Unleashing the Power of Probabilities: Soft Labels in Data Augmentation
Abstract
Context: Wildfire prediction is critical for mitigating natural disasters and protecting ecosystems. Traditional models often rely on complex labels, which may not capture the complexities and uncertainties inherent in real-world data.
Problem: Existing wildfire prediction models can struggle with overfitting and need more robustness, especially in noisy and ambiguous data environments.
Approach: This essay explores soft-label data augmentation, specifically MixUp, to enhance the generalization and robustness of wildfire prediction models. A synthetic dataset is generated, and a neural network is trained with MixUp augmentation. Feature importance is also analyzed using a Random Forest model.
Results: The augmented model achieves an accuracy of 98%, with a balanced performance across classes. The confusion matrix and classification report confirm the model’s reliability, and feature importance analysis identifies the most influential features.
Conclusions: Incorporating soft labels through MixUp significantly improves model performance and robustness. This approach offers a promising solution for enhancing wildfire prediction models, making them more adaptable to real-world data complexities.
Keywords: Soft Labels Data Augmentation; Wildfire Prediction Models; MixUp Technique; Machine Learning Robustness; Feature Importance Analysis.
