Abstract Denoising of spectra has been a great challenge in hyperspectral image analysis. Near-infrared hyperspectral images of milk powder, rice flour and soybean flour were acquired and denoising in the spectral domain were studied. Noise free spectra and noises were simulated based on sample pixel-wise spectra. The noisy spectra with signal to noise ratio (SNR) around 45 ​dB (similar to real pixel-wise spectra) were simulated. The simulated noisy spectra were preprocessed by traditional methods as moving average smoothing (MAS), Savitzky-Golay smoothing (SGS), wavelet transform (WT) and empirical mode decomposition (EMD). The basic denoising autoencoder (DAE-1) and the stacked DAE (DAE-2) were studied for denoising. The noisy spectra with SNR around 35 ​dB and 55 ​dB were further simulated to explore the effectiveness of DAE based methods. DAE-1 and DAE-2 performed better than the other methods, with higher SNR, lower mean squared error (MSE) and mean absolute error (MAE). The developed DAE methods were applied to real-world pixel-wise spectra with good performances. The overall results proved the feasibility of using DAE based methods for noise reduction in the spectral domain of hyperspectral images, and the DAE based methods have great potential to be extended to spectral denoising of other vibrational spectroscopy techniques.

中文翻译：

---

使用去噪自编码器方法在高光谱图像的谱域中降噪

摘要 光谱去噪一直是高光谱图像分析中的一大挑战。获取奶粉、米粉和豆粉的近红外高光谱图像，并研究了光谱域的去噪。基于样本像素光谱模拟无噪声光谱和噪声。模拟了信噪比 (SNR) 约为 45 dB（类似于真实像素级频谱）的噪声频谱。模拟的噪声谱通过移动平均平滑（MAS）、Savitzky-Golay平滑（SGS）、小波变换（WT）和经验模态分解（EMD）等传统方法进行预处理。研究了基本去噪自动编码器 (DAE-1) 和堆叠 DAE (DAE-2) 以进行去噪。进一步模拟了 SNR 约为 35 dB 和 55 dB 的噪声频谱，以探索基于 DAE 的方法的有效性。DAE-1 和 DAE-2 的性能优于其他方法，具有更高的 SNR、更低的均方误差 (MSE) 和平均绝对误差 (MAE)。开发的 DAE 方法应用于现实世界的像素级光谱，具有良好的性能。总体结果证明了在高光谱图像的光谱域中使用基于 DAE 的方法进行降噪的可行性，并且基于 DAE 的方法具有很大的潜力可以扩展到其他振动光谱技术的光谱去噪。