Lightweight models for weather identification

At present, the recognition of weather phenomena mainly depends on the weather sensors and the weather radar. However, large-scale deployment of meteorological observation equipment for intensive weather monitoring is difficult because it is expensive and difficult to maintain. Moreover, convolutional neural networks (CNNs) can also be used to identify weather phenomena, but existing methods require high computing power of equipment, making it difficult to deploy in practice. Therefore, designing a lightweight model that can be deployed in a small device with weak computing power is crucial for intensive weather monitoring. In this paper, we study the shortcomings of some existing lightweight models. By comparing the disadvantages of these models, a new lightweight model is proposed. In addition, considering the number of existing weather datasets are too small to meet real monitoring needs, so we produced a dataset with a more complex variety of weather phenomena. Through the experiments, the proposed method can save more than 25 times memory usage with only 1.55% accuracy lost compared with the best CNNs method which achieves state-of-the-art performance among the other lightweight models.