Prediction Of Sea Surface Temperature Field Based On Full Convolution Neural Network

From the bucket observation,the earliest SST observation method in the mid-19 th century,to the buoy observation of SST in the 1970 s,and to the use of satellite remote sensing technology to measure sea water temperature in the 1980 s,it can be known that the daily growth of SST data obtained before is less than that now,and the previous observation method requires a lot of manpower and the observation space is limited.With the development of science and technology,the method of observing SST has become more and more intelligent,the range of acquired data area has become larger and larger,and the spatial resolution has become higher and higher,and the data is growing exponentially.The data in the ocean has reached the level of big data,and more efficient research methods are needed to process and analyze the data.SST is one of the important parameters for the study of marine hydrological conditions.Accurate prediction of SST has important research significance for global climate and marine ecological environment.SST is an environmental factor that affects the temporal and spatial changes of marine fisheries.The development of marine economy cannot be separated from the utilization of fishery resources.SST also has an important impact on the generation,intensity change and movement path of tropical cyclones.Accurate prediction of SST has an important impact on people’s better early warning of natural disasters,and can reduce the economic losses caused by natural disasters.Deep learning provides a new research method for people to study ocean elements.Recently,the method of deep learning prediction based on pure data-driven neural network has been applied in the ocean neighborhood and other geoscientific neighborhoods.These methods are based on a large number of historical data to predict the next time state.The pure data-driven method can replace the performance of classical methods in many prediction tasks.The use of deep learning to mine information contained in massive ocean elements provides a good platform for humans to better understand and use ocean resources.Most of the previous researches on SST prediction are based on specific sites,without considering spatial factors.There is still a lot of room for improvement in the research on SST prediction of the overall sea area,especially in the improvement of the accuracy of SST prediction.This paper hopes to propose a model to improve the accuracy of SST prediction in the research area and provide reference for the research of SST prediction.The research contents of this paper are as follows:(1)Analysis and data set construction of the study area.On the one hand,analyze the location of the study area and the characteristics of the SST in the study area,and on the other hand,use scientific SST data to study.The study area of this paper is located in the eastern Pacific tropical unstable wave region,and the data of 16 years from 2006 to21 are selected for study.This research is about time series prediction.When analyzing data,it is mainly about time series characteristics of data.After the data is determined,it is necessary to preprocess the data,and at the same time,it is necessary to build a data set suitable for input into the neural network.The proportion of training set,verification set and test set is 6:1:1.The training set is the data from 2006-2017,the verification set is the data from 2018-2019,and the test set is the data from 2020-2021.(2)U-net neural network is proposed to predict the SST in the study area.Firstly,the composition of U-net neural network is introduced,and then the element structure is analyzed in detail.The function of each element is briefly explained.Based on the previous research on U-net neural network prediction,the U-net model of SST in the prediction area in this study is proposed.The original data is normalized to the range of[0,1] before training.Normalization can eliminate the influence of singular sample data on the experiment and improve the training speed of the model.Add Spatial Dropout2 D layer to the constructed model to simply improve the prediction effect of the model.The loss function used in the training process is mean square error,and the evaluation indicators of the model are root mean square error,mean absolute error and determination coefficient.(3)Modify U-net neural network to improve the accuracy of SST prediction.This study adds some elements to the existing U-net model to obtain a model for predicting SST with higher accuracy,and analyzes and compares the prediction results of the model.Modify the model mainly from two aspects: adding input layer and adding hole convolution.For adding input scale information,add the reduced image of the original image to the model pooling and upsampling input,and analyze the impact of adding location on the model prediction effect.Hole convolution is a module composed of adding hole convolution with the same void ratio or adding hole convolution with different void ratios.Analyze the prediction results of adding different number of hole convolutions with the same void ratio on the test set,and the training results of adding modules with different void ratios on the lowest layer of the encoder on the test set.Compare and analyze the effect of adding hole convolution on the model’s prediction of SST effect.(4)DMU-net model is proposed to predict SST.This model is a new model for predicting SST,which combines the two modification methods proposed previously,adding input layer and adding hole convolution.Compared with the previous model,this model has the best prediction effect.