Reduced Complexity Classifier Architecture for Edge Devices in Real-Time Wildfire Detection

This article presents a comprehensive approach to forest fire detection, starting with the construction of a deep learning model based on the MobileNet architecture and transfer learning. Post-training quantization is applied to make the model suitable for edge device implementation. It provides computational efficiency, reducing storage size by 64.75%, RAM usage by 58.38%, and inference time by 86.21%. Practical implementation is carried out, integrating the model into a low-cost Arduino Nano device in real time. The performance evaluation has included detailed metrics such as accuracy, recall, and F1-score, as well as tests with fire simulation videos. Despite limitations such as camera angle, the quantized model has demonstrated robust results, slightly outperforming the floating-point model. The proposal is relevant for forest fire detection and other areas, such as industries.