Class Weighting Approach For Handling Imbalanced Data On Forest Fire Classification Using EfficientNet-B1

Abstract

Listen

Translate article

Wildfires pose significant threats to ecosystems and human safety, necessitating effective monitoring techniques. Detecting forest fires based on images of forest conditions could be a breakthrough. But, the model built from imbalanced data leads to low accuracy. This research addresses the challenge of class imbalance in multi-class classification for forest fire detection using the EfficientNet-B1 model. This research explores the implementation of class weighting to enhance model performance, particularly focusing on minority classes namely: Fire, Smoke. A dataset of 7,331 training images, categorized into four classes. The results showed that employing the class weighting method achieved an accuracy of 90%. While training duration of 14 minutes and 45 seconds, outperforming the data augmentation method in terms of time efficiency. This study contributes to the development of more effective methods for forest fire monitoring and provides insights for future research in machine learning applications in environmental contexts.