Random forest (RF) regression algorithm was utilized for determination of protein content in fishmeal samples based on near-infrared (NIR) spectrometry. Considering the randomness of RF method, the optimized models were selected by tuning the two vital modeling parameters of the number of decision trees (ntree) and the number of split variables (nsv). The descending of Gini coefficient (G) is taken as the indicator performing the modeling importance of NIR valuables. It was used to select the informative wavelengths for NIR analysis of fishmeal, with an aim to improve the accuracy of quantitative models. According to statistical theory, we tried to select equivalent optimal model with relatively low computational complexity. The optimized RF model needed to construct 471 decision trees and randomly select 103 wavelength variables for node splitting when the decision trees grow. Simultaneously, 52 NIR informative wavelengths can be selected out according to the average of G descending values based on the trees in the forest. The equivalent optimized RF model output the root mean square error (RMSEv) and correlation coefficient (Rv) of validation set were 3.970% and 0.943, respectively. The optimized model was further evaluated by using the prediction samples that were excluded from modeling process, with the RMSEp of 5.271%, and the Rp of 0.906. Results showed that RF regression combined with G coefficients for wavelength selection is feasible and effective to improve the NIR predictive ability for quantitative determination of fishmeal protein.