光学卫星遥感数据在获取过程中易受云层干扰, 云区识别是光学遥感数据应用及分析的一个基础但重要的步骤, 高效的云区识别技术对节省数据收集成本和提高数据利用效率具有较强的现实意义.同态滤波算法是经典的基于单幅影像的云区识别方法之一, 该算法具有计算快速方便、云区检测精度较高的优点, 然而识别的云区范围极大程度取决于同态滤波器截止频率的位置.同态滤波截止频率通常采用经验值, 显然经验截止频率无法适应批量遥感数据的自动处理需求.针对以上问题, 本文通过建立输入影像频谱能量与截止频率的关系, 结合白度指数(Whiteness Index)和形态学算子, 实现对国产高分辨率光学卫星高分一号(GF-1)遥感数据的批量云区识别处理.与传统同态滤波方法相比, 该算法能根据影像频谱能量自适应判定同态滤波时采用的截止频率, 具有更强的适用性.通过对98景GF-1多光谱数据进行随机点人工目视标记精度检验, 精度检验结果表明该算法对云区有较好的检测效果, 总体识别精度达93.81%.该算法对GF-1遥感数据能进行批量化云区检测, 获得高精度的云区掩膜结果, 并有效降低高反射率地物造成的误识率.

Cloud detection for remote sensing imageries is a fundamental as well as significant step due to the inevitable existence of large amount of clouds in the optical remote sensing data. A highly efficient cloud detection approach is capable of saving data collection cost and improving data utilization efficiency. Homomorphic filtering algorithm is one of the most commonly methods that based on single-scene image for detecting clouds. This algorithm has the advantage of fast computation and high accuracy in cloud areas detection. However, the detected cloud areas are heavily dependent on the cut-off frequency of the homomorphic filter. The homomorphic filtering progress usually uses cut-off frequency with empirical value which might not be applicable to large amount of intricate input data. Therefore, this paper aims to construct the relationship between the image spectra power and the filter cut-off frequency. Based on the domestic high spatial resolution optical remote sensing data GF-1, this research makes the detection of clouds could be process to achieve a bulk deal. Our approach make the cut-off frequency self-adaptive changes rather than used empirical value when compared with the traditional homomorphic filtering, thus it could be able to meet more complicated scenarios. Further, the post-processing steps including whiteness index, spectral threshold, and morphological opening and closing operators are applied to coarse cloud mask to optimize results. We have tested on 98 GF-1 high resolution multispectral imageries, results indicated that our approach is capable of detecting cloud as well as haze areas with high accuracy of 93.81%. This novel self-adaptive method shows its great application potential for real-time and high efficient cloud detection, meanwhile reduced the error detection rates caused by high reflectance ground objects.