Image Sequence Prediction Based On LSTM And Tensor Completion

The impacts of climate change are all aspects,and people often take measures to intervene in the negative impacts.In recent years,there has been an increasing demand for rainfall prediction.Traditional rainfall prediction methods need to collect a large amount of physical information,which increases complexity,and overly regular prediction methods can lead to inaccurate predictions.With the development of deep learning,image processing has achieved results in many tasks.This paper proposes a method of using the image sequence prediction task to deal with the rainfall prediction task,and has achieved satisfactory results.Image sequence prediction is a spatio-temporal sequence prediction method for predicting one or more subsequent continuous pictures given a number of continuous pictures.It is one of the important topics in the field of computer vision.Image sequences have the characteristics of temporal and spatial feature changes and long time intervals,which brings great challenges to image sequence prediction.On the one hand,complex spatio-temporal feature changes are difficult to effectively capture and accurately predict;on the other hand,in long-term sequence prediction tasks,the recording time interval is longer,resulting in less data and weak data correlation.(1)Aiming at the problem that it is difficult to capture and accurately predict the spatial feature changes of short-term sequences,this paper proposes a motion-enhanced LSTM-based generative confrontation sequence prediction method,which considers the use of local feature enhancement to improve the ability of local information extraction and accurate prediction.In the traditional LSTM prediction,a generative confrontation network combined with an attention mechanism is added to improve the robustness of spatial feature extraction.Through experiments on the SRAD2018 radar echo data set and the Moving MNIST data set,the effectiveness of the motion enhancement model is verified.(2)For long-term sequences,the spatial distribution is variable,making it difficult to capture spatial features,and due to the long recording time interval,the amount of data obtained is small and the data correlation is weak.There is a small amount of sample training to obtain features.It is difficult to predict,and methods such as deep learning need to use a large number of data samples for training.Therefore,this paper proposes a long-term sequence prediction method based on tensor completion.In the case of a small number of samples and variable spatial structure,a sequence tensor completion method that minimizes the norm of the convolution kernel with non-negative constraints is adopted.Experiments on three-month meteorological data in Zigong City,Sichuan Province have verified the effectiveness of the algorithm,and the experiment has achieved a good prediction effect.