A scalable and cost-effective forest fire detection approach using deep transfer learning on a Raspberry Pi cluster

Due to the increasing frequency of forest fires and their rapid spread, early detection is critical for effective containment and mitigation. This paper proposes a cost-effective edge-based forest fire detection system that receives images from multiple sources (terrestrial cameras and UAVs) to predict and alert authorities of potential forest fires. The model of the system is built using transfer learning with MobileNetv2 on a realistic and diverse dataset, resulting in a lightweight CNN model that is further optimized by using quantization to reduce its size and improve the inference speed. The proposed model is deployed on an 8-node Raspberry Pi cluster, using Slurm and MPI to manage cluster task scheduling and parallel processing. The proposed system achieves 99.21% accuracy, precision, recall, and F1 score on a realistic test dataset containing vague real-world scenarios such as fog and sunset conditions, with an inference speed of 69 frames per second. These results, along with the system's autonomous and offline operation, cost effectiveness, power efficiency, and scalability, make it ideal for real-time forest fire monitoring at the edge, even in off-grid and remote areas.