Short-term Rainfall Prediction Using Radar Echo Image Based On Deep Learning

Short-term rainfall prediction aims to accurately and timely forecast the precipitation intensity of a certain region in the near future.It is an important research field in meteorology,which can provide early warnings for citizens,disaster reduction agencies,and government departments.Up to now,many methods have been proposed.Numerical prediction is an important method for rainfall prediction,which predicts the future precipitation intensity by solving partial differential equations in numerical models and can achieve relatively accurate long-term prediction.However,there are still some problems,such as high time complexity,which limits its application in near-term rainfall prediction.On the other hand,deep learning has made breakthroughs in many tasks in various fields and is considered an effective tool for approximating high-dimensional functions.The deep learning-based near-term rainfall prediction method is also an important approach,which can obtain better prediction results.Nevertheless,existing methods often suffer from the problem of rapid attenuation of spatiotemporal features and underestimation of heavy rainfall in high-echo conditions,especially when time elapses or the number of layers increases.To address these issues,this thesis introduces time-difference loss function and spatial difference loss function from the perspective of loss function and proposes a prediction-gated recurrent attention unit method from the perspective of model design.The experimental results show significant improvement over the state-of-the-art methods:1.In order to make fuller use of the spatiotemporal correlation of rainfall data,this thesis introduces time-difference loss function and spatial difference loss function and combines them into a new loss function.Considering that the impact of temporal and spatial correlations on prediction may be different,we set weighting factors to adjust the proportion of time-difference loss and spatial difference loss,making it more flexible.On the other hand,considering that Pred RANN is one of the advanced deep learning methods for near-term rainfall prediction,we apply the proposed loss function to its learning,which can lead to a more accurate model for near-term rainfall prediction.2.In some deep learning-based near-term rainfall prediction methods,the ST-LSTM module is an important approach.To reduce the parameter count of this module,this thesis proposes a new ST-GRU module.Moreover,by replacing the ST-LSTM module in Pred RANN with the ST-GRU module,the thesis presents a Pred GRAU method.In addition,this thesis applies time-difference and spatial difference loss functions to the Pred GRAU method.3.To evaluate the performance of the loss function and predictive gated recurrent attention unit method,we conduct experiments and analysis on the Radar CIKM dataset.The Radar CIKM dataset is the dataset used in the CIKM Analyti Cup 2017 competition,which covers a wide range of precipitation intensities.The results show that,under the use of time-difference and spatial difference loss functions,both Pred RANN and Pred GRAU outperform other methods in most evaluation metrics.The experiments demonstrate that,compared with Pred RANN,the proposed Pred GRAU has fewer parameters and achieves better performance in most evaluation metrics.