Location information-fused estimation method in relevance to tropical cyclone intensity

目的 精确估计热带气旋的强度有助于提升天气预报和预警的准确性。随着深度学习技术的不断发展，基于卷积神经网络（convolutional neural network，CNN）的方法已应用于强度估计任务中。然而，现有方法仍存在许多问题，例如无法充分利用不同波段的卫星图像信息、输入图像以热带气旋的定位为中心等限制，从而产生较大误差，影响实时估计的结果。针对以上问题，本文提出一种融合定位信息的强度估计网络IEFL（intensity estimation fusing location），提升强度估计的准确率。方法 模型采用双分支结构，能有效融合不同波段的图像特征，同时可以同步优化两个任务，达到互相促进的效果。此外，模型对强度估计任务做了定位的特征融合，将得到的定位特征图与强度特征图进行拼接，共同输出最后的强度结果，通过利用定位信息达到提升强度估计精度的目的。结果 本文在完成热带气旋强度估计的同时，可获取较好的热带气旋中心定位结果。收集了2015—2018年葵花-8卫星多通道图像用以训练模型，并在2019和2020年的数据上进行测试。结果表明，融合定位信息后模型的强度估计均方根误差为4.74 m/s，平均绝对误差为3.52 m/s。相比传统单一强度估计模型误差分别降低了7%和9%。结论 IEFL模型在不依赖定位准确率的同时，能够有效提升强度估计的准确率。;Objective A tropical cyclone can generate such severe weather condition like strong winds or heavy precipitation，as well as secondary disasters derived of floods，landslides，and mudslides. Tropical cyclones may often threaten survival contexts in related to coastal community. The intensity cyclones（TC）can be estimated accurately and it is beneficial for forecasting warning. Deep learning techniques-based convolutional networks（CNNs）methods have its optimal ability task apparently. However，CNN-based methods are still challenging problem insufficient use multi-channel satellite images，and the input images preferred centered on location cyclones. To resolve large errors real-time results. we develop a network called intensity-estimation-fusing-location（IEFL）to improve accuracy further. Method training data captured from Himawari-8 2015 2018 comparason with 2019 2020. dataset contains 42 028 5 229 testing images. First，the preprocessed remove non-TC cloud systems via clipping Then，the implementation augmentation strategy oriented optimize over-fitting enhance model robustness. Moreover， multiple channel analysis required reveal varied features TCs. Thus，a better combination developed through set up two-branch structure，which used fuse different effectively. Two sort tasks optimized simultaneously learnt mutually. In addition，the feed task-extracted into task. Specifically，their feature maps concatenated results generated following. experiment segmented two categories mentioned below：for first category，it focused only，and channels-related information fusion analyze information-fused impact relevance estimation. For second one，multi-channel integration selected integrated effect IEFL configurable Pytorch toolbox. resized 512×512 pixels training，the momentum parameter 0. 9，the rate 001，the batch size 5，and weight decay 10-4. stochastic gradient descent（SGD） procedure using an NVIDIA GTX TITAN XP device. loss function regression root mean square error（RMSE）. RMSE measure difference between ground truth predictable values intensity. smaller RMSE，the performance model. Furthermore，the recognized well. Therefore，the total sum loss. main contributions listed below：1）develop estimate location，called intensity-estimation-fusinglocation（IEFL）；2）validate intensities Himawari- 8 satellite，and 3）analyze each channel. Result non-location information-relevant error（RMSE）is 5. 08 m/s，in which 4. 74 m/s. Compared without task，the reduced by 7%. error-related comparative analyses carried out other six methods. traditional method deviation angle variance technique（DAVT），it increased about 27%. CNN-based methods，it 11% higher than network-tropical cyclone（CNNTC），8% net（TCIENet），as 4% classification net（TCICENet）. Conclusion terms fusion. This improving beyond accuracy. result shows that has potentials farther.