Abstract:

Recently, many deep learning methods are used in hyperspectral image (HSI) pixel classification task. However, spectral noise and spatial noise still exist in HSI simultaneously even with a series of preprocessing. Most of these deep learning methods only involve the noise filtering in spatial domain while paying no attention to that in spectral domain. So spectral noise is preserved in deep features after deep feature learning process, which   degrades the pixel classification accuracy. Given the excellent classification accuracy of spectral-spatial network (SSN), this paper proposes two deep feature learning models fused spectral noise filtering with it, called Saviztky-Galoy filtering SSN (SG-SSN) and median filtering SSN (MF-SSN), respectively. The models remove spectral noise by using Saviztky- Galoy filtering or median filtering on spectral domain first. Then, the denoised image is fed into SSN to learn deep features of HSI for its excellent classification effects. Experiments on two popular HSI datasets demonstrate the effectiveness of SG-SSN and MF-SSN, and classification result is measured by overall accuracy (OA), average accuracy (AA) and Kappa coefficient. How the appropriate layer number of model affects model performance is also discussed.

Key words: hyperspectral image, SG filter, median filter, spectral-spatial network (SSN)

摘要：

目前，大量深度学习方法被用于提取超光谱图像特征来完成像素点分类任务。这些超光谱图像虽然经过一系列前期处理，但是仍然包含很多噪声，噪声同时存在于光谱域和空间域。由于这些方法在提取图像深层特征时大多只进行了空间域去噪处理，没有进行光谱域噪声的滤除，最终获得的像素点深度特征中带有噪声，降低了分类精度。结合分类效果出色的空谱学习网络（spectral-spatial network，SSN），提出两种融合光谱域滤波的深层特征学习模型——Saviztky-Galoy滤波空谱学习网络（SG-SSN）和中值滤波空谱学习网络（MF-SSN）。两个模型首先分别对图像光谱域进行SG滤波和中值滤波，接着对滤波后的图像使用SSN学习图像深层特征。在两个标准数据集上，使用总体精度（overall accuracy，OA)、平均精度（average accuracy，AA)以及Kappa系数度量分类效果，发现融合光谱滤波的SG-SSN和MF-SSN模型能有效提升空谱学习网络的分类精度。利用实验比较和分析了深层特征学习中模型层数对模型性能的影响。

关键词: 超光谱图像, SG滤波, 中值滤波, 空谱网络（SSN）