A cloud and snow detection method of TH-1 image based on combined ResNet and DeepLabV3+

• Main Authors: ZHENG Kai, LI Jiansheng, YANG Jianfeng, OUYANG Wen, WANG Gaojie, ZHANG Xun
• Format: Article
• Language: zho
• Published: Surveying and Mapping Press 2020-10-01
• Series: Acta Geodaetica et Cartographica Sinica
• Subjects: satellite image; cloud snow detection; th-1; resnet; deeplabv3+
• Online Access: http://xb.sinomaps.com/article/2020/1001-1595/2020-10-1343.htm

Cloud and snow detection is an important part of satellite remote sensing image processing, and also a key step for its following analysis and interpretation. In this paper, a simultaneous cloud and snow detection method for satellite remote sensing images based on ResNet and DeepLabV3+ fully convolutional neural network is proposed. The ResNet50 backbone is used, and the DeepLabV3+ network structure is optimized and improved according to the characteristics of cloud and snow on TH-1 remote sensing image. The ELU activation function, Adam gradient descent method and Focal Loss function are used to speed up convergence and improve segmentation accuracy. The network is trained and tested with the cloud and snow image dataset of TH-1 satellite. The experimental results show that it has strong robustness compared with Otsu method, and the detection accuracy of proposed method exceeds FCN-8s and original DeepLabV3+ network, meanwhile the detection speed of proposed method is faster than original DeepLabV3+, which can be applied to a variety of different remote sensing images through some adjustment and has favorable application prospects.