Research on Meteorological Recognition Method Based on Improved Spatio-Temporal Two-Stream Network Model

Meteorological recognition, as an important application of meteorology, is a crucial foundation for ensuring human safety, agriculture, energy, and social development. It plays a key role in disaster warning, resource planning, and daily life decisions. However, the recognition of complex weather conditions such as rain and snow in challenging scenarios remains a significant challenge for existing models in terms of feature representation. In order to capture richer semantic features, we propose an improved spatio-temporal two-stream network model. This model combines 3D convolutional networks and an improved ConvNeXt network to better capture the temporal and spatial features of meteorological phenomena like rain and snow. In experiments, we compare our proposed algorithm with mainstream image and video classification models such as Swin Transformer and R3D. The results demonstrate that our algorithm achieves an accuracy of 94.8% on a multi-class meteorological video dataset constructed in this study, surpassing mainstream classification models. This result strongly validates the effectiveness and advancement of using the improved spatio-temporal two-stream network model for meteorological recognition.