王新念, 马毅, 刘荣杰, 等, 2024. 融合多尺度卷积和侧窗滤波的HY-1C CZI云检测方法[J]. 海洋科学进展, 42(1): 102-115. doi: 10.12362/j.issn.1671-6647.20220913001.
引用本文: 王新念, 马毅, 刘荣杰, 等, 2024. 融合多尺度卷积和侧窗滤波的HY-1C CZI云检测方法[J]. 海洋科学进展, 42(1): 102-115. doi: 10.12362/j.issn.1671-6647.20220913001.
WANG X N, MA Y, LIU R J, et al, 2024. HY-1C CZI cloud detection method based on multi-scale convolution and side window filtering[J]. Advances in Marine Science, 42(1): 102-115. doi: 10.12362/j.issn.1671-6647.20220913001
Citation: WANG X N, MA Y, LIU R J, et al, 2024. HY-1C CZI cloud detection method based on multi-scale convolution and side window filtering[J]. Advances in Marine Science, 42(1): 102-115. doi: 10.12362/j.issn.1671-6647.20220913001

融合多尺度卷积和侧窗滤波的HY-1C CZI云检测方法

HY-1C CZI Cloud Detection Method Based on Multi-Scale Convolution and Side Window Filtering

  • 摘要: 海洋一号C(HY-1C)卫星是中国首颗海洋水色业务卫星,其搭载的海岸带成像仪(Coastal Zone Imager, CZI)具有大幅宽、短重访周期的优势,可实现海洋和海岸带的大面积观测。作为光学传感器,CZI受云影响严重,准确识别云是CZI数据处理的关键,但是CZI缺少红外和短波红外等对云敏感的波段,云检测难度大。针对该问题,本文提出一种融合多尺度卷积和侧窗滤波的轻量化云检测方法,该方法通过多尺度卷积获取云的不同尺度特征,通过侧窗滤波突出边缘特征,减少椒盐噪声的影响,提升云边缘检测的精度。实验结果表明,本文所提出的方法可有效进行云检测,在云边缘提取方面表现较好,F1-score达92.77%,Kappa系数达0.89,与现有云检测方法相比优势明显,且模型训练速度快、参数量少,可为HY-1C CZI遥感影像处理提供有力支撑。

     

    Abstract: The HaiYang-1C (HY-1C) satellite is China’s first operational ocean color satellite. The Coastal Zone Imager (CZI) aboard on the HY-1C satellite, which has the advantages of wide range and short revisit period, has been widely used for large-scale ocean and coastal zone observation. As an optical sensor, CZI is seriously affected by clouds. The accurate detection of clouds is crucial for CZI data processing and application. However, CZI lacks cloud-sensitive bands such as infrared and short-wave infrared, which makes cloud detection difficult. So, this paper proposes a cloud detection method that combines multi-scale convolution and side window filtering. The multi-scale convolution is used to extract different scale features of clouds and the side window filtering is used to enhance the edge features and reduce the influence of image noise. The experimental results show that the proposed method can effectively detect cloud and performs well in cloud edge extraction, with F1-score of 92.77% and Kappa of 0.89. Compared with the currently existing cloud detection algorithms, the proposed method has obvious advantages of fast model training speed and less parameters, which will provide support for HY-1C CZI remote sensing image processing.

     

/

返回文章
返回