VIDEO-BASED WEATHER PHENOMENON RECOGNITION AND ITS APPLICATION RESEARCH
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.
solar energyimage recognitionvideo camerasneural networksweather phenomenasunshine duration
NETL
NSTL
万方数据
