激光探测对于获取云和气溶胶的垂直廓线, 研究大气中云和气溶胶的垂直分布特征以及对全球气候变化的影响意义重大。 而星载大气激光雷达云气溶胶分类算法的研究, 对于激光雷达数据的参数反演及应用极为重要。 针对激光条件下探测的云和气溶胶特有的光学信息和空间分布, 结合概率统计与机器学习算法, 提出了一种对于云/气溶胶、 云相态及气溶胶子类型识别的分类算法, 实现了星载激光雷达的大气特征层快速、 有效分类。 算法采用中国地区2016年CALIOP的观测数据作为样本数据, 主要由三部分组成: (1) 基于激光探测的云和气溶胶层不同的光学特性以及地理空间分布特征, 分别构建了云和气溶胶的γ532, χ, δ, Z和lat的五维概率密度函数, 以此为基础构建云气溶胶的分类置信函数, 并基于此实现了云和气溶胶类型的反演; (2) 选取支持向量机(SVM)作为随机朝向冰晶粒子(ROI)和水云分类的算法模型基础, 结合云层的γ532, χ, δ Z和云顶温度T的概率密度函数构建ROI, 水平朝向冰晶粒子(HOI)和水云的分类置信函数以修正SVM误分的特征层以及筛选出水云中少部分的HOI冰云, 获得云相态的分类结果; (3) 以各气溶胶子类型的光学以及空间分布特性为基础, 采用决策树策略的气溶胶子类型识别算法实现了对气溶胶子类型的区分, 完成气溶胶子类型的识别。 利用现有CALIOP观测结果作为样本数据构建分类数据库, 避免了对于地面以及航测数据的依赖, 而机器学习则大大简化了算法的实现过程, 使得云气溶胶分类更加高效。 算法结果与正交极化云气溶胶激光雷达垂直特征层分布数据(CALIPSO VFM)产品对比分析: 云层有98.51%一致性, 气溶胶有88.43%的一致性, 且白天比夜间一致性高。 对于云相态分类, 可以有效区分出水云和冰云, 其中二者水云一致性高达93.44%。 在气溶胶子类型反演结果中, 可以准确识别出大多数气溶胶特征层子类型。 霾、 沙尘以及晴空三种典型情况下的反演结果均与CALIOP VFM产品数据具有较好的一致性。 其中, 霾天的大部分煤烟型以及污染型(污染沙尘以及污染大陆) 气溶胶反演结果与VFM具有较好的一致性。 沙尘天也能够获得较好的沙尘以及污染沙尘的结果。 晴空为数不多的气溶胶层也取得了较为一致的结果。 对于实现的星载大气激光雷达特征层分类算法, 针对CALIOP激光测量的云气溶胶层的分类进行了重要的改进, 在保证一定精度的基础上, 简化了算法, 提高了数据处理的效率, 在下一步工作中, 将分别构建不同时段和季节的分类模型以及提高两种不同偏振特性的冰云和气溶胶子类型的分类精度。

LIDAR plays significant roles in monitoring the vertical distribution characteristics of clouds and aerosols and studying their impacts on the global climate change. For the space-born LIDAR, discrimination between clouds and aerosol is the first step of cloud/aerosol vertically optical property retrieve, and to a great extent, the retrieval precision depends on the accuracy of cloud and aerosol classification algorithm. Based on the optical and geographic characteristics of aerosols and clouds observed by LIDAR, in this study, the CALIOP aerosol and cloud products over China in the year of 2016 were trained as the sample sets. An effective cloud/aerosol classification algorithm was developed by combining the support vector machines (SVM) and decision tree methods. Our algorithm includes 3 parts: cloud and aerosol discrimination, ice-water cloud classification and aerosol subtype classification. (1) The cloud and aerosol were discriminated by the classification confidence functionsof 5-D probability density function (PDF) with parameters of γ532, χ, δ, Z and lat. (2) Randomly oriented ice (ROI) and water cloud were classified based onthe SVM. And by constructing the PDFs with γ532, χ, δ, Z and T, feature layers misclassified by SVM were corrected, and a small portion of the horizontally oriented ice (HOI) clouds were removed from the water clouds. (3) Based on the optical and geographic characteristics of aerosol subtypes, decision tree classification was used for the determination of aerosol subtypes. Our retrieval results showed a good agreement with the CALIOP VFM products. For the cloud and aerosol discrimination results, the consistency ratios between our retrieves and VFM products for aerosol and cloud are up to 98.51% and 88.43%, respectively. And the consistency ratios in the day are higher than those at night. For the cloud phase retrieval results, water clouds can be well separated, and the consistency ratio of water cloud between our retrieves and VFM products is as high as 93.44%. The consistency ratio of HOI is low due largely to the confusion between HOI and ROI. For the aerosol subtype classification, most aerosol subtypes could be well recognized by our algorithm. However, the consistency ratios of the mixed subtypes (e. g. polluted continental and polluted dust) between retrieval results and VFM products are relatively lower. Moreover, the cloud/aerosol, cloud phase and aerosol subtype classifications were also compared with the VFM products under three typical air conditions, i. e. haze, dust and clean. Under the haze condition, our results for most of the smoke aerosols agree quite well with the corresponding results from VFM. Under the duststorm condition, our algorithm can effectively discriminate the most of dust and polluted dust aerosols. For the clear day, our results for the few existing cloud and aerosol layers are quite consistent with the VFM results. This paper is an important improvement of the cloud and aerosol classification algorithms, which can simplify the processing and improve efficiency with satis factory accuracy. In the future work, we will build day/night and seasonal training sample sets, and consider more ice cloud phases and aerosol properties in the cloud/aerosol classification retrieval algorithm.