Forest Fire Image Classification via Hybrid Deep Learning and Stacking Ensemble Technique
In recent years, forest fires have increased significantly and in an intimidating manner worldwide, thereby affecting both the environment and human life. When they are overlooked and not detected at an early stage and in a short period of time, they will spread rapidly. Accurate and early fire dete...
Saved in:
| Published in: | Journal of Al-Qadisiyah for Computer Science and Mathematics Vol. 17; no. 3 |
|---|---|
| Main Author: | |
| Format: | Journal Article |
| Language: | English |
| Published: |
30.09.2025
|
| ISSN: | 2074-0204, 2521-3504 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In recent years, forest fires have increased significantly and in an intimidating manner worldwide, thereby affecting both the environment and human life. When they are overlooked and not detected at an early stage and in a short period of time, they will spread rapidly. Accurate and early fire detection and classification are crucial for disaster control and well-timed emergencies. To address this problem, a hybrid model of pre-trained embedding and diverse classifiers to extract features, detect, and classify fire images with optimized performance and stacking with different meta-learners to improve the reliability, is applied in this paper. Three types of pre-trained models, including InceptionV3, SqueezeNet, and DeepLoc, were used for feature extraction with different classifiers: Neural Network (NN), Decision Tree (DT), and XGBoost. The proposed model, which involves stacking multiple classifiers, provides an accurate, optimal, and efficient response for fire classification. A total of 1520 images were used in this study. The best performance was achieved by integrating InceptionV3 with NN, XGBoost, and Logistic Regression as a meta-learner, yielding 98.9% accuracy, 99.9% AUC, and 98.9% for each of (F1-score, Precision, and Recall) with 97.8% MCC, with 10 False Negatives and 7 False Positives in error types. This combination of deep and machine learning with stacking shows consistent progress across all experiments, and these results can enhance and improve the quality, leading to more dependable detection and classification in safety-critical applications. |
|---|---|
| ISSN: | 2074-0204 2521-3504 |
| DOI: | 10.29304/jqcsm.2025.17.32402 |