| Typhoon is a weather system with wide coverage and strong destructive power,which brings rain to the land and poses a huge threat to human life and property.Almost every year,coastal areas in China are plagued by typhoons,resulting in casualties and property losses.Timely forecasting and issuing typhoon warnings are important means to prevent typhoon disasters.Accurate location of typhoon centers is the basis for typhoon path prediction and is of great significance for typhoon prevention and disaster reduction.Minor errors in typhoon center location can cause significant deviation from typhoon path prediction,so it is necessary to accurately locate the typhoon center.Due to backward equipment and subjective errors,early observation methods such as radio,altitude observation,and radar led to inaccurate location of typhoon centers.The emergence of meteorological satellites has brought great convenience to this task,and meteorological researchers have begun to use infrared satellite cloud images taken by meteorological satellites to locate the typhoon center.Early typhoon center location methods based on infrared satellite cloud images can generally be classified as mathematical morphology methods,pattern matching methods,wind field analysis methods,and digital image processing methods.Most of these methods have achieved semi-automatic or automatic location of typhoon centers.However,these traditional methods are difficult to ensure high accuracy in the face of complex typhoon images,and the traditional methods have many steps and large computational complexity,resulting in low location efficiency.The deep learning method of artificial intelligence driven by big data can automatically adjust parameters through learning,and can also optimize models based on constantly updated neural network concepts,with the advantage of flexibly adjusting weights.Currently,the typhoon center location model based on deep learning has improved generalization,but there is still room for improvement in feature extraction capabilities.Due to the existence of interference information in typhoon images and the low efficiency of extracting spatial morphological features of typhoons using convolutional neural networks alone,it is difficult to extract typhoon features.In response to the above problems,this paper uses deep learning combined with attention mechanism to conduct research on typhoon center location.Before conducting the research,a set of typhoon center location dataset suitable for deep learning was constructed.Firstly,more than 8000 typhoon satellite images collected by the National Research Institute of Japan from 2015 to 2022 were collected,and the longitude and latitude coordinates of the typhoon center in the satellite images were converted into planar coordinates using the Mercator formula to complete data annotation.To reduce the computational complexity of the model,a bilinear interpolation method is used to reduce the image size from 512×512 compressed to 224×224.Next,this paper builds a neural network model for typhoon center location based on coordinate regression(TY-LOCNet).The main innovative work is as follows:(1)Firstly,a backbone network model is built using convolutional networks and residual networks to extract typhoon features,and then a channel attention mechanism is introduced to capture channel level information from typhoon features,enhancing the model’s attention to important channels,Then,the channel attention results are input into the coordinate attention mechanism to globally calibrate the typhoon location information,enabling the model to focus on the morphological structure of the typhoon in a larger area;(2)The feature fusion of the middle and high level outputs of the backbone network model is performed using a jump connection method to alleviate the phenomenon of fine granularity loss during the down sampling process.(3)A lightweight typhoon center location network was built to locate the typhoon center using coordinate regression.(4)The mean square error loss function failed to fuse the calculated coordinates,resulting in low location accuracy.Therefore,a distance loss function(DISTLoss)was proposed to improve the location accuracy of the model through distance regression.To verify the effectiveness of the model TY-LOCNet,comprehensive experiments were conducted on the proposed module,including comparative experiments and ablation experiments.In order to verify that the fused dual attention mechanism is helpful for the typhoon image feature extraction ability of this model,a comparative experiment was conducted with other attention mechanisms,and the experiment proved that the location accuracy of the fused dual attention mechanism is superior to other attention mechanisms.Comparing the proposed distance loss function with the mean square error loss function,the results show that the distance loss function has the advantage of rapid error reduction and smaller error during model training.Finally,two improved ablation experiments were performed using the control variable method,and the results showed that the improved method significantly improved the accuracy of the model.In order to verify the effectiveness of this model in locating typhoon centers,TY-LOCNet is compared with other depth learning based and classical methods.The experimental results show that the mean location error(MLE),mean absolute error(MAE),and detection speed of TYLOCNet are 7.3656 pixels,0.29 °,and 17 FPS,respectively,which are superior to other models.Typhoon center location model TY-LOCNet can provide real-time typhoon center location support for typhoon forecasting. |
