多通道特征融合Y型全卷积网络的对流云检测

Detection of convective cloud in Y-type full convolution network with multi-channel feature fusion

强对流天气具有生命周期短、突发性强、破坏性大等特点,并时常伴随着多种灾难性天气,给经济发展、环境保护、人民生命财产安全等带来巨大威胁。目前目视解译的卫星云图对流云检测方法依赖于人的经验和知识,存在难于界定对流云团边界、云图的多光谱信息利用不足、小尺度对流云易出现漏检与误检等问题。本文基于FY-2G卫星的红外1通道云图及水汽与红外通道的亮温差,并借鉴U-net网络在图像分割中所具有的精确定位能力,提出了一种新的多通道特征融合Y型全卷积网络的对流云检测方法。该方法将U-net网络改造成具有双路输入的Y型全卷积网络,并将红外1通道云图和亮温差图像分别作为Y型网络的两路输入,经过卷积及下采样处理,提取不同通道的特征信息;为了使网络具有精细的目标检测能力,Y型全卷积网络保留U-net网络的卷积及上采样结构,同时通过卷积和上采样将两个输入分支不同层次的特征图融合,从而实现一种多层次、多通道特征融合的对流云检测方法;不同层次特征图的可视化及其与融合特征图的对比,表明了所构造的Y型网络在利用云图不同通道特征信息中的有效性。实验结果表明,本文方法的对流云检测准确率为87.34%,精确率为89.77%,召回率为82.10%,F1-综合评价指标为84.82%,各项性能指标均优于基于DeconvNet、U-net等传统网络模型的对流云检测方法;与阈值法、亮温差法和SVM等传统对流云检测方法相比,本文方法不仅在对流云边缘界定及小尺度对流云的检测上具有明显优势,而且检测准确率和计算效率均得到了显著的提高。

Severe convective weather has the characteristics of short life cycle,sudden strong,destructive,and often accompanied by a variety of catastrophic weather,to economic development,environmental protection,people′s lives and property security and other great threats.At present,the visual interpretation of satellite cloud images is dependent on human experience and knowledge,and there are some problems such as difficulty in defining the boundary of convective cloud clusters,insufficient use of multi-spectral information in cloud images,and easy to miss and misdetect small scale convective clouds.In this paper,based on FY-2 G satellite′s infrared channel 1 cloud image and the bright temperature difference between water vapor and infrared channel,and referring to the accurate positioning ability of U-net network in image segmentation,a new convective cloud detection method based on multi-channel feature fusion Y-type full convolution network is proposed.In this method,U-net network is transformed into Y-type full convolution network with double-channel input,and infrared channel 1 cloud image and bright temperature difference image are taken as the two-channel input of Y-type network respectively.After convolution and down-sampling processing,characteristic information of different channels is extracted.In order to enable the network to have fine target detection ability,the Y-type full-convolution network retains the convolution and up-sampling structure of U-net network,and at the same time,the feature graphs of two input branches at different levels are fused through convolution and up-sampling,so as to realize a multi-level and multi-channel feature fusion convective cloud detection method.The visualization of feature maps at different levels and the comparison with fused feature maps show the effectiveness of the constructed Y-type network in utilizing feature information of different channels in cloud maps. The experimental results show that the accuracy,accuracy, recall and F1-measure in this paper are 87.34%,89.77%,82.10% and 84.82%,respectively.The performance indexes of the method in this paper are better than those in traditional network models such as DeconvNet and U-net.Compared with the traditional convective cloud detection methods such as threshold method,bright temperature difference method and SVM,the method in this paper not only has obvious advantages in the edge definition of the convective cloud and the detection of small scale convective cloud,but also has significantly improved the detection accuracy and computational efficiency.