Retrieval Of Near-sea Surface Air Temperature Based On Deep Learning

Near-sea surface air temperature is a vital important air parameter,but it is hard to be gotten.There are some correlation between near-sea surface air temperature and other related ocean-air parameters.Near-sea surface air temperature can be obtained by using these related ocean-air parameters through the method of retrieval.Recently,deep learning has made great breakthroughs in many fields as a hotspot of artificial intelligence.Deep learning brings many opportunities to the research of atmosphere and ocean.This paper aims at the retrieval of near-sea surface air temperature,with using deep learning to establish the models,and making sea surface temperature,wind speed and atmospheric pressure as inputs,and then relatively accurate near-sea surface temperature can be calculated.Firstly,the basic principles of artificial neural network and deep learning have been analyzed and introduced.It mainly includes structure models,setting of parameters,working modes,etc.,and math expressions and characteristics of neural network have been listed.By analyzing disadvantages of traditional neural network and the transformation from traditional neural network to deep neural network,the advantages of deep learning have been apparently.The basic structure and common algorithms of deep learning are also introduced.And then the principle and common means of retrieve are analyzed and summarized.Secondly,based on the analysis of commonly used oceanic and atmospheric observation data,the data set used in this project is introduced,and the data in the experimental area is pre-processed by screening,quality control and normalization.Then,the correlation of oceanic and atmospheric data is analyzed,and the correlation coefficient between near-sea surface temperature and SST,WS and SLP are obtained through experiments,which lays a foundation for the establishment of inversion model.Thirdly,based on the features of FCN,RNN and LSTM,and combined with the correlation between AT with SST,WS and SLP.There are three kinds of near-sea surface temperature retrieval methods have been designed,according to the features of the three networks.And three kinds of algorithms have been used to match the three networks.And then optimizations have been made for the defects of different methods.Finally,part of International Comprehensive Ocean-Atomsphere Datasets were used to improve the effectiveness of the methods that proposed in this paper.The three methods that proposed in this paper are better and more effective than traditional methods.And the method of LSTM and TBPTT is the best method.In conclusion,the near-sea surface temperature retrieval methods have been proposed in this paper can solve the obtain of near-sea surface air temperature,and the application of deep learning in the field of ocean and atmosphere has been innovated and expanded.