An intelligent fire point identification method fusing a physical optimization model and deep learning

Abstract

The application discloses an intelligent fire point recognition method fusing a physical optimization model and deep learning, and comprises the following steps: step 1, threshold dynamic processing is performed according to a physical mechanism model to obtain an optimized adaptive threshold physical mechanism model; step 2, a U-Net network model is constructed; step 3, a YOLOv5 model is introduced to detect highlighted targets; step 4, the optimized physical mechanism model and the U-Net network model are used to make a decision to recognize fire points, and remote sensing fire point preliminary screening results are obtained; and step 5, the highlighted targets recognized by the YOLOv5 model are used as error elimination on the remote sensing fire point preliminary screening results, so that the overall precision of the remote sensing fire point preliminary screening result recognition is improved; the problems of fire point false detection and missed detection caused by a fixed threshold are solved, various heat source errors in urban areas that seriously interfere with fire point detection for a long time are eliminated, and the overall precision is further improved; and the problems of poor universality caused by inherent system errors of traditional single thought algorithms are overcome.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire detection, and more particularly to an intelligent fire detection method that integrates a physical optimization model and deep learning. Background Technology

[0002] Satellite remote sensing is a crucial technology for monitoring real-time fires. Fire point identification algorithms based on remote sensing data mainly fall into two categories: radiative transfer physical models and deep learning models. While physical mechanism-based global fire point index methods and local context algorithms are relatively mature, the fixed thresholds for key indicators in traditional algorithms can lead to false positives and false negatives, especially in extreme conditions such as deserts, mountain shadows, and cloud shadows. These fixed thresholds exhibit significant drawbacks, generating numerous false positives and interfering with disaster emergency response. Therefore, adaptively and dynamically adjusting the thresholds in the physical model can effectively reduce various systematic errors and fundamentally improve the accuracy of fire point identification.

[0003] In comparison, physical models, based on radiative transfer mechanisms, offer stable detection results, but the radiative energy limits the effectiveness of fire detection. Deep learning, starting from data mining, achieves excellent fire detection, but lacks mechanistic constraints and often results in false positives on bright surfaces. Both existing approaches can achieve high accuracy, but single methods are no longer sufficient for innovative progress. Integrating their complementary advantages and fully utilizing the local, global, and abstract high-level attributes of images is a bottleneck that urgently needs to be overcome.

[0004] Furthermore, results from various remote sensing data sources and algorithms indicate that artificial heat sources on urban surfaces (such as solar reflectors, industrial mining areas, and self-built roofs made of special materials) can interfere with fire spot identification to some extent, leading to false alarms. While time-series data can be used to eliminate such errors, this approach is not suitable for single-scene imagery. Currently, there are no relevant publications or experimental reports on introducing deep learning object detection models to identify and eliminate such errors, but the applicant has verified the feasibility of this approach through significant preliminary experimental results. In the future, the close integration of deep learning and real-time remote sensing fire spot monitoring to achieve higher accuracy is an inevitable trend. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an intelligent fire point identification method that integrates a physical optimization model with deep learning.

[0006] The technical solution of the present invention is as follows:

[0007] A method for intelligent fire detection that integrates a physics optimization model and deep learning includes the following steps:

[0008] Step 1: Perform threshold dynamic processing based on the physical mechanism model to obtain an adaptive threshold and an optimized physical mechanism model;

[0009] Step 2: Construct a U-Net-based network model;

[0010] Step 3: Introduce the YOLO v5 model to detect highlighted targets;

[0011] Step 4: Based on the optimized physical mechanism model and U-Net network model, identify fire points and obtain the initial screening results of remote sensing fire points;

[0012] Step 5: Remove errors from the bright targets identified by the YOLO v5 model in the initial screening results of remote sensing fire points to improve the overall accuracy of the initial screening results of remote sensing fire points.

[0013] Preferably, step 1 includes the following sub-steps:

[0014] Sub-step 11: Use adaptive threshold optimization to optimize key indicators in the local context method to improve the flexibility between sliding windows in the algorithm;

[0015] Sub-step 12: Change the fixed threshold of the final judgment condition of the global fire index algorithm to an adaptive threshold to reduce system error.

[0016] Preferably, the specific processing procedure of step 2 is as follows:

[0017] The data used to train the remote sensing fire detection model is imbalanced. Therefore, by adopting a "blocking-fire detection-stitching" process for the input image, the inherent advantages of the U-Net network model can be brought into play.

[0018] Preferably, the building cluster highlighting target in step 3 is a comprehensive sample of morphological annotations, taking into account megacities, checkerboard-shaped urban areas, mountain towns, and remote villages.

[0019] Preferably, the specific calculation formula for the adaptive threshold optimization local context method in sub-step 11 is as follows:

[0020]

[0021]

[0022]

[0023]

[0024] δ′4>5K

[0025] Wherein, ΔT represents the difference between the brightness temperature of the band with a center wavelength of 4μm and the brightness temperature of the band with a center wavelength of 11μm; T4 represents the mean ΔT value of the effective pixels in the sliding window; T4 represents the brightness temperature of the 4μm center wavelength band. T represents the T4 mean of the effective cells in the sliding window; T 11 This indicates the brightness temperature in the 11μm band with a center wavelength. T represents the effective cells in the sliding window. 11 Mean; δ4 represents the average absolute deviation of the effective pixel center wavelength band of 4μm within the sliding window; δ 11 δ′4 represents the average absolute deviation of the effective pixel center wavelength band of 11μm in the sliding window; δ′4 represents the average absolute deviation of the background fire pixel center wavelength band of 4μm in the sliding window.

[0026] Preferably, the specific calculation formula for the global fire point index algorithm in sub-step 12 is as follows:

[0027]

[0028] Where ρ5 represents the reflectivity of the band with a center wavelength of 0.86μm, ρ6 represents the reflectivity of the band with a center wavelength of 1.60μm, ρ7 represents the reflectivity of the band with a center wavelength of 2.20μm, and k is a weighting parameter.

[0029] The beneficial effects of the intelligent fire detection method integrating physical optimization model and deep learning in this invention are as follows:

[0030] (1) For the first time, the YOLO v5 model detection model was used to eliminate errors such as bright urban buildings and artificial heat sources in remote sensing fire point identification.

[0031] (2) Adaptive threshold optimization is used to optimize the fire point identification process in the physical model, thereby reducing systematic errors such as sensor differences and sunlight environment.

[0032] (3) Construct an optimized physical mechanism model and a U-Net network model to make a decision on fire point identification. Combine the advantages of both to further reduce the false negative rate and false positive rate, and avoid the potential detection error of a single algorithm. Attached Figure Description

[0033] Figure 1 The local context algorithm diagram is used to optimize the physical mechanism model of this invention.

[0034] Figure 2 This is a diagram of the remote sensing fire point identification sub-model based on U-Net of the present invention.

[0035] Figure 3 This is a diagram of the urban highlighting error recognition sub-model based on YOLO v5 in this invention.

[0036] Figure 4 This is a flowchart of the fusion algorithm of the present invention.

[0037] Figure 5This is a Landsat 8OLI remote sensing image (762 band false color) of the present invention.

[0038] Figure 6 This is a comparison chart of the image fire detection algorithm performance of the present invention. Detailed Implementation

[0039] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0040] A smart fire detection method integrating physical optimization model and deep learning is proposed to solve the problem of false detection and false negative detection of fire points caused by fixed threshold and improve the robustness of the algorithm in extreme conditions (desert, Gobi, mountain shadow, cloud shadow, etc.).

[0041] A high-brightness target detection model for urban areas was established based on the YOLO v5 model. Various heat source errors in urban areas (such as solar reflectors, industrial mining areas, and self-built roofs made of special materials) that have long seriously interfered with fire point detection were eliminated, further improving the overall accuracy.

[0042] By combining the high utilization rate of remote sensing data by the optimized physical model with the advantages of deep learning in information mining, a fusion fire point identification algorithm is constructed to overcome the problems of poor universality caused by the inherent systematic errors of traditional single-approach algorithms.

[0043] Considering the shortcomings of both physical models and deep learning algorithms in fire detection, this algorithm, based on the optimization of the physical mechanism model, combines U-Net and YOLO v5 models to construct a remote sensing fire detection fusion algorithm. This approach complements the advantages of each method and can jointly eliminate single errors, improving accuracy and algorithm universality. In particular, it has a good effect on removing errors such as building highlighting and artificial heat sources that have long plagued remote sensing fire detection.

[0044] (1) Models based on physical mechanisms (see details) Figure 1 (Formulas 1 and 2) Threshold Dynamics. An adaptive dynamic threshold is proposed to optimize key indicators in the local context method, improving the flexibility between sliding windows in the algorithm. The fixed threshold in the final judgment condition of the global fire index algorithm (Formula 2) is changed to an adaptive threshold to reduce system error. Figure 1 For details of the formula within the box, please refer to Formula 1. ΔT represents the difference between the brightness temperature of the band with a center wavelength of 4μm and the brightness temperature of the band with a center wavelength of 11μm. T4 represents the mean ΔT value of the effective pixels in the sliding window; T4 represents the brightness temperature of the 4μm center wavelength band. T represents the T4 mean of the effective cells in the sliding window; T 11 This indicates the brightness temperature in the 11μm band with a center wavelength. T represents the effective cells in the sliding window. 11 Mean; δ4 represents the average absolute deviation of the effective pixel center wavelength band of 4μm within the sliding window; δ 11 δ′4 represents the average absolute deviation of the effective pixel center wavelength band at 11 μm within the sliding window; δ′4 represents the average absolute deviation of the background fire pixel center wavelength band at 4 μm within the sliding window. In Formula 2, ρ5 represents the reflectance at the center wavelength band of 0.86 μm, ρ6 represents the reflectance at the center wavelength band of 1.60 μm, ρ7 represents the reflectance at the center wavelength band of 2.20 μm, and k is a weighting parameter.

[0045]

[0046]

[0047]

[0048]

[0049] δ′4>5K Formula 1

[0050]

[0051] (2) Construct a deep learning fire detection model based on U-Net, such as Figure 2 As shown, the U-Net network has a simple structure, can fuse multi-scale features, and can segment each pixel to achieve higher accuracy. Since the remote sensing fire point data used for training is imbalanced, adopting a "blocking-fire point recognition-stitching" processing flow for the input image can more effectively leverage the inherent advantages of the U-Net network.

[0052] (3) Introducing the YOLO v5 model, such as Figure 3 As shown, the detection of highlighted targets in building clusters comprehensively considers morphologically labeled samples from megacities, checkerboard-shaped urban areas, mountainous towns, and remote villages, ensuring uniform selection across layers. This is the first attempt at using a high-precision target detection model to eliminate errors in remote sensing fire point identification, providing new insights for other scholars conducting related research.

[0053] (4) Overall fusion algorithm such as Figure 4The algorithm obtains high-precision fire point results. The optimized physical model and U-Net network intelligent decision-making system identify fire points, obtaining initial screening results for remote sensing fire points. High-brightness building targets identified by the YOLO v5 model are then removed from the initial screening results as errors, improving the overall accuracy of fire point identification. The algorithm integrates the advantages of each sub-model in fire point identification and error removal, while mitigating and supplementing their respective shortcomings, resulting in more stable performance, high fault tolerance, and good robustness.

[0054] Choose the fire in Shenyang, Liaoning on October 26, 2014 as... Figure 5 As shown in the test data, the deep learning part of the fusion algorithm inputs the data in blocks of size 256*256. The total number of samples for model training is 2040, and the ratio of training set:test set:validation set is 9:1:3.

[0055] In summary, this adaptive fusion algorithm is as follows: Figure 4 The results show good detection performance under extreme conditions such as the Gobi Desert, thin cloud shadows, and mountain shadows; it can effectively remove most false fire points such as bright buildings in urban areas and artificial heat sources, making up for the shortcomings of existing algorithms in this regard and greatly reducing the false alarm rate of fire point detection; it can detect fire point locations more accurately, improve the fire point detection hit rate, and obtain high-precision fire point remote sensing detection results, providing technical support for decision-makers to obtain accurate and timely fire point information.

Claims

1. A method for intelligent fire detection that integrates a physical optimization model and deep learning, characterized in that, Includes the following steps: Step 1: Perform threshold dynamic processing based on the physical mechanism model to obtain an adaptive threshold and an optimized physical mechanism model; Step 2: Construct a U-Net-based network model; Step 3: Introduce the YOLO v5 model to detect highlighted targets; Step 4: Based on the optimized physical mechanism model and U-Net network model, identify fire points and obtain the initial screening results of remote sensing fire points; Step 5: Remove errors from the bright targets identified by the YOLO v5 model in the initial screening results of remote sensing fire points to improve the overall accuracy of the initial screening results of remote sensing fire points. Step 1 includes the following sub-steps: Sub-step 11: Use adaptive threshold optimization to optimize key indicators in the local context method to improve the flexibility between sliding windows in the algorithm; Sub-step 12: Change the fixed threshold of the final judgment condition of the global fire point index algorithm to an adaptive threshold to reduce system error; The specific processing procedure for step 2 is as follows: The data used to train the remote sensing fire detection model is imbalanced, so a "blocking-fire detection-stitching" process is adopted for the input image to give full play to the inherent advantages of the U-Net network model. The building clusters highlighted in step 3 are samples of morphological annotations for megacities, checkerboard-shaped urban areas, mountain towns, and remote villages. The specific calculation formula for the adaptive threshold optimization local context method in sub-step 11 is as follows: ， in, This indicates that the center wavelength is 4. The brightness temperature of the band and the center wavelength are 11 The difference in brightness temperature across the bands; Represents the effective cells in the sliding window Mean; This indicates that the center wavelength is 4. Brightness temperature of the band; Represents the effective cells in the sliding window Mean; This indicates that the center wavelength is 11. Brightness temperature of the band; Represents the effective cells in the sliding window Mean; This indicates that the center wavelength of the effective pixel in the sliding window is 4. Mean absolute deviation of the band; This indicates that the center wavelength of the effective pixels in the sliding window is 11. Mean absolute deviation of the band; This indicates that the center wavelength of the background fire pixel in the sliding window is 4. Mean absolute deviation of the band; The specific calculation formula for the global fire point index algorithm in sub-step 12 is as follows: （ ）， in, The representative center wavelength is 0.

86. Band reflectivity, The representative center wavelength is 1.

60. Band reflectivity; The center wavelength is represented as 2.

20. Band reflectivity; k is the weighting parameter.

Images