Adaptive Learning with Logistic Regression for soil classification with LIBS

Laser-induced breakdown spectroscopy (LIBS) presents a promising avenue for real-time soil characterization, particularly with the integration of machine learning (ML) techniques to enhance its detection capabilities. However, the variability in soil matrix properties can lead to discrepancies in LIBS emission lines, challenging traditional ML algorithms that assume consistent distributions between training and test spectra. To address this challenge, we propose a novel domain adaptation approach leveraging logistic regression with class and pseudo-label weighting. We categorize classes based on transfer difficulties into hard, normal, and easy categories. Hard classes, with low predictive proportions due to misclassification, are assigned higher weights. In contrast, lower weights are assigned to easy classes to mitigate transfer imbalances. Furthermore, we introduce weighted pseudo-labels from hard classes misclassified into similar classes, obtained through few-shot learning based on similarity, to be incorporated into co-training alongside source-labeled samples. It refines the logistic regression model and boosts the test performance of hard classes. To highlight the effectiveness of the proposed method, it is tested with the Euro-Mediterranean Symposium on LIBS (EMSLIBS) contest dataset and compared against literature methods, including data calibration, adversarial training, and self-learning. Notably, our method outperforms these methods, achieving an impressive accuracy of 91.3%.