| Weather is an important component of human life,and different weather conditions determine different living locations and lifestyles.The advantages of traditional weather identification methods are high reliability,based on meteorology and remote sensing technology,with high scientific and reproducible qualities.However,these methods require a large amount of meteorological data and satellite remote sensing images,which are costly and require professional equipment and personnel to operate.At the same time,weather changes are often very rapid,and these methods cannot reflect the current weather conditions in a timely and accurate manner.The use of outdoor visual equipment such as cameras has the advantages of realtime performance and efficient acquisition of images.By combining traffic cameras and network cameras distributed throughout the city and using deep learning models to classify and recognize outdoor visual images obtained from cameras,the accuracy of realtime weather recognition can be improved in small-scale scenarios at a lower cost.Typically,meteorological images contain multiple aspects of information that may have complex interactions.A single deep learning model may not capture all this information,which can lead to overfitting,increased noise,or higher recognition error rates.To improve the predictive recognition ability of meteorological images,a multimodel ensemble method can be adopted.By combining multiple different deep learning models and utilizing their differences to compensate for the shortcomings of a single model,the accuracy of classification and recognition results can be improved.This article uses deep learning models as base learners for ensemble learning and combines them with adaptive evidence inference rules to study meteorological recognition algorithms.The main work is as follows:(1)A-ER(Adaptive-Evidential Reasoning)rule is proposed as an integration strategy for ensemble learning,addressing the issue of using the same evidential parameters for multiple models in traditional Evidential Reasoning(ER)rules.The model first pre-classifies the results of the base learners using the random forest algorithm,and then calculates the corresponding evidence weights and reliabilities for different categories.Then,the evidence category is judged before each evidence fusion,and different evidence parameters are adaptively assigned according to different categories.Finally,experiments demonstrate that this model can effectively improve image classification accuracy.(2)A-ER rule-based meteorological recognition model is proposed to address the limitations,robustness,and recognition accuracy issues that may arise from using a single deep learning model for recognizing meteorological images.For the first time,this method applies the ensemble learning model based on A-ER rules to the field of meteorological recognition,which improves the generalization ability of multiple different meteorological recognition base learning models by fusing their results.Experiments have shown that this method is effective in multi-category meteorological image recognition and has higher classification recognition accuracy compared to a single model.(3)Designed and implemented a meteorological recognition system based on A-ER rules.The system includes four modules: model training,parameter adaptation,meteorological recognition,and result visualization.By analyzing the requirements of each module of the system,the backend algorithms and the Windows display interface were developed and designed.Finally,the feasibility of the system was validated through example testing. |
