Seismic impedance inversion has been widely used to estimate subsurface properties. Conventional inversion assumes that seismic data are the convolution result of seismic wavelet and reflectivity, implying that seismic data are stationary when a constant wavelet is considered. However, seismic data are nonstationary because of noise contamination and attenuation during wave propagation, which means that the frequency spectrum of the seismic signal changes from shallow to deep formations. We have developed a time-variant wavelet extraction method by using a local-attribute-based spectral decomposition technique. Time-variant wavelets are generated according to the local frequency spectrum, which can be used to construct a time-variant wavelet kernel matrix. By using this time-variant kernel matrix, we can obtain a better correlation between synthetic and extracted seismograms than by using constant wavelet on a field data example. Using this example, we have also compared the time-variant and constant wavelets for inverting the field data to estimate subsurface acoustic impedance. Our results showed improved resolution and a better fit to well-log-measured impedance by using the time-variant wavelets.