CN-121982558-A - Satellite fire point detection method based on fusion of expert system and neural network

CN121982558ACN 121982558 ACN121982558 ACN 121982558ACN-121982558-A

Abstract

The invention discloses an on-satellite fire point detection method based on fusion of an expert system and a neural network, which comprises the steps of extracting and obtaining a fire point characteristic map library according to preset multi-source remote sensing data; the method comprises the steps of constructing a lightweight neural network, inputting preset multi-source remote sensing data into the lightweight neural network, outputting fire data by the lightweight neural network, fusing a fire characteristic map library to obtain fire type probability according to a kriging interpolation method on the fire data, and obtaining the fire type according to the fire type probability, the preset characteristic map confidence, the preset image detection confidence and the fire detection confidence. The intelligent detection method can be used for intelligent detection of the satellite fire points with high timeliness, high confidence and strong scene adaptability, so as to support high-frequency inspection and rapid early warning and improve the efficiency of the satellite in early detection of fire.

Inventors

WANG FUHAI
WANG CHENGLUN
LIU FENGJING
SHI JIAWEI
YUE RONGGANG
LI XIANG
CHEN CHEN
LI ZHUPENG
SUN RONGYANG
ZHAO LUMING
ZHANG GUANGYU

Assignees

中国空间技术研究院

Dates

Publication Date: 20260505
Application Date: 20251210

Claims (10)

1. An on-satellite fire point detection method based on fusion of an expert system and a neural network is characterized by comprising the following steps: extracting according to preset multi-source remote sensing data to obtain a fire point characteristic map library; Constructing a lightweight neural network, inputting preset multi-source remote sensing data into the lightweight neural network, and outputting fire point data by the lightweight neural network; and fusing the fire point characteristic map library to obtain the fire point type probability according to the fire point data by a Kriging interpolation method, and obtaining the fire point type according to the fire point type probability, the preset characteristic map confidence coefficient, the preset image detection confidence coefficient and the fire point detection confidence coefficient.
2. The method for detecting the fire point on the satellite based on the fusion of the expert system and the neural network according to claim 1 is characterized by further comprising the steps of obtaining a target detectability characterization model according to a background clutter quantification characterization model and a signal-to-clutter ratio calculation model, obtaining a detection band set according to the target detectability characterization model and target and background characteristic data, wherein all points of detection bands in the detection band set exceeding a set target signal-to-clutter ratio threshold are candidate bands, and selecting a point with the maximum signal-to-clutter ratio from the candidate bands, namely the optimal detection spectrum.
3. The method for detecting fire points on a satellite based on the fusion of an expert system and a neural network according to claim 1, wherein the fire point characteristic map library comprises color characteristics, target position information, spectral characteristics, texture characteristics and geometric characteristics, The color features are statistically characterized by using a color histogram, a color mean value, and color entropy; the target position information is obtained by inversion calculation of an image data coordinate system, the small target is represented by adopting a point coordinate form, and the large and medium targets are represented by adopting a series of positions of multiple points or planes, wherein the size of the target is small if the size is smaller than a preset size value, and the size of the target is large and medium if the size is larger than the preset size value; the texture features are local texture features of images in each grid unit are extracted by adopting a local binary pattern; The geometric features are local geometric features and global geometric features of the image are respectively extracted by adopting Speeded Up RobustFeatures operators and Evaluation of GIST operators.
4. The method for detecting fire points on a satellite based on fusion of an expert system and a neural network according to claim 3, wherein the spectral characteristics are obtained by the following formula: spe=[mean,std]; ; Wherein spe is a spectrum characteristic, mean is a band mean value, std is a band standard deviation, m is the number of pixels in a grid, The gray value of the j-th pixel in the grid is represented by j, which is the serial number of the pixel.
5. The method for detecting the fire point on the star based on the fusion of the expert system and the neural network, as set forth in claim 1, wherein the lightweight neural network uses a lightweight feature extraction network MobileNet V as a backbone network and introduces a depth separable convolution, a nonlinear activation function of the lightweight neural network uses a Mish activation function, The lightweight characteristic extraction network MobileNet V adopts an inverse residual error structure with a linear bottleneck to reduce the calculation overhead caused by excessive convolution kernel layers; the standard convolution is decomposed into a depth convolution and a point-by-point convolution by the depth separable convolution, wherein the depth convolution independently carries out space convolution on each input channel, and the point-by-point convolution is used for combining the depth convolution output; mish the activation function is mix (x) =x×tanh (1+e x ), where x is the input feature.
6. The method for detecting fire points on a star based on fusion of an expert system and a neural network as claimed in claim 1, wherein the fire points comprise coordinates of the fire points, confidence and probability of fire type.
7. The method for detecting fire points on a star based on fusion of an expert system and a neural network as claimed in claim 1, wherein the kriging interpolation method comprises: establishing a variation function model according to the attribute values and the space coordinates of the known points; Calculating the difference between the known points to obtain a semi-variation function; obtaining the space weight between the unknown point and the known point according to the attribute, the space coordinate and the semi-variation function of the known point; And predicting the unknown click through by using the attribute values and the spatial weights of the known points to obtain the attribute values.
8. The method for detecting fire points on a satellite based on fusion of expert system and neural network as claimed in claim 1, wherein the probability of fire point type is as follows Is obtained by the following formula: ; Wherein, the Representing the longitude and latitude of the fire point, For the n-th fire type, For the probability of the nth fire type, For a fire type of 1 st one, For the type 2 fire to be a type 2 fire, For a probability of the 1 st fire type, For a probability of the 2 nd fire type, Is the serial number of the fire point, Is the number of fires.
9. The method for detecting fire points on a satellite based on fusion of expert system and neural network as claimed in claim 1, wherein the fire point type is as follows Is obtained by the following formula: ; Wherein, the Representing the longitude and latitude of the fire point, In order to preset the confidence level of the characteristic map, Is the fire point Is used for determining the detection confidence of the test pattern, For the n-th fire type, For the fused probability for the nth fire type, For the probability of the nth fire type obtained by the neural network, For the probability of the nth fire type, Is the serial number of the fire point, Is the number of fires.
10. An on-satellite fire point detection system based on fusion of an expert system and a neural network is characterized by comprising: The first module is used for extracting and obtaining a fire point characteristic map library according to preset multi-source remote sensing data; the second module is used for constructing a lightweight neural network, inputting preset multi-source remote sensing data into the lightweight neural network, and outputting fire point data by the lightweight neural network; And the third module is used for fusing the fire point characteristic map library to obtain the fire point type probability according to the fire point data by a Kriging interpolation method, and obtaining the fire point type according to the fire point type probability, the preset characteristic map confidence coefficient, the preset image detection confidence coefficient and the detection confidence coefficient of the fire point.

Description

Satellite fire point detection method based on fusion of expert system and neural network Technical Field The invention belongs to the technical field of high-orbit imaging, and particularly relates to an on-satellite fire point detection method based on fusion of an expert system and a neural network. Background Disaster accidents, particularly large-scale forest fires and industrial fires (such as chemical plant explosions), pose a significant threat to human society, causing significant life and health losses and economic losses. Satellite remote sensing technology is a key means of fire detection by virtue of wide area coverage and real-time dynamic advantages, but most of current applications are post-disaster confirmation, lack of high-timeliness early warning capability and limit the effect of satellites in early intervention of fire. In the prior art, the traditional fire detection method based on threshold judgment has poor environmental adaptability, and is difficult to reject false fire targets such as high-reflection ground objects and the like, so that the false alarm rate is high. For example, fixed threshold detection schemes (such as the dynamic luminance temperature threshold method proposed by Ding et al, A WILDFIRE detection algorithm based on THE DYNAMIC brightness temperature threshold) are not robust enough under dynamic background changes (such as morning and evening alternation and seasonal variation), and detection accuracy fluctuates significantly. In recent years, deep learning algorithms have advanced in the field of fire detection, such as Gargiulo et al, which uses super-resolution CNN to process Sentinel-2 images (A CNN-based super-resolution technique for ACTIVE FIRE detection on Sentinel-2 data), improving small target detection capability, dell' Aglio et al, which further optimizes accuracy by combining an Active Fire Index (AFI) with a multi-scale convolution kernel （Active fire detection in multispectral super-resolved sentinel-2 images by means of sam-based approach）,. An example segmentation method is characterized in that pixel level segmentation is realized by a Mask R-CNN (Begum et al, mask R-CNN For Fire Detection), but the computation complexity is high, the on-board real-time processing requirement is difficult to meet, the data throughput is high when massive multispectral data are processed, and the timeliness cannot be guaranteed. Although the method improves the detection precision, the method still has the obvious defects that the data volume is huge due to multi-band combined use, the on-board processing timeliness is poor, the scene adaptability of the existing scheme is weak, the false alarm rate is not effectively controlled, and a mechanism of fusing expert knowledge is lacking, so that the method cannot adapt to dynamic background change. Disclosure of Invention The invention solves the technical problems of overcoming the defects of the prior art, providing the on-board fire detection method based on the fusion of the expert system and the neural network, which can perform on-board fire intelligent detection with high timeliness, high confidence and strong scene adaptability so as to support high-frequency inspection and rapid early warning and improve the effectiveness of satellites in early detection of fire. The technical scheme is that the on-satellite fire point detection method based on the fusion of the expert system and the neural network comprises the steps of extracting a fire point characteristic map library according to preset multi-source remote sensing data, constructing a lightweight neural network, inputting the preset multi-source remote sensing data into the lightweight neural network, outputting the fire point data by the lightweight neural network, fusing the fire point characteristic map library to obtain fire point type probability according to the fire point type probability, the preset characteristic map confidence, the preset image detection confidence and the fire point detection confidence, and obtaining the fire point type according to the fire point type probability, the preset characteristic map confidence and the fire point detection confidence. The on-satellite fire point detection method based on the fusion of the expert system and the neural network further comprises the steps of obtaining a target detectability characterization model according to the background clutter quantification characterization model and the signal-to-noise ratio calculation model, obtaining a detection band set according to the target detectability characterization model and target and background characteristic data, wherein all points of detection bands in the detection band set exceeding a set target signal-to-noise ratio threshold are candidate bands, and selecting the point with the largest signal-to-noise ratio from the candidate bands, namely the optimal detection spectrum. The on-satellite fire point detection method based on the fusion