A forest fire identification and prediction method based on a bidirectional coupling mechanism

Abstract

The application discloses a forest fire identification and prediction method based on a bidirectional coupling mechanism and belongs to the technical field of disaster prediction, and comprises the following steps: data preprocessing, fire source identification, fire intensity prediction, coupling optimization and phased application. The system innovatively constructs a bidirectional coupling mechanism of a full convolution network FCN and a Rothermel model, and realizes the collaborative optimization of fire identification and spread prediction. The FCN takes ResNet50 as a backbone network, dynamically adjusts a skip connection and an adaptive up-sampling coefficient, combines a cross-entropy and a Dice combined loss function, accurately outputs a fire area discretization boundary, and provides a reliable initial boundary for prediction. Meanwhile, the Rothermel physical rule is used to constrain the FCN to reduce the identification error, and a differentiated collaborative strategy is implemented for four stages (T1-T4) of the fire, so that the accuracy and stability of the fire identification and prediction in a complex scene are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of disaster prediction technology, specifically a forest fire identification and prediction method based on a two-way coupling mechanism. Background Technology

[0002] Forest fires pose a serious threat to the environment, causing not only direct economic losses but also profound impacts on regional ecological restoration and the global carbon cycle. Therefore, how to reduce disaster losses through efficient and accurate early fire detection technologies, coupled with scientific spread prediction modeling to support emergency decision-making, has become a critical issue that urgently needs to be addressed in the field of disaster prevention and mitigation.

[0003] Currently, many scholars both domestically and internationally have contributed to improving the efficiency and accuracy of forest fire identification and fire spread prediction. In the field of forest fire identification, Li Yi proposed a fast and efficient fire detection model based on MobileNetV3 and an anchorless structure. This model exhibits better performance and speed on public fire datasets and is suitable for real-time fire detection and embedded device applications. WuS proposed a multi-scale fire image detection method combining convolutional neural networks (CNN) and Transformers. In the field of fire spread prediction, M. Byari proposed a suitable multi-scale CA modeling method based on 3D geometric units, which is applied as part of the overall forecasting system to predict wildfire spread. Edigley et al. used a two-stage prediction scheme to adjust unknown parameters in the input data describing fire scenarios in forest fire simulation models to reduce the high uncertainty in the input data of fire scenarios.

[0004] Based on the above research, deep learning models have been widely used in the field of forest fire identification. However, due to limited training data, insufficient model generalization ability, and lack of interpretability (fires are often accompanied by complex phenomena such as smoke and changes in light), single deep learning models often perform poorly in practical applications. In the field of fire spread prediction, the Rothermel model, as the most commonly used and widely accepted physical model, can return the fire spread rate under specific conditions. However, this model requires high accuracy of input parameters, and its uncertainty can lead to the accumulation of prediction errors. Summary of the Invention

[0005] To overcome the problems mentioned in the background technology, this invention proposes a forest fire identification and prediction method based on a bidirectional coupling mechanism. This method deeply integrates a deep learning model (FCN) and a Rothermel physical model, allowing both to complement each other's advantages, thus achieving a comprehensive advantage of high fire identification accuracy and small fire prediction error.

[0006] This invention adopts the following technical solution: a forest fire identification and prediction method based on a two-way coupling mechanism, comprising the following steps:

[0007] A forest fire identification and prediction method based on a two-way coupling mechanism, characterized in that the method includes the following steps:

[0008] Fire source identification: Fire scene segmentation is achieved by using an improved fully convolutional network FCN. The FCN uses ResNet50 as the backbone network and integrates ConvBlock and IdentityBlock. It optimizes feature fusion by dynamically adjusting the number of skip connections and adaptive upsampling coefficients, and outputs the discretized boundary of the fire area.

[0009] Fire prediction: Based on NASA meteorological data and fire source identification results, the Rothermel model is used to predict the spread of fires;

[0010] Coupling optimization: Edge breakage repair is performed on the FCN output boundary, a collaborative mechanism between FCN and Rothermel model is established, the initial boundary of Rothermel model is dynamically corrected through FCN recognition results, and the FCN recognition error is constrained by Rothermel physical rules;

[0011] Phased application: Fire spread is divided into four phases: early ignition (T1), accelerated spread (T2), stable combustion (T3), and decay (T4). In the T1 phase, noise interference is reduced by FCN. In the T2-T3 phases, FCN texture features are combined with Rothermel dynamic parameter adjustment. In the T4 phase, FCN is used to identify the burn-out area to correct Rothermel prediction bias.

[0012] Furthermore, in the data preprocessing, a dedicated dataset containing aerial fire images and laboratory combustion data is constructed; the aerial images cover forest scenes, and the laboratory data includes 10 sets of outdoor steel wool burning experiments. The annotation is generated as a PASCALVOC format XML file using the LabelImg tool, and the dataset is divided into training and test sets in a ratio of 0.85:0.15.

[0013] Furthermore, the FCN includes a head module that reduces the feature map channels to the number of classes through a 1×1 convolutional layer, restores the original size through bilinear interpolation, and sets an auxiliary classifier to extract features from ResNet's Layer 3 to accelerate training.

[0014] Furthermore, the Rothermel model takes combustible density, wind speed, and slope parameters as input, and outputs the spread rate and burned area. The model can be simplified as follows: R represents the rate of fire spread; I R ζ represents the reaction intensity of the flame zone; ζ represents the dimensionless fire spread rate; Pb ε is the density of the combustible bed; ε is the effective thermal coefficient; Q ig It is pre-combustion heat, where the wind speed correction coefficient Φ W and slope correction factor Φ S Calculate using the following formula: , ,

[0015] Among them, U For wind speed, For fuel bed packing ratio, For the optimal packing ratio, For fuel bed density, C, B, E It is a constant. This is the slope correction factor. The slope angle, K 1 These are the coefficients for the forest fire model. propagation rate at different directional angles θ .

[0016] Furthermore, the morphological processing for edge fracture repair in the coupling optimization includes closing operations to repair small holes and expansion operations to fill edge fractures.

[0017] Furthermore, in the phased application module: in the early ignition stage T1, FCN suppresses star-shaped fire point identification noise through depth features, and the Rothermel model predicts the area of ​​the next stage based on circular diffusion features; from the diffusion acceleration stage T2 to the stable combustion stage T3, the flame area forms a continuous thermal radiation zone, FCN captures the high-frequency texture features of the flame, and the Rothermel model adjusts the predicted trajectory by combining real-time wind speed and terrain parameters; in the decay stage T4, FCN identifies the combustible material burnout area and corrects the prediction bias of the Rothermel model for the split fire zone.

[0018] Furthermore, the Rothermel model includes a parameter optimization unit that dynamically adjusts the wind speed correction coefficient using UKF. and slope correction factor .

[0019] Furthermore, the trajectories predicted by Rothermel for the four stages of fire spread are not only used as the final prediction results, but are also fed back into the FCN as spatial priors to guide the segmentation process of the image at the next moment.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] Based on traditional fire identification technology, this invention replaces the backbone network from VGG / GoogLeNet with ResNet, and constructs an FCN model that can adapt to various flame colors through dynamic feature fusion and auxiliary classifiers, effectively improving the robustness of the fire identification module.

[0022] This invention explores for the first time a two-way coupling mechanism between forest fire identification and fire spread prediction through the bidirectional collaboration of the Fire Control Network (FCN) and the Rothermel model: on the one hand, the initial boundary of the Rothermel model is dynamically corrected using the identification results of the FCN; on the other hand, the physical rules of the Rothermel model are used to inversely constrain the misidentified areas of the FCN. The two mechanisms complement each other, achieving linkage between identification and prediction, thereby significantly improving the accuracy of both fire identification and fire spread prediction.

[0023] This invention integrates fire identification with fire spread prediction and divides fire spread into four stages based on the characteristic differences of fire spread. It also develops adaptive processing methods for each stage. Furthermore, the trajectory predicted by Rothermel can not only be used as the prediction result but also be used as a spatial prior by being input back into FCN to guide the segmentation process of the image at the next moment or stage. This results in the construction of an identification and prediction scheme applicable to the entire life cycle of forest fires, which helps to support emergency decision-making and reduce disaster losses. Attached Figure Description

[0024] Figure 1 Comparison of FCN errors before and after correction using the Rothermel model. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the forest fire identification and prediction method of this invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0026] A method for identifying and predicting forest fires based on a two-way coupling mechanism, the method comprising the following steps: data preprocessing, fire source identification, fire intensity prediction, coupling optimization, and phased application.

[0027] Data preprocessing was performed to construct a dedicated dataset containing aerial fire images and laboratory combustion data. First, high-resolution aerial fire images were acquired to record the dynamic characteristics of flame propagation under different wind conditions. These images covered various scenes (62 images of coniferous forests, 48 ​​images of broadleaf forests, and 34 images of mixed mountain forests) and meteorological conditions (32 images of cloudy days, 86 images of sunny days, wind speeds 0-8 m / s), and included different flame colors (58 images of yellow and 42 images of red) and smoke interference samples (26 images). Second, an outdoor steel wool combustion experiment was conducted to collect combustion data, providing a foundation for multi-feature analysis of flames. Finally, a hybrid intelligent strategy was used to annotate the collected data (including aerial fire images and laboratory combustion data) to construct the dedicated dataset. Specific annotation steps included: initial target labeling using LabelImg (labeled "fire"), improving the U-Net network to achieve semi-automatic contour extraction, manual verification to ensure boundary accuracy, and generating XML format annotation files.

[0028] Fire source identification employs an improved fully convolutional network (FCN) to segment fire scenes and output discretized boundaries of the fire area. The improved FCN architecture innovatively uses ResNet50 as the backbone and also incorporates improvements to the FCN head module. The ResNet50 module, serving as the backbone of the fully convolutional network, is divided into five main stages, each containing a different number of residual blocks. These residual blocks are of two types: convolutional residual blocks (ConvBlocks) and identity residual blocks (IdentityBlocks). Convolutional residual blocks increase network depth, while identity residual blocks maintain dimensionality consistency of the feature maps. A batch normalization (BN) layer precedes each residual block to normalize the input data. In the five stages of ResNet50, Stage 1 contains a 7x7 convolutional layer and a max-pooling layer for initial feature extraction. Stages 2-5 contain 3, 4, 6, and 3 residual blocks respectively. The first residual block in Stage 2 contains a convolutional layer with a stride of 2 (for downsampling), while the first residual blocks in Stages 3-5 have a stride of 1. This residual structure effectively avoids the gradient vanishing problem. The FCN head module is a key component in the fully convolutional network for achieving pixel-level semantic segmentation. The main functions of the FCN head module are: reducing the dimension of the feature map channel number through a 1×1 convolutional layer and restoring the size of the feature map using bilinear interpolation. In this invention, an auxiliary classifier is introduced into the FCN head module to extract intermediate features from ResNet's Layer 3, accelerating model convergence through a multi-task learning mechanism and enhancing the gradient backpropagation effect. Meanwhile, a pre-training strategy of "ImageNet pre-training + COCO dataset fine-tuning" is adopted, combined with a loss function of cross-entropy loss and Dice loss, as well as data augmentation methods such as random rotation and scaling, to solve the problem of imbalance between flame and background pixels and improve robustness to multi-colored flames and complex environments.

[0029] Fire prediction, taking a wildfire in 2025 as an example, can be simplified as follows: R represents the rate of fire spread; I R ζ represents the reaction intensity of the flame zone; ζ represents the dimensionless fire spread rate; P b ε is the density of the combustible bed; ε is the effective thermal coefficient; Q ig It is pre-combustion heat, where the wind speed correction coefficient Φ W and slope correction factor Φ S Calculate using the following formula: , ,

[0030] Among them, U For wind speed, For fuel bed packing ratio, For the optimal packing ratio, For fuel bed density, C, B, E It is a constant. This is the slope correction factor. The slope angle, K 1 These are the coefficients for the forest fire model. propagation rate at different directional angles θ To improve the model's adaptability to complex terrain and meteorological conditions, a UKF-based parameter optimization unit was introduced into the fire prediction module. The system converts observed changes in the fire front position into spread rates and compares these with model predictions to construct an observation model. UKF dynamically updates wind speed and slope correction coefficients based on the residuals of each cycle, achieving adaptive matching between the model and the actual environment, thereby improving overall prediction accuracy and stability. Selecting the longitude and latitude of a fire, the Rothermel model predicts fire spread rates in different directions. Based on the directional spread characteristics and different stages of the fire, the fire spread area can be approximated as circular or elliptical. For calculating the predicted area area, to reduce systematic errors caused by insufficient area identification, a closed curve integral method is used to calculate the fire area, through symmetric difference... : The error between the predicted fire area and the actual fire area is obtained; to avoid the loss of calculation accuracy caused by non-smooth curves, Green's formula is used. Based on the shoelace formula This method is used to solve for the predicted fire area A and the actual fire area B, and to solve for the fitness function of the region area obtained by the two integration methods to evaluate the model stability. , Table 1 shows the errors of the Rothermel prediction model under the two approximation methods.

[0031] Table 1. Errors of the Rothermel prediction model under two approximation methods.

[0032]

[0033] The coupling optimization mechanism between the FCN and Rothermel models is the core of achieving complementary errors in the joint model. Its essence lies in the synergistic optimization of identification and prediction through optimization of the data interface layer and bidirectional feedback correction. Morphological post-processing is introduced at the data interface layer to expand and smooth the discretized boundaries of the FCN output, repairing pinholes and filling broken edges, thereby improving the continuity of the initial fire line and reducing the sensitivity of the Rothermel model to segmentation noise. Considering that flame edges often exhibit discontinuities due to smoke, light, and obstruction, the physical constraints of the Rothermel model (wind speed and slope limitations on the spread direction) are used to correct FCN misidentification areas. Furthermore, Rothermel estimates combustion intensity using fuel sample data; when the FCN identification area does not match the fuel distribution, the fire line position is recalculated through weighted averaging. This prevents the amplification of step-by-step errors while maintaining consistency with physical conditions. When the FCN mistakenly identifies a portion of the combustion zone as background due to smoke obscuring the view, the physical constraints provided by Rothermel compensate for this area through a boundary correction mechanism. Conversely, when the FCN captures a small fire point that the physical model fails to predict in time, the segmentation result can be fed back into Rothermel to drive the boundary update of the propagation equation. This forms a closed-loop, two-way coupled system that leverages both the fine-grained spatial recognition advantage of FCN and the physical reliability of Rothermel in temporal evolution.

[0034] The application is phased, and four typical time points are selected based on the flame area growth rate and motion characteristics: T1 (early ignition stage, flame area ratio <5%), T2 (diffusion acceleration stage, area ratio 15%-20%), T3 (stable combustion stage, area ratio 30%-35%), and T4 (decay stage, area ratio 25%-30%).

[0035] The T1 flame is characterized by scattered, star-shaped fire points with blurred edges, significantly affected by environmental noise and dynamic flickering. Traditional edge detection methods (such as the Canny operator) often suffer from broken flame contours due to noise, while FCN, relying on deep feature learning, reduces noise sensitivity to some extent. However, due to the small size of the fire points, the localization accuracy remains limited. In this stage, the flame region is concentrated and approximately circular. Based on the Rothermel model combined with experimental parameters and fire contours for physical constraints, the system effectively improves the recall rate for small target fire points. The predicted region output by Rothermel is transformed into a spatial prior mask, which is fed into the FCN along with the original image as an additional input channel in the next time step. This guides the FCN to pay more attention to small fire points within the predicted range in the next time step, reducing false detections in the background region and improving the recall rate and boundary coherence of small target fire points. In the T2 stage, the flame gradually forms a continuous thermal radiation band with obvious turbulent texture features and exhibits an accelerated diffusion trend driven by wind direction. At this stage, the feature extraction advantages of FCN are fully demonstrated. The pre-trained ResNet-50 model can effectively capture the high-frequency texture features of the flames, but factors such as occlusion and dense smoke may still cause edge detection errors or misjudgments. The shape of the flame region tends to be elliptical at this stage. Rothermel combines real-time updated fuel load, wind speed, and slope parameters to continuously correct the errors caused by image noise or occlusion in FCN. The prediction results are also converted into a probability prior map, in which regions closer to the center of the ellipse are given higher prior probabilities. This prior map is multiplied element-wise with the deep feature map of FCN, which is equivalent to embedding physical prediction information into the spatial attention mechanism, guiding FCN to pay more attention to the effective area within the prediction range during the next step of segmentation, avoiding missed detections and over-detections caused by noise or occlusion. The T3 fire enters a stable spread state, with clear flame region boundaries, complete thermal radiation zones, and the influence of terrain slope on fire propagation gradually becomes significant. FCN primarily utilizes high-frequency texture features (such as turbulent wave patterns) extracted by ResNet-50 to capture flame dynamics; while the Rothermel model observes that the curvature change coefficient of the flame spread region begins to decrease, and the fire gradually transitions from an irregular shape to an approximately circular shape, spreading in multiple directions. By feeding back fuel and environmental parameters in real time, Rothermel effectively suppresses error accumulation caused by fluctuations in FCN recognition accuracy, achieving high-precision prediction of the stable spread stage. The prediction results are simultaneously fed back to FCN as a spatial prior, making the segmentation process in the next moment more stable under physical constraints, effectively avoiding error accumulation. In T4, after the combustibles in the central area of ​​the fire gradually burned out, the flame area began to split into multiple irregular shapes, and the fire spread rate slowed down significantly. At this time, FCN identifies the burned-out area and marks it as a "non-combustion zone" by using a grayscale threshold (<50), and uses a 128×128 multi-scale sliding window scan to accurately locate scattered fire points.Because the Rothermel model struggles to identify burn-out regions at the microscopic scale, its prediction error is relatively large. Therefore, through the synergistic compensation of FCN and Rothermel, the model maintains high overall accuracy. Simultaneously, the results feedback serves as a local prior, weighting and amplifying these local regions in the input of the next time step, thereby compensating for Rothermel's shortcomings at the microscopic level.

[0036] The initial recognition errors of the FCN neural network were 60%, 36%, 29%, and 31%, respectively, while the prediction errors corrected by the Rothermel model were significantly reduced to 30%, 7.2%, 4.95%, and 7.05%. Figure 1 As shown in Table 2, adding the FCN fire identification network to the Rothermel model to correct for the fire area improves its prediction accuracy. Under the same conditions, the prediction accuracy using FCN is 5% to 20% higher than that without FCN, as shown in Table 2.

[0037] Table 2: Comparison of the accuracy of FCN-Rothermel predictions and traditional Rothermel model predictions

[0038]

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A forest fire identification and prediction method based on a two-way coupling mechanism, characterized in that, The method includes the following steps: Data preprocessing, and dataset construction; Fire source identification: Fire scene segmentation is achieved by using an improved fully convolutional network FCN. The FCN uses ResNet50 as the backbone network and integrates ConvBlock and IdentityBlock. It optimizes feature fusion by dynamically adjusting the number of skip connections and adaptive upsampling coefficients, and outputs the discretized boundary of the fire area. Fire prediction: Based on NASA meteorological data and fire source identification results, the Rothermel model is used to predict the spread of fires; Coupling optimization: Edge breakage repair is performed on the FCN output boundary, a collaborative mechanism between FCN and Rothermel model is established, the initial boundary of Rothermel model is dynamically corrected through FCN recognition results, and the FCN recognition error is constrained by Rothermel physical rules; Phased application: Fire spread is divided into four phases: early ignition (T1), accelerated spread (T2), stable combustion (T3), and decay (T4). In the T1 phase, noise interference is reduced by FCN. In the T2-T3 phases, FCN texture features are combined with Rothermel dynamic parameter adjustment. In the T4 phase, FCN is used to identify the burn-out area to correct Rothermel prediction bias.

2. The forest fire identification and prediction method based on a bidirectional coupling mechanism according to claim 1, characterized in that, In the data preprocessing, a dedicated dataset containing aerial fire images and laboratory combustion data was constructed. The aerial images cover forest scenes, and the laboratory data includes 10 sets of outdoor steel wool burning experiments. The annotation was generated as a PASCALVOC format XML file using the LabelImg tool. The dataset was divided into training and test sets in a ratio of 0.85:0.

15.

3. The forest fire identification and prediction method based on a bidirectional coupling mechanism according to claim 1, characterized in that, The FCN includes a head module that reduces the feature map channels to the number of classes through a 1×1 convolutional layer, restores the original size through bilinear interpolation, and sets up an auxiliary classifier to extract features from ResNet's Layer 3 to accelerate training.

4. The forest fire identification and prediction method based on a bidirectional coupling mechanism according to claim 1, characterized in that, The Rothermel model takes combustible density, wind speed, and slope parameters as inputs and outputs the spread rate and burned area. The model is simplified as follows: R represents the rate of fire spread; I R ζ represents the reaction intensity of the flame zone; ζ represents the dimensionless fire spread rate; P b It is the density of the combustible material bed; ε is the effective heat coefficient; Q ig It is pre-combustion heat, where the wind speed correction coefficient Φ W and slope correction factor Φ S Calculate using the following formula: , , Among them, U For wind speed, For fuel bed packing ratio, For the optimal packing ratio, For fuel bed density, C, B, E It is a constant. This is the slope correction factor. The slope angle, K 1 These are the coefficients for the forest fire model. propagation rate at different directional angles θ Rothermel predicts the fire spread rate in each direction. Based on the spread characteristics, the fire area is approximately elliptical or circular. The fire area is calculated using closed curve integrals and the shoelace formula.

5. The forest fire identification and prediction method based on a bidirectional coupling mechanism according to claim 1, characterized in that, The morphological processing for edge fracture repair in the coupling optimization includes closing operations to repair small holes and expansion operations to fill edge fractures.

6. The forest fire identification and prediction method based on a bidirectional coupling mechanism according to claim 1, characterized in that, The phased application steps In the early ignition stage T1, FCN suppresses noise in star-shaped fire point identification through depth features, and the Rothermel model predicts the area of ​​the next stage based on circular diffusion features. During the diffusion acceleration stage (T2) to the stable combustion stage (T3), a continuous thermal radiation zone is formed in the flame area. FCN captures the high-frequency texture features of the flame, and the Rothermel model combines real-time wind speed and terrain parameters to adjust the predicted trajectory. During the decay phase T4, FCN identifies the burnt-out area of ​​combustibles and corrects the prediction bias of the Rothermel model for the split fire zone.

7. A forest fire identification and prediction method based on a two-way coupling mechanism according to claim 1 or 4, characterized in that, The Rothermel model includes a parameter optimization unit that dynamically adjusts the wind speed correction coefficient using UKF. and slope correction factor .

8. The forest fire identification and prediction method based on a bidirectional coupling mechanism according to claim 6, characterized in that, The trajectories predicted by Rothermel for the four stages of fire spread are not only used as the final prediction results, but also used as spatial priors by being input back into the FCN to guide the segmentation process of the image at the next moment.

Images