A novel depth-homography model for Infrared (IR) and Visible (RGB) images registration and YOLOv8n-CoordAttn detection model for wildfire detection are presented. In low-light and smoke-occluded conditions, fire detection using IR images performs better than RGB images, while RGB images may still complement IR information as heat radiation around the fire makes the fire boundary blurry in the thermal imagery. Hence, fire detection based on image fusion between IR and RGB images is a more reliable approach. For image alignment between these two modalities, camera calibration is widely used, while in this work, an innovative depth-homography model as a simpler and yet precise alternative is presented, which estimates the homography matrix for an arbitrary depth with which the image alignment is conducted. Moreover, YOLOv8n-CoordAttn is presented, where YOLOv8n is augmented with Coordinate Attention modules. This detection model predicts bounding boxes of fire spots based on multispectral IR-RGB images, aiming to improve accuracy while still conducting inference in real-time. Also, outdoor flight tests using a DJI M300 UAV equipped with an H20T camera system in daytime and nighttime are carried out to gather IR-RGB datasets for training and evaluating the depth-homography and YOLOv8n-CoordAttn detection models, whose video demonstration is available at https://www.youtube.com/watch?v=Hq6X-FcUVss.

Wildfires pose escalating threats to ecosystems, infrastructure, and human safety. This paper presents an integrated autonomous UAV-based wildfire suppression system designed to execute multiple fire-spot extinguishing missions efficiently. The proposed framework assumes prior detection of wildfire locations and comprises two main modules: a trajectory planning algorithm using dynamic programming and a Linear Quadratic Regulator (LQR) controller for optimal trajectory tracking. The dynamic programming algorithm minimizes the flight distance of the multiple fire-spots firefighting trajectory. The LQR controller is developed based on a linearized model of the quadrotor UAV and ensures accurate trajectory tracking. The effectiveness of the proposed system is demonstrated through MATLAB simulations and validated via outdoor experiments using the DJI M300 UAV platform equipped with a custom-designed water-dropping mechanism. Results confirm the feasibility of the system in suppressing multiple wildfire spots autonomously.

Wildfires threaten ecosystems and communities and require rapid detection and effective suppression. This paper presents an autonomous leader-follower UAV system for real-time wildfire detection and suppression. A leader UAV employs deep learning and thermal imaging to detect fires accurately, even under low-visibility conditions, achieving 95.2% detection accuracy. It integrates visual and thermal data to refine fire coordinates, which are relayed to a ground station. The ground station directs a follower UAV, equipped with a dual-tank water payload, to execute targeted water drops with 0.2 m precision along optimized suppression paths. By separating detection and suppression roles, the system enhances mission endurance and payload capacity. This coordinated approach enables rapid and accurate wildfire containment, offering a scalable solution for early-stage fire management.

This paper presents a novel, precise, and fast framework for autonomous aerial forest fire fighting using unmanned aerial vehicles (UAVs) to extinguish a line of fires by most efficiently utilizing the on-board camera. Autonomous aerial firefighting algorithms using UAVs have been proven to be promising in early wildfire suppression. However, UAVs/drones have limited payloads, which do not allow them to carry as much retardant as fixed-wing aircraft do. Hence, firefighting drones should release the retardant in a way that accurately extinguishes wildfire and efficiently suppresses the line of fires. In this work, a DJI M300 RTK drone mounted on which a RGB camera and a 3D-printed water tanker are mounted and utilized to extinguish a line of fires in an outdoor experimental environment. A line of fires has been set up using firepits, whose GPS locations are known. The optimal path along which a drone should approach and release the retardant to extinguish the fire line is calculated using the RANSAC algorithm, as well as the in-motion dropping mission starting point. Nonetheless, due to errors in the drone's onboard GPS sensor, the drone does not exactly position itself at the starting point. In fact, in the proposed method, the drone uses on-board camera images to adjust itself and get aligned with the retardant-releasing line as it is supposed based on prediction, followed by approaching the fire line and releasing the retardant. The testing results show efficient and accurate suppression of fire spots whose video verification is provided at https://www.youtube.com/watch?v=ZP2KoxtwAsg.

In this paper, rethinking the application of deep neural network models in aerial image-based wildfire perception, a gradient-weighted class activation mapping (Grad-CAM) scheme is studied and tested on the model of you only look once (YOLO) for the drone-image-based wildfire detection. Grad-CAM helps the model to be more explainable and understandable. It also helps engineers and fire-fighters to view the production of target layer(s) of the wildfire detection model. Such a benefit is similar as higher-weighted segmentation models. Therefore, Grad-CAM could be a tool to help designing less-weighted wildfire detection models. It could be also used to support fire-fighters labeling the suspect wildfire area and make decision-making process more efficiently.