基于残差网络的地基云图识别方法研究

Research on ground-based cloud recognition method based on residual network

地基云的精细化识别对气候预测和气象研究具有重要的意义。针对目前地基云识别准确率低、泛化性差、不利于边缘化部署的问题,提出了基于残差网络的地基云图识别模型,命名为GBcNet。设计的模型由1个卷积层、2个池化层、5个残差块以及1个全连接层构成,利用第1个卷积层和第1个池化层初步提取特征信息并降低特征图大小,通过残差块提取更多的特征信息,同时抑制网络的过拟合和梯度消失,最后利用另1个池化层降低特征图的大小,并通过全连接层输出识别结果。利用数据集对模型进行训练和测试,实验结果表明,GBcNet模型对数据集的综合平均准确率达到了96.02%,11种类别地基云的识别精确率均在93%~99%,且具有更好的泛化性,单个类别和整体识别性能均优于其他模型。进一步采用SWIMCAT数据集对模型进行实验,综合识别准确率达99.7%,证明模型对地基云图识别具有普适性。模型结构简单,相较于其他模型,更有利于边缘化部署。

The fine identification of ground-based clouds is of great significance for climate prediction and meteorological research.Aiming at the current problems of low accuracy of ground-based cloud recognition,poor generalization,and detrimental deployment of marginalized deployment,a ground-based cloud map recognition model based on residual network is proposed,named GBcNet.The designed model consists of one convolutional layer,two pooling layers,five residual blocks and one fully connected layer,using the first convolutional layer and the first pooling layer to initially extract feature information and reduce the feature map size,and extract more feature information through the residual block,while suppressing the overfitting and gradient disappearance of the network,and finally using another pooling layer to reduce the size of the feature map,and finally output the recognition results through the fully connected layer.The experimental results show that the comprehensive average accuracy of GBcNet model on the dataset reaches 96.02%,and the recognition accuracy of 11 categories of ground-based clouds is between 93%and 99%,and has better generalization,and the recognition performance of single category and overall is better than that of other models.Furthermore,the SWIMCAT dataset was used to experiment with the model,and the comprehensive recognition accuracy reached 99.7%,which proved that the model was universally applicable to the recognition of ground-based cloud maps.The model has a simple structure,which is more conducive to marginalized deployment than other models.