Xu Ming Zhong

Ground-airborne frequency-domain electromagnetic (GAFDEM) data are often severely degraded by complex multi-source noise, including broadband environmental interference, low-frequency motion disturbance, and non-stationary field contamination, which significantly reduce signal quality and hinder subsequent interpretation. This study develops an SNR-Stacking ensemble denoising framework for recovering effective electromagnetic responses under mixed noise conditions. The method combines multi-scale time–frequency feature extraction with heterogeneous base learners and introduces a physically interpretable signal-to-noise ratio (SNR)-guided adaptive weighting strategy within a two-stage stacking architecture. In the first stage, complementary predictors are fused according to their reconstruction reliability, and in the second stage a meta-learner is used to further refine the residual structure of the denoised signal. Experiments on simulated GAFDEM data show improved waveform reconstruction, spectral consistency, and signal preservation compared with conventional filtering baselines, individual learning models, and conventional stacking. Field application suggests that the simulation-trained model can suppress broadband and impulsive interference without field-data retraining while improving amplitude fidelity at representative frequencies. The reconstructed amplitudes remain consistently closer to the ground reference than the noisy observations, with retention ratios ranging from 0.957 to 0.991 at the analysed representative frequencies. These results support SNR-guided stacking as a compact and interpretable strategy for improving GAFDEM signal quality before quantitative interpretation.