Abstract The detection and segmentation of clouds in images taken by ground based cameras is of utmost importance for a large number of applications including all-sky imager based nowcasting systems which optimize solar power plant operation, calculation of the global irradiance, estimation of the cloud base height and support of optical satellite downlink operations. Many approaches to segment clouds in camera images are published. However, comparisons of different approaches are not frequently conducted. Here, we address this question by benchmarking six different cloud segmentation algorithms on images taken by an off-the-shelf surveillance camera. The six different algorithms include (1) a color-channel threshold-based algorithm, (2) a Clear Sky Library (CSL) based approach, (3) a region growing algorithm, (4) the Hybrid thresholding algorithm (HYTA), and a (5) novel, HYTA-based development named HYTA+. Furthermore, (6) a deep convolutional neural network (FCN) is adapted via transfer learning to this problem. The segmentation results of algorithms (1) to (5) are compared to 829 manually segmented reference images. The segmentation algorithms are benchmarked on a test dataset which is divided into 16 meteorological categories. These categories cover different Linke turbidity values, solar positions and cloud cover situations. Results show that three out of the six presented segmentation methods (CSL, HYTA+ and FCN) achieve overall accuracy values above 90%. These approaches outperform the other methods and correctly segment images with a higher consistency. Fixed threshold based methods, as the multicolor criterion, HYTA or the region growing algorithm fail under certain meteorological conditions. The FCN based segmentation (6) is tested on 160 images where it delivers the best overall pixel-by-pixel accuracy of 97.0%.

中文翻译：

---

地基全天成像仪六种云分割算法的基准测试

摘要 地基相机拍摄的图像中云的检测和分割对于大量应用至关重要，包括基于全天成像仪的临近预报系统优化太阳能发电厂的运行、计算全球辐照度、估计云底。光卫星下行链路操作的高度和支持。发布了许多在相机图像中分割云的方法。然而，不经常进行不同方法的比较。在这里，我们通过对现成的监控摄像机拍摄的图像进行六种不同的云分割算法进行基准测试来解决这个问题。六种不同的算法包括（1）基于颜色通道阈值的算法，（2）基于晴空库（CSL）的方法，（3）区域增长算法，(4) 混合阈值算法 (HYTA)，以及 (5) 名为 HYTA+ 的新型基于 HYTA 的开发。此外，（6）深度卷积神经网络（FCN）通过迁移学习适应了这个问题。将算法（1）至（5）的分割结果与829幅手动分割的参考图像进行比较。分割算法在一个测试数据集上进行了基准测试，该数据集分为 16 个气象类别。这些类别涵盖了不同的林克浊度值、太阳位置和云量情况。结果表明，提出的六种分割方法中的三种（CSL、HYTA+ 和 FCN）实现了 90% 以上的整体准确度值。这些方法优于其他方法，并以更高的一致性正确分割图像。基于固定阈值的方法，作为多色标准，HYTA 或区域增长算法在某些气象条件下会失效。基于 FCN 的分割 (6) 在 160 张图像上进行了测试，它提供了 97.0% 的最佳整体逐像素精度。