For the problem that deep convolutional neural network models are difficult to deploy at the edge due to their large number of parameters and high hardware device requirements, this paper researched on the weather re-cognition algorithm based on light-weight neural network with the application scenario of highway surveillance images. The light-weight neural network model MobileNet was first analyzed theoretically, and the difference between the deeply separable convolution operation and the standard convolution operation was analyzed in terms of the number of parameters and the number of computations. At the same time, a weather recognition dataset based on highway surveillance images was collected and labeled. Based on this, models including several light-weight neural networks were built and trained for comparison experiments, and the experimental results verified the advantages of MobileNet in terms of recognition accuracy, speed and number of model parameters. In addition, this paper explored the feature representation of MobileNet as well as the inter-class separability and intra-class clustering of features by the visualization algorithm t-SNE in terms of both class activation analysis and feature distribution, and the results further supported the above analysis.