An Analysis of Pre-trained Models Versus Custom Deep Learning Models for Forest Fire Detection

• Main Authors: Partheepan, Shouthiri, Sanati, Farzad, Hassan, Jahan
• Format: Conference Proceeding
• Language: English
• Subjects: Analytical models; classifier; convolutional neural network; custom model; Data models; Deep learning; Feature extraction; Forestry; pre-trained model; Safety; Transfer learning
• ISSN: 2474-154X
• DOI: 10.1109/ITNAC59571.2023.10368557

The detection of forest fires is crucial for human and environmental safety due to their catastrophic impacts. Utilizing machine learning (ML) and deep learning (DL) enhances forest fire management by analyzing data, enabling early warning systems, monitoring fire behaviour, and optimizing resource allocation. The DL algorithms provide a significant influence in the context of forest fire detection. The result of many research studies in the recent past is evident that it is becoming increasingly important in future detection tasks. The majority of research studies use DL models that have been pre-trained. Nevertheless, an investigation should examine how custom models perform compared to pre-trained models in detection tasks. This study aims to compare the performance of the simple CNN model with that of the pre-trained DL model and also compare the pre-trained model as a feature extractor for the custom model. Five CNN pre-trained models, such as VGG16, ResNet50, MobileNetV2, Xception, and Inception were used for our study. The Xception model achieved high accuracy in both scenarios as a pre-trained model and feature extractor with custom layers for fire detection. Also, the result proved that combining CNN models as a feature extractor with custom layers performed well compared with the custom and pre-trained models.