面向云团时空预测的多尺度门控融合模型OA
Multi-scale Gated Fusion Model for Spatiotemporal Prediction of Cloud Clusters
传统云图时间序列预测方法在刻画小尺度云团的快速生消过程与大尺度云系的缓慢移动及其跨尺度作用方面存在不足,导致预测结果中出现云团结构模糊和不确定性增加的问题.为了有效解决多尺度时空动态性特征捕捉问题,本文提出一种基于多尺度异步门控特征融合的卫星云图预测模型MAG-ConvLSTM(Multi-scale Asynchronous Gated Conv-LSTM).该方法基于ConvLSTM架构,引入分层异步更新机制,通过三步状态更新策略适配大、中、小不同尺度云团运动速度差异,增强对多尺度云团时间演变特征的捕获能力.在ConvLSTM单元中集成空间多尺度特征提取模块,实现对不同尺度云团空间演变特征的同时提取.最后设计时空门控特征融合机制,动态调节历史信息与当前输入信息的融合比例,缓解预测结果随时间模糊化的问题.实验采用葵花8号L1数据集,空间分辨率为5 km,时间分辨率为10 min,TBB_13红外通道.实验结果表明,MAG-ConvLSTM模型在SSIM和PSNR这2项指标上相比基线模型ConvLSTM分别提升了12.87%和7.62%,RMSE降低了18.44%,同时也优于SwinLSTM等若干当前先进模型.该模型能准确预测云团的生消演变过程,为复杂气象条件下的云图精准预测提供了有效的解决方案.
Traditional cloud image time series prediction methods exhibit limitations in capturing the rapid generation and dissi-pation processes of small-scale cloud clusters as well as the slow movement of large-scale cloud systems and their cross-scale in-teractions,leading to issues such as blurred cloud structures and increased uncertainty in prediction results.To effectively ad-dress the challenge of capturing multi-scale spatiotemporal dynamic features,a satellite cloud image prediction model based on multi-scale asynchronous gated feature fusion model MAG-ConvLSTM(Multi-scale Asynchronous Gated ConvLSTM)is pro-posed.This method builds upon the ConvLSTM architecture and introduces a hierarchical asynchronous update mechanism.Through a three-step state update strategy,it adapts to the varying motion speeds of large,medium,and small-scale cloud clus-ters,enhancing the ability to capture temporal evolution features of multi-scale cloud systems.The model integrates a spatial multi-scale feature extraction module within the ConvLSTM unit to simultaneously extract spatial evolution features of cloud clus-ters at different scales.Finally,a spatiotemporal gated feature fusion mechanism is designed to dynamically adjust the fusion ra-tio between historical information and current input information,mitigating the issue of prediction blurring over time.Experi-ments were conducted on the Himawari-8 L1 dataset with a spatial resolution of 5 kilometers and a temporal resolution of 10 min-utes,focusing on the TBB_13 infrared channel.Experimental results demonstrate that the proposed MAG-ConvLSTM model achieves improvements of 12.87%and 7.62%in SSIM and PSNR,respectively,and decreased 18.44%in RMSE compared to the baseline ConvLSTM model,and also outperforms several state-of-the-art models such as SwinLSTM.The model accurately predicts the generation,dissipation,and evolution processes of cloud clusters,providing an effective solution for precise cloud image prediction under complex meteorological conditions.
张传远;张海清;唐鑫;杨孟辑;王晓雍;于曦
成都信息工程大学软件工程学院,四川 成都 610225||智能公差设计与检测四川省工程研究中心,四川 成都 610255成都信息工程大学软件工程学院,四川 成都 610225||智能公差设计与检测四川省工程研究中心,四川 成都 610255成都理工大学工程技术学院,四川 乐山 614000法国里昂第二大学,法国 里昂 69500青海省海北藏族自治州气象局,青海 海北 812200成都大学斯特灵学院,四川 成都 610106
信息技术与安全科学
卫星云图时间序列多尺度动态建模异步更新门控特征融合
satellite cloud imagetime seriesmulti-scale dynamic modelingasynchronous updatinggated feature fusion
《计算机与现代化》 2026 (9)
55-64,10
四川省科学技术厅项目(2025YFHZ0219)青海省气象局"揭榜挂帅"科技项目(QXGS2023-01)
评论