This study is an extensive comparison of the predictive performance of a bagging neural network (BANN), partial least squares (PLS) regressor, and a kernel-based, nonlinear PLS (KPLS) regressor, given experimentally obtained CO2-N2-Ar plasma emission spectra. The spectra, 3,62,31 in total, were obtained from controlled gas mixtures with varying CO2 concentrations fed to a stripline split-ring resonator microplasma source and recorded with an UV-NIR, 2nm resolution spectrometer. The regression methods’ dependence on (i) number of observations used in training, (ii) preprocessing steps, and (iii) feature selection (in this case wavelength ranges) was evaluated by training and testing 60–66 models per method, each with a unique combination of the aforementioned configuration options, and each trained 4 times with random train-test splits. To compare the models a custom metric that compounds R2, Pearson correlation, and a weighted root mean squared relative error was used. The results show that the BANN models outperform both PLS and KPLS, reaching a peak score of 0.873, with the others getting 0.561 and 0.581, respectively, using the (−∞, 1] metric. The top performing BANN model was trained without any feature selection or preprocessing, these steps were, however, required for both the best PLS and KPLS models. In a wider perspective, the results show that BANNs are not only suitable as in-place replacements for PLS-based methods, but increase regression model prediction accuracy on low resolution spectra to such an extent that they offer modelling of previously unattainable, nonlinear information contained in emission spectra datasets.