Based on the random frog algorithm, a window-based random frog algorithm is proposed. With a continuous window instead of a single wavelength point, the proposed algorithm improves the optimization accuracy of the original random frog algorithm and reduces the iteration times of the algorithm, thus improving the convergence speed. The results of the blood samples show that compared with the full-band results, the root mean square error of prediction (RMSEP) of the built partial least square model decreases by 47.9%, and the correlation coefficient of the prediction set, R_p, increases by 4.07%, proving the validity of the proposed algorithm. The regression analysis of the characteristic wavelengths selected by the three mainstream algorithms and the window-based random frog algorithm is carried out, which demonstrates the superiority of the improved algorithm in selecting the characteristic wavelength.