Seismic coherence is commonly used to delineate structural and stratigraphic discontinuities. We generally use full-bandwidth seismic data to calculate coherence. However, some seismic stratigraphic features may be buried in this full-bandwidth data but can be highlighted by certain spectral components. Due to thin-bed tuning phenomena, discontinuities in a thicker stratigraphic feature may be tuned and thus better delineated at a lower frequency, whereas discontinuities in the thinner units may be tuned and thus better delineated at a higher frequency. Additionally, whether due to the seismic data quality or underlying geology, certain spectral components exhibit higher quality over other components, resulting in correspondingly higher quality coherence images. Multispectral coherence provides an effective tool to exploit these observations. We have developed the performance of multispectral coherence using different spectral decomposition methods: the continuous wavelet transform (CWT), maximum entropy, amplitude volume technique (AVT), and spectral probe. Applications to a 3D seismic data volume indicate that multispectral coherence images are superior to full-bandwidth coherence, providing better delineation of incised channels with less noise. From the CWT experiments, we find that providing exponentially spaced CWT components provides better coherence images than equally spaced components for the same computation cost. The multispectral coherence image computed using maximum entropy spectral voices further improves the resolution of the thinner channels and small-scale features. The coherence from AVT data set provides continuous images of thicker channel boundaries but poor images of the small-scale features inside the thicker channels. Additionally, multispectral coherence computed using the nonlinear spectral probes exhibits more balanced and reveals clear small-scale geologic features inside the thicker channel. However, because amplitudes are not preserved in the nonlinear spectral probe decomposition, noise in the noisier shorter period components has an equal weight when building the covariance matrix, resulting in increased noise in the generated multispectral coherence images.