首页 > , Vol. , Issue () : -
从遥感图像中快速准确地提取水体信息对水资源和洪涝灾害监测有重要意义。本文对多维密集连接深度卷积神经网络（DenseNet）进行改进，应用于高分一号卫星数据进行洪泽湖流域的水体识别，并采用多种评价指标，与ResNet、VGG、HRNet等神经网络模型和归一化差异水体指数法（Normalized Difference Water Index，NDWI）等的水体识别结果进行对比。结果表明，ResNet、VGG、HRNet、DenseNet等深度卷积神经网络，在遥感水体识别方面均显著优于传统的水体指数方法。本文在经典的DenseNet网络结构中增加了上采样过程和跳层连接，其水体识别效果明显优于传统的ResNet、VGG等网络结构；DenseNet网络与最新提出的HRNet网络识别的水体区域较为接近，但识别精度指标与训练效率优于HRNet。
The rapid and accurate extraction of water information from remote sensing images is of great significance to the monitoring of water resources and flooding disaster. In this paper, we have proposed a network based on the Densely Connected Convolutional Networks (DenseNet), and used GF-1 images to identify water in the Hongze Lake area. A variety of evaluation indexes were used to compare with the water identify results of the classical neural networks of ResNet, VGG, HRNet and traditional water index method (NDWI). The results show that the deep convolutional neural networks of ResNet, VGG, HRNet and DenseNet, are far superior to the traditional water index method in the remote sensing water identify. Based on the classic DenseNet structure, the up-sampling process and the skip connection are added; and the new structure shows significantly improved efficiency in water identification than the traditional ResNet and VGG network structures; the water recognition regin of the DenseNet network is close to the newly proposed HRNet network, but the evaluation indexs and training efficiency are better than HRNet.