摘要
针对目前遥感影像云层检测中出现的漏检、误检等问题,提出了一种基于生成式对抗网络的遥感影像云检测方法。在生成网络中,考虑了云本身的形态特征,使用全卷积神经网络对图像中的云进行检测。为了获得更为精细、准确的检测结果,使用对抗网络对检测结果进行判别,并将结果反馈给生成网络,使之调整网络参数,优化检测结果。以GF-1号原始遥感影像进行检测实验,并与传统的Otsu方法和全卷积神经网络进行对比,结果表明该方法检测准确率较高,漏检现象较少。
In order to avoid missing or errors during remote sensing image cloud detection, we proposed a cloud detection method based on generative adversarial networks(GANs). In generative network, we took into account the morphological feature of clouds, and adopted the fully convolutional neural network(FCNN) to detect the clouds in the images. In order to obtain more elaborate and accurate results, we used adversarial network for judgment. The feedback of results to adversarial network made it possible to adjust network parameters and optimize detection results. We selected the original remote sensing images of GF-1 satellite for experiment, and compared with FCNN and the traditional Otsu method. The conclusion shows that the said detection method has high accuracy and less missing detection.
出处
《地理空间信息》
2018年第5期19-22,共4页
Geospatial Information
基金
安徽省地理信息智能技术工程研究中心创新资助项目(皖发改创新 (2017)587号)
