基于知识引导的遥感影像融合方法

Knowledge-based remote sensing image fusion method

遥感图像融合技术能够将含有互补信息的多源图像进行融合,从而得到内容更丰富、光谱质量更高的图像,是遥感应用的关键和基础。针对遥感影像融合过程中容易出现的光谱失真和空间结构失真问题,该文以注意力机制为基础,利用归一化植被指数(normalized difference vegetation index,NDVI)和归一化水指数(normalized difference water index,NDWI)作为先验知识,构建了基于知识引导的遥感影像融合模型(remote sensing image FuseNet,RSFuseNet)。首先,基于高通滤波能够充分提取边缘纹理细节的优势,构建了高通滤波模块提取全色图像高频细节信息;其次,提取多光谱图像的NDVI和NDWI信息;再次,构建自适应挤压激励(squeeze-and-excitation,SE)模块,对输入特征进行重标定;最后,将自适应SE模块与卷积单元结合,对输入特征进行融合处理。以高分六号遥感影像作为数据源,选择施密特正交化(gram-schmidt,GS)、主成分分析(principal component analysis,PCA)、深度网络结构的图像融合网络(a deep network architecture for pan-sharpening,PanNet)、以卷积神经网络为基础的融合网络(pansharpening by convolutional neural networks,PNN)模型作为对比模型开展实验,实验结果表明:该文提出模型的峰值信噪比(peak signal to noise ratio,PSNR)指标(40.5)和结构相似性(structural similarity,SSIM)指标(0.98)均优于对比模型,表明这一方法在遥感影像图像融合方面具有明显的优势。

The remote sensing image fusion technology can combine multi-source images containing complementary information to obtain images with richer content and higher spectral quality,thus it is the key and foundation of remote sensing applications.Aiming at the problems of spectral distortion and spatial structure distortion that are prone to occur in the process of remote sensing image fusion,the knowledge-based remote sensing image FuseNet(RSFuseNet)was constructed based on the attention mechanism and using normalized difference vegetation index(NDVI)and normalized difference water index(NDWI)as prior knowledge.Firstly,considering that the high-pass filtering can fully extract edge texture details,a high-pass filtering module was constructed to extract high-frequency details of panchromatic images.Secondly,NDVI and NDWI were extracted from multi-spectral images.Then,an adaptive squeeze-and-excitation(SE)module was constructed to recalibrate the input features.Finally,the adaptive SE module was combined with the convolution unit to perform fusion processing on the input features.The experiment was conducted using Gaofen 6 remote sensing image as the data source,and selecting Gram-Schmidt(GS)transformation,principal component analysis(PCA),a deep network architecture for pan-sharpening(PanNet),and pansharpening by convolutional neural networks(PNN)models as comparative models.The experimental results show that the peak signal to noise ratio(PSNR)index(40.5)and the structural similarity(SSIM)index(0.98)of the RSFuseNet model are better than those of comparative models,indicating that the method in this study has obvious advantages in remote sensing image fusion.