Seismic waveform feature extraction is a critical task in seismic exploration, as it directly impacts reservoir prediction and geological interpretation. However, large-scale seismic data and nonlinear relationships between seismic signals and reservoir properties are challenging for traditional machine learning methods. To address these limitations, this paper proposes a novel framework combining Convolutional Neural Network (CNN) and Uniform Manifold Approximation and Projection (UMAP) for seismic waveform feature extraction and analysis. The UMAP-CNN framework leverages the strengths of manifold learning and deep learning, enabling multi-scale feature extraction and dimensionality reduction while preserving both local and global data structures. The evaluation experiments, which considered runtime, receiver operating characteristic (ROC) curves, embedding distribution maps, and other quantitative assessments, illustrated that the UMAP-CNN outperformed t-distributed stochastic neighbor embedding (t-SNE), locally linear embedding (LLE) and isometric feature mapping (Isomap). A case study in the Ordos Basin further demonstrated that UMAP-CNN offers a high degree of accuracy in predicting coal seam thickness. Furthermore, our framework exhibited superior computational efficiency and robustness in handling large-scale datasets.

Due to the limitations of seismic data resolution, conventional inversion methods are commonly limited by the band-limited nature of seismic data, noise interference, and insufficient resolution for subtle geological targets. This study proposes a continuous wavelet transform (CWT)-assisted multiscale stepwise learning (MSL) workflow for seismic impedance inversion. The proposed workflow transforms seismic traces into time–frequency representations using CWT and progressively learns different frequency components through convolutional modules with different kernel sizes. A seismic forward-modeling constraint is further introduced to improve the consistency between the predicted impedance and the observed seismic response at non-well locations. The network takes a single seismic trace represented by CWT coefficients in the 1–80 Hz frequency range and a corresponding low-frequency initial impedance model as inputs, and outputs the acoustic-impedance trace at each time sample. The main methodological novelty is the progressive learning of different frequency components using frequency-dependent convolutional kernels. The method is evaluated on a 3D seismic dataset from the Sichuan Basin, China, using leave-one-well-out validation involving seven wells. In addition, we provide detailed ablation experiments and quantitative analysis. Across all seven blind-well tests, the proposed method consistently achieves the highest mean correlation coefficient (0.8509) and the lowest mean MSE (0.0636) among the compared methods. These results indicate that the proposed workflow can effectively integrate multiscale seismic information and provide effective support for seismic impedance inversion.