Yin et al. (2025) Meteorological observation research based on an improved EfficientNetV2 model

Identification

• Journal: Environmental Modelling & Software
• Year: 2025
• Date: 2025-12-16
• Authors: Houshang Yin, Yu Cao, Linlin Liu, Dan Chen, Qiong Zhang
• DOI: 10.1016/j.envsoft.2025.106835

Research Groups

• School of Information and Control Engineering, Liaoning Petrochemical University, Fushun, China
• School of Artificial Intelligence and Software, Liaoning Petrochemical University, Fushun, China

Short Summary

This study proposes a novel deep learning model, EfficientNetV2-CBAM-PANet, to improve meteorological image recognition in complex weather scenarios by enhancing feature extraction and robustness. The model achieved a recognition accuracy of 97.6% on a self-constructed dataset, demonstrating strong classification capability across various weather conditions.

Objective

• To address challenges in meteorological image recognition under complex weather conditions, specifically insufficient feature extraction, inadequate utilization of scale information, and poor robustness to interference, by developing an improved deep learning model.

Study Configuration

• Spatial Scale: Not explicitly defined, but implicitly local to regional as it focuses on image-based weather phenomena.
• Temporal Scale: Real-time or near real-time for meteorological observation and classification.

Methodology and Data

• Models used: Improved EfficientNetV2 model, integrated with the Convolutional Block Attention Module (CBAM) and the Path Aggregation Network (PANet) structure. Transfer learning was applied using a pretrained EfficientNetV2 model.
• Data sources: A self-constructed dataset of meteorological images.

Main Results

• The proposed EfficientNetV2-CBAM-PANet model achieved a recognition accuracy of 97.6% on the self-constructed dataset.
• The model demonstrated strong classification capability across various weather conditions, outperforming some existing deep learning models in accuracy.

Contributions

• Proposes a novel deep learning architecture (EfficientNetV2-CBAM-PANet) specifically designed to overcome limitations in meteorological image recognition, such as insufficient feature extraction and poor robustness in complex weather.
• Enhances training efficiency and accuracy through the strategic use of transfer learning with EfficientNetV2.
• Integrates the CBAM attention mechanism to improve the model's perception of critical meteorological features.
• Incorporates the PANet structure to enable effective multi-level feature fusion, addressing issues related to scale information utilization.
• Achieves a high recognition accuracy (97.6%), providing a robust solution for weather image classification and offering insights for future forecasting research.

Funding

• Not specified in the provided text.

Citation

@article{Yin2025Meteorological,
  author = {Yin, Houshang and Cao, Yu and Liu, Linlin and Chen, Dan and Zhang, Qiong},
  title = {Meteorological observation research based on an improved EfficientNetV2 model},
  journal = {Environmental Modelling & Software},
  year = {2025},
  doi = {10.1016/j.envsoft.2025.106835},
  url = {https://doi.org/10.1016/j.envsoft.2025.106835}
}