A Deep Learning-Based Denoising Method for Airborne Electromagnetic Data

Airborne electromagnetic (AEM) measurements are vulnerable to mixed environmental noise, and rapid signal decay hampers weak transient recovery. We propose WL-CNN-BiLSTM, integrating dual-stage wavelet filtering, parallel multi-scale convolutions, residual skip connections, bidirectional recurrent units, and a hybrid reconstruction objective for complex TEM waveforms. On the archived 20,000-trace benchmark’s 3000-trace test subset, it achieved an aggregate physical-domain SNR 17.53 dB (RMSE 1.48 × 10−10 T; PCC 0.989), versus 11.49 dB for PC-BiLSTMNet, 11.76 dB for validation-selected Median filtering, and 13.04 dB for VMD; the per-trace SNR P10 was −3.41 dB (minimum −25.45 dB). These archived results characterize stored numerical targets, not independently verified AEM responses. On an independent 1000-trace blind test, a separately trained layered-earth configuration using training-only scalers achieved a median SNR of 5.91 dB versus 0.39 dB (Median) and −0.05 dB (VMD), but poorer lower-tail SNR (P10 −43.93 dB) and a median of NRMSE 7.402 over 101–1001 μs. Proprietary iFTEM-II windows showed waveform changes versus the engineering curve, but unavailable proxy provenance/calibration made this qualitative evidence, not field-accuracy validation. A matched p = 0 comparison, three-seed ablations, complete training-only-scaler retraining, and out-of-distribution diagnostics delineate distributional dependencies; temporal feature weighting is an implemented design component, not a demonstrated late-time enhancer. It shows aggregate reconstruction gains within the benchmark and characterizes dataset- and protocol-dependent limitations.