Seismic denoising is a crucial step in data processing to understand geological structures. However, during seismic data acquisition, the seismic data is distorted by random and erratic noise. The noise may lead to the misinterpretation of geological structures. Sparse coding is a fundamental step in traditional dictionary learning and is computationally expensive due to its combinatorial nature using orthogonal matching pursuit (OMP). To address this issue, compressive sampling matching pursuit (CoSaMP) is used to provide efficient sparse coefficients. Initially, the data were segmented into two parts based on the signal-tonoise ratio (SNR) of each patch, which is calculated individually. Patches containing SNR≥1 contained more signal features, and those containing SNR<1 contained less signal features. In second stage, we utilized two parallel adaptive dictionaries, the adaptive SNR-guided patch partitioning strategy enables the two dictionaries to learn distinct feature distributions from seismic data. The primary dictionary D1 learns coherent and high-energy seismic reflections from high-SNR patches, whereas the secondary dictionary D2 captures low-amplitude, discontinuous, and noise-corrupted structural patterns from low-SNR patches. We effectively eliminated erratic noise using Markov subsampling with the Huber criterion. In the third stage, K-SVD is performed for residual noise attenuation. Both dictionaries learn signal features and combine them. MDL-KSVD is applied to two 2D and two 3D synthetic and two field datasets. The numerical results demonstrate that our method provides a lower mean squared error (MSE) and a higher SNR for denoised seismic data than existing methods. Moreover, the visualization of the denoised seismic sections clearly shows the geological features.