Prediction system based on Visible-NIR Imaging had tested in various cases. This kind of prediction system excels at cases which is hard to do inspection by human eyesight. This ability is due to lots of available features or wavelengths (>100 features). However, that much of features are not always worth the prediction system performance. In this study, a genetic algorithm is used as wavelength selection methods for enhancing Visible-NIR Imaging prediction system. Prediction system would focus on the prediction of carotenoid content green amaranth (Amaranthus sp.) leaf. This study used 20 leaves of green amaranth. Image of each amaranth leaf acquired at 400-1000 nm. For each leaf, four regions of interest (ROI) is selected for determination of carotenoid content. Measurement of the reference value is done by using the Sims-Gamon method. Image of amaranth leaf then processed through correction, segmentation, and extraction. The regression model is built for predicting carotenoid content by using partial least square regression (PLSR) algorithm. Without wavelength selection, prediction system has the performance of 0.584 for R
and 0.0169 for RMSE. Prediction system with implemented wavelength selection only use 89 of 224 wavelengths and has the performance of 0.878 for R
and 0.01 for RMSE Results of this study showed that prediction system performance significantly improved by implementing the genetic algorithm as wavelength selection.