基于视频的天气现象识别及其应用研究

VIDEO-BASED WEATHER PHENOMENON RECOGNITION AND ITS APPLICATION RESEARCH

针对传统天气现象观测因采用专用设备导致布设和维护成本高、获取难的问题,提出一种利用广泛布设的视频实现对天气现象观测的方法。通过将因特网获取的16327张天气现象图片放入深度神经网络中训练,建立一个天气现象分类的预训练模型。在此基础上加入上海徐家汇和洋山港气象站2021年视频图像,对模型进行微调,开发基于站点视频的天气现象识别模型。利用2022年1—10月份的视频图像数据对模型进行检验,模型识别结果的F1评分分别为0.74和0.67,而人工识别结果分别为0.67和0.61,表明所建立的模型性能接近或优于人眼识别的效果。通过例举两个应用案例,证明该项技术具有较好的应用前景。

In this study a newly method is proposed to identify weather phenomena with widely dispersed video data to address the issues of high deployment and maintenance costs by traditional weather phenomenon observation equipment.In this method,a deep neural network is firstly used to train 16327 weather phenomenon photographs from internet for building a pre-training categorization model,and then a fine-tuning procedure is applied with those video photographs from two meteorological observation station to improve the identifying accuracy.The model was testeel by using the video image data from January to October 2022.The results indicate that the performance of the proposed model is comparable to or superior to that of human eye recognition with the F1 scores of 0.74 and 0.67 for the model recognition in the two different station compared with 0.67 and 0.61 for the manual recognition.A cases analysis shows that the proposed model is mostly perfect for identifying sunny and fog but unsatisfactory for rainfall.Further,two application cases show that the model can be used to retrieve the sunshine duration and identify weather phenomena in real time using the online video images.Most important,this method can be employed as a substituted scheme for the present weather phenomena instrument to achieve low-cost weather information observation.