基于高分遥感数据和深度学习的石冰川自动提取研究

石冰川是以冰岩混合物为基础形成的一类具有舌状堆积纹理的冰缘地貌，了解其分布和变化对于寒区环境研究具有重要价值，遥感技术的发展为石冰川的识别提供了有效的手段。针对石冰川发育地的偏远和调查的困难，以及其光谱特征的微弱性，提出了一种基于深度学习的石冰川识别方法，以ResNet作为训练网络，得到石冰川的图像分类模型，以国产高分一号遥感影像作为实验数据，在念青唐古拉山西段展开了应用，共识别出石冰川96条。验证结果表明：该方法具有较高的识别精度（98.72%的总体精度、89.48%的生产精度和81.77 %的用户精度），证明该方法能够有效地识别石冰川，并为在大区域开展石冰川的调查和分析提供了基础。

Automatically Identifying Rock Glacier based on Gaofen Satellite Image and Deep Learning

Rock glacier is a geomorphological landform with ligule accumulation texture that formed from mixture of ice and debris. Investigating the distribution of rock glaciers can provide effective information for studying the environment and climate change in cold regions. The development of remote sensing technology provides an effective way for identifying rock glacier. However， its execution is difficult due to the large area of rock glacier distribution as well as the similarity between rock glacier and its surroundings in spectral surface reflectance. Comparing to the traditional visual interpretation approach， this paper presented a more effective method for automatically identifying rock glaciers in high-resolution images. The method was implemented by integrating the Deep Learning development framework to build the model through interactive training from the ResNet network. The model was then applied to identify rock glaciers in GaoFen-1 images that collected in West Nyainqentanglha Mountains， where are rich of rock glaciers. Gaofen-1 images were used as the satellite data source， and 96 rock glaciers were identified in the West Nyainqentanglha Mountains. Accuracy of the results were assessed by comparing to human interpreted data， and it reported 98.72% Overall Accuracy， 89.48% Producer's Accuracy， and 81.77% User's Accuracy， suggesting that the presented method is very effective for identifying rock glaciers， and it provides a potential capability for mapping the distribution of rock glacier in large areas.