采用改进全卷积网络的“高分一号”影像居民地提取

针对现有的高分辨率遥感图像居民地信息提取精度和效率不够高的问题，提出了一种基于改进全卷积网络的“高分一号”（GF-1）遥感影像居民地提取方法。首先，通过专业的目视解译制备大量居民地训练样本；然后，将预训练过的深度卷积神经网络进行全卷积网络的改造，并以具有多尺度卷积核的Inception模块代替由全连接层改造的卷积层，达到减小网络模型参数量、增加特征表达能力的目的；最后，用制作好的高分辨率遥感图像居民地数据集进行训练和验证，生成可直接进行居民地信息提取的全卷积网络。实验结果表明，基于改进全卷积网络的方法可以实现精确有效的居民地信息提取，Kappa系数超过94%。

Extraction of residential areas in GF-1 remote sensing images based on improved fully convolutional network

The existing residential areas extraction methods suffer from low accuracy and efficiency. To overcome these problems,an improved fully convolutional network based residential areas extraction method is proposed. Firstly,large number of training samples are prepared by professional visual interpretation,and then a pre-trained deep convolution neural network is transformed into a fully convolutional network. Additionally,in order to reduce the amount of parameters and improve feature expression ability of the network,the convolutional layers transformed from fully connected layers are replaced with Inception module. Finally,the improved fully convolutional network is trained by the dataset prepared before. The experiment reveals that the proposed method achieves automatic,effective extraction of residential area information,and Kappa Coefficient is raised up to above 94%.