Forest region satellite fire point monitoring method based on deep learning

Abstract

The invention relates to a forest region satellite fire point monitoring method based on deep learning, and the method comprises the steps: collecting a multispectral image of a target forest region through a satellite remote sensing system, and carrying out the preprocessing of the multispectral image; feature information extraction is carried out on the preprocessed multispectral image, the extracted feature information is classified by using a convolutional neural network, and a classification feature vector is obtained; fire source areas are determined according to the classification feature vectors, and a first fire source area set is obtained; acquiring environment data, and acquiring a second fire source region set through a support vector machine model according to the environment data and the preprocessed multispectral image; according to the first fire source area set and the second fire source area set, real fire point pixels are obtained; and real fire point pixels are mapped to geographic coordinates of the target forest region, missing data are filled by adopting a spatial interpolation technology, and final fire point position distribution is obtained. According to the method, the accuracy and reliability of forest fire detection under the complex terrain condition are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of fire monitoring technology, and in particular to a deep learning-based satellite fire detection method for forest areas. Background Technology

[0002] Forest fires are one of the major threats to global ecological security, and their early and accurate identification is crucial for timely fire suppression, reducing economic losses, and protecting the ecological environment.

[0003] Satellite remote sensing technology has become a core means of monitoring forest fires, capturing changes in fire conditions through multi-source images. However, current identification methods that rely on traditional spectral thresholds or simple image processing cannot effectively cope with the complex interference of the forest environment, resulting in early fire signals often being missed or misjudged.

[0004] The main drawback of these methods when processing multi-band satellite images is that they ignore the dynamic correlation between features in different spectral ranges. For example, the diffusion pattern of smoke in visible light and the radiation intensity of high temperature in infrared should complement each other to verify the authenticity of fire points, but due to the lack of integration, identification bias occurs. Especially in densely vegetated areas, smoke reflection is confused with surface heat sources, causing frequent false positive alarms and consuming valuable response resources.

[0005] The core technical challenge in forest area satellite image recognition lies in the failure to effectively integrate the unique features of multispectral bands with the interaction with the geographical environment. This interaction stems primarily from the occlusion of satellite images by terrain slope and vegetation type distribution, causing infrared high-temperature signals to attenuate on steep slopes, while visible light smoke signals are interfered with by increased reflection due to canopy density. This band attenuation further amplifies the geographical occlusion effect, resulting in inconsistent pixel intensity distributions for the same fire point in different areas. For example, in dense coniferous forests, the infrared brightness temperature of a real fire source may be below the threshold, while the smoke reflection of neighboring pixels is abnormally high, creating a contradiction of feature mismatch.

[0006] Therefore, how to accurately integrate multispectral band features with forest terrain and vegetation distribution in real-time satellite images to achieve unified determination of fire point pixels under obstructed environments has become a key issue in improving the reliability of early identification of forest fires. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based satellite fire detection method for forest areas, which improves the accuracy and reliability of forest fire detection under complex terrain conditions and provides effective technical support for forest fire early warning.

[0008] To achieve the above objectives, the present invention provides the following solution:

[0009] A deep learning-based method for monitoring satellite fire points in forest areas includes:

[0010] Multispectral images of the target forest area are acquired using a satellite remote sensing system, and the multispectral images are preprocessed.

[0011] Feature information is extracted from the preprocessed multispectral image, and a convolutional neural network is used to classify the extracted feature information to obtain classification feature vectors;

[0012] The fire source region is determined based on the classification feature vector, and a first set of fire source regions is obtained.

[0013] Acquire environmental data, and obtain a second set of fire source regions based on the environmental data and the preprocessed multispectral image using a support vector machine model;

[0014] Based on the first set of fire source regions and the second set of fire source regions, obtain the actual fire point pixels;

[0015] The actual fire point pixels are mapped to the geographic coordinates of the target forest area, and spatial interpolation technology is used to fill in the missing data to obtain the final fire point location distribution.

[0016] Optionally, preprocessing the multispectral image includes:

[0017] Wavelet transform is used to remove atmospheric interference from the multispectral image to obtain a preliminary denoised multispectral image;

[0018] The radiation distortion of the preliminary denoised multispectral image is determined based on the number of spectral bands. If the radiation distortion exceeds a preset threshold, histogram equalization is used to correct the preliminary denoised multispectral image to obtain a radiation-corrected multispectral image.

[0019] The radiometrically corrected multispectral image is processed by polynomial transformation to eliminate spatial analytical offset, and a geometrically corrected multispectral image is obtained as the preprocessed multispectral image.

[0020] Optionally, feature extraction from the preprocessed multispectral image includes:

[0021] Based on the preprocessed multispectral image, the visible light band is obtained;

[0022] Based on the visible light band, smoke diffusion features and high-temperature radiation features are extracted to obtain a fused image.

[0023] Optionally, a convolutional neural network can be used to classify the extracted features to obtain classification feature vectors, including:

[0024] The fused image is input into a convolutional neural network, which outputs a classification feature vector. The convolutional neural network includes an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer connected in sequence. The input layer receives the fused image. The convolutional layers extract local features of the data using convolutional kernels of different sizes. The pooling layers downsample the feature maps output by the convolutional layers. The fully connected layers fully connect the features output by the pooling layers to achieve feature fusion. The output layer outputs a classification feature vector. The categories of the classification feature vector include normal vegetation, smoke areas, high-temperature areas, and background areas. The classification feature vector contains the probability value of each pixel belonging to each category.

[0025] Optionally, determining the potential fire source area based on the classification feature vector includes:

[0026] For pixels divided into the smoke region and the high temperature region, the matching degree of smoke diffusion characteristics and high temperature radiation characteristics is calculated. If the matching degree is higher than a preset value, it is determined as the potential fire source region.

[0027] Optionally, obtaining a second set of potential fire source areas based on the environmental data and the preprocessed multispectral image using a support vector machine model includes:

[0028] Acquire terrain occlusion data for the target area. If the occlusion angle is greater than the critical angle value, perform occlusion correction processing on the band features.

[0029] The corrected band features are input into a support vector machine model to obtain a second set of potential fire source regions, wherein the support vector machine model establishes a nonlinear mapping relationship between band features and fire source probabilities.

[0030] Optionally, mapping the actual fire point pixels to the geographic coordinates of the target forest area includes:

[0031] The actual fire point pixels are mapped to the forest area geographic coordinate system through pixel mapping relationships;

[0032] Using a pre-defined coordinate transformation model and combining it with georeferenced information from remote sensing images, the geographic coordinates corresponding to the fire point pixels are determined, and a set of fire point geographic coordinates is obtained.

[0033] Optionally, spatial interpolation techniques are used to fill in missing data to obtain the final fire point location distribution, including:

[0034] The missing data is identified from the set of fire point geographic coordinates. Spatial autocorrelation analysis is used to determine whether there is data loss due to occlusion of geographic coordinate points, and the missing data area is obtained.

[0035] The Kriging interpolation algorithm is used to calculate spatial interpolation weights based on the set of fire point geographic coordinates around the missing data area, fill in the missing data in the occluded area, and obtain the complete set of fire point geographic coordinates.

[0036] If outliers still exist in the completed set of fire point geographic coordinates, the distance-weighted average method is used to smooth out the outliers by combining the spatial distribution characteristics of neighboring coordinates, and a smoothed set of fire point geographic coordinates is obtained.

[0037] By using density clustering algorithms, spatial clustering analysis is performed on the smoothed set of fire point geographic coordinates to determine the dense areas of fire point location distribution and obtain the final fire point location distribution.

[0038] The beneficial effects of this invention are as follows: This invention solves the key problem of low accuracy in early identification of forest fires under complex terrain conditions. Addressing the difficulty of misjudgment caused by terrain obstruction and vegetation interference in traditional fire point detection, this invention acquires multispectral images of forest areas through satellite remote sensing, integrates visible light smoke diffusion characteristics with infrared high-temperature radiation characteristics, employs convolutional neural networks for feature classification and matching to identify potential fire source areas, uses a support vector machine model to obtain the actual fire point pixels under complex environments, and finally utilizes spatial interpolation technology to compensate for missing data due to obstruction, accurately locating the geographical coordinates of the fire point. This invention significantly improves the accuracy and reliability of forest fire detection under complex terrain conditions, providing effective technical support for forest fire early warning. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This invention provides a deep learning-based method for monitoring satellite fire points in forest areas. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] like Figure 1 As shown in the figure, this embodiment proposes a deep learning-based satellite fire detection method for forest areas, including:

[0044] Multispectral images of the target forest area were acquired using a satellite remote sensing system, and the multispectral images were preprocessed.

[0045] Feature information is extracted from the preprocessed multispectral image, and a convolutional neural network is used to classify the extracted feature information to obtain classification feature vectors;

[0046] The fire source region is determined based on the classification feature vector, and the first set of fire source regions is obtained.

[0047] Acquire environmental data, and based on the environmental data and preprocessed multispectral images, obtain the set of second fire source regions through a support vector machine model;

[0048] Based on the first set of fire source regions and the second set of fire source regions, obtain the actual fire point pixels;

[0049] The actual fire point pixels are mapped to the geographic coordinates of the target forest area, and spatial interpolation technology is used to fill in the missing data to obtain the final distribution of fire point locations.

[0050] Further preprocessing of the multispectral images includes:

[0051] Wavelet transform was used to remove atmospheric interference from the multispectral image, resulting in a preliminary denoised multispectral image.

[0052] The radiation distortion of the preliminary denoised multispectral image is determined based on the number of spectral bands. If the radiation distortion exceeds a preset threshold, histogram equalization is used to correct the preliminary denoised multispectral image to obtain a radiation-corrected multispectral image.

[0053] Spatial analytical offset is eliminated by processing the radiometrically corrected multispectral image through polynomial transformation, and the geometrically corrected multispectral image is obtained as the preprocessed multispectral image.

[0054] Specifically, the Landsat-8 satellite can acquire data in 11 spectral bands, including visible light, near-infrared, and shortwave infrared, with a spatial resolution of 30 meters, effectively covering large areas of forest. This multispectral imaging technology reflects the differences in the spectral characteristics of vegetation through different bands, providing rich spectral information for subsequent forest type identification and health monitoring. Wavelet transform processing of multispectral images to remove atmospheric interference is a key preprocessing step. Specifically, water vapor and aerosols in the atmosphere can cause image blurring and spectral distortion. Wavelet transform can decompose the image into different frequency components, and thresholding can remove high-frequency noise.

[0055] In one embodiment, when the difference in radiance between the visible light band and the near-infrared band exceeds a preset threshold of 15%, it indicates a significant radiation distortion. At this point, histogram equalization is used to adjust the grayscale distribution of each band to a similar range, ensuring radiation consistency across different bands. This processing method can eliminate the effects of uneven sensor response and variations in solar altitude angle. Polynomial transform processing for geometric correction is an effective method for eliminating spatial analytical offset.

[0056] Furthermore, feature extraction from the preprocessed multispectral image includes:

[0057] The visible light band is obtained from the preprocessed multispectral image;

[0058] Based on the visible light band, smoke diffusion features and high-temperature radiation features are extracted to obtain a fused image.

[0059] Furthermore, a convolutional neural network is used to classify the extracted features, obtaining classification feature vectors including:

[0060] The fused image is input into a convolutional neural network (CNN), which outputs a classification feature vector. The CNN consists of an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer connected in sequence. The input layer receives the fused image. The convolutional layers extract local features of the data using convolutional kernels of different sizes. The pooling layers downsample the feature maps output by the convolutional layers. The fully connected layers fully connect the features output by the pooling layers to achieve feature fusion. The output layer outputs a classification feature vector. The categories of the classification feature vector include normal vegetation, smoke areas, high-temperature areas, and background areas. The classification feature vector contains the probability value of each pixel belonging to each category.

[0061] Specifically, model training includes: training the constructed deep learning model using labeled satellite image data of forest areas, with the labeled data including category information; during training, using the cross-entropy loss function as the optimization objective, and employing stochastic gradient descent (SGD) and its variants such as Adagrad, Adadelta, and Adam to optimize the model parameters, continuously adjusting the model's weights and biases to minimize the difference between the model's predictions on the training dataset and the labeled results; controlling the convergence speed and stability of the training process by setting appropriate hyperparameters such as learning rate and batch size; and employing early stopping to avoid overfitting, stopping training when the model's performance on the validation set no longer improves.

[0062] In the Convolutional Neural Network (CNN) model, convolutional layers use alternating 3×3 and 5×5 kernels to extract features from different receptive fields; pooling layers employ max pooling with a 2×2 kernel size and a stride of 2; fully connected layers contain two hidden layers, the first with 512 nodes and the second with 256 nodes. The attention mechanism combines channel attention and spatial attention; channel attention calculates channel weights through global average pooling and the two fully connected layers, while spatial attention calculates spatial weights through convolutional operations. These two weighted sums are then fused onto the original feature map.

[0063] Furthermore, determining potential fire source areas based on classification feature vectors includes:

[0064] For pixels divided into smoke areas and high-temperature areas, the matching degree of smoke diffusion characteristics and high-temperature radiation characteristics is calculated. If the matching degree is higher than the preset value, it is determined as a potential fire source area.

[0065] Specifically, based on the preprocessed and post-processed multispectral image, a first image is obtained. The visible light band is extracted from the first image to obtain the smoke diffusion and diffusion path, resulting in a second image. The infrared band is extracted from the first image to obtain the high-temperature radiation and radiation intensity, resulting in a third image. If the smoke diffusion in the second image exceeds a preset threshold, a convolutional neural network is used to process the second and third images to obtain a fourth image. Preliminary classification is performed using the fourth image to determine the classification vector. The location of the high-temperature radiation is determined based on the classification vector, resulting in a fifth image. The diffusion path and radiation intensity are obtained from the fifth image to determine the final classification result.

[0066] For example, when extracting the visible light band from the first image, the system focuses on the spectral response in the 400 to 700 nanometer wavelength range. Smoke diffusion detection is achieved by analyzing the spatial distribution changes of pixel grayscale values; when the grayscale difference between adjacent pixels is less than 15 digit levels, it usually indicates the presence of smoke obstruction. The diffusion path is determined based on the gradient direction of the smoke edge, forming a continuous path trajectory by connecting pixels with similar spectral characteristics.

[0067] In one possible implementation, infrared band extraction primarily targets the 8-14 micrometer thermal infrared band, which is effective in capturing surface temperature anomalies. High-temperature radiation detection sets a temperature threshold of 45 degrees Celsius; areas exceeding this temperature are marked as potential fire points or heat sources. Radiation intensity is quantified using a digital numerical DN value; a DN value exceeding 200 indicates a significant thermal radiation anomaly in the area.

[0068] For example, when the smoke diffusion area in the second image exceeds 20% of the total image area, the convolutional neural network processing mechanism is triggered. The network adopts a multi-scale feature fusion architecture, processing both visible light and infrared band information simultaneously. Specifically, the network extracts multi-scale features using convolutional kernels of different sizes (3×3 and 5×5), and then effectively integrates the information from the two bands through a feature fusion layer to generate a fourth image containing rich spatial and spectral information. It should be noted that the initial classification process uses a pixel-level classification method, dividing the image into four categories: normal vegetation, smoke areas, high-temperature areas, and background areas. The classification vector contains the probability value of each pixel belonging to each category, with a vector dimension of 4×1. When the probability value of a high-temperature area exceeds 0.8, the pixel is identified as a high-temperature radiation location.

[0069] In one embodiment, the fifth image is generated by simultaneously extracting the direction angle of the diffusion path and the peak distribution of radiation intensity. Diffusion path analysis considers wind direction and topographic factors, assessing the diffusion trend by calculating the continuity and consistency of the path. Radiation intensity analysis focuses on the spatial clustering of hotspots and temperature gradient changes; when the clustering index exceeds 0.75 and the temperature gradient is greater than 2 degrees Celsius per meter, the system determines the presence of an active heat source. The final classification result comprehensively considers the smoke diffusion range, high-temperature radiation intensity, and spatial distribution characteristics, achieving accurate identification and location of abnormal conditions in forest areas.

[0070] In one possible implementation, cosine similarity calculation assesses the matching degree by comparing the directional consistency of the high-temperature radiation feature vector and the smoke diffusion feature vector. When the cosine similarity value of the two feature vectors reaches 0.85 or higher, it indicates that both high temperature and smoke diffusion exist at that location, and the system determines that there is a fire source risk. This calculation method can effectively eliminate false alarms caused by simply high-temperature equipment or harmless smoke.

[0071] Furthermore, based on environmental data and preprocessed multispectral images, a set of second potential fire source regions is obtained using a support vector machine model, including:

[0072] Acquire terrain occlusion data for the target area. If the occlusion angle is greater than the critical angle value, perform occlusion correction processing on the band features.

[0073] The corrected band features are input into the support vector machine model to obtain the second set of potential fire source regions. The support vector machine model establishes a nonlinear mapping relationship between the band features and the fire source probability.

[0074] Specifically, terrain occlusion data is acquired through a digital elevation model (DEM). The system collects elevation information for each grid point within the monitoring area, and combines this with lidar scanning to obtain precise surface undulation data. For example, an area is marked when its slope exceeds 25 degrees and its vegetation density index is 0.45. This is because steep slopes facilitate the upward spread of fire, while moderate vegetation density provides both combustible material and adequate ventilation.

[0075] In one possible implementation, terrain occlusion correction compensates for spectral distortion caused by obstacles such as mountains and buildings. When the occlusion angle is greater than 30 degrees, the system adjusts the band reflectivity accordingly based on the geometric characteristics of the obstruction and the lighting conditions. Specifically, the near-infrared reflectivity of the obstructed area decreases by 15% to 25%, while the thermal infrared band may exhibit temperature underestimation. A support vector machine model is trained on a large amount of known fire source sample data to establish a complex mapping relationship between band features and fire source probability. The model can identify that the thermal infrared band contributes the most to fire source detection, with a weight coefficient set to 0.4; the near-infrared band has a weight of 0.25; the visible red band has a weight of 0.2; and the remaining bands have weights of 0.15.

[0076] Furthermore, mapping the actual fire point pixels to the geographic coordinates of the target forest area includes:

[0077] By mapping the pixel relationships, the real fire point pixels are mapped to the forest area geographic coordinate system;

[0078] Using a pre-defined coordinate transformation model and combining it with georeferenced information from remote sensing images, the geographic coordinates corresponding to the fire point pixels are determined, and a set of fire point geographic coordinates is obtained.

[0079] Specifically, establishing pixel mapping relationships requires combining the orbital parameters and imaging geometry model of the remote sensing satellite. Taking Landsat-8 satellite data as an example, each pixel corresponds to a 30m x 30m actual area on the ground. Through an affine transformation model, the pixel coordinates are converted into latitude and longitude information in the geographic coordinate system. When the image covers a forest area, the fire point pixel located at row 500, column 300 may correspond to the geographical location of 45.2356 degrees north latitude and 126.7891 degrees east longitude.

[0080] Furthermore, spatial interpolation techniques are used to fill in the missing data and obtain the final fire point location distribution, including:

[0081] Identify missing data from the set of fire point geographic coordinates, use spatial autocorrelation analysis to determine whether there is data loss due to occlusion of geographic coordinate points, and obtain the missing data area;

[0082] The Kriging interpolation algorithm is used to calculate spatial interpolation weights based on the set of fire point geographic coordinates around the missing data area, fill in the missing data in the occluded area, and obtain the complete set of fire point geographic coordinates.

[0083] If outliers still exist in the completed set of fire point geographic coordinates, the distance-weighted average method is used to smooth out the outliers by combining the spatial distribution characteristics of neighboring coordinates, and a smoothed set of fire point geographic coordinates is obtained.

[0084] By using density clustering algorithms, spatial clustering analysis is performed on the smoothed set of fire point geographic coordinates to determine the dense areas of fire point location distribution and obtain the final fire point location distribution.

[0085] In one possible implementation, spatial autocorrelation analysis identifies occlusion areas by calculating the Moran index. When the fire point density in a certain area differs significantly from the surrounding area and the Moran index is below 0.1, the system determines that the area has missing data due to cloud or smoke occlusion. For example, in a 10 square kilometer monitoring area, 15 fire points should normally be detected, but only 3 are actually detected; the system identifies occlusion in this area. For instance, when processing missing data, the Kriging interpolation algorithm selects known fire points within a 2-kilometer radius of the missing point as references. By calculating spatial distance weights, fire points closer to the missing point have higher weights. When there are 4 known fire points around the missing point, the system calculates the interpolation result for the missing location based on the spatial location and fire intensity value of each point, thereby completing the fire point information for the occluded area.

[0086] In one embodiment, a distance-weighted average method is used to smooth out outliers. When the location coordinates of a fire point are more than 500 meters away from the average distance of its eight neighboring points, the system identifies it as an outlier. By calculating the weighted average coordinates of this point and its neighbors, where the distance weights use an inverse distance weighting method, the coordinates of the outlier are ultimately adjusted to a reasonable range. Specifically, the density clustering algorithm sets a minimum neighborhood radius of 200 meters and a minimum number of points of 5. When a region contains more than 5 fire points within a 200-meter radius, the system marks that region as a densely populated fire area. Through this clustering analysis, the main fire distribution areas within the forest area can be identified, providing important spatial distribution information for fire fighting.

[0087] In this embodiment, a gridded analysis method is used to divide the forest area's geographical coordinate system into spatial grids based on the distribution of fire points. The fire point density within each grid is calculated to obtain a density map of fire point location distribution. For example, the gridded analysis divides the entire forest area into regular 100m × 100m grids. The number of fire points within each grid is counted, and the fire point density value is calculated. When a grid contains 3 fire points, the fire point density of that grid is 300 per square kilometer. Through this density calculation method, a fire point location distribution density map covering the entire forest area is finally generated, intuitively showing the spatial distribution characteristics and intensity variation patterns of the fires.

[0088] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A forest satellite fire point monitoring method based on deep learning, characterized in that, The method comprises the following steps: Collecting a multispectral image of a target forest area by a satellite remote sensing system, and preprocessing the multispectral image; Extracting feature information from the preprocessed multispectral image, classifying the extracted feature information by using a convolutional neural network, and obtaining a classification feature vector; Determining a fire source area according to the classification feature vector, and obtaining a first fire source area set; Obtaining environmental data, and obtaining a second fire source area set by using a support vector machine model according to the environmental data and the preprocessed multispectral image; Obtaining a real fire point pixel according to the first fire source area set and the second fire source area set; Mapping the real fire point pixel to geographical coordinates of the target forest area, filling in missing data by using a spatial interpolation technique, and obtaining a final fire point position distribution. 2.The deep learning-based forest satellite fire point monitoring method of claim 1, wherein, The preprocessing of the multispectral image comprises the following steps: Removing atmospheric interference from the multispectral image by using a wavelet transform, and obtaining a preliminary denoising multispectral image; Judging radiation distortion of the preliminary denoising multispectral image according to the number of spectral bands, and correcting the preliminary denoising multispectral image by using histogram equalization if the radiation distortion exceeds a preset threshold, thereby obtaining a radiation correction multispectral image; Processing the radiation correction multispectral image by using a polynomial transform to eliminate spatial resolution deviation, and obtaining a geometric correction multispectral image as the preprocessed multispectral image. 3.The deep learning-based forest satellite fire point monitoring method of claim 1, wherein, The feature information extraction from the preprocessed multispectral image comprises the following steps: Obtaining a visible light band according to the preprocessed multispectral image; Extracting smoke diffusion features and high-temperature radiation features according to the visible light band, and obtaining a fusion image. 4.The deep learning-based forest satellite fire spot monitoring method of claim 3, wherein, The classification of the extracted features by using a convolutional neural network to obtain a classification feature vector comprises the following steps: Inputting the fusion image into a convolutional neural network, and outputting the classification feature vector, wherein the convolutional neural network comprises an input layer, a plurality of convolutional layers, a pooling layer, a fully connected layer and an output layer connected in sequence; the input layer receives the fusion image, the convolutional layers extract local features of data by using convolutional kernels of different sizes, the pooling layer down-samples feature maps output by the convolutional layers, the fully connected layer fully connects features output by the pooling layer to realize feature fusion, and the output layer outputs the classification feature vector; the classification of the classification feature vector includes normal vegetation, smoke area, high-temperature area and background area, and the classification feature vector contains probability values of each pixel belonging to each category.

5. The deep learning-based forest satellite fire spot monitoring method of claim 4, wherein, Determining a potential fire source area according to the classification feature vector comprises the following steps: For the pixel points divided into the smoke area and the high-temperature area, calculating the matching degree of the smoke diffusion features and the high-temperature radiation features, and determining the potential fire source area if the matching degree is higher than a preset value. 6.The deep learning-based forest satellite fire spot monitoring method of claim 1, wherein, Obtaining a second potential fire source area set by using a support vector machine model according to the environmental data and the preprocessed multispectral image comprises the following steps: Obtaining terrain occlusion data of the target area, and performing occlusion correction processing on the band features if the occlusion angle is greater than a critical angle value; Inputting the corrected band features into a support vector machine model to obtain the second potential fire source area set, wherein the support vector machine model establishes a nonlinear mapping relationship between the band features and the fire source probability. 7.The deep learning-based forest satellite fire spot monitoring method of claim 1, wherein, Mapping the real fire point pixel to geographical coordinates of the target forest area comprises: Mapping the real fire point pixel to a forest geographical coordinate system through a pixel mapping relationship; Using a preset coordinate conversion model, combined with geographical reference information of remote sensing images, determining geographical coordinates corresponding to the fire point pixel, and obtaining a fire point geographical coordinate set. 8.The deep learning-based forest satellite fire spot monitoring method of claim 7, wherein, Filling in missing data using spatial interpolation technology to obtain the final fire point position distribution comprises: Identifying missing data from the fire point geographical coordinate set, using spatial autocorrelation analysis to determine whether there is data missing caused by occlusion in the geographical coordinate point, and obtaining a missing data region; Using a Kriging interpolation algorithm, based on the fire point geographical coordinate set around the missing data region, calculating spatial interpolation weights, filling in missing data in the occluded region, and obtaining a completed fire point geographical coordinate set; If there are still abnormal points in the completed fire point geographical coordinate set, then using a distance weighted average method, combined with the spatial distribution characteristics of adjacent coordinate points, smoothing the abnormal points, and obtaining a smoothed fire point geographical coordinate set; Through a density clustering algorithm, performing spatial clustering analysis on the smoothed fire point geographical coordinate set to determine the dense area of the fire point position distribution, and obtaining the final fire point position distribution.

Images