The seismic frequency spectrum provides a useful source of information for reservoir characterization. For a seismic profile presented in the time-space domain, a vector of the frequency spectrum can be generated at every sampling point. Because the spectrum vectors at different time-space locations have different variation features, I attempt for the first time to exploit the variation pattern of the frequency spectrum for reservoir characterization, and test this innovative technology in prediction of coalbed methane (CBM) reservoirs. The prediction process implicitly takes account of the CBM reservoir factors (such as viscosity, elasticity, cleat system, wave interference within a coal seam, etc.) that affect the frequency spectrum, but strong amplitudes in seismic reflections do not necessarily show any influence in clustering analysis of spectral variation patterns. By calibrating these variation patterns quantitatively with CBM productions in well locations, we are able to characterize the spatial distribution of potential reservoirs.