毕业生知识库

强对流天气是由大气强烈的垂直运动引发的雷暴、对流性大风、短时强降水、龙卷风等灾害性天气，是仅次于热带气旋、地震与洪涝的世界第四大灾害性天气。近年来，我国强对流天气频发，造成重大财产损失与人员伤亡，及时准确地识别强对流云对于强对流天气的监测与预警具有重要的应用价值。受限于观测手段、观测能力以及对大气强对流运动的热力学和动力学机理认识的不足，强对流云的识别依然是极具挑战的研究问题。气象卫星观测时空范围广，相比雷达观测可更早地捕捉对流初生的信号，因而成为重要的观测手段。2021年6月，我国风云四号B星（FY-4B）发射升空，其提供的更高时空分辨率观测数据进一步提升了对中小尺灾害天气的观测能力。本文基于FY-4B数据，在从定性到定量的综合集成方法论指导下，采用知识与数据协同驱动的机器学习方法，开展强对流云的可解释性特征提取、真值标签学习和端到端识别模型的研究，研制卫星观测的全天候强对流云识别系统。论文的主要工作与创新点归纳如下：

（1）针对单次观测卫星快照的强对流云识别问题，本文提出了一种基于云微观物理特征和语义分割深度神经网络的卫星快照强对流云识别方法。根据强对流云的深厚、云顶呈冰云相态和云顶高的垂直结构特性，提取云光学厚度、云顶相态和云顶高度微观物理特征图。在FY-4B快速成像仪（GHI）数据集上的实验结果表明，该方法相比原始多光谱图直接输入语义分割网络和传统的云顶亮温阈值法都显著提升了识别性能，调和准确率（F₁）达到0.4434。

（2）针对实际业务中存在的强对流云真值标签的获取难题，本文分析了标签设计的正、反问题，即何种含噪标签能使机器学习推断出真实标签？以及如何生成此类含噪标签并结合机器学习得到真实标签？并提出了一种基于从定性到定量综合集成的强对流云真值标签学习方法，利用专家对于强对流云空间分布知识生成一系列随机定性猜想标签，后通过多个数据驱动的深度学习模型集成消除标签噪声和观测依赖，估计出强对流云真值标签。实验结果表明，该方法学得的强对流云标签消除了业务常用的雷达组合反射率≥35dBZ判识准则导致的大部分不确定性和噪声，提升了标签质量，且利用该标签训练的模型大幅提高了强对流云的识别性能，在卫星快照识别任务上F₁达到0.7611。

（3）针对单次卫星观测难以准确识别强对流云的垂直运动状态问题，本文提出了一种基于图像集合选择性融合时相特征的卫星观测序列强对流云识别方法，基于卫星快照识别结果提取云顶降温量和云顶表观变动（分形维数）时相特征，据此选取历史时间窗内的关键帧，通过卷积网络融合关键帧与当前帧实现强对流云的时相信息自动挖掘与运动状态识别，有效避免了对流云的目标跟踪难题和复杂大气运动的时序特征建模难题。实验结果表明，融合云顶降温量或云顶表观变动时相特征的强对流云识别结果都要优于单次观测卫星快照的识别结果（F₁分别高出0.03和0.02），图像集合关键帧的选取与融合相比时序特征融合不仅在识别效果上有提升，而且提高了推理效率，能够达到分钟级的实时监测识别要求。

（4）针对卫星观测可见光与近红外通道无法在夜间使用的问题，本文基于FY-4B卫星先进静止轨道辐射成像仪（AGRI）数据和GHI数据分别设计了夜间强对流云识别方法。实验结果表明，基于FY-4B AGRI构建的夜间强对流云识别模型相比基于FY-4B GHI构建的模型识别结果更优（F₁高出0.09）。本文将夜间识别模型应用至超强台风“轩岚诺”外围强对流云的识别任务中，验证其模型良好的泛化性能，并研制了卫星观测的全天候强对流云识别系统，白天采用FY-4B GHI可见光与近红外通道识别模型，夜间采用FY-4B AGRI多红外通道识别模型。

Severe convective weather is a kind of disastrous weather caused by strong vertical movement of the atmosphere, resulting in thunderstorm, convective gale, short-term strong precipitation and tornado. It stands as the world’s fourth most disastrous weather event following tropical cyclone, earthquake and flood. In recent years, severe convective weather has occurred frequently in China, causing major property losses and casualties. Timely and accurate recognition of severe convective clouds has important practical value for monitoring and warning of severe convective weather. Due to the limitations of current observation methods and capabilities, as well as the lack of understanding of the thermodynamic and dynamic mechanisms of convection in the atmosphere, the recognition of severe convective clouds is still a challenging problem. The meteorological satellite observation has a wide range of time and space, and can capture the signals of the beginning of convection earlier than the radar observation, so it becomes an important observation method of convection. In June 2021, Fengyun-4B satellite (FY-4B) was launched by China. The data with higher spatial and temporal resolution provided by FY-4B further improved the observation ability of medium and small scale disaster weather. Based on FY-4B satellite data, under the guidance of metasynthesis, we propose a machine learning method driven by knowledge and data to carry out the research on the interpretable feature extraction, ground truth label learning and end-to-end recognition model of severe convective clouds, and develop an all-weather severe convective cloud recognition system from satellite observation. The main contributions and innovations of this thesis are summarized as follows:

(1) Aiming at the problem of severe convective cloud recognition in a single satellite snapshot, we propose a severe convective cloud recognition method based on cloud microphysical characteristics and semantic segmentation neural network. We extract microphysical feature maps of cloud optical thickness, cloud top phase, and cloud top height based on the vertical structural characteristics of severe convective clouds, such as depth, ice cloud phase, and cloud top height. The experimental results on the FY-4B Geo High-speed Imager (GHI) dataset show that our method significantly improves the recognition performance compared to directly inputting multispectral images or using traditional thresholding method based on the cloud top brightness temperature, and the harmonic accuracy (F₁) reaches 0.4434.

(2) In view of the difficulty of obtaining the ground truth label of severe convective cloud in practical applications, we theoretically analyze the forward and inverse problems of ground truth label design, that is, what kind of noisy label can make machine learning infer ground truth label? And how to generate such noisy labels and combine them with machine learning to obtain ground truth labels? We propose a severe convective cloud ground truth label learning method based on metasynthesis. It first generates a series of random qualitative conjecture labels using experts’ knowledge of the spatial distribution of severe convective clouds, then eliminate label noise and observation dependencies through the integration of multiple data-driven deep learning models to estimate the ground truth labels of severe convective clouds. The experimental results show that the severe convective cloud label learned by this method eliminates most of the uncertainties caused by the radar combination reflectivity≥35dBZ recognition criteria commonly used in practical applications and improves the label quality. Moreover, the model trained by these labels greatly improves the recognition performance of severe convective clouds, achieving F₁ of 0.7611 in satellite snapshot recognition task.

(3) To address the issue of difficulty in accurately identifying the vertical motion state of severe convective clouds in a single satellite snapshot, we propose a severe convective cloud recognition method based on the selective fusion of time-phase features of satellite sequence based on image sets. Based on the satellite snapshot recognition results, time-phase features such as cloud top cooling and cloud top apparent change (fractal dimension) are extracted, and the key frame in the historical time window is selected accordingly. Then we use convolution network to fuse the key frame and the current frame to realize the automatic mining of the time-phase information and the recognition of the motion state of severe convective clouds, which effectively avoids the difficulties of the temporal tracking of the convective clouds and the temporal feature modeling of the complex atmospheric motion. Experiments show that the recognition results of severe convective clouds based on the time-phase features of cloud top cooling or cloud top apparent change are better than the recognition results of a single satellite snapshot (F₁ is 0.03 and 0.02 higher respectively), and the selection and fusion of key frames in the image set not only improves the recognition accuracy, but also enhances the inference efficiency, enabling real-time monitoring and recognition at a minute-level.

(4) In view of the problem that the visible and near-infrared channels observed by satellites cannot be used at night, methods for identifying severe convective clouds at night are designed based on the Advanced Geosynchronous Radiation Imager (AGRI) data and GHI data of FY-4B satellite respectively. Experiments show that the nighttime severe convective cloud recognition model based on FY-4B AGRI is better than the recognition model based on FY-4B GHI (F₁ is 0.09 higher). Finally, we apply the model to the task of recognizing severe convective clouds in the periphery of the super typhoon Hinnamnor to expand the application scenario of the model, and develop a prototype system of all-weather severe convective cloud recognition observed by satellite. The FY-4B GHI visible and near-infrared channel recognition model is used in the daytime, and the FY-4B AGRI multi-infrared channel recognition model is used at night.