Research On Tropical Cyclone Monitoring And Forecasting Methods Based On Deep Learning

A tropical cyclone is a deep and rapidly rotating low-pressure system originating in the tropical or subtropical oceans and is a general term for a non-frontal vortex with organized convection and deterministic circulation.The appearance of tropical cyclones is always accompanied by high winds and heavy precipitation,causing significant human and economic losses to countries during landfall.Therefore,it is of great importance to improve the accuracy of tropical cyclone forecasting for disaster prevention and mitigation.Tropical cyclones mostly occur in distant seas.Meteorological radars and ground stations cannot provide real-time monitoring data.Thus,meteorological satellites become a powerful tool for monitoring and forecasting tropical cyclones.Nowadays,the amount of satellite data has increased exponentially,so it is difficult to handle such a huge amount of data by traditional analysis methods.In recent years,with the development of artificial intelligence technology,especially deep learning,more and more researchers have applied deep learning methods to various studies of tropical cyclones.Although deep learning methods have achieved certain results,they are still in the preliminary exploratory and experimental stage in tropical cyclone applications.In this paper,the performance of satellite data in tropical cyclone monitoring and forecasting tasks is explored and analyzed using deep learning methods.And four aspects are studied according to the stages of tropical cyclone development,i.e.,tropical cloud cluster forecasting before tropical cyclone generation,center localization after tropical cyclone generation,intensity estimation,and real-time forecasting of tropical cyclone tracks.The specific contents and conclusions of this paper are as follows.(1)A tropical cyclogenesis forecast model based on multi-factor combination is constructed by fusing the reanalysis data and satellite images.The satellite can provide real-time monitoring data,but the information source is relatively single;the reanalysis data can reflect the condition of the whole atmosphere,which provides richer historical information for tropical cyclogenesis forecasting,but lacks quantitative analysis.Therefore,in this paper,based on the satellite infrared images,we fuse eight forecast factors from the reanalysis data.To make full use of the spatial information of the forecast factors,we adopt 2-dimensional and 3-dimensional convolutional networks to extract the deep convolutional features of the forecast factors in the infrared images and reanalysis data,respectively.And then,we fuse the learned deep features through the fully connected layer to learn the relationship between the fused features and the probability of tropical cyclogenesis.Considering the different dimensions of the factors,a separate convolutional branch is given to each factor and the influence of different factors on the accuracy is explored.The experimental results show that the infrared image,geopotential height and relative vorticity outperform the other forecast factors;we also explore the effect of combining multiple factors on the forecast performance and conclude that the best forecast performance can be obtained by combining the infrared image with five factors: geopotential height,relative humidity,relative vorticity,10 m wind speed and mean sea level pressure.(2)A RepVGG convolutional neural network model for locating the center of a tropical cyclone is proposed.The training phase of this model adopts two kinds of residual structures and adds multiple paths for gradient flow,which enhances the robustness of the model by an idea similar to the integration of multiple sub-networks;meanwhile,the inference phase converts all network layers into 3×3 convolutions by parameter reconstruction,thus speeding up the localization inference.In order to explore the effects of different bands and their combinations on localization,this paper produced 42,292 training samples by data augmentation based on 2015-2019 Himawari-8 satellite data(in 5 bands of 3.9,6.2,10.4,12.3and 13.3 ?m).Firstly,the localization error of Rep VGG model on the 5 bands was tested,and the results showed that the smallest error was 42.8 km at 10.4 ?m.Secondly,to verify the validity of the model,the performance of Rep VGG and other 7 convolutional neural network models on 10.4 ?m was compared.Finally,this paper performs a band combination to explore which bands are more suitable for center localization.The results show that the 10.4,12.3 and13.3 ?m band combinations have the smallest error of 36.7 km.(3)A tropical cyclone intensity estimation model,TCIENet,is proposed.The model introduces the idea of Siamese networks and takes the satellite infrared image pairs as model inputs.The loss function of the model contains two components,one is to reduce the difference between the intensity estimate and the ground truth at the current moment,and the other is to reduce the difference between the intensity values estimated by two CNN branches to ensure the intensity estimate is continuous in time.In addition,to explore the factors influence on the model intensity estimation error,the correlation between precipitation and estimation error is also calculated and analyzed in this paper.In the experiments,the model with an input image size of 60 × 60(radius of about 450 km)performs best.The effects of WVA and regression approach are similar,while the inclusion of temporal information leads to a greater improvement in the intensity estimation results of the model.The RMSE and MAE for the whole test set are 5.13 and 4.03 m/s,respectively.The intensity estimation error does not always increase with tropical cyclone intensity.The RMSE and MAE for TS are the lowest,and the RMSE and MSE for STS are the highest.Except for TD,all other categories of tropical cyclones have a tendency to be underestimated.In addition,the correlation coefficient between intensity estimation bias and precipitation intensity is 0.19.It is difficult to precisely analyze the relationship between them due to the lack of sufficient samples.(4)This paper proposed a TCTF-GAN model to forecast tropical cyclone tracks.TCTFGAN predicts satellite images by mining the trend of cloud systems to obtain tropical cyclone tracks.In this paper,10.4 and 6.2 ?m band images of Himawari-8 satellite from 2015 to 2019 with a temporal resolution of 1 hour are collected.The occurrence of tropical cyclones is a small probability event,so only a small portion of the satellite images have the presence of tropical cyclones.To train the model effectively,a two-stage training method is used: first,all the satellite images are used for pre-training to obtain the movement of the cloud system;then the images containing tropical cyclones are used to fine-tune the pre-training model.In order to emphasize the prediction accuracy of tropical cyclone itself,the local loss is added in the finetuning stage.And its influence on the prediction results is explored by changing the local loss weights.The above training procedure is applied to both the infrared band(IR)at 10.4 ?m and the water vapor band(WV)at 6.2 ?m.The results show that the track prediction error of the WV channel is smaller than that of the IR channel,and the error is about 46 km for 6 h.In addition,the prediction results of different local loss weights and different bands are averaged together.It is found that the track mean errors of different prediction periods are reduced,and the prediction results are more stable.