Comparative Study of CNN and MobileNetV2 for Forest Fire Detection with Data Augmentation

Abstract

Forest fires pose a significant environmental threat, and quick, reliable detection is essential to reduce damage. This paper compares lightweight custom Convolutional Neural Network (CNN) with a fine-tuned MobileNetV2 for automated forest fire detection. Both models were trained on a Kaggle dataset containing 5,050 labeled forest images. The images underwent preprocessing that included resizing and normalization, and they were improved through extensive data augmentation like rotation, zoom, shifting, and flipping. The CNN achieved an accuracy of 92% with an F1-score of 0.90 for detecting fire. In comparison, MobileNetV2 achieved a higher accuracy of 95% with an F1-score of 0.94, cutting false negatives by 39%. Both models reached a ROC-AUC of 0.99, but MobileNetV2 showed better sensitivity, making it more suitable for real-time monitoring systems like drones and satellites. The results show that while CNN is robust with fewer false positives, MobileNetV2 strikes a better balance between precision and recall, making it the preferred choice for forest fire detection.