Application of ConvLSTM Neural Networks in Forest Fire Early Warning Systems for Vietnam

Abstract

Forest fires pose a significant threat to Viet Nam, particularly during the dry season. This study investigates a ConvLSTM-based deep learning architecture for early fire detection. Unlike single-frame or threshold-based methods, ConvLSTM jointly models spatial features (via convolutional layers) and temporal dependencies (via LSTM units). We utilize a publicly available dataset of 999 fire and non-fire images. To apply ConvLSTM to static images, we construct temporal sequences using a sliding window over augmented variants. The proposed model achieves 98.3% accuracy, 98.1% precision, 96.8% recall, and a 98.1% F1-score on the test set. These results are compared with standalone CNN and LSTM models. Limitations include the limited dataset size, class imbalance (75% fire images), and the lack of Viet Nam-specific data. Future work should focus on larger, region-specific datasets and real-time deployment on edge devices.