Abstract:Based on field experiments at Nanyue Mountain Meteorological Station and Huaihua National Reference Climatological Station in Hunan Province, the camera images of snow cover weather phenomena were collected minutely from January to March in 2018. The convolution neural network technology is employed for modelling and training using the experimental pictures at Nanyue Station, and the results are examined by the testing pictures of Nanyue Station and Huaihua Station. Furthermore, based on deep learning, the environmental layout requirements of snow cover image identification are discussed. The main conclusions are as follows. Recognition accuracy at Nanyue Station is 99.23%, omission rate is 0.49%, false identification rate is 0.28%, and the recognition result at daytime is better than at night. The accuracy increases to approximately 99.99% with high stability during the formation stage of snow cover, and it decreases with snow melting. There are a few false cases when the snow cover on the ground is in the early stage of snow cover formation and about to melt. There are occasional misjudgment because of background pollution of fog and rime. The test results of Huaihua Station are similar to those for Nanyue Station, with an accuracy of 97.78%, a false identification rate of 1.92% and 0.3% missed, but less stable. This is because of two reasons. First, the data of Huaihua were not used for modelling, and then the cameras were not well fixed leading to bad images. The test results show that the artificial intelligence identification model established in this paper can extract the key features of snow cover in different stages, and the identification result is well. In addition, the false accept and omission can be further eliminated by including the meteorological element conditions and judging the consistency of the snow co ver. This method can provide important technical support for the automatic observation of snow cover weather phenomena.