基于Sentinel-2波段/产品的图像云检测效果对比研究

高时空分辨率遥感影像对精细尺度土地利用和土地覆盖变化研究具有重要意义，然而云噪声的存在给影像的解译和分析带来了一定的挑战，因此云噪声检测作为一项基础性工作在影像解译与分析过程中扮演了非常重要的作用。QA60产品被广泛推荐为Sentinel-2卫星影像的常规云检测产品，然而，我们最近的研究发现基于QA60产品的云检测通常会出现明显的云噪声漏检测现象。为探索提高Sentinel-2卫星影像云噪声检测效果的方法，基于Google Earth Engine（GEE）平台，结合Sentinel-2卫星影像2A级（L2A）数据的2个云相关波段（B1和B9）以及4个产品（QA60、AOT、MSK_CLDPRB和SCL产品），设计相应分割算法，并以典型区为案例，从影像波段特性、云微物理学等角度分析了相关波段/产品云检测结果的空间分布格局及差异，并借助定量化指标对云检测效果进行评价。结果表明：①在云检测算法方面，B1和B9波段采用的动态阈值分割算法稳健性较好，检测结果能在一定程度上拟合其波段特性，并合理地表征相应波段的云噪声；②从云检测空间分布看，AOT产品效果较差，B9波段和QA60产品云检测可靠性较低，B1、SCL、MSK_CLDPRB 3个波段/产品的云检测潜力较强；③从评价结果看，B1波段的云检测效果最佳，对云噪声的敏感度高于其他云相关波段/产品，查准率、查全率、准确率和F1分数均大于0.90，稳健性最好。本文验证了气溶胶（B1）波段对云检测的精确性、稳健性和敏感程度，有望为进一步优化常规云检测算法提供新参考。

Comparsions of Image Cloud Detection Effect based on Sentinel-2 Bands/Products

Fine spatial and temporal resolution remote sensing images are of great significance for the study of fine scale land use and land cover change. However， the presence of cloud noise poses challenges to image interpretation and analysis. Therefore， cloud detection plays a very important role in image interpretation and analysis. The QA60 has been widely recommended as the cloud detection product for Sentinel-2 （S2） images. However， our recent research found that there was a obvious cloud noise omission in the cloud detection results based on the QA60 product. To improve the ability of the cloud noise detection in S2 satellite images， this study developed cloud segmentation algorithms based on two cloud-related bands （B1 and B9） and four products （QA60， AOT， MSK_CLDPRB and SCL products） of 2A-level （L2A） data with the help of Google Earth Engine （GEE） platform. By taking three typical regions as cases， we investigated the spatial patterns and differences of different cloud detection results from the perspectives of image band characteristics and cloud microphysics. Further， we also evaluated accuracies of the different cloud detection reults. Results showed that： （1） From the perspective of cloud detection algorithm， the dynamic threshold segmentation algorithms used in B1 and B9 bands presneted the good robustness. And the detection results could largely match characteristics of corresponding bands and reasonably captured the cloud patterns. （2） For the spatial distributions of cloud， a relatively poor performance was observed in AOT product. The reliabilities of B9 band and QA60 product were relatively low. By contrast， the cloud detection potentials of B1， SCL and MSK_CLDPRB were much stronger. （3） B1 band gave the best cloud detection effect， and its sensitivity was much stronger than those of other cloud-related bands/products. Also user’s accuracy， product’s accuracy， overall accuracy and F1 score were all greater than 0.90， implying the robustest performance. This study estimated the accuracy， robustness and sensitivity of B1 （i.e. aerosol band） for cloud detection of S2 images. These findings are expected to provide some new references for further optimizing the cloud detection of satellite images.