In recent years, forest fires have become increasingly frequent and destructive, posing significant challenges to global ecosystems and human safety. Traditional monitoring methods, such as satellite remote sensing and ground patrols, often suffer from limitations like delayed response, high costs, and restricted coverage. To address these issues, we propose an innovative approach leveraging quadrotor unmanned aerial vehicles (UAVs) equipped with computer vision systems for real-time fire detection and precise localization. This method integrates multi-sensor data and advanced algorithms to enhance the accuracy and efficiency of forest fire management. In this paper, we detail the system architecture, mathematical models, and experimental validation of our quadrotor-based solution, emphasizing its potential to revolutionize fire monitoring strategies.
The core of our system revolves around a quadrotor UAV platform, chosen for its stability, maneuverability, and ability to operate in diverse terrains. A quadrotor’s design allows for precise control and hovering capabilities, making it ideal for capturing high-resolution imagery and sensor data in forested areas. We equipped the quadrotor with a suite of sensors, including an infrared thermal imager, a visible-light camera, and a LiDAR module, all synchronized via an onboard computer. The hardware configuration is summarized in Table 1, which outlines the key components and their specifications. This multi-sensor approach enables the quadrotor to detect both visible flames and hidden hotspots, even in low-visibility conditions such as smoke or darkness.
| Component | Specifications | Function |
|---|---|---|
| Infrared Thermal Imager | 30 Hz frame rate, high sensitivity | Detects heat signatures and hidden fires |
| Visible-Light Camera | 20 MP, 1920×1080 resolution | Captures visual imagery for fire identification |
| LiDAR Module | Range up to 1200 m | Measures distance and smoke concentration |
| Onboard Computer | NUC8i5 model | Processes data in real-time |
| GPS Module | High-precision positioning | Provides UAV location and timing data |
Forest fire identification begins with data acquisition from the quadrotor’s sensors. The infrared thermal imager captures heat distributions, where higher pixel values indicate potential fire points. We employ a histogram-based analysis to distinguish fire pixels from background elements. Let $M$ represent the total number of pixels in an image, $m_1$ the count of fire pixels (typically in the range of 233–255), and $m_2$ the count of non-fire pixels (0–35). A fire is detected if the condition $m_1 + m_2 > 0.85M$ is satisfied. The threshold $Z$ for pixel classification is determined using Otsu’s method, which maximizes inter-class variance. Fire regions are then highlighted by marking their edges, as shown in the processing pipeline.
For early smoke detection, the LiDAR module emits laser beams that interact with aerosol particles in smoke. The backscattered signal is analyzed to derive smoke properties. The depolarization ratio $\delta$ is calculated as:
$$ \delta = k \frac{P_S}{P_P} $$
where $k$ is the system gain ratio, $P_S$ is the vertical signal intensity, and $P_P$ is the horizontal signal intensity. This ratio helps distinguish spherical smoke particles (common in early fires) from non-spherical interferents like water droplets. The LiDAR data undergoes inversion processing to estimate parameters such as smoke concentration and diffusion velocity, enabling early warning capabilities. The integration of these sensors on a quadrotor ensures comprehensive coverage and reduces false alarms.
The localization of fire points utilizes a binocular vision approach, where two quadrotors capture synchronized images from different positions. This method overcomes the depth limitations of monocular systems by leveraging parallax. We define a right-handed 3D coordinate system with the origin at a reference point on the ground: the x-axis points east, the y-axis north, and the z-axis vertically. The positions of the two quadrotors, denoted as $A(X_1, Y_1, Z_1)$ and $B(X_2, Y_2, Z_2)$, are obtained via GPS. Their projected coordinates on the x-y plane are $A'(X_1, Y_1)$ and $B'(X_2, Y_2)$. The fire point $P(X_3, Y_3, Z_3)$ is located using azimuth and elevation angles from each quadrotor: $\angle D_1$ and $\angle C_1$ for quadrotor A, and $\angle D_2$ and $\angle C_2$ for quadrotor B.
The projected coordinates $X_3$ and $Y_3$ are derived from the following equations based on tangent relationships:
$$ X_3 + Y_3 \tan(\pi – \angle D_1) = Y_1 \tan(\pi – \angle D_1) + X_1 $$
$$ Y_3 \tan(\angle D_2 – \pi) – X_3 = Y_2 \tan(\angle D_2 – \pi) – X_2 $$
Solving this system yields $X_3$ and $Y_3$. The elevation $Z_3$ is computed using laser rangefinder data for distances $AP$ and $BP$:
$$ Z_3 = Z_1 – AP \cdot \tan(\angle C_1) $$
This model ensures accurate 3D localization, which is converted to geographic coordinates for practical use. The quadrotor’s agility allows for optimal positioning to minimize errors, and the use of multiple quadrotors enhances reliability through data fusion.
To validate our system, we conducted a series of experiments in a controlled forest environment, simulating fire scenarios. The quadrotor UAV used was a M300RTK model, capable of flying at speeds up to 8 m/s and altitudes up to 40 m. We performed 10 test runs, varying flight height and speed, to assess localization accuracy. The actual fire point coordinates were (40.164196°, 116.039578°, 55.200 m). Table 2 summarizes the experimental parameters and computed fire point coordinates, while Table 3 details the errors in longitude, latitude, and elevation.
| Test Run | Height (m) | Speed (m/s) | Longitude (°) | Latitude (°) | Elevation (m) |
|---|---|---|---|---|---|
| 1 | 10 | 2 | 40.164146 | 116.039498 | 55.156 |
| 2 | 17 | 2 | 40.164198 | 116.039542 | 55.211 |
| 3 | 23 | 2 | 40.164277 | 116.039677 | 55.259 |
| 4 | 27 | 2 | 40.164176 | 116.039436 | 55.251 |
| 5 | 30 | 2 | 40.164252 | 116.039451 | 55.180 |
| 6 | 35 | 2 | 40.164344 | 116.039834 | 55.120 |
| 7 | 40 | 2 | 40.164015 | 116.039245 | 55.351 |
| 8 | 30 | 4 | 40.164281 | 116.039714 | 55.275 |
| 9 | 30 | 6 | 40.164063 | 116.039322 | 55.098 |
| 10 | 30 | 8 | 40.163947 | 116.039205 | 55.314 |
| Test Run | Longitude Error (°) | Latitude Error (°) | Elevation Error (m) |
|---|---|---|---|
| 1 | 1.24 × 10-6 | 6.89 × 10-7 | 7.97 × 10-4 |
| 2 | 4.98 × 10-8 | 3.10 × 10-7 | 1.99 × 10-4 |
| 3 | 2.02 × 10-6 | 8.53 × 10-7 | 1.07 × 10-3 |
| 4 | 4.97 × 10-7 | 1.22 × 10-6 | 9.23 × 10-4 |
| 5 | 1.39 × 10-6 | 1.09 × 10-6 | 3.62 × 10-4 |
| 6 | 3.68 × 10-6 | 2.21 × 10-6 | 1.45 × 10-3 |
| 7 | 4.50 × 10-6 | 2.87 × 10-6 | 2.73 × 10-3 |
| 8 | 2.12 × 10-6 | 1.17 × 10-6 | 1.36 × 10-3 |
| 9 | 3.31 × 10-6 | 2.21 × 10-6 | 1.85 × 10-3 |
| 10 | 6.20 × 10-6 | 3.21 × 10-6 | 2.06 × 10-3 |
The results demonstrate that the quadrotor system achieves high precision, with maximum errors of $6.20 \times 10^{-6}$° in longitude, $3.21 \times 10^{-6}$° in latitude, and $2.73 \times 10^{-3}$ m in elevation. Average errors were $2.50 \times 10^{-6}$° for longitude, $1.58 \times 10^{-6}$° for latitude, and $1.28 \times 10^{-3}$ m for elevation. We observed that lower flight heights (below 30 m) and slower speeds (around 2 m/s) minimized errors, as they improve sensor stability and data quality. The reliability of the quadrotor-based localization was further confirmed by analyzing standardized residuals, which followed a normal distribution with over 95% of values within the range $[-2, 2]$, indicating robust performance.
In conclusion, our research presents a comprehensive framework for forest fire monitoring using quadrotor UAVs integrated with computer vision. The quadrotor platform offers unparalleled flexibility in accessing remote areas, while the multi-sensor fusion and binocular localization model ensure accurate fire detection and positioning. This approach addresses key limitations of existing methods, such as delayed response and limited coverage, by providing real-time, high-precision data. Future work will focus on integrating this system with fire spread prediction models to enhance proactive fire management. The quadrotor technology, combined with advanced algorithms, holds great promise for improving forest fire prevention and control strategies worldwide.
The mathematical models and experimental data underscore the effectiveness of the quadrotor UAV in environmental monitoring. By continuously refining the sensor configurations and algorithms, we aim to further reduce errors and expand applications to other domains, such as urban fire safety or agricultural monitoring. The quadrotor’s adaptability makes it a valuable tool in the era of smart forestry and disaster response.