In seismic data processing, the denoising of earthquake data plays a crucial role, with the enhancement of the signal-tonoise ratio serving as a foundational step for subsequent high-quality processing and precise interpretation. Presently, methods for seismic data denoising have undergone extensive development, with dictionary learning methods demonstrating unique advantages. The current classical K-singular value decomposition (K-SVD) dictionary learning algorithm exhibits issues such as partial loss of the original signal in denoising results and less-than-ideal computational efficiency. In order to further optimize these problems, a dictionary learning method based on the sequential generalized K-means (SGK) algorithm for seismic data denoising was proposed. Initially, blocks of random positions were extracted from the sample data, with blank blocks being removed to initialize the dictionary. Subsequently, in the dictionary learning phase, the most recent sparse representation dictionary was adaptively constructed based on the inherent features of seismic data. Following this, the learned dictionary was employed to denoise seismic data blocks containing noise. The denoised blocks were then subjected to averaging, and the image blocks were reconstructed, ultimately achieving the denoising of seismic data. The denoising performance of the proposed method was validated through experiments using synthetic and real seismic data. The evaluation encompasses aspects such as signal-to-noise ratio, computational efficiency, and the preservation of effective signals.