Automatic Observation Experiments of Cloud Amounts and Cloud Forms Based on the Image Recognition

Abstract

This paper uses the instruments independently developed by Hunan Meteorological Station to carry out automatic observation experiments for cloud amounts and cloud forms in Changsha and Xilinhot. By comparing the manual and automatic observation results at 270 observation times from June to July in 2019, the feasibility of the automatic observation for cloud amounts and cloud forms based on image recognition is discussed. Meanwhile, suggestions and the basis for instrument improvement are provided in this paper. The main conclusions are as follows. During the test, the instrument runs stably; compared with those in manual observations, the consistency rates of the automatic observation results for cloud amounts and cloud forms are 70.8% and 77.1%, respectively; the consistency rates are over 90% when precipitation occurs. Most automatic observation deviations of the cloud form are concentrated in recognizing cloud genera. Being able to provide favorable supports for weather forecasts and services, the automatic observation results are highly applicable. The simple datasource and imbalanced classification of training samples are the main reasons for the low consistency rates of automatic observation results in some cases, and the image distortion of the fish-eye camera also has a great influence on the cloud-amount recognition results. The automatic observation results could be improved by enriching training sample sets, as well as optimizing the image processing and recognition algorithm in the future.