A Meteorological Landscape Identification Algorithm: Based on Improved CNN Combined with Multi-level Feature Fusion

The meteorological landscape is a vital part of tourism viewing and a substantial representation of meteorological conditions, but adequate monitoring and recording tools are currently missing. This study proposes an algorithm model based on an updated CNN network combined with multi-level feature fusion for identifying typical meteorological sceneries. First, five popular meteorological landscape datasets are created, and four image features are employed as shallow features: dark channel, brightness, average gradient, and contrast. Simultaneously, a CNN(Convolutional Neural Network) model ResNet34-F is provided as a medium-deep semantic information extractor, which has a better ability to extract middle-level semantic information from images. Finally, a better training approach and loss function are employed to further enhance the model based on the application background of this paper. With a 96.5 percent accuracy, this model can recognize five frequent meteorological landscapes: Glaze ice, Rime, Sunrise, Glory(optical phenomena), and Sea of clouds. The ablation experiment demonstrates the importance of the proposed improvement in this research and the improved accuracy of our method is higher than that of adding the attention mechanism to CNN.