Research On Tropical Cyclone Intensity Estimation Based On Deep Learning And Fusion Of Environmental Field Informatio

Tropical cyclones(TC)are weather phenomena with great destructive power,which have a significant impact on the global climate system due to their unique shape and intense circulation of storm systems.As global warming and sea-level rise continue,the intensity and frequency of TC are increasing.Therefore,accurate estimation of TC intensity is crucial for preventing disasters.However,traditional TC intensity estimation methods have many limitations.For example,the widely used Dvorak method relies on human observation of satellite images to assess intensity,which is subject to subjective errors.Meanwhile,the distribution of TC intensity datasets exhibits a ”long-tail distribution” pattern,and directly using deep learning models without relevant processing will lead to a weakened generalization ability of the model,resulting in underestimation of large typhoons and overestimation of small typhoons.Additionally,TC are non-rigid objects that exhibit diverse forms and wide TC wind speed fluctuations,and simple convolutional neural networks are unable to cover their variability.Moreover,traditional deep learning models have not considered the influence of environmental factors on TC development and intensity,which can easily cause misjudgment of TC intensity.Therefore,to address these issues,this paper focuses on the following research work for TC intensity estimation:(1)To address the subjectivity and long-tail distribution issues in the Dvorak technique,we propose a dynamic balance convolutional neural network intensity estimation model(TCDBNet)utilizing satellite imagery.The model consists of two branches: one for training on the raw data distribution and another using an inverse sampler to extract data for balanced training.By designing a balance factor parameter,the model dynamically adjusts,gradually shifting its focus from features learned from the raw data distribution to those acquired from balanced training,thereby reducing the underestimation of strong typhoons and overestimation of weak typhoons to improve intensity estimation accuracy.Finally,a visual analysis of TCDBNet is conducted,revealing that the feature learning process of the model is generally similar to that of Dvorak,thus demonstrating the reliability of the proposed method.Experimental results indicate that the accuracy of the model is enhanced by 2.18 kt compared to Dvorak and by 1.13 kt relative to the state-of-the-art deep learning technique(CNN-TC).(2)To address the issue of convolutional neural networks lacking global spatial relationships between pixels in local convolutions for TC intensity estimation tasks,and being only capable of capturing limited contextual information,resulting in insufficient spatial context modeling and an inability to fully express the rotational invariance and symmetric structure of TC,an improved deep learning-based TC intensity estimation model(TCICVIT)is proposed based on the work in(1).By combining rotation-equivariant convolution and Transformer architecture,the model captures both local and global spatial contextual information,achieving more accurate intensity estimation.Additionally,environmental field information is incorporated into the model,helping it capture the influence of external factors on TC intensity,further enhancing the accuracy of the estimation.Experimental results show that,compared to models utilizing similar deep learning methods,the accuracy of the proposed model is improved by1.26 kt.