Introduction: The Challenge of Class Imbalance
In the field of meteorological monitoring, the transition from expensive specialized sensors to ubiquitous surveillance cameras offers a cost-effective path toward high-resolution weather data. However, a significant hurdle remains: class imbalance. In many regions, common weather conditions like "sunny" or "cloudy" occur with high frequency, while critical events such as "snowy" or "hazy" days are rare, often accounting for less than 2% of annual data.
Standard machine learning models are inherently "accuracy-hungry," often optimizing for the majority class at the expense of the minority. A model that predicts "sunny" 100% of the time in a 98% sunny dataset achieves high accuracy but fails its primary purpose of detecting rare weather events. This paper proposes a robust framework to solve this imbalance through generative data augmentation and ensemble learning.
The GAN-ENN Ensemble Framework
The authors introduce a multi-stage pipeline designed to balance datasets and improve classification stability. The architecture integrates generative modeling, feature extraction, and a voting-based classification committee.
1. Data Augmentation via DCGAN
Rather than simple oversampling—which can lead to overfitting by duplicating existing images—the framework employs a Deep Convolutional Generative Adversarial Network (DCGAN). The GAN learns the underlying distribution of minority classes to generate entirely new, synthetic images of rain, snow, or haze, effectively expanding the diversity of the training set.
2. Feature Extraction and Data Cleaning
To process these images, the team utilizes a pre-trained VGG16 network to extract high-level feature vectors. Recognizing that GANs can occasionally produce distorted or unrealistic samples, the authors apply the Edited Nearest Neighbors (ENN) rule. This algorithm identifies and removes "noisy" generated samples whose labels contradict the majority of their neighbors in the feature space, ensuring only high-quality data informs the model.
3. Ensemble Classification
To ensure robust performance, the framework avoids reliance on a single algorithm. Instead, it uses an ensemble of four distinct classifiers: k-Nearest Neighbors (KNN), Support Vector Machines (SVM), Random Forest (RF), and AdaBoost. The final weather classification is determined by a majority vote among these models.
Experimental Results
The framework, designated as +GAN+ENN, was evaluated across three datasets: the Multi-class Weather Image (MWI) set, the larger Multi-class Weather Dataset (MWD) with 65,000 images, and a laboratory-curated self-made dataset.
Key findings include:
- Superior Accuracy: On the MWD dataset, the ensemble reached an accuracy of 99.16%, a significant leap from the 89.32% achieved using original imbalanced data.
- The Cleaning Advantage: The inclusion of the ENN cleaning step consistently outperformed raw GAN augmentation. By removing "unfriendly" neighbors, the model avoids the confusion caused by distorted synthetic images.
- Ensemble Stability: While SVM proved to be the strongest individual classifier, the voting method provided the highest overall performance, catching nuances in lighting and shadows that single models might miss.
Conclusion and Future Directions
The study demonstrates that combining generative models with intelligent data cleaning and ensemble learning can effectively overcome the limitations of skewed real-world data. While the system excels at generating high-quality haze and fog images, the authors note that visually similar conditions—such as light rain versus light snow—remain a challenge for current generative models. Future work will focus on refining these generative techniques to better distinguish between visually complex minority classes.
Summary updated