G-CNN模型在浓雾天气形势识别中的应用

天气形势图特征分类是实现浓雾智能在线预报的关键因素之一,为了探索高效、正确的天气形势图分类模型,一种基于方向滤波器的深度卷积神经网络(Gabor-Convolutional Neural Network,G-CNN)模型被提出,模型通过训练已建立的浓雾天气形势图-雾型关系数据集,建立形势图纹理特征与雾型之间非线性映射关系进而实现天气形势图智能化识别。G-CNN模型:利用Gabor滤波器对输入的天气形势图纹理特征进行强化;利用2个卷积-池化层及2个全连接层的CNN架构拟合天气形势图与雾型之间的映射关系。利用江苏地区2010至2016年浓雾天气形势图-雾型关系数据集,从中随机取样70%作训练集,余下30%样本作为测试集的情形下,训练、测试建立的模型,并针对3个惯用的评价指标:准确率(Probability of Detection,POD)、虚警率(False Alarm Rate,FAR)及临界成功指数(Critical Success Index,CSI)对模型进行评价。试验结果显示POD、FAR及CSI分别为0.86、0.11及0.77,指标值表明模型具有高度的有效性和正确性。

Application of G-CNN Model in Foggy Weather Situation Recognition

Correct classification of fog weather situation map is one of key factors to carry out a fog intelligent online recognition model, in order to efficiently and correctly recognize fog type（ToF） according to Weather Situation Map（WSM）, a Gabor-Convolutional Neural Network（G-CNN） fog intelligent online recognition model is proposed, within training built WSM-ToF relation dataset through G-CNN to build a non-linear map relationship between WSM texture feature and the classification label based on mankind forecasting. In theory, Gabor filter is firstly applied to enhance texture feature of input image by G-CNN. Secondly, in order to build a non-linear map relationship between the input and the output, it adopts G-CNN architecture with 6-layers to carry out non-linear map relationship, in which output layer is taken as the predicted fog label for input. With WSM-ToF relation dataset of Jiangsu region from 2010 to 2016, 70% samples are randomly taken as training dataset and the rest 30% is taken as test dataset, and the model is trained, tested and evaluated based on the three common indicators：Probability of Detection（POD）, False Alarm Rate（FAR） and Critical Success Index（CSI）, experimental results show that POD, FAR and CSI are reached at 0.86, 0.11 and 0.77, and these evaluation values comfirm the proposed model is highly correct and efficient.