Application of Convolutional Neural Networks For Wild Fire Detection

Early detection of wildfires is a critical modern-day challenge for first responders and emergency services. The existing systems for wildfire detection include video surveillance with high-end cameras using pattern recognition and satellite-based monitoring. These systems are very expensive which makes them beyond the reach of medium to small scale users. In a prior work, we proposed a low-power and low-cost wildfire monitoring system based on distributed sensor networks. The system combined a camera with humidity and smoke sensors. A major challenge in the implementation of that system was to classify the difference between camera images with and without the fire/smoke. The latest developments in deep learning offer a viable solution to process the camera images and detect the presence of wildfire/smoke using convolutional neural networks (CNNs). We use a very large wildfire image database to train and validate the CNN classifiers and analyze their classification accuracy. The results indicate that CNNs can detect presence of fire/smoke in wildfire images with very high accuracy. We plan to continue this work and integrate the CNN based classification method into our proposed wildfire detection system.