Abstract:

To solve the problem of unrepresentative selection of road features and low segmentation accuracy under complex terrain conditions, a high resolution remote sensing road extraction method based on the convolutional neural network（PPMU-net） is proposed. Firstly, the Gaofen-2 satellite spectral information of the three channels and the corresponding topographic information（slope, aspect and dem） are integrated into multi-feature to synthesize the remote sensing image of the six channels. Secondly, the multi-feature remote sensing images are cut and the images containing roads are screened by Convolutional Neural Network（CNN）. Finally, the remote sensing images containing only roads are sent to PPMU-net for training, and the road extraction model of high-resolution remote sensing images is constructed. Compared with U-Net and PSPNet, the proposed method can achieve better results in the extraction of high-resolution remote sensing roads and improve the accuracy of road segmentation under complex terrain conditions.

Key words: Gaofen-2 satellite, PPMU-net Neural Network, multi-features fusion, complex terrain

摘要：

针对复杂地形条件下道路特征选取不具代表性，分割精度低的问题，提出了一种基于卷积神经网络（PPMU-net）的高分辨率遥感道路提取的方法。将3通道的高分二号光谱信息与相应的地形信息（坡度、坡向、数字高程信息）进行多特征融合，合成6通道的遥感图像；对多特征的遥感图像进行切割并利用卷积网络（CNN）筛选出含道路的图像；将只含道路的遥感图像送进PPMU-net中训练，构建出高分辨率遥感图像道路提取模型。在与U-net神经网络、PSPnet神经网络相比时，所提的方法在对高分辨率遥感道路提取时能够达到较好的效果，提高了复杂地形条件下道路分割的精度。

关键词: 高分二号, PPMU-net神经网络, 多特征, 复杂地形