| Cloud plays an important role in weather forecasting,meteorological disaster prevention,and ecological environment monitoring,therefore accurate cloud classification is of great significance.Traditional cloud classification methods mainly use manual methods such as statistical observation and threshold screening,which cannot meet the requirements of real-time classification.In recent years,with the development of deep learning,using remote sensing data and neural network technology for cloud classification has become a research hotspot.In the current cloud classification task,the existing deep neural network methods have poor fusion with remote sensing data,and have not fully exerted the performance of meteorological satellites.In addition,there are significant differences in data distribution between different satellites,making it time-consuming and laborious to obtain comprehensive and fine-grained cloud classification labels in different domains.How to train a high-performance model in the domain of complete meteorological satellite data annotation,and apply this model to different satellite data domains to save computational resources and reduce annotation costs,has become an urgent practical problem.Therefore,this paper focuses on the research work of meteorological satellite image cloud classification technology based on domain transfer:1.In view of the characteristics of meteorological satellite images,this paper designs channel dimension feature fusion,multi-scale feature fusion,and pixel-level attention mechanism modules,which innovatively achieve nearly real-time classification of 10 types of clouds.At the same time,visualization results are provided to improve the readability and visualization effect of cloud classification results.2.1n response to the current problem of domain differences caused by channel defaults between different satellites,this paper extracts data features of meteorological satellites,weakens domain differences through adversarial learning,and achieves style consistency transfer,thereby solving the practical problem of visual interpretation.At the same time,the feasibility of domain transfer in the field of meteorological satellites is verified,providing strong support for the construction of the final cloud classification domain transfer model in this paper.3.In response to the more complex cloud classification domain transfer problem compared to visual interpretation,this paper adds a content consistency module design based on style consistency transfer to improve the generalization ability of cloud classification models and better utilize meteorological satellite data to achieve more universal cloud classification tasks.In summary,the meteorological satellite image cloud classification technology based on domain transfer proposed in this paper has important application value and scientific significance,and can provide better support for fields such as weather forecasting,meteorological disaster prevention,and ecological environment monitoring. |
