The Cloud-type Classification Research And Its Application For The New Generation Geostationary Satellite Himawari-8

Clouds play a pivotal role in influencing the dynamics of climate and weather conditions at both a local and global scale,as varying cloud types have different radiative effects on the surface-atmosphere system of the Earth.The automatic,accurate detection and classification of clouds is useful for a multitude of hydrologic,climatic,and atmospheric applications.Convective clouds such as cumulonimbus clouds are associated with atmospheric instability,turbulence and thunderstorms,and as such,the continuous monitoring of clouds is vitally important for the purposes of weather forecasting and air traffic control.In order to effectively improve the ability to ensure aviation safety and efficiency in South China,it is necessary to develop objective and quantitative cloud-type and rainfall retrieval algorithm with high spatial and temporal resolution throughout the day.Many effective and accurate schemes,such as relatively simple threshold-based(or parallelepiped)methods,texturebased approaches and traditional mathematical statistics-based pattern recognition algorithms have been introduced into the remote sensing community for cloud detection and classification from satellite images.With the rapid development of technological innovations,sophisticated machine learning(ML)methods,such as the fuzzy logic method,artificial neural networks,support vector machine classifiers and random forests(RF)classifiers,have recently proven to be successful in cloud detection and classification.ML algorithms can be very versatile and have great potential;however,the results and efficiency of these algorithms are highly dependent on training samples and the overall quality of the data usedThe new generation of geostationary satellites such as Himawari-8,which has been put into operation in recent years,provides more detailed spatiotemporal resolution and more spectral channels.Whether it is beneficial to the retrieval of cloud-type and rainfall is worthy of further studyBased on the spectral characteristics of Himawari-8 for cloud,this thesis studies how to combine the cloud profiles of CloudSat to retrieve the cloud types,and how to combine the rain gauge to retrive the rainfall intensity.The comparison with independent data shows that the cloud classification and rainfall distribution results obtained in this paper are reasonable and reliable.The algorithm is fast and very suitable for operational business.However,the algorithm is underestimated for heavy rain,and needs to be improved in future research.The conclusions are summarized as follows:1.With the cloud scenario products from CloudSat,we developed a high spatiotemporal resolution cloud type classification procedure for Himawari-8 multispectral datasets using maximum likelihood estimation(MLE)and RF classification.The training and classification procedures were processed independently,and both algorithms provided cloud type results with a good performance.Validation indicated that the use of the visible(VIS)channel significantly improved the cloud type identification capabilities,while the use of three or more channels simultaneously resulted in considerable improvements over the use of bispectral combinations.The best performance of MLE is improved from 78.95%at night to 87.38%during the day,and the RF performance is increased from 85.23%to 94.23%.In general,combined classifiers with three or more channels are more accurate than bispectral classifiers.For RF at night-time,the accuracy is increased by 6%,and for MLE at daytime,the accuracy is increased by nearly 10%.The comparison among different classifiers also revealed that RF was more sensitive than MLE to the quality and distribution of the training data.2.The distribution of cloud types in the spectral space is autonomously clustered.Therefore,it is reasonable for MLE to perform cluster adjustment and correction on training samples based on statistical probability.After retraining the RF using MLE-based clustered samples,we produced two more reasonable and efficient classifiers that can be used during the day and night(i.e.,MRD and MRN).The comparison with the National Oceanic and Atmospheric Administration(NOAA)cloud products proves that this method can give full play to the advantages of more spectral channels of Himawari-8,and also has higher tolerance for the errors that may exist in the samples themselves.Importantly,once trained,the RF classifiers were computationally more efficient than the MLE classifiers,as the former were essentially composed of multiple logical judgments or decision trees.Thus,MRD and MRN are appropriate for operational use when rapid results are needed.3.The rainfall retrieval of Himawari-8 is achieved by combining the measurements of rain gauge,and is verified with the GPM(Global Precipitation Measurement)precipitation product.Since the data used for modeling and verification is of high spatial and temporal resolutions(satellite imagery maintains 0.025°resolution information),which ensures the rainfall retrieval with high accuracy with this method.This algorithm is conducted in two steps:first,rainfall and nonrainfall areas are determined,and then the rain rate at the pixel of rainfall area is retrieved,thus excluding the interference of nonrainfall pixels on rain rate retrieval process.This thesis combines the results of cloud classification to further analyze the relationship between rainfall samples and cloud types,and realizes the machine learning algorithm to rechieve the rainfall rank.In this way,a classifier suitable for daytime and nighttime is obtained.The rechieval results are consistent with the rainfall distribution of the GPM.However,due to the distribution characteristics of the sample size and the characteristics of the rainfall,it is easy to cause the rainfall rank to be weak when using the criterion based on the average value of the sample rainfall.