The invention provides a weight-based local cross-correlation time frequency domain phase inversion method via which time frequency domain phase information of seismic data is introduced into a cross-correlation objective function. The method is called the weight-based local cross-correlation time frequency domain phase inversion method. The introduction of the phase information alleviates the dependence of full waveform inversion on an initial velocity model. At the same time, weighting factors are added in the time frequency domain objective function, and anti-noise ability and stability ofthe inversion can be greatly increased. In a low frequency band weight-based local cross-correlation time frequency domain phase inversion method, a good initial velocity model can be obtained and canbe then used in a high frequency band weight-based local cross-correlation time frequency domain phase inversion method, and high resolution inversion results can be obtained. Tests in which low frequency components and strong Gaussian background noise are missing show that the weight-based local cross-correlation time frequency domain phase inversion method is advantaged by strong anti-noise ability, capability of being independent of the initial velocity model and the like.