Early Wildfire Detection and Identification in High Depth-of-Field Scenarios Using LWIR Thermal Imaging From Ground-Based Systems

Abstract

This article presents a novel hybrid framework that integrates spatial and temporal processing techniques for detecting and identifying incipient fires. By combining thermal imaging with a core detector, based on an object detection model, and a secondary detector, leveraging temporal features, the framework significantly enhances the detection of thermal anomalies and the identification of fires using raw long-wave infrared (LWIR) thermal imaging. The framework was tested with different core detectors, trained using the thermal anomaly (TA) dataset, on the fire's latent activity monitoring and evaluation through thermography (FLAME-T) dataset, achieving improvements in mean average precision (mAP) and F1 scores of up to 35.9% and 20.9%, respectively, with the addition of the secondary detector. Although these improvements introduced higher processing times, the framework demonstrated its capability to maintain high detection accuracy even on a resource-constrained platform like the Raspberry Pi 5. The proposed novel identification algorithm achieves high classification accuracy for early fires at a significant depth of field, with accuracies of up to 0.913 and identification times of approximately 2 ms, making it suitable for edge applications. The code and data needed to replicate this work are available at: https://github.com/AntonioIDeTIC/IFTH.