PyroFocus: A Deep Learning Approach to Real-Time Wildfire Detection in Multispectral Remote Sensing Imagery

Mark Moussa, Andre Williams, Seth Roffe, Douglas C Morton; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 2855-2864

Rapid and accurate wildfire detection and quantification are crucial for emergency response and environmental management. In airborne and spaceborne missions, it is useful for real-time detection algorithms to not only classify whether the observed area consists of no fire, active fire, or post-fire, but also to discriminate among varying fire intensities. Hyperspectral and multispectral thermal imagers provide sufficient spectral information to address these challenges; however, high data dimensionality and limited onboard resources complicate real-time processing. With wildfires becoming more common, the need for real-time detection is crucial for successful wildfire management and damage mitigation. It is therefore imperative for onboard solutions to achieve both high accuracy and low inference latency at minimal computational cost. In this work, we present several contributions divided into two main parts. First, we systematically evaluate multiple cutting-edge deep learning architectures, such as custom Convolutional Neural Networks (CNN) and Transformer-based models for multi-class fire classification. Second, we propose a novel adaptation of a two-stage pipeline, which we call PyroFocus, consisting of classification of fires and fire radiative power (FRP) regression or image segmentation, designed to significantly reduce inference times and computational requirements for onboard wildfire detection in airborne and spaceborne applications. We evaluate model accuracy, inference latency, and computational cost to identify the most effective solution for resource-constrained environments. Our study leverages a dataset from NASA's MODIS/ASTER Airborne Simulator (MASTER), which provides data similar to that of an in-house next-generation fire detection sensor, NASA's Compact Fire Imager (CFI). Experimental results indicate that the proposed two-stage pipeline provides favorable trade-offs among inference speed and accuracy, demonstrating strong potential for edge deployment in future wildfire monitoring missions.