques. This approach offers a promising pathway toward integrating multimodal spectral data within probabilistic or likelihood ratio-based frameworks following comparative examinations of paint.