RANDOM FOREST PARA A CLASSIFICAÇÃO DE ESPECTROS LIBS: TÉCNICAS DE NORMALIZAÇÃO E SELEÇÃO DE VARIÁVEIS

Abstract

Recent research highlights the use of the LIBS (Laser-Induced Breakdown Spectroscopy) technique for material classification across various fields. In most cases, this classification is carried out using machine learning algorithms. In this regard, other studies show the effectiveness of combining normalization, feature selection, and dimensionality reduction methods with machine learning classifiers, including Random Forest. This study presents a protocol that employs normalization, feature selection, and dimensionality reduction techniques in conjunction with Random Forest for classifying firearm residue from non-toxic ammunition. The spectra were normalized using Standard Normal Variate (SNV), and then Principal Component Analysis (PCA) and the Boruta technique were applied in a permuted fashion to reduce and select features, aiming to reduce data dimensionality and train the Random Forest classification model. While training, the data were tested using internal validation through leave-one-out cross-validation, and the Random Forest hyperparameters were optimized using Grid Search to achieve the best possible accuracy for the model. This optimal result demonstrated 100% accuracy for both internal and external validation, highlighting the effectiveness of the developed protocol.