基于先验地表反射率数据库支持的动态阈值云检测算法(UDTCDA)可以显著提高卫星数据的云检测精度。为进一步提高其在波段相对较少的高空间分辨率卫星数据云检测应用中的精度,改进了UDTCDA中先验地表反射率数据与待检测卫星数据的空间匹配方法。与原方法使用重采样达到空间分辨率一致不同,该方法根据待检测影像高空间分辨率的特点,采用逐像元空间地理坐标配准的方法与真实地表反射率数据进行配准,然后进行云像元检测。该方法保留了高分辨率影像空间分辨率的优势,可以有效降低空间重采样造成的像元信息丢失。分别使用资源3号、高分1号、高分2号和高分4号高分辨率卫星数据开展云检测实验。通过遥感目视解译的方法对结果进行精度验证,并与UDTCDA云识别结果进行对比。结果表明,改进后的算法能以较高的精度识别不同高分辨率卫星影像中的云,总体精度可达到93.92%,对于碎云和薄云具有整体较高的识别精度,漏分误差和错分误差分别低于10.40%和9.57%。

With the support of a pre-calculated land surface reflectance database, the universal dynamic threshold cloud detection algorithm (UDTCDA) can significantly improve the cloud detection accuracy of satellite data. To further improve its precision in the application of cloud detection for high spatial-resolution satellite data with relatively few bands, we improve the spatial matching method between the prior surface reflectance and the satellite observed reflectance. Different with the directly resample method in the UDTCDA, the pixel-by-pixel registration method is adopted to realize the matching between the satellite image and surface reflectance image. This approach preserves the spatial resolution advantage of high resolution images, and effectively reduces the loss of pixel information caused by spatial resampling. Four high-resolution satellite data, namely ZY-3, GF-1, GF-2 and GF-4, are used in cloud detection experiments. The cloud detection results of the improved UDTCDA are verified by the visual interpretation cloud results, and compared with the original UDTCDA cloud results. Results show that the improved algorithm can accurately identify different kinds of clouds in different high-resolution satellite images with an average accuracy of 93.92%. Especially for the broken and thin clouds, the accuracy is significantly improved with overall low omission and commission errors less than 10.40% and 9.57%, respectively.