Using genetic algorithms to select wavelengths in near-infrared spectra: Application to sugar content prediction in cherries

The best parameter for evaluating fruit taste quality is sugar content. This parameter can be nondestructively measured by near-infrared spectrometry. Models that give the best performances are based on classical partial least-squares regression (PLSR) applied over the whole spectrum. Selecting the best wavelengths improves the precision and the robustness of these models. Genetic algorithms (GAs) applied to the results of a cross-validation make this selection in a stochastic way. The influence of GA tuning and of the cross-validation parameters is discussed. An application of this technique on cherry samples decreases the cross-validation error by 20% and divides the prediction error by more than 3.