Wildfires have become increasingly extreme and frequent, caused by climate change and human activities, and posing considerable threats to ecosystems, air quality, and public health. This survey provides a comprehensive review of state-of-the-art computer vision techniques for both wildfire detection and emissions quantification, covering 27 detection studies and 16 emissions studies published between 2021 and 2024. We explore methodologies that include convolutional neural networks (CNNs), transfer learning, You Only Look Once (YOLO), and hybrid deep learning models, leveraging diverse datasets such as satellite imagery, UAV feeds, and thermal infrared data. The surveyed detection methods achieved accuracy rates ranging from 79.4% to 99.9%, with some real-time models achieving over 15 FPS on edge devices. Emission quantification approaches, including those using burn-area maps and fire radiative power (FRP), support near real-time environmental monitoring. Importantly, the accuracy of emission estimates is tightly coupled with upstream detection and segmentation performance, with uncertainties propagating across the detection-emission pipeline. By integrating multispectral imaging and real-time data processing, these methods enhance the speed and accuracy of fire detection, which is vital for effective wildfire management. Outputs from computer vision-based wildfire detection, including burned area, fire perimeter, and fire radiative power, are key inputs for emission models. As a result, uncertainties at the detection stage can propagate into emission estimates and influence air-quality and climate assessments. However, challenges such as high computational demands, environmental variability, and data quality issues remain. Beyond technical performance, the reviewed approaches have direct implications for operational wildfire management, air quality assessment, and policy-relevant decision-making under real-world deployment constraints. The study also highlights future research directions, including optimizing models for scalability, improving sensor fusion techniques, and addressing the computational limitations of current approaches.
Emerging Trends in Wildfire Detection Through the Lens of Computer Vision and Wildfire Emission Quantification: A Comprehensive Survey
Das K.;Flesca S.;
2026-01-01
Abstract
Wildfires have become increasingly extreme and frequent, caused by climate change and human activities, and posing considerable threats to ecosystems, air quality, and public health. This survey provides a comprehensive review of state-of-the-art computer vision techniques for both wildfire detection and emissions quantification, covering 27 detection studies and 16 emissions studies published between 2021 and 2024. We explore methodologies that include convolutional neural networks (CNNs), transfer learning, You Only Look Once (YOLO), and hybrid deep learning models, leveraging diverse datasets such as satellite imagery, UAV feeds, and thermal infrared data. The surveyed detection methods achieved accuracy rates ranging from 79.4% to 99.9%, with some real-time models achieving over 15 FPS on edge devices. Emission quantification approaches, including those using burn-area maps and fire radiative power (FRP), support near real-time environmental monitoring. Importantly, the accuracy of emission estimates is tightly coupled with upstream detection and segmentation performance, with uncertainties propagating across the detection-emission pipeline. By integrating multispectral imaging and real-time data processing, these methods enhance the speed and accuracy of fire detection, which is vital for effective wildfire management. Outputs from computer vision-based wildfire detection, including burned area, fire perimeter, and fire radiative power, are key inputs for emission models. As a result, uncertainties at the detection stage can propagate into emission estimates and influence air-quality and climate assessments. However, challenges such as high computational demands, environmental variability, and data quality issues remain. Beyond technical performance, the reviewed approaches have direct implications for operational wildfire management, air quality assessment, and policy-relevant decision-making under real-world deployment constraints. The study also highlights future research directions, including optimizing models for scalability, improving sensor fusion techniques, and addressing the computational limitations of current approaches.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
