密集连接网络在SAR与多光谱影像融合中的应用

Application of densely connected network in SAR and multispectral image fusion

为克服单一卫星传感器成像的不足,提出了基于密集连接网络的合成孔径雷达(SAR)与多光谱影像的融合算法。首先分别对SAR影像与多光谱影像进行预处理,使用双三次插值法重采样到同一空间分辨率下,然后使用密集连接网络来分别提取影像的特征图,并采用区域能量最大的融合策略将深度特征进行融合,将融合图像输入到预训练的解码器中进行重构,获得最终融合影像。实验采用哨兵1号SAR影像、Landsat-8影像和高分1号卫星影像进行验证,并与基于成分替换、基于多尺度分解和基于卷积神经网络的代表性方法进行对比试验。实验结果表明,基于密集连接网络的融合算法在多尺度结构相似度指标上的精度高达0.9307,在其他多种评价指标上也都优于其他融合算法,SAR影像的细节信息和多光谱影像的光谱信息得到很好的保留。

To overcome the shortcomings of single satellite sensor imaging,a fusion algorithm for synthetic aperture radar(SAR)and multispectral images based on densely connected networks is proposed herein.Firstly,the SAR and multispectral images are preprocessed separately,and the bicubic interpolation method is used to resample the same spatial resolution.Then,the densely connected network is used to extract the feature maps of the image separately,and the fusion strategy with the largest regional energy is used to combine the depth features.The fused image is input to a pre-trained decoder for reconstruction to obtain the final fused image.The experiment uses Sentinel-1 SAR images,Landsat-8 images,and Gaofen-1 satellite images for verification and draws comparisons with methods based on component substitution,those based on multiscale decomposition,and those based on deep learning.Experimental results indicate that the accuracy of the fusion algorithm based on densely connected networks in terms of the multiscale structural similarity index is as high as 0.9307,and it is better than other fusion algorithms in terms of other evaluation indexes.Detailed information of SAR images and multispectral images are well preserved.