Abstract: Conventional Butterworth low-pass filter design commonly relies on empirically selected stopband ratios and filter orders, making it difficult to accommodate spectral variations among seismic traces. To address this problem, a spectral-feature-driven adaptive parameter design method is proposed. This method constructs a spectral feature vector that characterizes effective low-frequency components, interference in the band adjacent to the cutoff frequency, and high-frequency noise distribution. During synthetic-sample generation, a normalized composite cost function is established to incorporate waveform fidelity, passband stability, stopband suppression, and filter-order complexity, and is used to generate supervised labels for the stopband ratio and filter order. During synthetic-sample calibration and validation, random forest models are trained to predict the stopband ratio and a filter-order prior, and a robust local search is then performed around this prior. In field seismic data processing, the offline-calibrated models are used to recommend filtering parameters, and reference-free posterior quality control is conducted using the correlation coefficient in the low-frequency target band, the relative variation in root-mean-square amplitude within that band, and the high-frequency residual-energy ratio. Synthetic experiments show that the proposed method achieves an average signal-to-noise ratio gain of 15.08 dB, waveform fidelity of 90.84%, and stopband attenuation of 33.72 dB, while reducing the proportion of cases reaching the prescribed filter-order boundary from 46.3% to 2.4%. Application to two field datasets reduces the energy ratios in the high-frequency interference evaluation bands by 83.5% and 87.0%, respectively, while the correlation coefficients in the low-frequency target bands reach 0.974 and 0.968. The results indicate that the method is suitable for active-source seismic records in which effective low-frequency components and high-frequency random noise or narrowband interference are reasonably separable in the frequency domain. Low-frequency coherent noise overlapping the effective signal band still requires attenuation in combination with other methods.