运用U-Net++模型的遥感影像多尺度云检测

Multi-scale Cloud Detection in Remote Sensing Image Based on U-Net++Model

针对传统方法将建筑物、水体误识别为云以及提取边缘轮廓不连续等问题,提出了一种结合多尺度语义分割模型的云检测方法。首先对遥感影像各个波段进行标注、裁剪、数据增广等处理,形成多个数据集;再将数据集输入U-Net++网络进行训练、直至收敛,并利用深度监督方法进行剪枝以减少待保存参数,得到各波段的云检测模型;最后通过多模型加权融合获取总体精度更高、主观效果更优的云检测结果。实验结果表明,该方法的总体精度达到91.04%、约比OTSU法高6%,且改善了提取云区轮廓效果。

Aiming at the problem of misidentifying building and water body as cloud by traditional method and the extracted edge contour discontinuity,we proposed a cloud detection method based on multi-scale semantic segmentation model.Firstly,we marked,cut and data augmented each band of remote sensing image to form multiple data sets.Then,we sent the data set to the U-Net++network for training until it converged,and carried out pruning through in-depth supervision to reduce the parameters to be saved,so as to obtain the cloud detection models of each band.Finally,we obtained the cloud detection results with higher overall accuracy and better subjective effect by multi-model weighted fusion.The experimental results show that the overall accuracy of this method is 91.04%,which is about 6%higher than that of OTSU,and the method improves the effect of cloud contour extraction.