融合两种空间光谱分辨率影像的语义分割网络

A Semantic Segmentation Network Fusion of Two Images with Different Spatial and Spectral Resolutions

针对遥感影像的语义分割领域中处理两种不同空间和光谱分辨率的影像时,通常采用传统的图像融合方法融合两种影像作为语义分割网络的输入,之后进行地表覆盖类型分类会导致效率和精度低这一问题,提出一种能输入两种不同空间和光谱分辨率遥感影像的语义分割网络,实现端到端的遥感影像融合与地表覆盖分类。首先,采用光谱特征编码器融合两种影像的光谱信息;然后,分别采用骨干网络扩大特征图的感受野、空间特征编码器保留遥感影像的空间细节信息;最后,采用解码器融合提取的特征图生成分割结果。以广东省广州市周边高分二号卫星影像作为研究区域。实验结果证明,与采用最邻近扩散(nearest neighbor diffusion pan sharpening,NNDiffuse)法融合研究区域的全色影像和多光谱影像之后,再输入到目前最优的语义分割网络Deeplab v3+相比,所提出的语义分割网络的总体分类精度和平均交并比分别提高了3.99%和5.26%,验证了该网络的有效性。

In the field of remote sensing image semantic segmentation,when processing two kinds of images with different spatial and spectral resolutions,the traditional image fusion method is usually used to fuse the two images as the input of semantic segmentation network to classify the land cover types.This method has low efficiency and accuracy.To solve this problem,a semantic segmentation network which can input two kinds of remote sensing images with different spatial and spectral resolutions is proposed to realize end-to-end remote sensing image fusion and land cover classification.Firstly,spectral feature coder is used to fuse the spectral information of two kinds of images,and then backbone network is used to expand the receptive field of feature map,and spatial feature encoder is used to retain the spatial details of remote sensing image.Finally,the decoder is used to fuse the extracted feature map to generate segmentation results.In this paper,the GF-2 satellite image around Guangzhou city,Guangdong province,is used as the research area.After fusing the panchromatic and multispectral images of the study area,input them into the current optimal semantic segmentation network Deeplab v3+.Experimental results show that the overall accuracy and mean intersection over union of the proposed semantic segmentation network are improved by 3.99%and 5.26%,respectively,which verifies the effectiveness of the network.