Denoising of standoff ultraviolet Raman spectra under complex noise

Accurate remote detection and identification of unknown substances under complex noise interference have long been a major research focus worldwide. Raman spectroscopy offers the unique advantage of providing molecular fingerprint information of target substances under noncontact conditions. However, complex noise sources, such as background light, detector noise, and long-distance transmission attenuation, can significantly reduce the distinguishability of weak spectral peaks. To address the difficulties in obtaining clean labels, the complexity of mixed noise, and the limited generalization capability of parametric filtering methods in remote measurements, an improved Noise2Void method incorporating a spectrally second-derivative-regularized loss, termed NN2V, is proposed. This method extends blind-spot learning from two-dimensional images to one-dimensional spectra. By masking the central sampling point, identity mapping is avoided, while the second-derivative-regularized loss constrains the local curvature of spectral lines, thereby suppressing high-frequency noise while preserving peak positions and intensities. In this study, a distance-simulation model based on the coupling of near-field measured signals and background noise was further constructed, and a 266 nm ultraviolet remote Raman transmitting-receiving system was established. Experimental validation was performed on solid calcium carbonate, sodium nitrate solution, trace hydrogen, and trace methane samples. Compared with Savitzky-Golay filtering, wavelet transform, and two deep learning-based methods, NN2V achieved a peak-intensity preservation ratio of 0.98 and improved the SNR of the trace hydrogen spectrum from 16.04 to 34.68 dB. The simulation and 10 m experimental results demonstrate that, under the current optical path, gating, and acquisition conditions, the proposed method can enhance the identifiability of weak Raman peaks in complex noise backgrounds. This study provides a feasible solution for denoising remote ultraviolet Raman spectra lacking clean labels under complex noise conditions.