Deep Learning Regression Algorithm Based On Image Sequence And Its Application Research

Image time series data is the very important data in the real world,it's often high-dimensional and large-scale in size.Image sequence prediction is often used in areas such as autonomous driving,and gain good progress.With the rapid development of observation equipment innovation and storage technology in the meteorological field,data collection become more easier,which drives the growth of demand in processing plenty of radar echo images.The conventional meteorological forecasting model relies on optical flow extrapolation and statistical analysis to obtain predictions of image sequences and estimation of rainfall.However,the success of these methods in most of process is limited because the lack of ability to model time series and accuracy decreases rapidly with rapid changes in intensity and pattern.it is challenging to establish an effective representations of spatiotemporal model for generating images in the meteorological field.Recent advances in deep learning on sequence modeling have brought new ideas to solve this problem.We carried out a weather prediction model based on deep learning and its application research to image prediction and estimation of rainfall based on it in the meteorological field.The optical flow method cannot model rapid and non-linear movements,recently,a method based on ConvLSTM is developed,however,it tends to yield blurring extrapolation images and fails to multi-modal and skewed intensity distribution,to overcome the shortcomings,we proposed a Generative Adversarial GRU(GA-GRU)model in this paper,.After a detailed research on the background knowledge of image data and actual business needs in the field,a model was designed.The model adopts an end-to-end structure,which mainly includes two networks.One network adopts a sequence-to-sequence framework to generate sequence images,and the other network corrects the generated images according to real images to make the predicted images more accurate and real.The network was trained alternately.the method of this paper is compared with the method used in the current business at the application level,which can realize the second-level real-time image sequence prediction,which verifies that the model is effective for modeling the atmospheric spatiotemporal law and learning the representations.The ultimate goal of predicting image sequences is to predict heavy rainfall accurately and establish a regression model from image to rainfall.The solution of this problem at home and abroad relies on statistical and statistical correction methods at present,the prediction of heavy rainfall is often unsatisfactory considering the weak modeling ability.Therefore,this paper we proposes three rainfall regression model based on recurrent neural network,convolutional neural network and C-BiGRU model.Considering the factors affecting heavy rainfall and the complexity of the genesis,a datasets was organized and constructed for rainfall-related image data,perform a model with automatic feature extraction based on CNN,a model with statistical feature extraction based on BiGRU,and a model use feature from CNN model as input base on C-BiGRU.Using traditional machine learning methods such as random forest and XGBoost to model at the same time,and finally comparing these methods above,it is proved that the deep learning model has a good performance in the modeling of nonlinear complex systems.