Abstract: To address the empirical selection of stopband parameters, the lack of spectral prior information in filter-order determination, and the scale-dependent efficiency of multi-trace seismic filtering, a spectrum-feature-driven adaptive parameter design method is proposed for Butterworth low-pass filters. First, a spectral feature vector is constructed using the normalized cutoff frequency, low-frequency energy ratio, adjacent-band energy ratio near the cutoff frequency, high-frequency energy ratio, spectral centroid, spectral bandwidth, and spectral edge-energy ratio. Then, a normalized composite evaluation function considering waveform fidelity, passband stability, stopband suppression, and order complexity is established, and parameter labels are generated by traversing candidate stopband ratios and filter orders. Finally, random forest regression models are used to predict the stopband ratio and the filter-order prior, respectively, and the final filter order is determined using a locally robust search strategy. Experiments on synthetic seismic signals show that the proposed stopband-parameter prediction and order-prior-based local search strategy achieves an average SNR gain of 15.08 dB, a waveform fidelity of 90.84%, and a stopband attenuation of 33.72 dB, outperforming fixed-stopband, empirical-stopband, and simple binary-search strategies. Ablation experiments indicate that the boundary-triggering rate of filter order is reduced to approximately 2.4% after introducing the order-prior-based local search, thereby improving the stability of order selection. CPU/GPU efficiency tests under different data scales show that GPU processing with data transfer is not necessarily faster than vectorized CPU processing for small-scale data, whereas GPU parallelism becomes more advantageous in large-scale multi-trace batch processing. Application to real seismic data demonstrates that the proposed method can effectively suppress high-frequency interference while preserving the waveform in the target low-frequency band. The results indicate that the proposed method improves the adaptivity and stability of filter-parameter selection under complex spectral conditions while maintaining the interpretability of Butterworth filter parameters.