A real-time forest fire and smoke detection system using deep learning

Large parts of the world's forests are threatened by fires. These fires happen continuously every month around the globe. They are very costly to society and cause serious damage to the ecosystem. This raises the necessity to build a detection system to intervene early and take action. Fire and smoke have various colours, textures, and shapes, which are challenging to detect. In the modern world, neural networks are used extensively in most fields of human activities. For the detection of fire and smoke, we suggest a deep learning technology using transfer learning to extract features of forest fire and smoke. We used a pre-trained Inception-ResNet-v2 network on the ImageNet dataset to be trained on our dataset which consists of 1,102 images for each fire and smoke class. The classification accuracy, precision, recall, F1-Score, and specificity were 99.09%, 100%, 98.08%, 99.09%, and 98.30%, respectively. This model has been deployed on a Raspberry Pi device with a camera. For real-time detection, we used the Open CV library to read the camera stream frame by frame and predict the probability of fire or smoke.