Landsat影像冰川提取的上下文感知语义分割网络法

针对冰川提取存在云阴影、山体阴影、结冰湖泊等同物异谱、同谱异物导致难以有效区分的问题，设计了一种用于冰川提取的上下文感知深度学习语义分割网络。首先引入resnet50作为基准编码网络，以实现冰川特征提取的精度和效率平衡，其次针对现有语义分割网络存在上下文信息学习不足方面，设计了包括空洞卷积组块和最大池化组块的上下文信息提取层，以便更好地提取冰川的上下文信息。选择多景样本数据和验证数据的多源遥感影像进行试验，与现有基于特征指数的冰川提取方法、其他深度学习语义分割网络方法进行定性和定量对比，结果表明本文网络方法在结冰湖面等误提取，阴影的漏提取，以及提取结果完整性等方面，具有较好的效果，验证了本文方法的有效性与稳健性。

Landsat image glacier extraction based on context semantic segmentation network

According to the glacier characteristics of remote sensing image, a context-aware deep learning semantic segmentation network for glacier extraction is proposed based on the glacier characteristics of remote sensing image. Firstly, resnet50 is introduced as the feature extraction network to achieve the accuracy and efficiency balance of glacier feature extraction. Secondly, the context-information learning of the existing semantic segmentation network is designed. The context information including the dilated-convolutional block and the max-pooled block is designed to better extract the context information of the glacier. Multiple remote sensing trained images and tested images are selected for experiment, which is qualitatively and quantitatively compared with the existing glacier feature index extraction method and other semantic segmentation network methods. The results show that the network method in the frozen lake surface, the leakage of the mountain shadow, cloud shadow and the integrity of the extraction results have a good effect, which verifies the effectiveness and robustness of the proposed method.