Edge-AI Framework for Fire Detection in Wildland-Urban Interface using TinyML

Wildfires in the Wildland-Urban Interface (WUI) pose a significant threat to public safety and property, indicating the need for advanced early detection systems that are not only cost-effective and low-maintenance but also capable of operating in remote locations. This study addresses this need by developing an embedded wildfire detection system optimized for the WUI, using Convolutional Neural Network (CNN) models deployed on resource-constrained hardware. The approach involves building a balanced image dataset encompassing diverse wildfire scenarios, training and evaluating multiple CNN models on this dataset, in order to select the most effective model that adheres to the memory constraints of an Arduino Nano 33 BLE Sense platform. The models were tested and compared using metrics such as global accuracy, true positive accuracy, the area under the ROC curve (AUC-ROC), and inference time. The MobileNetV2 model was selected, achieving a global accuracy of 91.4% and an AUC-ROC of 0.93 while maintaining an inference time of 992 ms. Additionally, a custom enclosure was designed to protect the hardware from environmental factors, thereby ensuring its functionality and durability in real-world deployment scenarios.