Ground-based Cloud Recognition Based on Lightweight Neural Network

In view of the current low efficiency of cloud identification and the lack of recognized and rigorous open cloud dataset, the largest open cloud dataset HBMCD was constructed by professional equipment combined with manual annotation and transfer learning. It conformed to the standards of the international meteorological organization under the guidance of professionals. On the basis of this, the basic unit by means of depth separable convolution and expansion convolution was constructed, and the light cloud image classification model LCCNet was established by repeating different basic units. Through enough comparative experiments, it is proved that LCCNet not only has low level at parameters and operation complexity, but also has a high accuracy (97.35%) for HBMCD data set, which provides a possibility for equipment integration and practical application.