Abstract
<jats:p>Earth science relies on long, continuously recorded time-series signals such as seismic waveforms, volcanic tremors, and geomagnetic and gravity records. These signals are contaminated by sensor noise, environmental microseism, sampling jitter, and increasingly by deliberate data tampering, while labeled data remain scarce and station-specific. Despite the mathematical similarity between geoscientific time series and the speech, audio, and video signals tackled by recent deep-learning advances, no existing earth-science framework jointly addresses enhancement, integrity verification, and quality assessment in a single end-to-end model. We propose GeoFIRING-Net, a unified framework that systematically transfers eight recent advances into a single backbone tailored to one-dimensional three-component seismic traces: filtered feature recycling from FIRING-Net, local and non-local attention from CASENet, adversarial fine-tuning from Vsegan, progressive reinforcement learning from Vq-insight, deep watermarking from Omniguard, static-to-dynamic transfer from LVE-S2D, visual incontext learning from VICL, and weak-to-strong generalization from WTS. The shared encoder branches into three task-specific heads that jointly perform seismic waveform enhancement, watermark-aided tamper localization, and progressive-RL quality assessment. We also introduce two new benchmarks, GeoTamper and GeoQA-5K, for seismic data integrity verification and expert-aligned quality assessment. On STEAD and INSTANCE, GeoFIRING-Net achieves 12.37 dB and 12.05 dB output SNR at 0 dB input, outperforming the strongest baseline by 1.33 dB and 1.27 dB respectively, while reaching 90.7 percent accuracy on tamper localization and 0.815 SRCC on quality assessment against human expert ratings. Ablations confirm that every transferred module contributes positively, with the filtered-recycling backbone being the most critical. The results validate cross-domain method transfer as a practical route for earth-science signal processing.</jats:p>