%0 Journal Article
%T 基于弱监督与光谱指数的哨兵2号图像裸土提取<br> Bare Soil Extraction from Sentinel-2 Images with Weakly Supervised and Spectral Index
%A 赵燚海
%J Journal of Sensor Technology and Application
%P 37-45
%@ 2331-0243
%D 2024
%I Hans Publishing
%R 10.12677/JSTA.2024.121005
%X 利用哨兵2号图像进行裸土区动态制图对环境管理和生态恢复具有重要意义。深度学习技术已经彻底改变了土地利用与土地覆盖分类的方法，包括裸露土地区域的映射。然而，目前的深度学习方法存在两个主要问题，即标记成本高和模型性能较差的问题。在本文中，我们开发了一种新的深度语义分割网络O-Net来解决当前的问题。O-Net具有典型的编码器–解码器结构，其中编码器和解码器都可以实例化为一个特定的全卷积网络，其中编码器用于预测裸土面积的提取，解码器用于重建输入图像斑块的裸土指数。我们基于弱监督(如不完全或不准确的标签)来训练网络参数。在我们的注释浙江数据集上的实验表明，所提出的方法可以实现较先进的性能。<br />
Dynamic mapping of bare soil areas using sentinel-2 images is important for environmental management and ecological restoration. Deep learning technology has revolutionized the approaches of Land Use and Land Cover classification including the mapping of bare soil areas. However, there are two main problems with current deep learning methods, namely the high cost of labeling and poor model performance. In this paper, we develop a new deep semantic segmentation network called O-Net for tackling the current problems. The O-Net has a typical encoder-decoder structure, among which both the en-coder and the decoder can be instantiated as a specific fully convolutional network, where the en-coder is used to predict the extraction of bare soil areas and the decoder is used to reconstruct the bare soil indices of the input image patches. We train the network parameters based on weak supervision such as incomplete or inaccurate annotations. Experiments over our annotated Zhejiang Dataset demonstrate that the proposed method can achieve more advanced performance.
%K 裸土提取，多光谱影像，深度学习，弱监督，裸土指数<br> Bare Soil Extraction
%K Multispectral Images
%K Deep Learning
%K Weak Supervision
%K Bare Soil Index
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=78740