高分五号可见短波红外高光谱影像云检测研究

Cloud Detection for GF-5 Visible-Shortwave Infrared Advanced Hyperspectral Image

高分五号(GF-5)卫星上荷载的可见短波红外高光谱相机(AHSI)能够同时获取330个谱段的光谱信息,对大气和陆地进行综合高光谱观测,能有效获取地物的精确信息。云的存在会对遥感影像造成污染,为了提高GF-5数据的利用率,本文结合AHSI的地物高光谱特性,研究多种下垫面背景下的云检测方法。对得到的1级产品,利用产品配套的定标系数以及光谱响应函数文件,得到各波段的大气顶层表观反射率数据。使用多种典型地物与云像元进行表观反射率的对比后发现,厚云与其他类型的像元在可见光波段具有显著差异。高光谱数据由于波段宽度窄,易受到噪声的影响,因此在进行厚云像元判定时,使用多个窄波段数据进行等效计算,得到对应的宽波段表观反射率,在此基础上使用简单的检测阈值可以将厚云筛选出来。之后使用卷云波段,筛选出潜在的薄云像元。高亮地表作为薄云检测的重点研究对象,检测时极易与薄云造成混淆,为了将薄云区域与高亮地表进行有效区分,统计不同波段之间表观反射率比值的变化,将薄云与易造成误判的高亮区域进行对比,确定最优判定波段与阈值。为了验证算法的精度,对多景AHSI影像进行目视解译,勾选出云像元区域作为基准数据。实验结果表明,本文所提方法的云检测总体精度可达91%以上,可以准确区分云与晴空区域,实现高精度的高光谱遥感影像云检测。

The visible-shortwave infrared advanced hyperspectral imager(AHSI)loaded on the GF-5 satellite can acquire information about 330 spectral bands,which facilitates the derivation of land surface properties by deploying the hyperspectral observations of both atmosphere and land surface.However,cloud contamination in remotely sensed images often limits its application.To improve the availability of the GF-5 data,this study proposed a cloud detection approach that can be applied to various situations using the hyperspectral data proved by GF-5 AHSI.The apparent reflectance at the top of the atmosphere was firstly derived with the Level-1 product by the associated radiometric calibration coefficients and spectral response function.We found that the thick cloud pixels in the images can be effectively distinguished from other land cover types at the visible spectral region after the comparison of their apparent reflectance.The broadband apparent reflectance derived with the corresponding narrow bands was used to detect thick cloud pixels,which can eliminate the impact of noise associated with the narrow-band data.On this basis,the thick clouds can be screened out using simple detection thresholds.We then obtained the candidate thin cloud pixels using the cirrus cloud band.As thin cloud pixels were generally confused with high-albedo pixels,the distinction between these two features was studied by comparing the band ratios in various combinations.The thin cloud pixels were finally detected based on the optimal band combination and the corresponding threshold.Furthermore,we adopted the visual interpretation of cloud pixels to evaluate the performance of our algorithm using several GF-5 AHSI images.The cloud pixels can be well distinguished from clear sky pixels with an accuracy of over 91%,which indicates that our approach can be used to accurately detecting clouds for hyperspectral remote sensing images.