论文阅读：弱监督语义分割、遥感

论文阅读：Weakly Supervised Deep Learning for Segmentation of Remote Sensing Imagery

Wang, S.; Chen, W.; Xie, S.M.; Azzari, G.; Lobell, D.B. Weakly supervised deep learning for segmentation of remote sensing imagery. Remote Sensing 2020, 12, 207.

文章定义了弱监督源：（1）单个地理标记点组成的label （2）图像级标签，证明了在每个图像的单个标记像素上训练的U-Net，以及转移到分割的U-Net图像分类器可以胜过像素级算法，例如逻辑回归，支持向量机和随机森林。

也就是说效果比一般的机器学习算法好，但是并未提及和强监督学习的对比，文章提到“While the high performance of neural networks is well-established for large datasets, our experiments indicate that U-Nets trained on weak labels outperform baseline methods with as few as 100 labels. ”意思就是说这个效果比标签少的强监督要好？

文章这样生成了两种不同的弱标签：

按照我的理解©就是