We propose an improved opposition-based self-adaptive differential evolution (IOSaDE) algorithm for multi-parameter optimization in vibrational hybrid femtosecond/picosecond coherent anti-Stokes Raman scattering (CARS) thermometry. This new algorithm self-adaptively combines the advantages of three mutation schemes and introduces two opposite population stages to avoid premature convergence. The probability of choosing each mutation scheme will be updated based on its previous performance after the first learning period. The IOSaDE method is compared with nine other traditional differential evolution (DE) methods in simulated spectra with different simulation parameters and experimental spectra at different probe time delays. In simulated spectra, both the average and standard deviation values of the final residuals from 20 consecutive trials using IOSaDE are more than two orders of magnitude smaller than those using other methods. Meanwhile, the fitting temperatures in simulated spectra using IOSaDE are all consistent with the target temperatures. In experimental spectra, the standard deviations of the fitting temperatures from 20 consecutive trials decrease more than four times by using IOSaDE, and the errors of the fitting temperatures also decrease more than 18%. The performance of the IOSaDE algorithm shows the ability to achieve accurate and stable temperature measurement in CARS thermometry and indicates the potential in applications where multiple parameters need to be considered.