In order to recover weak signal that is submerged in heavy background noise, a weak signal recovery method based on three-dimensional curvelet transform and generalised cross validation (GCV) is presented. Their method includes three stages: firstly, three-dimensional dataset was decomposed into sub-volumes by three-dimensional curvelet transform, and then Graphics Processing Unit (GPU) is used to improve the speed of parallel processing of sub-volumes. Secondly, the GCV criterion and genetic algorithm were used to improve the signal-to-noise ratio (SNR) of the processed signal. Finally, according to different distribution between the effective signal energy and the noise energy, adaptive filter is used to enhance the recovered weak signal. Furthermore, to verify the availability of the method, a wedge of simulation data and 100 groups of three-dimensional seismic data for testing are analysed, the stratigraphic structure information is much clearer in the processed seismic data than in the original data. The results show that the SNR of the recovered data is improved by 3 dB and the band of frequency is increased by 100 Hz. Their method is four to five times faster than a recovery method based on CPU processing.