基于W-Net的高分辨率遥感卫星图像分割

W-Net-Based Segmentation for Remote Sensing Satellite Image of High Resolution

遥感卫星图像因其实时性和客观性可以提供准确的地物位置信息,而在农业生产和环境保护等领域得到广泛的运用。针对海量的遥感卫星图像难以识别的问题,文中使用基于卷积神经网络的图像分割方法提取遥感图像中典型土地光谱信息和空间信息来识别遥感卫星图像。首先,通过裁剪遥感图像数据集和标注数据生成实验数据,对数据中的类别进行统计,使用过采样处理数据不平衡的问题;然后,在U-Net网络中添加自上而下的特征金字塔结构,并且结合全局上下文模块,提出名为W-Net的网络结构进行训练;最后,使用影像重叠策略对大尺寸的遥感卫星图像进行识别。与3种流行的语义分割网络、2种遥感卫星图像分割专用的网络进行对比,文中提出的W-Net在“2017CCF卫星影像的AI分类与识别竞赛”卫星图像识别模块中取得了74.7%的平均重叠度、95.1%的分类精度,在Massachusetts建筑物分割模块中取得了69.6%的查准率和79.9%的查全率,其分割准确率和平均重叠度在6种网络中均为最高。实验表明,使用特征金字塔结构和全局上下文模块能够提升语义分割网络的分割准确率,该方法用于遥感卫星图像分割是可行的。

Remote sensing satellite image,which can provide accurate location information because of instantaneity and objectivity,has been widely used in agricultural production,environmental protection and other fields.In order to solve the difficulties in recognizing the mass remote sensing satellite images,the image segmentation method based on convolutional neural networks was used to extract the typical land spectral information and spatial information in the remote sensing image and to identify the remote sensing satellite image.Firstly,the experimental data was generated by clipping remote sensing ima-ge data and annotating data.The categories in the data were counted and oversampling was used to deal with the problem of data imbalance.Then,a new image segmentation network named W-NET was proposed to train the data by adding a top-down feature pyramid structure to the U-Net framework combining with global context module.Finally,large-scale remote sensing satellite images were recognized by image overlap strategy.Compared with three popular semantic segmentation networks and two special networks for remote sensing satellite image segmentation,our method achieves 74.7%mean IoU and 95.1%accuracy in 2017 AI classification and recognition contest of CCF satellite images,and achieves 69.6%precision and 79.9%recall in Massachusetts building segmentation task.W-Net has the highest accuracy and mean IoU among the six networks.Experimental results show that the feature pyramid structure and global context module can improve the segmentation accuracy of the semantic segmentation network,and this method is feasible for remote sensing satellite image segmentation.