激光与光电子学进展, 2020, 57 (20): 201011, 网络出版: 2020-10-17
结合双流3D卷积和监控图像的降水临近预报 下载: 687次
Precipitation Nowcasting Based on Dual-Flow 3D Convolution and Monitoring Images
图像处理 3D卷积 室外监控图像 光学图像 降水临近预报 image processing 3D convolution outdoor monitoring image optical imag precipitation nowcasting
摘要
针对大部分降水临近预报产品无法兼顾高覆盖率、高准确率及低成本的问题,提出一种基于室外监控图像和深度神经网络能预报未来1 h降水强度的方法。设计双流3D卷积神经网络来提取图像降雨信息的高维特征。该网络在低计算代价下自适应产生局部信息,并通过双损失函数从整体和局部统筹网络,提取降雨信息的时间特性和空间特性。实验结果表明,在降水强度预报领域,基于双损失函数的神经网络优于单损失函数。所提网络的误警率、命中率、临界成功指数、准确率在多数情况下优于其他模型。在模型效果可视化方面,所提网络能有效提取降水图像的特征信息。所提降水临近预报方法有能力进行精细且低成本的降水临近预报。
Abstract
At present, most of the precipitation nowcasting production is unable to consider the problems of high coverage, high accuracy, and low cost. Therefore, we herein propose a method based on outdoor monitoring images and deep neural network to forecast the rainfall intensity in the next 1 h. We design a dual-flow 3D convolutional neural network to extract high-dimensional features of rainfall information in images. The local information is adaptively generated at a low computational cost, and the temporal and spatial characteristics of rainfall information are extracted by the proposed network which integrates the whole network and the local network using a double loss function. The experimental results show that the neural network based on the dual loss function is better than that based on the single loss function in precipitation intensity forecasting. Percent of doom, false alarm rat, critical success index, and the accuracy of the proposed network are better than those of other models in most cases. In terms of visualization of the model effect, the proposed network can effectively extract the feature information of the precipitation images. Therefore, the proposed precipitation nowcasting method is capable of fine and low-cost precipitation prediction.
杨素慧, 林志玮, 赖绍钧, 刘金福. 结合双流3D卷积和监控图像的降水临近预报[J]. 激光与光电子学进展, 2020, 57(20): 201011. Suhui Yang, Zhiwei Lin, Shaojun Lai, Jinfu Liu. Precipitation Nowcasting Based on Dual-Flow 3D Convolution and Monitoring Images[J]. Laser & Optoelectronics Progress, 2020, 57(20): 201011.