Ground fog detection and analysis with Machine Learning (GFog-ML)

雾和低层云（LST）对辐射平衡至关重要，因此对地球气候系统至关重要。缺乏对其时空动态以及间接气溶胶效应的认识是IPCC第五次评估报告（AR5）最新预估存在不确定性的一个主要原因。同时，雾/地表温度对交通安全和空气质量造成问题，这需要可靠和可操作的临近预报程序。特别是正确区分LST和地面雾将意味着在操作临近预报方面向前迈进了一大步，而且也将成为使用最新一代的操作气象卫星系统长期监测雾/LST动态的基础。当前项目的主要目标是基于欧洲同步气象卫星第二代（MSG）系统及其机载的SEVIRI / HRV（旋转增强可见光和红外成像仪/高分辨率可见光）仪器，开发一种新的昼/夜分辨地表温度和地面雾的技术。这项新技术将首次用于生成LST/地面雾（雾小时）及其特性（有效半径、光学深度）的10年数据集。该项目基于现成的LST检测处理链，这是马尔堡地面真相和分析站用于训练/验证的独特数据集，并将利用机器学习技术随机森林（RF）计算3公里和1公里分辨率的LST数据集，地面雾（后者仅在白天分辨率）和雾属性。该综合数据集首次实现了业务气象学所需的地面雾小时的空间明确映射。对生成的时间序列进行气候学分析的重点是探测过去十年来由于本地（土地利用和气溶胶浓度）和外地（大气环流）形成因子的变化而引起的区域/局部地面雾/地表温度的变化。

Fog and Low Stratus (LST) are of utmost importance for the radiation balance and thus, the earth's climate system. The lack of knowledge on its spatio-temporal dynamics in combination with the indirect aerosol effect is one main reason for uncertainties in the latest IPCC projections of the Fifth Assessment Report (AR5). At the same time, fog/LST are causing problems for traffic security and air quality which requires reliable and operational nowcasting procedures. Particularly a proper distinction between LST and ground fog would mean a big step forward in operational nowcasting but would also form the basis for long-term monitoring of fog/LST dynamics with the newest generation of operational weather satellite systems. The main aim of the current project is to develop a novel daylight/night technique for the discrimination of LST and ground fog based on the European geostationary Meteosat Second Generation (MSG) system with its SEVIRI / HRV (Spinning Enhanced Visible and Infrared Imager / High Resolution Visible) instrument onboard. The novel technique will be used to generate for the first time a 10 years data set of LST/ground fog (fog hours) and its properties (effective radius, optical depth). The project is based on the readily available processing chain of LST detection, a unique data set for training/validation from the Marburg Ground Truth and Profiling Station, and will make use of the machine learning technique Random Forest (RF) to calculate 3 and 1 km resolved datasets of LST, ground fog (the latter resolution during daylight only) and fog properties. The comprehensive dataset enables for the first time the spatial-explicit mapping of ground fog hours as required by operational meteorology. The climatological analysis of the generated time series will focus on the detection of changes in regional/local ground fog / LST occurrence over the last decade due to autochtonous (land used and aerosol concentration) and changes in allochthonous (atmospheric circulation) formation factors.