Research On Automatic Detection And Intelligent Forecast Methods Of Tropical Cyclones Based On Machine Learning

The tropical cyclone(TC)is one of the most destructive natural disasters,causing severe economic losses and casualties every year.Due to the variability of morphology and the complexity of evolution,there are various challenges in TC identification and prediction.Meteorological satellites can provide observations of the entire TC process,which provide data for the objective description of TCs using intelligent learning algorithms.The development of TC is affected by large-scale weather patterns.Numerical weather prediction(NWP)can provide environmental information affecting TCs.Based on satellite and NWP data,TC estimating and forecasting models are studied using machine learning algorithms and computer vision technologies,including TC cloud identification,center location,intensity estimation,precipitation estimation,extrapolation,and path prediction.The specific work contents are as follows:(1)Accurate TC clouds identification and precise center positions are the basis for TC forecasting.A multi-scale feature extraction and fusion network is constructed to identify TC clouds.The multi-level features that reflect the global and detailed characteristics of TCs are extracted.Then two-level fusion is used to provide object boxes of different scales to adapt to the TC scales while providing central candidate positions.Within the range of TC clouds,the candidate centers based on mechanism characteristics are obtained through image segmentation,contour extraction,circle fitting,and the construction of differential features.The tropical cyclone center is finally located after comprehensive decision-making.This strategy achieves a higher TC clouds detection and a lower center location deviation.(2)The point estimation and interval estimation of TC intensity are realized.First,based on the prior knowledge of the characteristics of TCs on satellite infrared images,four groups of intensity-related features are proposed: eye feature,circle feature,texture feature,and time feature.Then,the random forest model is used to achieve intensity regression,that is,point estimation.Using the random forest as the underlying algorithm,a local weight cross conformal prediction framework is constructed.The synergy model can not only realize the intensity regression but also control the model’s error rate and achieve the interval estimation under a given significance level.(3)Precipitation estimation and forecasting are one of the most challenging subjects in meteorological operations due to the complexity of precipitation.Convolutional neural networks with strong nonlinear performance capabilities are used to solve problems in precipitation.First,a digital elevation module characterizing geographic information and a spatial transformation module reflecting the drift between high clouds and ground precipitation are added in the primary encoding-decoding network to estimate precipitation.Then,a 3-dimensional(3D)encoding-decoding network that can process spatiotemporal data is used to extrapolate precipitation.These two models can estimate and forecast the detailed structure of precipitation more accurately.(4)TC has the complex 3D structure,and the surrounding atmospheric environment is an essential factor affecting the development of TCs.Therefore,the accurate path prediction should be based on the precise description of its structure and surrounding atmospheric environment.A hybrid model combining 3D convolutional neural network(3DCNN)and gated recurrent unit(GRU)is proposed for trajectory prediction.3DCNN is used to explore the potential relationship between environmental variables and TC motion.GRU is used to convert path prediction to a spatiotemporal sequence prediction problem.In addition,a post-processing algorithm is proposed to suppress the unreasonable jumps in model output and further improve the accuracy of path prediction.