基于全国开展的降水天气现象平行观测工作期间积累的数据基础，收集并整理降水现象仪自动观测数据和人工观测数据，解析降水现象仪产生的雨滴谱数据文件，依据降水粒子大小和下落速度分级以及累计时段内粒子数量，生成与之对应的二维图像，同时结合降水现象仪平行观测期间的人工观测数据，利用深度学习图像分类技术进行训练，建立降水现象自动识别模型，完成不同降水现象的自动识别。模型训练集和验证集平均准确率均达到86%，雨的测试准确率达到62.7%，雨和毛毛雨的总识别率达到89.1%，雪的测试准确率达到93%，说明利用深度卷积神经网络对雨滴谱数据生成的雨滴图进行自动降水现象识别方案可行。
Based on the data accumulated during
the parallel observation of precipitation weather phenomena carried out
throughout the country, collect and sort out the automatic observation data and
manual observation data of the precipitation phenomenometer, analyze the
raindrop spectrum data file generated by the precipitation phenomenometer to
generate the corresponding two-di- mensional image. At the same time, combined
with the artificial observation data during the parallel observation period of
the precipitation phenomenon instrument, the deep learning image classification technology is used for training,
and the automatic recognition model of precipitation phenomenon is established to complete the automatic
recognition of different precipitation phenomena. The average accuracy of the
model training set and validation set is 86%, the test accuracy of rain is
62.7%, the total recognition rate of rain and drizzle is 89.1%, and the test
accuracy of snow is 93%, indicating that the scheme of automatic precipitation
recognition based on the raindrop pattern generated by the deep convolutional
neural network is feasible.
Uchida, K., Tanaka, M. and Okutomi, M. (2018) Coupled Convolution Layer for Convolutional Neural Network. Neural Networks, 105, 197-205. https://doi.org/10.1016/j.neunet.2018.05.002
Akilan, T., Wu, Q.M.J., Yang, Y., et al. (2017) Fusion of Transfer Learning Features and Its Application in Image Classification. 2017 IEEE 30th Canadian Conference on Electrical and Computer Engineering (CCECE), Windsor, 30 April-3 May 2017, 1-5. https://doi.org/10.1109/CCECE.2017.7946733
