A Study On Cloud Phase Detection Algorithm Based On Multispectral Threshold And Machine Learning Methods

Cloud has a significant role in regulating energy balance between the earth and atmosphere.Different phases of cloud are of different absorption and scattering properties,and the thermal kinetic process accompanied by their transformation will directly affect the formation and evolution of the weather system.Understanding the cloud phase is not only conducive to weather monitoring and forecasting,but also beneficial to artificial weather modification.Currently,the basic theory of retrieving cloud phase by satellite remote sensing data is mature.However,it is difficult to detect supercooled water cloud with temperatures between233 and 273 K.Therefore,we develop cloud phase algorithms based on multispectral threshold method and machine learning method,and focus on the detection of supercooled water cloud and mixed-phased cloud.The two algorithms are used for visible infrared imaging radiometer(VIIRS)onboard on NPP and Advanced Geosynchronous Radiation Imager(AGRI)onboard on FY-4A.This study introduces a multispectral threshold cloud phase detection algorithm used for VIIRS observations.The reflectance difference between the 1.61 and 2.25 ?m channels(i.e.,?R)is introduced as a key criterion for detection and is recommended for use in operational algorithms.Additionally,the brightness temperature difference(BTD)between 8.5 and 11 ?m channels and cloud top temperature(CTT)are synergistically combined with ?R to better distinguish cloud phase.Theoretical simulation results demonstrate that water and ice cloud are well separated in the ?R-BTD space domain.These channels have different scattering and absorption characteristics for different cloud phases,which is the basis for the development of cloud phase algorithm.Encouraged by the successful separation,one simple but efficient equation is derived from the use of cloud phase from the active remote sensing product collocated with VIIRS data,as a cloud phase detection algorithm.We quantitatively evaluate the results of our algorithm and the current operational VIIRS cloud optical phase product.For clouds with temperature between 233 and 273 K,it is shown that our algorithm can correctly detect over 91% of supercooled water clouds and 85% of ice cloud pixels,substantially better than accuracies found in VIIRS product(85% and 77%).A cloud phase detection algorithm based on machine learning is developed.Cloud phase from space radar and lidar measurements is used for the development and validation of the algorithm.The algorithm is developed for daytime and all-time model and improves the accuracy of detecting mixed-phased clouds.Using independent verification results from the radar and lidar measurements,the algorithm outperformes the current official AGRI cloud phase product.The algorithm detection accuracy of cloud mask,water,ice and mixed-phased clouds are 0.95?0.91?0.88?0.82.The results of the all-time algorithm are less accurate than the daytime model.In the future,we will improve the algorithm and apply these two cloud phase algorithms to the spectral imagers with similar channels.