Forest Fire Detection Leveraging Hybrid Convolutional-Recurrent Models

Forest fires pose serious environmental and economic risks across tropical, temperate, and boreal regions. Traditional detection methods are often limited in accuracy and adaptability, motivating the use of deep learning for automated solutions. While Convolutional Neural Networks (CNNs) have shown promise, fewer studies have systematically examined hybrid models combining CNN feature extraction with recurrent layers such as Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). This study compares CNN-MLP, CNN-RNN, CNN-LSTM, and CNN-GRU architectures on a public forest fire dataset, evaluating classification performance and computational efficiency. Results show that CNN-GRU offers the best trade-off, closely matching CNN-MLP in accuracy while requiring fewer resources. CNN-LSTM provides stable performance, whereas CNN-RNN underperforms and needs refinement. Computational analysis further indicates that CNN-MLP is the the most resource intensive models with over 1 millions parameter. These findings highlight CNN-GRU as a strong candidate for real-time forest fire detection, balancing accuracy and efficiency, and suggest future exploration of adaptive thresholds and transformer-based approaches.