摘要
现有深度学习方法在预测台风时没有考虑其特征内化损失问题,难以全面捕捉台风结构变化。为此,本文提出一种基于多尺度特征融合的时空注意力长短期记忆网络(MSTA-LSTM)方法。引入特征增强模块加强台风特征信息,通过跳跃连接缓解编解码过程中的台风细节特征损失,同时在时空长短期记忆网络(ST-LSTM)单元中利用卷积块注意力模块优化信息传递,最后通过反卷积调整不同尺度的解码输出,融合后输出结果。使用“葵花8号”卫星获取的东亚—东南亚太平洋沿岸地区的台风云图数据集开展验证和消融实验,该数据集包含16个台风过程的训练集和3个台风过程的测试集。与其他网络相比,MSTA-LSTM网络的均方根误差、峰值信噪比和结构相似性指数指标分别为42.76、16.38和0.4817,有效提高了台风云图预测的准确性。
Existing deep learning methodologies for typhoon prediction overlook the issue of internal feature loss,hindering a comprehensive capture of the intricate structural changes within typhoons.To address this problem,this paper introduces a Multiscale Feature Fusion Spatiotemporal Attention Long Short-Term Memory Network(MSTA-LSTM).Initially,a feature enhancement module is incorporated to strengthen typhoon feature information.Then,the loss of typhoon-specific details is mitigated through skip connections during the encoding and decoding processes.Simultaneously,the Convolutional Block Attention Module within the Spatiotemporal Long Short-Term Memory(ST-LSTM)units is used to optimize information transmission.Finally,decoded outputs from different scales are adjusted through deconvolution and fused to produce the final output.Validation and ablation experiments are conducted using a dataset of typhoon cloud maps obtained by Himawari-8 for the Pacific coastal regions from East Asia to Southeast Asia,which contains a training set of about 16 typhoon processes and a test set of 3 typhoon processes.Compared to other networks,the MSTA-LSTM network demonstrates improvements in the accuracy of typhoon cloud map prediction,with root mean square error,peak signal-to-noise ratio,and structural similarity index metric reaching 42.76,16.38,and 0.4817,respectively.
作者
程勇
钱坤
王军
渠海峰
李伟
杨玲
韩晓东
刘敏
CHENG Yong;QIAN Kun;WANG Jun;QU Haifeng;LI Wei;YANG Ling;HAN Xiaodong;LIU Min(Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)
出处
《海洋预报》
北大核心
2025年第2期89-98,共10页
Marine Forecasts
基金
国家自然科学基金(41975183、41875184)。
关键词
时间序列预测
多尺度特征
时空长短期记忆网络
注意力机制
time series prediction
multiscale features
Spatiotemporal Long Short-Term Memory
attention mechanism
