深度U-Net网络在遥感山地冰川边界分割中的应用

Applied of deep U-Net in remote sensing mountainglacier boundary segmentation

冰川变化会对当地的气候环境、水资源环境产生重要影响,随着遥感技术的发展,通过遥感图像进行冰川提取成为相关研究的主要手段,相比于人工目视解释法会出现的耗时长、效率低、主观因素大等问题,深度学习有着一定的优势。该文基于传统U-Net语义分割网络进行冰川分割,但因受限于冰川训练集缺失,真彩色图像在冰川地区进行分割会有较大的干扰,无法凸显冰川的特征,冰川分割效率较低。因此,利用冰川的矢量数据,基于Landsat 8遥感卫星图像,建立成对的假彩色冰川分割训练集,充分利用遥感多波段图像的优势,强化冰川特征信息。同时,通过添加不同波段组合的假彩色图像,丰富冰川的分割信息,并利用Inception v1深度学习模块将两种特征信息进行融合,提升冰川分割的准确性。实验结果表明,所提方法可以有效分割出冰川范围,相比于其他深度学习方法,分割准确性有了一定的提高。

Glacier change has always been the key direction of Geographical research and has an important impact on the local climate environment and water resources environment.No matter what direction of research,the first problem is to extract the range of glaciers.With the development of remote sensing technology,extracting the range of glacier from remote sensing images has become the first way of related research.Compared with the artificial visual interpretation method which has long time-consuming,low efficiency and large subjective factors,deep learning has certain advantages.This paper uses the traditional deep learning semantic segmentation network to segment glaciers.However,because of lack of glacier training dataset,the traditional u-net semantic segmentation network has great interference in the segmentation of true color images in glacier areas,which cannot highlight the characteristics of glaciers and has low efficiency in glacier segmentation.Therefore,through the vector data of glaciers and based on Landsat 8 remote sensing satellite images,we establish glacier segmentation training dataset,which can make full use of the advantages of remote sensing multi band images and strengthen the glacier feature information.At the same time,by adding additional color images combined with different bands,it enriches the glacier segmentation information.And it will improve the accuracy of glacier segmentation by using two feature information in deep learning module.The experimental results show that the proposed method can segment the glacier range,which is better than other depth learning methods.