一种改进U-net网络的新增建设用地提取方法

A Method of New Construction Land Extraction Based on Improved U-net Network

针对传统的新增建设用地提取主要依赖于人工目视解译,任务繁重,人力耗费过大,而现有的全卷积神经网络提取方法存在特征表达能力不够,易引起过拟合的问题,提出了一种改进U-net网络的高分辨率遥感影像新增建设用地提取方法。基于高分二号影像并结合历史土地利用变更调查成果构建新增建设用地样本数据集;同时,采用新型的激活函数、批标准化以及退化学习率的方法进行网络设计,以防止过拟合。在下采样的过程中加入空洞卷积的算法扩大感受野以感受更多的地物信息,提取更详细的地物特征。结果表明,本研究方法提取新增建设用地的F1值达到了0.88,明显优于FCN与U-net的结果,在新增建设用地的高精度自动提取和业务化应用上具有较高潜力。

In view of the problems that the traditional extraction of newly-added construction land mainly depends on manual investigation and visual interpretation,with heavy tasks and excessive labor consumption,and the existing full convolution neural network extraction method has insufficient feature expression capability and is easy to cause over-fitting,this paper proposes an extraction method of newly-added construction land from high-resolution remote sensing images by improving U-net.It is based on the image of GF-2 and combined with the investigation results of historical land use changes.In the network design of this study,a new method of activation function,batch normalization and degraded learning rate is used to prevent the problem of overfitting,and the algorithm of Atrous convolution is added to the process of sampling to expand the receptive field to feel more material information and to extract more geographical features.The experimental results show the F1 value of new construction land extracted by this method reaches 0.88,which is significantly better than that of FCN and U-net,and it has high potential in the high-precision automatic extraction and business application of new construction land.