Forestry Fire Monitoring Method Based on Super-Resolution Image Processing

Abstract

This invention relates to a forestry fire monitoring method based on super-resolution image processing, belonging to the technical field of forestry fire monitoring using super-resolution image processing. It includes the following steps: acquiring forestry fire monitoring data and preprocessing the data; constructing a forestry fire monitoring model based on super-resolution image processing, the model including a forestry fire feature capture backbone network, a blurred texture feature reconstruction module, a high-sensitivity color reconstruction module, a global feature aggregation module, and an output terminal; training the model using the preprocessed data; optimizing the model parameters through a loss function during training to obtain a trained model; and inputting the image to be detected into the trained model to obtain the forestry fire monitoring result. This invention can enhance the ability to capture pixel features at image edges, suppress edge blurring in reconstructed images, and reduce false alarms caused by complex background interference.

Description

Technical Field

[0001] This invention belongs to the field of forestry fire monitoring technology based on super-resolution image processing, and specifically relates to a forestry fire monitoring method based on super-resolution image processing. Background Technology

[0002] Currently, forest fire monitoring mainly employs target detection technology and image semantic segmentation networks. By analyzing infrared and visible light images collected by satellites / UAVs, it can achieve fire location, smoke identification, and prediction of fire spread. However, this type of technology has shortcomings. For example, forest fires are small in the early stages, with flames and smoke accounting for a very small percentage of the total image pixels. Existing models lack the ability to extract features from small targets, resulting in a high false negative rate. Furthermore, forests contain interference such as mountain fog, clouds, and sunset sunlight. These interferences in complex backgrounds are highly similar to the visual features of flames and smoke, leading to a surge in false alarms.

[0003] Super-resolution forestry fire monitoring technology upscales low-resolution images to high-resolution images using image reconstruction networks such as ESRGAN and RCAN. Simultaneously, it reconstructs initial fire points from the original low-resolution images to a resolvable scale, enhancing the sensitivity of the monitoring model to fire points, smoke, and other fire characteristics. This effectively addresses the problem of missed features caused by the limited receptive field in traditional target monitoring models. During super-resolution image processing, methods such as PCD alignment are used to reconstruct high-resolution images from multimodal low-resolution data (visible light, thermal infrared, and shortwave infrared) while preserving their spectral characteristics, generating physically consistent forestry fire reconstruction images, which helps improve the accuracy of forestry fire monitoring models.

[0004] However, there are still some problems with the application of super-resolution image processing technology in the field of forestry fire image monitoring. The forestry fire images reconstructed by mainstream image reconstruction perceptual networks such as RCAN and Transformer will show distortion. In particular, the presence of semi-transparent smoke textures and strong color saturation of flames in the scene will lead to poor reconstruction ability of super-resolution image processing methods for key features of smoke background and blurry flame edges, thus interfering with the detection of relevant key features in real forestry fire images by monitoring models.

[0005] Therefore, the present invention provides a forestry fire monitoring method based on super-resolution image processing to solve the above problems. Summary of the Invention

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] This invention provides a forestry fire monitoring method based on super-resolution image processing, comprising the following steps:

[0008] Acquire forest fire monitoring data and preprocess the data;

[0009] A forestry fire monitoring model based on super-resolution image processing is constructed. The forestry fire monitoring model includes a forestry fire feature capture backbone network, a fuzzy texture feature reconstruction module, a high-light color reconstruction module, a global feature aggregation module, and an output terminal. The model is trained using preprocessed data.

[0010] During training, the model parameters are optimized using a loss function to obtain a well-trained model;

[0011] The image to be detected is input into the trained model to obtain the forest fire monitoring results.

[0012] Furthermore, in the forestry fire feature capture backbone network, the low-resolution images of forestry fires... After shallow feature extraction, the first feature map is obtained. ; the first feature map go through Layer normalization operation yields the second feature map. The second feature map After the SplitReshape operation, the second feature map is... Divided into The size of the sub-features is used to obtain the first sub-feature map. Second sub-feature map Third sub-feature map Fourth sub-feature map Traditional convolutional rectangular windows cannot effectively capture edge pixel features in an image, leading to blurred or distorted edges in the subsequent reconstructed image. Therefore, a rectangular sliding window design is used for feature capture. The rectangular sliding window is used to calculate the correlation between sub-features using attention. The first sub-feature map... Second sub-feature map The formula for calculating the correlation between them is as follows:

[0013] ,

[0014] ,

[0015] in, , , These represent the first query vector, the first key vector, and the first value vector, respectively. , , These represent the first learnable projection matrix, the second learnable projection matrix, and the third learnable projection matrix, respectively. Represents the first attention feature map; This represents the softmax activation function; Indicates the scaling factor; This represents the characteristic matrix multiplication operation; similarly, it calculates... , The correlation between them yields the second attention feature map. ;calculate The correlation between the two is used to obtain the third attention feature map. ;calculate , The correlation between them yields the fourth attention feature map. The three attention feature maps are concatenated and fused to obtain the third feature map. .

[0016] Furthermore, in the forestry fire feature capture backbone network, the first feature map go through Global average pooling operation yields pooled feature maps. The pooling feature map After convolution kernel size is Depth separable convolutional layers The activation function yields the first pooling feature map. The first pooling feature map After convolution kernel size is Depth-separable convolutional layers and The activation function is used to obtain the second pooling feature map. The second pooling feature map With the third feature map Perform element-wise addition and output the main features. .

[0017] Furthermore, in the fuzzy texture feature reconstruction module, the backbone features go through Batch normalization layer, to obtain the fourth feature map Fourth feature map After convolution kernel size is The convolution kernel is used to obtain the first convolutional feature map. The fourth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the second convolutional feature map. The second convolutional feature map With the fourth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the third convolutional feature map. ;

[0018] The fourth feature map With the third convolution feature map After element-by-element addition, the fifth feature map is obtained. The fifth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the fourth convolutional feature map. The fourth convolutional feature map With the fifth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the fifth convolutional feature map. ;

[0019] The fourth feature map With the fifth convolution feature map After element-by-element addition, the sixth feature map is obtained. The sixth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the sixth convolutional feature map. The sixth convolutional feature map With the sixth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the seventh convolutional feature map. .

[0020] Furthermore, in the blurred texture feature reconstruction module, the first convolutional feature map is... Third convolution feature map Fifth convolution feature map 7th convolution feature map Perform feature concatenation to obtain the seventh feature map. The seventh feature map After convolution kernel size is convolution kernel, Batch normalization layer and The activation function yields the eighth convolutional feature. The eighth convolutional feature With the fourth feature map Perform element-wise addition and then... Layer normalization operation, Activation function to obtain blurred texture reconstruction features .

[0021] Furthermore, in the high-light-sensitivity color reconstruction module, the blurred texture reconstruction feature After convolution kernel size is Pointwise convolution operation, Batch normalization layer and The activation function yields the first reconstructed feature map. The first reconstructed feature map After convolution kernel size is The pointwise convolution operation yields the second reconstructed feature map. The second reconstructed feature map go through The operation uniformly divides it into first reconstructed sub-features along the channel dimension. Second Reconstruction Sub-features and the third reconstructed features .

[0022] Furthermore, in the high-light-sensitivity color reconstruction module, the first reconstructed sub-feature After convolution kernel size is The depth of the separable convolutional layer is used to obtain the first convolutional reconstruction sub-feature map. The first convolutional reconstruction of the sub-feature map After convolution kernel size is The dilated convolutional layer yields the second convolutional reconstructed sub-feature map. The second convolutional reconstruction sub-feature map After convolution kernel size is The pointwise convolutional layers yield the third convolutional reconstructed sub-feature map. The first reconstructed sub-feature After convolution kernel size is The depthwise separable convolutional layer is processed and then combined with the third convolutional layer to reconstruct the sub-feature map. Perform element-wise addition to obtain the third reconstructed feature map. Similarly, the second reconstructed sub-feature Execution as the first reconstruction sub-feature The same process yields the fourth reconstructed feature map. The third reconstructed feature Execution as the first reconstruction sub-feature The same process yields the fifth reconstructed feature map. ;

[0023] The third reconstruction feature map Fourth Reconstruction Feature Map and the fifth reconstruction feature map go through The feature concatenation operation yields the eighth feature map. The eighth feature map Compared with the first reconstructed feature map After element-wise addition and convolution with a kernel size of [missing value], Pointwise convolutional layers Activation function to obtain the high-sensitivity color reconstruction feature map .

[0024] Furthermore, in the global feature aggregation module, the fuzzy texture reconstruction features After convolution kernel size is Depth separable convolutional layers Batch grouping into one layer and Activation function, output convolutional reconstruction feature map The high-sensitivity color reconstruction feature map After convolution kernel size is Depth separable convolutional layers Batch grouping into one layer and Activation function to obtain convolutional recombination feature map The convolutional reconstruction feature map and convolutional recombination feature maps The fifth attention feature map is obtained by fusing the features through a self-attention mechanism. The fifth attention feature map go through Global average pooling operation, kernel size is Convolutional layers and Activation function to obtain global aggregated features .

[0025] Furthermore, in the output, the global aggregation feature After convolution kernel size is Convolutional layers Activation function and Fully connected layer, outputting forest fire image monitoring results. .

[0026] Furthermore, the loss function adopts a variant of the binary cross-entropy loss function combined with the sigmoid activation function: the BCEWithLogitsLoss loss function, as an optimization method for model training; this avoids the mandatory constraint of label mutual exclusion imposed by the traditional multi-class cross-entropy loss.

[0027] The advantages of this invention are:

[0028] This invention proposes a forestry fire monitoring method based on super-resolution image processing. By constructing the FFMT-Net model, it effectively solves the problems of poor reconstruction ability of translucent smoke texture, blurred flame edges, and color distortion in existing super-resolution technologies for forestry fire monitoring. Specifically, the feature capture backbone network employs a rectangular sliding window attention mechanism to enhance the capture ability of pixel features at image edges, significantly suppressing edge blurring in the reconstructed image. The blurred texture feature reconstruction module and the high-sensitivity color reconstruction module improve the clarity reconstruction ability of smoke texture and the fidelity of flame color saturation through multi-scale deep separable convolution and feature fusion strategies, respectively, effectively reducing false alarms caused by complex background interference. The global feature aggregation module achieves deep collaboration between texture and color features. Experimental results show that the method achieves an overall accuracy, precision, and F1 score of 93.67%, 92.30%, and 87.91% on the Forest Fire dataset, significantly improving the accuracy and reliability of early fire identification. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0030] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0031] Figure 2 The results of a comparative experiment on forest fire monitoring are presented in this invention.

[0032] Figure 3 The image shows a forest fire image processed by the method of this invention compared with the original image. Detailed Implementation

[0033] 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.

[0034] Example 1

[0035] In this embodiment, as Figure 1 As shown, this invention provides a forestry fire monitoring method based on super-resolution image processing, the specific steps of which include:

[0036] S1. Acquire forest fire monitoring data and preprocess the data: The specific construction and processing of the dataset are as follows: During model training, image data from the publicly available Forest Fire database is selected, and images containing fire, smoke, and normal forest scenes are collected and summarized. First, using the OpenCV image processing library, all images are uniformly cropped or scaled to the fixed size required by the model to ensure input consistency. Second, to improve the model's generalization ability and expand the data volume, various data augmentation techniques are applied to the training set images, including random horizontal / vertical flipping and brightness / contrast fine-tuning. After data augmentation, the entire augmented dataset is randomly divided into training, testing, and validation sets in a ratio of 7:2:1 to ensure data distribution consistency. Finally, the batch size is set to 128 during training, testing, and validation.

[0037] S2. Construct a forestry fire monitoring model based on super-resolution image processing (FFMT-Net model). The forestry fire monitoring model includes a forestry fire feature capture backbone network, a fuzzy texture feature reconstruction module, a high-light color reconstruction module, a global feature aggregation module, and an output terminal. The model is trained using preprocessed data. The proposed forestry fire monitoring method uses the designed FFMT-Net model to monitor forestry fires.

[0038] The specific construction steps of the proposed model are as follows:

[0039] The proposed FFMT-Net forest fire monitoring model uses a batch size of 128 during training. The learning rate is dynamically set with an initial learning rate of 0.001. Every 20 iterations, the learning rate is reduced to half of the original learning rate. The number of training iterations is set to 256.

[0040] The specific construction method of the forest fire feature capture backbone network is as follows: First, low-resolution images of forest fires are... The input to the convolution kernel size is Convolutional layers and batch normalization layers as well as A shallow feature extraction is performed using a nonlinear activation function to obtain the first feature map. The first feature map Spatial resolution ( The number of channels remains unchanged, but the number of channels becomes 128. The calculation formula involved in this process is:

[0041] ,

[0042] in, Indicates the kernel size as Convolution operations; Indicates the batch normalization layer; express Nonlinear activation function; the first feature map go through Layer normalization operation yields the second feature map. The second feature map After the SplitReshape operation, the second feature map is... Divided into The size of the sub-features is used to obtain the first sub-feature map. Second sub-feature map Third sub-feature map Fourth sub-feature map Traditional convolutional rectangular windows cannot effectively capture edge pixel features in an image, leading to blurred or distorted edges in the subsequent reconstructed image. Therefore, a rectangular sliding window design is used for feature capture. The rectangular sliding window is used to calculate the correlation between sub-features using attention. The first sub-feature map... Second sub-feature map The formula for calculating the correlation between them is as follows:

[0043] ,

[0044] ,

[0045] in, , , These represent the first query vector, the first key vector, and the first value vector, respectively. , , These represent the first learnable projection matrix, the second learnable projection matrix, and the third learnable projection matrix, respectively. Represents the first attention feature map; This represents the softmax activation function; Indicates the scaling factor; This represents the characteristic matrix multiplication operation; similarly, it calculates... , The correlation between them yields the second attention feature map. ;calculate The correlation between the two is used to obtain the third attention feature map. ;calculate , The correlation between them yields the fourth attention feature map. The three attention feature maps are concatenated and fused to obtain the third feature map. The formula is expressed as follows:

[0046] ,

[0047] in, Indicates feature concatenation operation;

[0048] First feature map go through Global average pooling operation yields pooled feature maps. The pooling feature map After convolution kernel size is Depth separable convolutional layers The activation function yields the first pooling feature map. The first pooling feature map After convolution kernel size is Depth-separable convolutional layers and The activation function is used to obtain the second pooling feature map. The second pooling feature map With the third feature map Perform element-wise addition and output the main features. The formula for the above process is expressed as follows:

[0049] ,

[0050] ,

[0051] in, This represents an element-wise addition operation; express Activation function; express The depth of the separable convolutional layer; This represents the GeLU activation function; express The depth of the separable convolutional layer; This indicates a global average pooling operation.

[0052] The fuzzy texture feature reconstruction module is constructed as follows: the backbone features go through Batch normalization layer, to obtain the fourth feature map Fourth feature map After convolution kernel size is The convolution kernel is used to obtain the first convolutional feature map. The fourth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the second convolutional feature map. The second convolutional feature map With the fourth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the third convolutional feature map. The formula for the above process is expressed as follows:

[0053] ,

[0054] ,

[0055] ,

[0056] in, Indicates the kernel size as Pointwise convolution operation;

[0057] The fourth feature map With the third convolution feature map After element-by-element addition, the fifth feature map is obtained. The fifth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the fourth convolutional feature map. The fourth convolutional feature map With the fifth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the fifth convolutional feature map. The formula for the above process is expressed as follows:

[0058] ,

[0059] ,

[0060] ,

[0061] in, Indicates the kernel size as Pointwise convolution operation; Indicates the kernel size as Depth-separable convolution operations;

[0062] The fourth feature map With the fifth convolution feature map After element-by-element addition, the sixth feature map is obtained. The sixth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the sixth convolutional feature map. The sixth convolutional feature map With the sixth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the seventh convolutional feature map. The formula for the above process is expressed as follows:

[0063] ,

[0064] ,

[0065] ,

[0066] in, Indicates the kernel size as Pointwise convolution operation; Indicates the kernel size as Depth-separable convolution operations;

[0067] The first convolutional feature map Third convolution feature map Fifth convolution feature map 7th convolution feature map Perform feature concatenation to obtain the seventh feature map. The seventh feature map After convolution kernel size is convolution kernel, Batch normalization layer and The activation function yields the eighth convolutional feature. The eighth convolutional feature With the fourth feature map Perform element-wise addition and then... Layer normalization operation, Activation function to obtain blurred texture reconstruction features The formula for the above process is expressed as follows:

[0068] ,

[0069] ,

[0070] ,

[0071] in, Presentation layer normalization operation.

[0072] The specific construction method of the high-gloss color reconstruction module is as follows: the blurred texture reconstruction feature After convolution kernel size is Pointwise convolution operation, Batch normalization layer and The activation function yields the first reconstructed feature map. The first reconstructed feature map After convolution kernel size is The pointwise convolution operation yields the second reconstructed feature map. The second reconstructed feature map go through The operation uniformly divides it into first reconstructed sub-features along the channel dimension. Second Reconstruction Sub-features and the third reconstructed features The formula for the above process is expressed as follows:

[0073] ,

[0074] ,

[0075] ,

[0076] in, Indicates the kernel size as The pointwise convolution operation.

[0077] First Reconstruction Sub-feature After convolution kernel size is The depth of the separable convolutional layer is used to obtain the first convolutional reconstruction sub-feature map. The first convolutional reconstruction of the sub-feature map After convolution kernel size is The dilated convolutional layer yields the second convolutional reconstructed sub-feature map. The second convolutional reconstruction sub-feature map After convolution kernel size is The pointwise convolutional layers yield the third convolutional reconstructed sub-feature map. The first reconstructed sub-feature After convolution kernel size is The depthwise separable convolutional layer is processed and then combined with the third convolutional layer to reconstruct the sub-feature map. Perform element-wise addition to obtain the third reconstructed feature map. The formula for the above process is expressed as follows:

[0078] ,

[0079] ,

[0080] in, Indicates the kernel size as The depth of the separable convolutional layer; Indicates the kernel size as Hollow convolutional layers; Indicates the kernel size as Pointwise convolutional layers;

[0081] Similarly, the second reconstructed sub-feature Execution as the first reconstruction sub-feature The same process yields the fourth reconstructed feature map. The third reconstructed feature Execution as the first reconstruction sub-feature The same process yields the fifth reconstructed feature map. Specifically, sub-features After convolution kernel size is The depth of the separable convolutional layer is used to obtain the feature map. ; feature map After convolution kernel size is The feature map is obtained from the dilated convolutional layer. ; feature map After convolution kernel size is The pointwise convolutional layers yield feature maps. Sub-features After convolution kernel size is The depth of the separable convolutional layer yields feature maps and feature maps. Perform element-wise addition to obtain the feature map. The calculation formulas involved in the above process are as follows:

[0082] ,

[0083] ,

[0084] Sub-features After convolution kernel size is The depth of the separable convolutional layer is used to obtain the feature map. ; feature map After convolution kernel size is The feature map is obtained from the dilated convolutional layer. ; feature map After convolution kernel size is The pointwise convolutional layers yield feature maps. Sub-features After convolution kernel size is The depth of the separable convolutional layer yields feature maps and feature maps. Perform element-wise addition to obtain the feature map. The calculation formulas involved in the above process are as follows:

[0085] ,

[0086] ,

[0087] The third reconstruction feature map Fourth Reconstruction Feature Map and the fifth reconstruction feature map go through The feature concatenation operation yields the eighth feature map. The eighth feature map Compared with the first reconstructed feature map After element-wise addition and convolution with a kernel size of [missing value], Pointwise convolutional layers Activation function to obtain the high-sensitivity color reconstruction feature map The formula for the above process is expressed as follows:

[0088] ,

[0089] ,

[0090] in, Indicates the kernel size as The pointwise convolutional layer.

[0091] The specific construction process of the global feature aggregation module is as follows: the fuzzy texture reconstruction features After convolution kernel size is Depth separable convolutional layers Batch grouping into one layer and Activation function, output convolutional reconstruction feature map The high-sensitivity color reconstruction feature map After convolution kernel size is Depth separable convolutional layers Batch grouping into one layer and Activation function to obtain convolutional recombination feature map ;

[0092] The convolutional reconstruction feature map and convolutional recombination feature maps The fifth attention feature map is obtained by fusing the features through a self-attention mechanism. The query in the attention mechanism Weights learned through backpropagation during model training AND feature map Feature matrix multiplication operation, key in attention mechanism Sum Weights learned through backpropagation during model training and The matrix multiplication process ultimately uses The activation function computes the attention feature map of the fused features. The fifth attention feature map go through Global average pooling operation, kernel size is Convolutional layers and Activation function to obtain global aggregated features The formula for the above process is expressed as follows:

[0093] ,

[0094] ,

[0095] ,

[0096] ,

[0097] ,

[0098] in, , , This represents the second query vector, the second key vector, and the second value vector; , , These represent the fourth, fifth, and sixth learnable projection matrices, respectively.

[0099] The specific construction method of the output end is as follows: the global aggregation feature After convolution kernel size is Convolutional layers Activation function and Fully connected layer, outputting forest fire image monitoring results. .

[0100] S3. During training, the model parameters are optimized using a loss function to obtain a well-trained model.

[0101] The method proposed in this invention employs a variant of the binary cross-entropy loss function combined with the Sigmoid activation function, namely the BCEWithLogitsLoss loss function, as an optimization method for model training. By calculating the cross-entropy between the predicted probability of each label and the true label (0 or 1) separately, and finally averaging the loss of all labels, it avoids the mandatory constraint of label mutual exclusivity imposed by the traditional multi-class cross-entropy loss.

[0102] S4. Input the image to be detected into the trained model to obtain the forest fire monitoring results.

[0103] Example 2

[0104] In this embodiment, to verify the effectiveness of the forestry fire monitoring method based on super-resolution image processing proposed in this invention in forestry fire monitoring, a comparative experiment was conducted between the method proposed in this invention and existing forestry fire monitoring methods. The results of the image monitoring comparison experiment demonstrate the effectiveness of the proposed forestry fire monitoring method based on super-resolution image processing.

[0105] In the comparative experiments, three existing forestry fire monitoring methods in this field were used as benchmark models to compare with the proposed forestry fire monitoring method based on super-resolution image processing. ResNet-108, Vision Transformer (ViT), and Swin-Transformer were selected as benchmark models. ResNet-108 utilizes a residual connection structure, making it stable, reliable, and relatively easy to optimize in image feature extraction; however, its receptive field is relatively limited. ViT breaks the locality constraints of traditional CNNs through a global self-attention mechanism, significantly improving the model's understanding of global image information; however, its computational complexity is high. Swin-Transformer introduces a shift window mechanism, which, while retaining the advantages of global modeling of Transformer, significantly reduces computational complexity and achieves hierarchical feature representation through local window calculation, enabling it to process high-resolution images more efficiently; however, this model requires a large number of training samples.

[0106] In the comparative experiments, this invention selected overall accuracy (Acc), precision (Pre), and F1 score (F1-score) as relevant indicators for evaluating the experiments: Accuracy (Acc) reflects the overall proportion of the model correctly classifying all samples (including fire images and non-fire images), and is a direct reflection of the model's basic performance; Precision (Pre) focuses on the results predicted by the model as "fire", calculating the proportion of real fires; while F1 score (F1-score) is the harmonic mean of precision (Pre) and recall (i.e., the proportion of real fires successfully detected by the model). This indicator comprehensively balances the model's ability to avoid false positives and false negatives. F1 score provides the most comprehensive single performance metric and is the core indicator for judging the practicality of the model.

[0107] In the comparative experiment of forest fire monitoring methods based on super-resolution image processing, the experimental results are shown in Table 1. The results show that the proposed method outperforms the three benchmark models: the overall accuracy of the proposed method is 93.67%, the precision is 92.30%, and the F1 score is 87.91%, which is 3.54%, 3.18%, and 2.83% higher than the best benchmark model Swin-Transformer (Acc 90.13%, Pre 89.12%, F1 85.08%), respectively. Specifically, the traditional convolutional network ResNet-108 performed the weakest due to its limited feature extraction capabilities (Accuracy Acc 78.22%, Precision Pre 76.62%, F1 score 74.73%). The proposed method enhances the fine-grained features of forest fire images through super-resolution image processing technology, comprehensively improving the overall performance while maintaining high accuracy. It significantly reduces the risk of false fire alarms and improves the reliability of early fire identification, fully verifying the technical advantages and practical value of the proposed method.

[0108] Table 1 shows the comparison results of validating the proposed method on the dataset.

[0109]

[0110] In the comparative experiment of the proposed forestry fire monitoring method based on super-resolution image processing, the comparative experimental results of forestry fire monitoring are as follows: Figure 2As shown, real-time monitoring of forest fires is often accompanied by dense smoke and fire. Therefore, this invention monitors smoke and fire in comparative experiments. Experimental results show that when forest fires are small and accompanied by smoke, the proposed forest fire monitoring method based on super-resolution image processing achieves the best results compared to the three benchmark models. Specifically, in the case of a shrub-grassland fire in scenario one, where the flames are small and the smoke is light, the proposed method identifies an 88% probability of a fire and a 96% probability of smoke generation in the image, which is an improvement of 13% and 9% respectively compared to the best benchmark model, the Swin-Transformer. In forest fire identification scenarios where trees obscure the flames, the proposed monitoring method identifies an 86% probability of a fire and a 94% probability of smoke generation in the image, which is also the best result among the three benchmark models. Figure 3 The image is compared with the original image after being processed by the proposed forestry fire monitoring method based on super-resolution image processing. The comparison shows that the specific features such as smoke and flames are clearer in the processed forestry fire image compared with the original image.

[0111] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are 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 forestry fire monitoring method based on super-resolution image processing, characterized in that, Includes the following steps: Acquire forest fire monitoring data and preprocess the data; A forestry fire monitoring model based on super-resolution image processing is constructed. The forestry fire monitoring model includes a forestry fire feature capture backbone network, a fuzzy texture feature reconstruction module, a high-light color reconstruction module, a global feature aggregation module, and an output terminal. The model is trained using preprocessed data; the specific steps include: In the forestry fire feature capture backbone network, low-resolution images of forestry fires The input to the convolution kernel size is Convolutional layers and batch normalization layers as well as A shallow feature extraction is performed using a nonlinear activation function to obtain the first feature map. ; the first feature map go through Layer normalization operation yields the second feature map. The second feature map After the SplitReshape operation, the second feature map is... Divided into The size of the sub-features is used to obtain the first sub-feature map. Second sub-feature map Third sub-feature map Fourth sub-feature map The correlation between sub-features is calculated using a rectangular sliding window attention algorithm. (First sub-feature map) Second sub-feature map The formula for calculating the correlation between them is as follows: ， ， in, , , These represent the first query vector, the first key vector, and the first value vector, respectively. , , These represent the first learnable projection matrix, the second learnable projection matrix, and the third learnable projection matrix, respectively. Represents the first attention feature map; This represents the softmax activation function; Indicates the scaling factor; This represents the characteristic matrix multiplication operation; similarly, it calculates... , The correlation between them yields the second attention feature map. ;calculate The correlation between the two is used to obtain the third attention feature map. ;calculate , The correlation between them yields the fourth attention feature map. The three attention feature maps are concatenated and fused to obtain the third feature map. The first feature map After global average pooling, the pooled feature map is obtained. The pooling feature map After convolution kernel size is Depth separable convolutional layers The activation function yields the first pooling feature map. The first pooling feature map After convolution kernel size is Depth-separable convolutional layers and The activation function is used to obtain the second pooling feature map. The second pooling feature map With the third feature map Perform element-wise addition and output the main features. ; In the fuzzy texture feature reconstruction module, the backbone features After the batch normalization layer, the fourth feature map is obtained. Fourth feature map After convolution kernel size is The convolution operation yields the first convolutional feature map. The fourth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the second convolutional feature map. The second convolutional feature map With the fourth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the third convolutional feature map. The fourth feature map With the third convolution feature map After element-by-element addition, the fifth feature map is obtained. The fifth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the fourth convolutional feature map. The fourth convolutional feature map With the fifth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the fifth convolutional feature map. The fourth feature map With the fifth convolution feature map After element-by-element addition, the sixth feature map is obtained. The sixth feature map After convolution kernel size is The pointwise convolution operation, with a kernel size of The depthwise separable convolution operation yields the sixth convolutional feature map. The sixth convolutional feature map With the sixth feature map After element-by-element addition and then The activation function and kernel size are The convolutional layer yields the seventh convolutional feature map. ; the first convolutional feature map Third convolution feature map Fifth convolution feature map 7th convolution feature map Perform feature concatenation to obtain the seventh feature map. The seventh feature map After convolution kernel size is convolution kernel, Batch normalization layer and The activation function yields the eighth convolutional feature. The eighth convolutional feature With the fourth feature map Perform element-wise addition and then perform layer normalization. The activation function is used to process the data to obtain the blurred texture reconstruction features. ; In the high-light color reconstruction module, the blurred texture reconstruction feature After convolution kernel size is Pointwise convolution operation, Batch normalization layer and The activation function yields the first reconstructed feature map. The first reconstructed feature map After convolution kernel size is The pointwise convolution operation yields the second reconstructed feature map. The second reconstructed feature map go through The operation uniformly divides it into first reconstructed sub-features along the channel dimension. Second Reconstruction Sub-features and the third reconstructed sub-features The first reconstructed sub-feature After convolution kernel size is The depth of the separable convolutional layer is used to obtain the first convolutional reconstruction sub-feature map. The first convolutional reconstruction of the sub-feature map After convolution kernel size is The dilated convolutional layer yields the second convolutional reconstructed sub-feature map. The second convolutional reconstruction sub-feature map After convolution kernel size is The pointwise convolutional layers yield the third convolutional reconstructed sub-feature map. The first reconstructed sub-feature After convolution kernel size is The depthwise separable convolutional layer is processed and then combined with the third convolutional layer to reconstruct the sub-feature map. Perform element-wise addition to obtain the third reconstructed feature map. Similarly, the second reconstructed sub-feature Execution as the first reconstruction sub-feature The same process yields the fourth reconstructed feature map. The third reconstructed feature Execution as the first reconstruction sub-feature The same process yields the fifth reconstructed feature map. The third reconstructed feature map Fourth Reconstruction Feature Map and the fifth reconstruction feature map go through The feature concatenation operation yields the eighth feature map. The eighth feature map Compared with the first reconstructed feature map After element-wise addition and convolution with a kernel size of [missing value], Pointwise convolutional layers Activation function to obtain the high-sensitivity color reconstruction feature map ; In the global feature aggregation module, the fuzzy texture reconstruction features After convolution kernel size is Depth separable convolutional layers Batch grouping into one layer and Activation function, output convolutional reconstructed feature map The high-sensitivity color reconstruction feature map After convolution kernel size is Depth separable convolutional layers Batch grouping into one layer and Activation function to obtain convolutional recombination feature map The convolutional reconstruction feature map and convolutional recombination feature maps The fifth attention feature map is obtained by fusing the features through a self-attention mechanism. The fifth attention feature map After global average pooling, the kernel size is... Convolutional layers and Activation function to obtain global aggregated features ; During training, the model parameters are optimized using a loss function to obtain a well-trained model; The image to be detected is input into the trained model to obtain the forest fire monitoring results.

2. The forestry fire monitoring method based on super-resolution image processing according to claim 1, characterized in that, In the output, the global aggregation feature After convolution kernel size is Convolutional layers The activation function and fully connected layer output the forest fire image monitoring results. .

3. The forestry fire monitoring method based on super-resolution image processing according to claim 2, characterized in that, The loss function uses a variant of the binary cross-entropy loss function combined with the sigmoid activation function: BCEWithLogitsLoss loss function, as an optimization method for model training.

Images