多源星载光学遥感数据的薄云雾检测与去除研究

The Earth observation using spaceborne optical sensors can be affected by clouds, and remotely sensed optical data collected may not reveal the true spectral characteristics of a ground target. Thus, clouds adversely impact the data quantity, quality of radiometric calibration, and possible scientific researches and applications. Since clouds are spatiotemporally heterogeneous, their identification and removal in optical data are not entirely resolved yet. Algorithms to remove thin clouds using the single image is able to overcome the limitation of algorithms to replace cloud-contaminated data with cloud-free data acquired by the same or different sensors at another time. Thus, the former algorithms are much more valuable than the latter ones because errors caused by multi-temporal or multi-sensor differences are avoided. In this proposal, focusing on thin clouds, and their identification and removal, and using a single remotely sensed dataset, we develop the theoretical framework for the cloud identification and removal. Under the framework and based on physical characteristics of thin clouds, the mechanism of the interaction between the cloud components and electromagnetic radiation is studied, and the imaging model for remotely sensed optical data developed. Then, factors that are the most sensitive to thin clouds and fogs are delineated for remote sensors from various satellite platforms. After the identification and removal of the clouds and fogs, the objective is achieved. In addition, in the request for proposal of 2017 National Natural Science Foundation of China, one of the key research initiatives is the development of new theories and methods for quantitative remote sensing modeling and model-inversion. Our proposed cloud-removal research will definitely contribute to the initiative because remote sensing data contaminated by clouds may become usable. The data-pool for the development of the initiative is expanded after the cloud removal.

光学遥感卫星在对地观测时易受到大气中云雾的影响，获取的数据无法真实地反映地物的光谱信息，影响光学遥感数据的数量、辐射定标质量及科研和应用范围。由于云具有时空异质性，遥感数据中的薄云检测及去除是一个未解决的难题。基于单幅图像的薄云去除算法能克服多时相替代去除云算法存在的局限性，避免了多时相或多传感器数据差别带来的误差，具高的研究意义和实用价值。因此我们以薄云为研究对象，构建基于单幅图像的多源星载遥感数据云检测及去除算法理论框架，在该框架下，以云的物理特性为基础，阐明其各组成成分与电磁辐射的相互作用机理，研发有云条件下的遥感数据成像模型，针对多源星载传感器系统参数各自的特点分别提取对薄云雾敏感的特征因子，识别及去除云，解决光学数据受云影响而难以直接应用的难题。同时，2017年国家自然科学基金委将重点支持定量遥感建模和参数反演新理论和新方法，本项目的算法、成果可为其进一步拓展可用、有效的数据源。

云雾的存在极大的影响了光学遥感图像数据的质量与应用，如何检测并去除光学遥感影像中的云成分具有高理论和实际应用的意义，因此，我们研究薄云雾条件下的遥感影像成像模型，构建薄云雾的参数反演及去除的理论框架，提供快速、准确的光学遥感影像的薄云雾检测与去除方法。.通过分析大量光学图像可见光波段的光谱特征，我们发现在无云条件下可见光波谱范围内相邻波段数据之间具有很强的线性统计关系，并对云雾十分敏感。因此利用该特征，结合简化的辐射传输模型进行了薄云雾大气表观反射率反演方程的推演，实现了利用图像自身统计特征反演薄云雾的大气表观反射率。同时把这种薄云雾模拟算法引入了其它去云算法，模拟结果能够一定程度上代替卷云探测波段，扩大算法的应用范围。另外、从多个方面改进了基于辐射传输模型的去云算法。首先简化了辐射传输模型，将薄云去除简化为求卷云波段与各可见光波段中云成分的系数关系的问题；其次对模拟出的云成分，进行取值、滤波操作，得到纠正后的云成分，提高了算法的精度，改善了原算法中的颜色失真问题；最后探究了不同土地利用和土地覆盖类型对于算法效果的影响，提出将研究区域分为植被、裸土和水体分别应用算法。.对于不依赖图像自身统计经验模型的情况，研究了基于变换的影像薄云检测和去除方法，提出的薄云雾模拟算法能够代替独立成分分析（Independent Component Analysis, ICA）算法所需的卷云探测波段。基于卷云探测波段的反射率图像在经过去云之后应对趋于零的先验知识，保留该成分中的地物信息并生成一个新的独立成分，推导了待去云波段的去云方程以及参数求解方法，提高了ICA算法的去云结果。我们还提出了一种基于变换的慢特征分析算法（Slow Feature Analysis, SFA），实现了Landsat-8数据1-7波段的薄云去除，其去云性能显著且稳定；并对慢特征分析算法做了进一步分析，引入了无云辅助影像来改善SFA算法在没有卷云波段时的应用效果，扩大了算法的应用范围。