Forest fire severity inversion method based on dual-temporal feature fusion and boundary constraint

Abstract

The application relates to a forest fire severity inversion method based on dual-time-phase feature fusion and boundary constraint. The method comprises the following steps: according to the upsampling turn of the decoding process, inputting target difference features and each candidate difference feature into a boundary perception decoding branch in sequence, guiding a severity regression decoding branch layer by layer through a cascaded knowledge guiding module, and outputting final regression decoding features and final boundary decoding features; inputting the final regression decoding features into an uncertainty perception and confidence module to output a differential normalized burning rate continuous numerical map and an uncertainty confidence map, and determining a fire spot edge map based on the final boundary decoding features. In the decoding stage, the severity regression branch is guided layer by layer, the candidate difference features and the target difference features of different scales are fully utilized, the spatial information loss in the downsampling process is relieved, and the adaptability to complex terrain scenes is improved.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a method for inverting the severity of forest fires by fusing dual temporal features and boundary constraints. Background Technology

[0002] With the continuous development of remote sensing observation technology and Earth monitoring capabilities, information acquisition methods centered on remote sensing technology have been widely applied in fields such as ecological environment monitoring, disaster assessment, and resource management. Among these, forest fires, as a typical sudden natural disaster, have a significant destructive effect on the structure and function of ecosystems. Severity assessment after a forest fire is crucial for formulating ecological restoration strategies, evaluating the impact of carbon cycles, and carrying out post-disaster management.

[0003] Traditional techniques use pixel-level classification to assess the severity of forest fires, outputting discrete severity categories, such as high, medium, or low severity for a particular forest area. However, this approach suffers from problems such as blurred fire boundaries and low accuracy in severity inversion. Summary of the Invention

[0004] Based on this, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for forest fire severity inversion that can improve the accuracy of severity inversion by fusing dual temporal features and boundary constraints.

[0005] Firstly, this application provides a method for inverting forest fire severity based on dual-temporal feature fusion and boundary constraints. The method includes: Acquire dual-temporal remote sensing images of the target forest area before and after the fire, and develop a forest fire severity prediction model; The dual-temporal remote sensing images include pre-disaster remote sensing images and post-disaster remote sensing images; the forest fire severity prediction model sequentially includes two coding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge guidance module, and an uncertainty perception and confidence module. According to the downsampling rounds of the encoding process, the pre-disaster remote sensing images and post-disaster remote sensing images are respectively input into each of the encoding branches, and at least two candidate difference features are output through each of the encoding branches, and the target difference features are output through the cross-attention change perception module. According to the upsampling rounds of the decoding process, the target difference features and each of the candidate difference features are sequentially input into the boundary-aware decoding branch. The severity regression decoding branch is guided layer by layer by the cascaded knowledge guidance module, and the final regression decoding features and the final boundary decoding features are output. The final regression decoding features are input into the uncertainty perception and confidence module, which outputs the differential normalized combustion rate continuous numerical map and uncertainty confidence map corresponding to the dual-temporal remote sensing image, and determines the fire spot edge map corresponding to the dual-temporal remote sensing image based on the final boundary decoding features.

[0006] Secondly, this application also provides a forest fire severity inversion device based on dual-temporal feature fusion and boundary constraints. The device includes: The acquisition module is used to acquire dual-temporal remote sensing images of the target forest area before and after the fire, as well as a forest fire severity prediction model. The dual-temporal remote sensing images include pre-disaster remote sensing images and post-disaster remote sensing images; the forest fire severity prediction model sequentially includes two coding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge guidance module, and an uncertainty perception and confidence module. The encoding module is used to input the pre-disaster remote sensing image and the post-disaster remote sensing image into each of the encoding branches according to the downsampling rounds of the encoding process, output at least two candidate difference features through each of the encoding branches, and output target difference features through the cross-attention change perception module. The decoding module is used to input the target difference feature and each candidate difference feature into the boundary-aware decoding branch in sequence according to the upsampling rounds of the decoding process, and guide the severity regression decoding branch layer by layer through the cascaded knowledge guidance module, and output the final regression decoding feature and the final boundary decoding feature. The prediction module is used to input the final regression decoding features into the uncertainty perception and confidence module, output the differential normalized combustion rate continuous numerical map and uncertainty confidence map corresponding to the dual-temporal remote sensing image, and determine the fire spot edge map corresponding to the dual-temporal remote sensing image based on the final boundary decoding features.

[0007] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above method steps.

[0008] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the above method steps.

[0009] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the above method steps.

[0010] The aforementioned method, apparatus, computer equipment, storage medium, and computer program product for forest fire severity inversion based on dual-temporal feature fusion and boundary constraints acquire dual-temporal remote sensing images of the target forest area before and after a fire, as well as a forest fire severity prediction model. The dual-temporal remote sensing images include pre-fire and post-fire remote sensing images. The forest fire severity prediction model sequentially includes two encoding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge guidance module, and an uncertainty perception and confidence module.

[0011] According to the downsampling rounds of the coding process, pre-disaster and post-disaster remote sensing images are input into each coding branch. Each coding branch outputs at least two candidate difference features, and the target difference feature is output through the cross-attention change perception module. By extracting multi-level features from pre-disaster and post-disaster remote sensing images through dual coding branches and combining the target difference features output by the cross-attention change perception module, the changes in land cover before and after the fire can be captured more accurately, thereby improving the quantitative inversion accuracy of the differentially normalized burn rate continuous numerical map.

[0012] Following the upsampling rounds of the decoding process, the target difference features and each candidate difference feature are sequentially input into the boundary-aware decoding branch. The severity regression decoding branch is then guided layer by layer by a cascaded knowledge-guided module, outputting the final regression decoding features and the final boundary decoding features. The boundary-aware decoding branch, in conjunction with the cascaded knowledge-guided module, enables the model to focus on flame edge information while regressing severity, ultimately outputting a flame edge map. This effectively overcomes the problems of blurred boundaries and discontinuous segmentation in traditional methods, thereby improving the accuracy of severity inversion. The final regression decoding features are input into the uncertainty perception and confidence module, which outputs a differentially normalized continuous numerical map of the combustion rate and an uncertainty confidence map corresponding to the dual-temporal remote sensing images. Based on the final boundary decoding features, the flame edge map corresponding to the dual-temporal remote sensing images is determined.

[0013] This method employs an uncertainty perception and confidence module that outputs an uncertainty confidence map. On one hand, this map provides users with the reliability of the prediction results for each pixel, assisting decision-makers in identifying high-risk or ambiguous prediction areas and further improving the accuracy of severity inversion. On the other hand, the cascaded knowledge guidance module guides the severity regression branch layer by layer during the decoding stage, fully utilizing candidate and target difference features at different scales to mitigate spatial information loss during downsampling and enhance adaptability to complex terrain scenarios. Attached Figure Description

[0014] Figure 1This is a flowchart illustrating a forest fire severity inversion method based on dual-temporal feature fusion and boundary constraints in one embodiment. Figure 2 This is a flowchart illustrating a method for determining differential features in one embodiment; Figure 3 This is a diagram of the architecture of a dual-temporal multi-task fire area identification network model in one embodiment; Figure 4 This is a schematic diagram of the dual-temporal cross-attention change sensing module structure in one embodiment; Figure 4 (a) is a schematic diagram of the data flow between modules in the dual-temporal cross-attention change perception module; Figure 4 (b) is a schematic diagram of the edge prior module; Figure 4 (c) is a schematic diagram of the fusion gate structure; Figure 4 (d) is a schematic diagram of the prior encoder; Figure 4 (e) is a schematic diagram of the edge coding structure; Figure 5 This is a schematic diagram of the structure of a multi-scale gated cascaded knowledge guidance module in one embodiment; Figure 5 (a) is a schematic diagram of the data flow between modules in the multi-scale dual-gated cascaded knowledge guidance module; Figure 5 (b) is a schematic diagram of the boundary prior module; Figure 5 (c) is a schematic diagram of the structure of local edge branches and void edge branches; Figure 5 (d) is a schematic diagram of the edge blending module and the detail projection module; Figure 5 (e) is a schematic diagram of the space gate structure; Figure 5 (f) is a schematic diagram of the passageway door; Figure 5 (g) is a schematic diagram of the fusion module; Figure 6 This is a structural diagram of the uncertainty perception and confidence module in one embodiment; Figure 6 (a) is a schematic diagram of the data flow in the uncertainty perception and confidence module; Figure 6 (b) is a schematic diagram of the structure of the convolution-batch normalization-activation module; Figure 6 (c) is a schematic diagram of the compression-excitation module; Figure 7 A visualization of different models on a test set in one embodiment; Figure 7 (a) is the evaluation result diagram of the semantic change detection method; Figure 7 (b) is an evaluation effect diagram corresponding to the farmland boundary delineation method; Figure 7 (c) shows the evaluation results of the very high resolution image change detection method; Figure 7 (d) shows the evaluation results for image change detection; Figure 7(e) and Figure 7 (f) is an evaluation effect diagram corresponding to the method in this embodiment; Figure 8 This is a structural block diagram of a forest fire severity inversion device with dual-temporal feature fusion and boundary constraints in one embodiment. Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] In one embodiment, such as Figure 1 As shown, a method for forest fire severity inversion based on dual-temporal feature fusion and boundary constraints is provided, including: S101, acquire dual-temporal remote sensing images of the target forest area before and after the fire, as well as a forest fire severity prediction model.

[0017] The dual-temporal remote sensing images include pre-disaster and post-disaster images. The forest fire severity prediction model consists of two coding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge-guided module, and an uncertainty perception and confidence module.

[0018] The target forest area can be any area within the forest, and there are no restrictions here.

[0019] Dual-temporal remote sensing images can be a pair of images of the same area, including remote sensing images before and after the fire.

[0020] Pre-disaster remote sensing images are images acquired before a fire occurs, when the ground surface is under normal vegetation cover. Characteristics of pre-disaster remote sensing images may include: high reflectivity in the near-infrared band; absence of burned areas, carbonization, or withered vegetation; and spectral characteristics of land features such as roads, water bodies, and bare soil consistent with normal conditions.

[0021] Post-disaster remote sensing images are images acquired after a fire occurs when the ground surface is in an abnormal state. Characteristics of post-disaster remote sensing images may include: a significant decrease in near-infrared reflectance and an increase in short-wave infrared reflectance (charring and exposed soil); partial burning and gradual transition pixels in the boundary areas of fire spots; and possible interference such as high-temperature anomalies in the thermal infrared band (fire spot pixels) and smoke obscuring the images.

[0022] The forest fire severity prediction model can be a prediction model trained separately for the fire severity inversion task. This prediction model can be used to invert the degree of damage to the forest area corresponding to the input remote sensing image, obtain the specific severity value, and the smooth fire spot edge map, so as to facilitate the formulation of ecological restoration strategies, assessment of carbon cycle impact, and post-disaster management.

[0023] The forest fire severity prediction model consists of two encoding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge guidance module, and an uncertainty perception and confidence module.

[0024] The coding branches utilize spatial-spectral features, difference enhancement features, or temporal context from remote sensing images before and after the fire. For example, a Siamese network structure can be used to directly compare the differences between two temporal phases. By setting two coding branches, asymmetric features can be incorporated, with one branch focusing on pre-fire land cover types (such as forest type and topography) and the other on post-fire spectral changes (such as carbonization index and vegetation mortality). Furthermore, the two coding branches can be used to extract distinguishing features before and after the fire, more accurately capturing changes in land cover before and after the fire and improving the accuracy of subsequent predictions.

[0025] The Edge Decoder branch can be used to extract the boundary structure of a fire zone.

[0026] The severity regression decoder branch can be used to output continuous values, such as differentially normalized burn rate continuous numerical plots.

[0027] The Bi-Temporal Cross-Attention Fusion (BCAF) module can be used to establish long-distance dependencies between two time phases and automatically focus on the areas where changes are most dramatic before and after a fire.

[0028] The Cascaded Knowledge-Guided Module (CKGM) can be used to process the information output from the boundary branch and then input it into the severity regression branch, while preventing gradient contamination of the boundary branch by the regression task.

[0029] The Uncertainty-Aware Confidence Guidance (UACG) module can be used to estimate the reliability of each prediction and output an uncertainty graph.

[0030] The dual-temporal images are input into specific severity numerical maps, confidence maps, and fire spot edge maps, which are then output to form a complete end-to-end deep learning model. No manual feature design or post-processing steps are required, making it easy to deploy automatically in fire emergency monitoring systems.

[0031] S102, according to the downsampling rounds of the encoding process, the pre-disaster remote sensing images and post-disaster remote sensing images are input into each encoding branch respectively, and at least two candidate difference features are output through each encoding branch, and the target difference features are output through the cross-attention change perception module.

[0032] The downsampling round can represent the number of downsampling rounds in the coding branch. When the downsampling round is 3, a coding branch can be set with one basic convolutional module and three spatial-frequency joint visual state space modules.

[0033] Candidate difference features can be the output differences between the underlying convolutional modules of the two coding branches.

[0034] The target difference features can be the features output by the cross-attention change perception module.

[0035] Specifically, the process involves inputting pre-disaster and post-disaster remote sensing images into each coding branch, outputting the first coding feature through the basic convolutional module in each coding branch, inputting each first coding feature into the spatial-frequency joint visual state space module, and outputting the second, third, and fourth coding features of each coding branch sequentially according to the downsampling rounds of the coding process, inputting each fourth coding feature into the cross-attention change perception module to output the target difference feature, and determining candidate difference features based on the differences of each first coding feature, each second coding feature, and each third coding feature.

[0036] First, it captures spectral changes caused by fire at different scales, avoiding noise from direct differential processing. Second, the cross-attention module focuses on the regions that have actually changed, suppressing invariant interferences such as seasons and lighting, and outputting more discriminative target difference features, while retaining candidate difference features from each layer of the encoder for subsequent decoding.

[0037] S103, according to the upsampling rounds of the decoding process, the target difference features and each candidate difference feature are sequentially input into the boundary-aware decoding branch. The severity regression decoding branch is guided layer by layer through the cascaded knowledge guidance module, and the final regression decoding features and the final boundary decoding features are output.

[0038] The upsampling rounds represent the number of times the decoding branch (boundary-aware decoding branch and severity regression decoding branch) is upsampled in the decoding branch. When the upsampling rounds are 2, the boundary-aware decoding branch can be configured with 3 boundary-aware decoders, and the severity regression decoding branch can be configured with 3 severity regression decoders.

[0039] Specifically, when the upsampling count is 0, the boundary-aware decoding branch includes a boundary-aware decoder, and the severity regression decoding branch includes a severity regression decoder. Target differential features and candidate differential features can be input into the boundary-aware decoder to output the corresponding features. The cascaded knowledge-guided module can process these features and input the output features into the severity regression decoder in the severity regression decoding branch, outputting the final boundary decoding features.

[0040] For example, the target difference features and candidate difference features are concatenated through channels and then input into the boundary-aware decoding branch to output the final boundary decoding features; the final boundary decoding features are input into the cascaded knowledge guidance module to output the cascaded output features; the cascaded output features, target difference features, and candidate difference features are input into the severity regression decoding branch to output the final regression decoding features.

[0041] First, boundary features are injected layer by layer into the severity regression branch, causing the predicted dNBR values ​​to change steeply at the fire spot edges (clear boundaries) while remaining smooth inside the patch. Second, differential features from different downsampling layers are fused in rounds, preserving both high-level semantics (overall combustion intensity) and low-level details (irregular fire line contours), thus improving adaptability to complex terrain.

[0042] S104 inputs the final regression decoding features into the uncertainty perception and confidence module, outputs the differential normalized combustion rate continuous numerical map and uncertainty confidence map corresponding to the dual-temporal remote sensing images, and determines the fire spot edge map corresponding to the dual-temporal remote sensing images based on the final boundary decoding features.

[0043] Among them, the differentially normalized combustion rate continuous numerical map can be a continuous numerical index calculated based on pre-disaster and post-disaster remote sensing images, reflecting the severity of the fire in each pixel.

[0044] The uncertainty confidence map provides a confidence / uncertainty estimate for the differentially normalized combustion rate prediction for each pixel, characterizing the reliability of the model's prediction results for that location.

[0045] The fire edge map extracts a binary or probability map of the boundary contour of the burned area, distinguishing the transition zone between the burned and unburned areas.

[0046] First, the differentially normalized burn rate continuous numerical map provides a quantitative burn severity for each pixel (mapped to light / moderate / severe levels). Second, the uncertainty confidence map marks unreliable prediction areas (such as smoke obscuration and mixed pixels), enhancing decision-making credibility. Third, the fire spot edge map directly extracts the boundaries of burned areas, supporting burned area statistics and patch segmentation.

[0047] The above-mentioned method for forest fire severity inversion based on dual-temporal feature fusion and boundary constraints acquires dual-temporal remote sensing images of the target forest area before and after the fire, as well as a forest fire severity prediction model. The dual-temporal remote sensing images include pre-fire and post-fire remote sensing images. The forest fire severity prediction model includes two encoding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge guidance module, and an uncertainty perception and confidence module.

[0048] According to the downsampling rounds of the coding process, the pre-disaster remote sensing images and post-disaster remote sensing images are input into each coding branch respectively. Each coding branch outputs at least two candidate difference features, and the target difference features are output through the cross-attention change perception module. By extracting multi-level features from pre- and post-disaster remote sensing images through dual-branch processing, and combining the target difference features output by the cross-attention change perception module, the changes in land cover before and after the fire can be captured more accurately, thereby improving the quantitative inversion accuracy of the differential normalized burn rate continuous numerical map.

[0049] According to the upsampling rounds of the decoding process, the target difference features and each candidate difference feature are sequentially input into the boundary-aware decoding branch. The severity regression decoding branch is guided layer by layer through the cascaded knowledge guidance module, and the final regression decoding features and the final boundary decoding features are output. The boundary-aware decoding branch, in conjunction with the cascaded knowledge-guided module, enables the model to focus on flame edge information while regressing severity, ultimately outputting a flame edge map. This effectively overcomes the problems of blurred boundaries and discontinuous segmentation in traditional methods, thereby improving the accuracy of severity inversion. The final regression decoding features are input into the uncertainty perception and confidence module, which outputs a differentially normalized continuous numerical map of the burn rate and an uncertainty confidence map corresponding to the dual-temporal remote sensing images. Based on the final boundary decoding features, the flame edge map corresponding to the dual-temporal remote sensing images is determined.

[0050] This method employs an uncertainty perception and confidence module that outputs an uncertainty confidence map. On one hand, this map provides users with the reliability of the prediction results for each pixel, assisting decision-makers in identifying high-risk or ambiguous prediction areas and further improving the accuracy of severity inversion. On the other hand, the cascaded knowledge guidance module guides the severity regression branch layer by layer during the decoding stage, fully utilizing candidate and target difference features at different scales to mitigate spatial information loss during downsampling and enhance adaptability to complex terrain scenarios.

[0051] In one embodiment, the encoding branch includes a basic convolutional module and multiple spatial-frequency joint visual state space modules. For example... Figure 2 As shown, step S102 specifically includes: S201, input the pre-disaster remote sensing images and post-disaster remote sensing images into each coding branch respectively, and output the first coding feature through the basic convolutional module in each coding branch.

[0052] Specifically, such as Figure 3 The diagram shows that pre-disaster remote sensing images are input into the coding branch, and the first coded feature is output through the convolutional module in the coding branch. Post-disaster remote sensing images are also input into the coding branch, and the first coded feature is output through the convolutional module in the coding branch. The difference between the two first coded features is then determined as candidate difference features.

[0053] S202, input each first coding feature into the spatial-frequency joint visual state space module, and output the second coding feature, third coding feature and fourth coding feature of each coding branch in sequence according to the downsampling round of the coding process.

[0054] The space-frequency joint visual state space module can include a Fourier frequency domain branch and a spatial state gating branch. Specifically, the first encoded feature can be input into the space-frequency joint visual state space module. Through the Fourier frequency domain branch, the first encoded feature is subjected to Fourier transform and the frequency domain feature is re-splittered to output the frequency domain branch feature. The first encoded feature is then input into the spatial state gating branch. Through the spatial state gating branch, the first encoded feature undergoes separable convolution and activation processing to obtain the spatial branch feature. The frequency domain branch feature and the spatial branch feature are concatenated by channels to obtain the second encoded feature.

[0055] Furthermore, the second encoded feature is input into the second space-frequency joint visual state space module, and through processing by Fourier frequency domain branch and spatial state gating branch, the third encoded feature is obtained. The third encoded feature is input into the third space-frequency joint visual state space module, and through processing by Fourier frequency domain branch and spatial state gating branch, the fourth encoded feature is obtained.

[0056] S203, input each fourth encoding feature into the cross-attention change perception module, and output the target difference feature.

[0057] The fourth coding feature can be the feature output by the third spatial-frequency joint visual state space module in the coding branch. It can be understood that the two coding branches correspond to two fourth coding features: one is the fourth coding feature corresponding to the remote sensing image during the fire, and the other is the fourth coding feature corresponding to the remote sensing image after the fire.

[0058] like Figure 4 The diagram shown is a structural schematic of the cross-attention change sensing module. The cross-attention change sensing module can be a bi-temporal cross-attention fusion (BCAF) module. It is designed for the bi-temporal characteristics of forest fire remote sensing. It reconstructs the traditional general attention mechanism into a "baseline-current status" retrieval model with physical logic. It calibrates the state in the fire by calibrating the pre-fire image and couples the underlying physical difference prior to enhance the combustion signal.

[0059] Specifically, the fourth coded feature (pre-fire feature) corresponding to the pre-fire remote sensing image is used. The fourth coding feature (fire feature) corresponding to the post-fire remote sensing image. The input is fed into this module, which utilizes asymmetric projection based on Transformer. Pre-fire features are used. As a query, based on post-fire features As the key and value, we obtain the result through convolutional mapping. Modeling channel relationships and local texture information. Expanding the spatial dimension into Cross-attention weights are represented as , This represents the dimension of each attention head. This leads to the cross-attention features. , Represents the output mapping function, where the cross-attention features are in Figure 4 The output of the ReLU activation function on the upper side of (a).

[0060] Furthermore, a two-difference physical prior is constructed. To preserve the varying amplitude and direction information, signed differences are calculated separately. and absolute difference After concatenating the two features along the channel dimension, a double prior feature is obtained through convolutional projection. Among them, double prior features in Figure 4 (a) is the output of the prior encoder.

[0061] To enhance the ability to represent the boundaries of changing regions, the module further extracts boundary priors from the absolute difference features. First, the mean value of each channel is calculated to obtain the single-channel change response map. Then, the horizontal gradients are calculated using the Sabel operator. and vertical gradient The boundary magnitude map can be represented as ,in To prevent small constants from causing numerical instability, boundary features are obtained through convolutional mapping. In this model, boundary features are output by the Edge Prior encoder. Simultaneously, to assist in boundary supervision, a boundary prediction head is introduced to output the boundary response. .

[0062] Furthermore, the cross-attention features calculated above... Double prior features and boundary features The graphs are stitched together, and an adaptive weighted graph is generated using a gating function. ,in This represents the Sigmoid activation function. The gate weights are used to dynamically adjust the contribution of the cross-attention branch.

[0063] The final fusion result is expressed as ,in This indicates element-wise multiplication, with the last term being a lightweight residual connection used to preserve original contextual information and improve training stability. The final module outputs features, namely the target difference features. , that is Figure 3 In .

[0064] First, it constructs a classical physical quantity (difference) for remote sensing change detection through double difference physical priors, without relying on pure black box learning; it can distinguish vegetation reduction caused by burning. Second, it extracts boundary priors and explicitly extracts boundary information of the changed area (burnt area outline, discontinuous boundary); based on operators, it reduces the number of model parameters; and by setting ε, it avoids division by zero or square root instability in gradient calculation.

[0065] S204, candidate difference features are determined based on the differences of each first coding feature, each second coding feature, and each third coding feature.

[0066] Specifically, feature difference processing can be performed on the differences among the first coding features, the second coding features, and the third coding features to obtain multiple candidate difference features. The difference of the first coding feature corresponds to the first candidate difference feature, the difference of the second coding feature corresponds to the second candidate difference feature, and the difference of the third coding feature corresponds to the third candidate difference feature.

[0067] In this embodiment, multi-level features of pre- and post-disaster remote sensing images are extracted by dual-branching, and target difference features are output by the cross-attention change perception module. This enables more accurate capture of changes in land cover before and after the fire, thereby improving the quantitative inversion accuracy of the differential normalized combustion rate continuous numerical map.

[0068] In one embodiment, the space-frequency joint visual state space module includes a Fourier frequency domain branch and a spatial state gating branch; the step S202 above, which inputs the first encoded feature into the space-frequency joint visual state space module and outputs the second encoded feature, specifically includes: performing batch normalization on the first encoded feature to obtain normalized features; within the Fourier frequency domain branch, performing a two-dimensional real-number fast Fourier transform on the normalized features and re-splitting the frequency domain features to output frequency domain branch features; within the spatial state gating branch, performing separable convolution and activation processing on the normalized features to obtain spatial branch features; and concatenating the frequency domain branch features and spatial branch features through channels to obtain the second encoded feature.

[0069] Among them, the Spatial-Frequency Vision State Space Block (SF_VSS Block) simultaneously covers the global frequency domain response and local spatial topology of fire remote sensing images, improving the model's ability to represent complex fire areas.

[0070] Specifically, the first coding feature can be The features processed by batch normalization are represented as follows: The normalized features are then fed into the Fourier frequency domain branch and the spatial state gating branch for processing and output.

[0071] Within the Fourier frequency domain branch, a two-dimensional real-number Fast Fourier Transform is performed on the normalized features, mapping the features from the spatial domain to the frequency domain. Since the frequency domain features are themselves in complex form, their corresponding real and imaginary parts are extracted and concatenated along the channel dimension to construct a real-valued representation that can be learned by convolution. To learn the reweighting relationships between different frequency components and enhance the frequency domain feature selection capability, i.e. ,in, This indicates a frequency domain convolution mapping operation.

[0072] The weighted frequency domain features are re-split to form real and imaginary parts, and then restored to a complex spectrum. The modulated frequency domain features are as follows: ,in, The imaginary unit is represented. A two-dimensional inverse Fourier transform is performed on the reconstructed complex spectrum to remap it back to the spatial domain, yielding an enhanced frequency domain response. To control the feature dimensions, we employ... Convolution performs channel compression on the frequency domain branch output, outputting frequency domain branch features. .

[0073] Within the spatial state gating branch, for normalized features Depthwise separable convolutions are used to extract local neighborhood topological relationships and spatial continuity features. To simulate the information flow and selection mechanisms in state-space modeling, spatial features are mapped to main branch features. and gated branch features The gated branch features are processed using the SiLU activation function to construct smooth dynamic gate weights, and spatial state selection is achieved through element-wise multiplication. Spatial branching features .

[0074] After obtaining the frequency domain branching features and spatial branching features, the frequency domain branching features are... and spatial branching features By stitching along the channel dimension, we obtain information that jointly represents global spectral information and local spatial structure information. The spliced ​​dual-domain features are then processed... After convolution, the features of the inter-branch information interaction domain are integrated to obtain the output increment of the module. The fused enhanced features are added to the original input features (first encoded features) to preserve low-level details while enhancing gradient propagation capabilities, resulting in the second encoded features. .

[0075] In this embodiment, firstly, a two-dimensional Fast Fourier Transform (FFT) is used to transform features from the spatial domain to the frequency domain. A point in the frequency domain also depends on the global information of the entire image, enabling the model to capture large-scale smoke plume diffusion or regional spectral shifts caused by fire. Secondly, separable convolutions are used to slide along the spatial dimension. Compared to standard convolutions, this requires fewer parameters but can effectively capture the local texture of fire patches and the sharp edges of burned areas. An activation function is used as a gating signal, acting as a dynamic filter to retain important features (such as high-intensity burning areas) and suppress irrelevant features (such as unchanging bare soil or shadows), achieving adaptive filtering of spatial information. Thirdly, frequency domain processing prevents the loss of high-frequency information during downsampling (aliasing), while spatial branching ensures pixel-level positioning accuracy. The combination of these two provides a more discriminative feature representation for subsequent change detection.

[0076] In one embodiment, in S103 above, the target difference feature and each candidate difference feature are sequentially input into the boundary-aware decoding branch, and the severity regression decoding branch is guided layer by layer by the cascaded knowledge guidance module to output the final regression decoding feature and the final boundary decoding feature. Specifically, this includes: inputting the target difference feature and candidate difference feature into the boundary-aware decoding branch after concatenating them through channels, and outputting the final boundary decoding feature; inputting the final boundary decoding feature into the cascaded knowledge guidance module to output the cascaded output feature; and inputting the cascaded output feature, the target difference feature, and the candidate difference feature into the severity regression decoding branch to output the final regression decoding feature.

[0077] Specifically, the target difference features and candidate difference features are concatenated through channels and then input into the boundary-aware decoding branch to output the final boundary decoding features.

[0078] Among them, such as Figure 3 As shown, the target difference features and candidate differential features , and After channel concatenation, the data is input into the boundary decoder in the boundary-aware decoding branch, and the final boundary decoding feature is output. Figure 3 (Output of the middle boundary decoder).

[0079] Specifically, the final boundary decoding features are input into the cascaded knowledge guidance module, and the cascaded output features are output.

[0080] Among them, such as Figure 5 The diagram shows the structure of a cascaded knowledge-guided module, officially named the Multi-Scale Dual-Gated Cascaded Knowledge-Guided Module (CKGM). Specifically... Figure 5 (a) includes a boundary prior module, local edge branches and hole edge branches, an edge blending module and a detail projection module, a space gate, a passage gate, and a blending module. Furthermore, the structure of the boundary prior module is as follows: Figure 5 As shown in (b), the structures of local edge branches and void edge branches are as follows: Figure 5 As shown in (c), the structures of the edge blending module and the detail projection module are as follows: Figure 5 As shown in (d), the structure of the space gate is as follows: Figure 5 As shown in (e), the structure of the passageway door is as follows: Figure 5 As shown in (f), the structure of the fusion module is as follows: Figure 5 As shown in (g). It should be noted that Conv is a convolutional module, BN is a batch normalization module, SiLU, Sigmoid, and ReLU are all activation functions, and AvgPool is an average pooling layer.

[0081] The cascaded knowledge-guided module first performs multi-scale context modeling on the boundary-aware decoding branch to capture boundary structure patterns at different scales. To reduce the back interference of the regression branch on the boundary branch, the final boundary decoding features are... Gradient separation is performed to obtain the boundary features of the input. (correspond Figure 5 The expression for the boundary features in the text is: ; Furthermore, based on boundary features Determine the basic boundary features ,Right now Figure 4 The features output by the edge prior module in (a), i.e. Figure 4 (b) The final output features. To simultaneously capture local boundary details and boundary context with a larger receptive field, based on the basic boundary features... Constructing ordinary convolution branches ,Right now Figure 5 (a) Features output by the local edge branch module, and features based on the basic boundary features dilated convolution branch ,Right now Figure 5 (a) Features output by the hollow cavity edge branch module. The three features are concatenated and fused to obtain multi-scale boundary context features. ,Right now Figure 4 The output of the channel splicing operation on the left side of (a) is... Figure 4 (d) Output of the prior encoder Figure 4 (e) The output of edge coding, and Figure 4 (a) The output of the upper ReLU activation function is used as the input for the channel concatenation operation. Based on this, the features after channel concatenation are fed into... Figure 4 The fusion gate in (c) is further processed.

[0082] Furthermore, by combining spatial gate and channel gate mechanisms, the features of the severity regression decoding branch are jointly guided and dynamically modulated to enhance the network's response capability to boundary-sensitive regions. Spatial gate is defined as... Channel gating is defined as To avoid gating only suppressing features in one direction, residual modulation is used, allowing gating to both enhance and suppress features. A spatial modulation term is constructed. and channel modulation terms Therefore, the regression features guided by dual gating are expressed as follows: .in, For severity regression decoding branch features (i.e. Figure 3 (output of the regression decoder).

[0083] Furthermore, the multi-scale dual-gated cascaded knowledge-guided module enhances the structural detail representation in the severity regression branch, projecting boundary context features into explicit boundary details. ,Right now Figure 5 The output of the mid-detail projection model is used, and explicit boundary details are injected into the severity regression branch. Spatial gating is utilized. By filtering detailed features, the injected boundary information is more focused on the boundary-related areas.

[0084] Furthermore, based on the guided regression features and explicit boundary details Then, the two are concatenated along the channel dimension and deeply fused using a convolutional module. To avoid excessive interference from priors in the regression branch representation at the boundary branches, the module introduces learnable residual coefficients. The knowledge injection amplitude is adaptively controlled. The final module outputs cascaded feature outputs. .

[0085] Specifically, the cascaded output characteristics are obtained. Then, the cascaded output features will be... Target Difference Characteristics and candidate differential features ( , and The input is fed into the severity regression decoding branch, and the output is the final regression decoding feature. Figure 3 The output of the third regression decoder in the middle.

[0086] After obtaining the final regression decoding features, firstly, these features can be input into the uncertainty perception and confidence module to output a differentially normalized continuous numerical map of the combustion rate and an uncertainty confidence map corresponding to the dual-temporal remote sensing images. Secondly, the fire spot edge map corresponding to the dual-temporal remote sensing images can be determined based on the final boundary decoding features.

[0087] Firstly, such as Figure 6 The diagram shows the structure of the Uncertainty-Aware Confidence Guidance (UACG) module. This module enables the severity regression decoding branch to model the region of change, boundary details, and prediction uncertainty. Specifically... Figure 6 (a) includes multiple convolution-batch normalization-activation modules (ConvBNAct) and compression-activation modules (SE). Furthermore, the structure of the convolution-batch normalization-activation module is as follows: Figure 6As shown in (b), a convolution-batch normalization-activation module includes, in sequence, convolution (Conv), batch normalization (BN), and activation function (SiLU). A schematic diagram of the compression-activation module is shown below. Figure 6 As shown in (c), the compression-activation module includes, in sequence, an average pooling function AvgPool, a convolutional function Conv, an activation function SiLU, another convolutional function Conv, and an activation function Sigmoid.

[0088] UACG first performs preprocessing on the input information to refine shared features. Let the input features (final regression decoding features) be... Where B, C, H, and W represent batch size, number of channels, height, and width, respectively. A shared convolutional block is used to initially refine the input features, yielding the basic feature representation. ,in express The convolutional transformation consists of convolution, batch normalization, and SiLU activation. This preprocessing provides a stable feature base for subsequent multi-scale modeling.

[0089] Furthermore, multi-scale modeling is used to extract local and contextual information. UACG employs a two-branch structure to refine features at multiple scales. The first branch is the local branch, i.e. The second branch is a context branch, that is... The outputs of the two branches are concatenated along the channel dimension, and a unified feature representation is obtained through fusion convolution. .

[0090] Furthermore, after multi-scale feature modeling is completed, to enhance the role of important semantic channels, an SE channel attention mechanism is introduced to recalibrate the features, specifically expressed as follows: ,in This refers to a channel attention function composed of global average pooling and two layers of channel mapping. This function adaptively adjusts the weights of each channel based on the global context, ultimately improving feature representation. Meanwhile, crucial details from the decoding stage are retained in the current input features; residual connections are introduced to process the final decoded features. and enhanced features After superposition, the main features are obtained. .

[0091] To enhance the model's response capability to boundary regions, UACG constructs a boundary-aware gating mechanism. Let the output of the boundary-aware decoding branch be... Scale alignment is performed to obtain Where Interp represents bilinear interpolation, used to interpolate boundary features Adjust to main feature Consistent spatial resolution. Secondly, boundary features are represented through channel projection. Mapping to a compact representation space If a boundary prior exists, then it is compared with the main feature. splicing as gating input If the purpose is to provide boundary priors, then Based on this, a boundary-aware gated map is obtained through a gating generation network. .

[0092] Furthermore, to simultaneously complete dNBR regression prediction and uncertainty estimation, a mean branch and a log-variance branch are introduced for decoupled modeling. The mean prediction branch input is constructed. Log-variance branch input Mean predictions are obtained using two independent prediction heads. That is, the differential normalized combustion rate continuous numerical plot and the log-variance prediction. This refers to the uncertainty confidence plot. To avoid training instability caused by excessively large or small log-variance predictions in heteroscedastic regression, interval clipping is performed to obtain... ,in and These represent the preset minimum and maximum values, respectively. This constraint effectively suppresses outliers and improves training stability and prediction accuracy.

[0093] Secondly, based on the boundary probability map obtained by Sigmoid activation of the final boundary decoding features, the fire spot edge map corresponding to the dual-temporal remote sensing image is obtained.

[0094] In this embodiment, firstly, by fusing deep semantic difference features (target difference features) and shallow detail difference features (candidate difference features), the boundary-aware decoding branch can simultaneously utilize global change information and local edge information to generate a more accurate and continuous fire spot edge map, avoiding boundary breaks or positioning offsets. Secondly, the cascaded knowledge guidance module performs multi-level refinement and transformation on the boundary features, converting them into structural prior knowledge that the regression branch can understand, effectively bridging the feature gap between the boundary detection and severity regression tasks. Thirdly, the regression decoding branch simultaneously receives boundary prior knowledge (cascaded output) and multi-scale difference features (target + candidate), ensuring that the dNBR predicted value remains smooth and consistent within the fire spot, while forming a steep change at the boundary, significantly improving the spatial consistency and edge sharpness of the combustion severity inversion.

[0095] In one embodiment, the training process of the forest fire severity prediction model includes: acquiring multiple remote sensing biphasic images of the forest before and after a fire; selecting target band images from each remote sensing biphasic image, including near-infrared narrowband images with a center wavelength of 865nm and short-wave infrared images with center wavelengths of 1610nm and 2190nm; filtering target occlusion rate images from each target band image that meet the cloud occlusion rate filtering conditions; performing resolution unification processing on the target occlusion rate images corresponding to different target bands to obtain forest fire biphasic images; training the forest fire severity prediction model based on the forest fire biphasic images, the corresponding continuous value regression labels, and the boundary-aware mask; and outputting the forest fire severity prediction model when the loss function meets the model iteration termination condition.

[0096] Among them, the near-infrared narrowband band (865nm) characterizes the changes in scattering of vegetation mesophyll cells and the degree of damage to the vertical structure of vegetation caused by fire. In the short-wave infrared band, 1610nm is highly sensitive to moisture loss from the vegetation canopy, while 2190nm shows a significant increase in reflectance for post-fire ash, carbonaceous material, and exposed drought background.

[0097] By acquiring dual-temporal image data, systematic errors introduced by topographic shadows, seasonal phenological changes, and differences in sensor observation angles can be reduced.

[0098] One of the screening criteria for cloud cover rate is that the cloud cover rate is below 5%.

[0099] Specifically, bilinear interpolation can be used to unify the resolution of target occlusion rate images corresponding to different target bands, resulting in a forest fire dual-temporal image with a unified spatial resolution of 20m.

[0100] Specifically, based on the forest fire dual-temporal images, the corresponding continuous value regression labels and boundary-aware masks, the forest fire severity prediction model is trained. When the loss function satisfies the model iteration termination condition, the forest fire severity prediction model is output. The training set for the model is composed of two-phase images of forest fires, corresponding continuous value regression labels, and boundary-aware masks.

[0101] To address issues such as sample imbalance, poor cross-task structural integration, and data noise interference in fire monitoring tasks, a dual-temporal multi-task fire area identification network is constructed, which includes a joint loss function consisting of boundary segmentation, structural consistency constraints, and an uncertainty preheating mechanism.

[0102] The continuous value regression labels in the training set are used to calculate the Gaussian negative log-likelihood loss. .

[0103] Boundary-aware masks in the training set Used to calculate boundary segmentation loss .

[0104] The gradient structure between the continuous regression labels in the training set and the dNBR predicted by the model is used to calculate the structural consistency loss. .

[0105] Firstly, regarding the boundary segmentation loss function The boundary segmentation loss uses the Dice loss function, and the intersection-over-union ratio (IoU) is used to directly measure the spatial overlap between the predicted boundary and the true boundary. The formula for this metric is: ; in, Boundary probability map representing the network's pre-output ( Figure 3 (firemark edge image obtained after sigmoid activation). Represents the actual binary boundary label. It is a smooth term that avoids a denominator of 0.

[0106] Secondly, to establish strong physical coupling between the boundary segmentation and severity regression branches, a structural consistency loss based on image gradient is designed. The formula is as follows: ; ; in and The horizontal and vertical gradient components of the graph are represented by the following formula: , . and It is a fixed Sobel operator in the horizontal and vertical directions. It is a continuous severity mean plot.

[0107] Third, for the severity regression branch, a Gaussian negative log-likelihood loss is introduced. To model data uncertainty. The formula for calculating this indicator is: ; For the current training round, This corresponds to the preheating threshold. When the model is in the preheating phase, This will require the model to use mean squared error (MSE) for deterministic fitting, while also incorporating... Regular terms (Where, alpha is the weight of the regularization term, which applies an L2 constraint to the log-variance of the model's predictions) to reduce the prediction variance. Once the model possesses basic representational capabilities, At this point, the variance constraint is removed, and a Bayesian loss calculation with log-variance decay is performed.

[0108] After obtaining the Gaussian negative log-likelihood loss Boundary segmentation loss Structural consistency loss Next, the multi-task joint objective loss function for: ; in, , , For weight hyperparameters.

[0109] If the loss function (multi-task joint objective loss function) satisfies the model iteration termination condition, the forest fire severity prediction model is output.

[0110] In one embodiment, the method described in S101-S104 above is used as an example for experimentation. The experiment uses pre-fire and in-fire dual-temporal remote sensing images as input, and simultaneously performs two tasks: continuous value regression of fire severity (dNBR) and fire boundary segmentation. To ensure fairness in comparison, all comparison methods use the same input data and training strategy. The regression task uses mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean bias error (MBE), coefficient of determination (R²), and Pearson correlation coefficient (PearsonR) as evaluation indicators; the segmentation task uses precision, recall, F1 score, intersection-over-union ratio (mIoU), overall accuracy, specificity, balanced accuracy (BalACC), false positive rate (FPR), and false negative rate (FNR) as evaluation indicators.

[0111] As shown in Table 1, the method of this embodiment is significantly better than the existing methods in all error indicators, and also significantly improves the coefficient of determination and Pearson correlation coefficient. This indicates that this embodiment can more accurately characterize the continuous change characteristics of fire severity and significantly improve regression accuracy and stability.

[0112] Table 1. Performance comparison of multi-task models in fire severity regression tasks. ; As shown in Table 2, the method in this embodiment achieves the best results in key indicators such as F1 score and mean intersection-union ratio (mIoU). At the same time, the false positive rate (FPR) and false negative rate (FNR) are significantly reduced, indicating that this embodiment can effectively reduce false positives and false negatives while maintaining the integrity of the boundary, thereby improving the accuracy of fire boundary identification.

[0113] Table 2 Performance comparison of multi-task models in fire boundary segmentation tasks ; As shown in Table 3, the method of this embodiment is significantly better than the existing methods in all error indicators. At the same time, it significantly improves the coefficient of determination and Pearson correlation coefficient, indicating that the method of this embodiment can not only significantly reduce the prediction error, but also fit the changing trend of the real dNBR more accurately. It is superior to other comparative models in terms of regression accuracy and stability.

[0114] Table 3. Illustrative comparison of the performance of different methods on the dNBR regression task. ; As shown in Table 4, the method in this embodiment achieved the best test performance in all evaluation index dimensions, which can extract the fire boundary area more accurately, suppress false detection of background area, reduce the problem of missed detection of fire boundary, and finally obtain stable and accurate segmentation results.

[0115] Table 4. Illustrative performance comparison of different methods on boundary segmentation tasks ; Depend on Figure 7 The visualizations shown depict the different models on the test set, specifically in the evaluation of the dNBR prediction task and the boundary segmentation task. Figure 7 (a) shows the evaluation results of the semantic change detection method (BGSNet). Figure 7 (b) is the evaluation result diagram of the farmland boundary delineation method (BSINet). Figure 7 (c) shows the evaluation results of the very high-resolution image change detection method (SiamNestedUNet). Figure 7 (d) shows the evaluation results of image change detection (STANet). Figure 7 (e) and Figure 7 (f) shows the evaluation results of the method (BTMT-BurnNet and Label) in this embodiment. The method in this embodiment can restore the fire boundary morphology relatively completely and maintain stable spatial consistency in high-severity areas and boundary transition areas.

[0116] In this embodiment, multispectral remote sensing images before and during a fire are used as input. A unified preprocessing and precise registration dual-temporal data system is constructed to achieve consistent representation of multi-source information. Based on this, a multi-task modeling mechanism for severity and boundary labels is designed, and a deep network framework including a dual-branch encoder and dual-decoding structure is constructed. In the encoding stage, a spatial-frequency joint visual state space module (SF_VSS) is introduced to achieve collaborative modeling of frequency and spatial information. In deep feature fusion, a cross-attention change perception module (BCAF) is used to enhance the representation of dual-temporal difference features. In the decoding stage, a cascaded knowledge guidance module (CKGM) is used to dynamically guide and finely modulate the boundary information for severity regression. Simultaneously, an uncertainty perception and confidence assessment module (UACG) is used to model the uncertainty and evaluate the confidence level of the prediction results. Experimental results show that the method in this embodiment can effectively improve the accuracy and boundary representation capability of fire severity inversion in complex scenarios, and enhance the reliability and interpretability of the results.

[0117] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0118] Based on the same inventive concept, this application also provides a dual-temporal feature fusion and boundary constraint forest fire severity inversion device for implementing the aforementioned dual-temporal feature fusion and boundary constraint forest fire severity inversion method. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more dual-temporal feature fusion and boundary constraint forest fire severity inversion device embodiments provided below can be found in the limitations of the dual-temporal feature fusion and boundary constraint forest fire severity inversion method described above, and will not be repeated here.

[0119] In one embodiment, such as Figure 8 As shown, a forest fire severity inversion device based on dual-temporal feature fusion and boundary constraints is provided, comprising: an acquisition module 801, an encoding module 802, a decoding module 803, and a prediction module 804, wherein: The acquisition module 801 is used to acquire dual-temporal remote sensing images of the target forest area before and during the fire, as well as a forest fire severity prediction model. The dual-temporal remote sensing images include pre-disaster remote sensing images and during-disaster remote sensing images; the forest fire severity prediction model includes two coding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge guidance module, and an uncertainty perception and confidence module. The encoding module 802 is used to input the pre-disaster remote sensing image and the remote sensing image during the disaster into each encoding branch according to the downsampling round of the encoding process, output at least two candidate difference features through each encoding branch, and output the target difference feature through the cross-attention change perception module. The decoding module 803 is used to input the target difference features and each candidate difference feature into the boundary-aware decoding branch in sequence according to the upsampling rounds of the decoding process. The severity regression decoding branch is guided layer by layer through the cascaded knowledge guidance module, and the final regression decoding features and the final boundary decoding features are output. The prediction module 804 is used to input the final regression decoding features into the uncertainty perception and confidence module, output the differential normalized continuous numerical map of the combustion rate and the uncertainty confidence map corresponding to the dual-temporal remote sensing images, and determine the fire spot edge map corresponding to the dual-temporal remote sensing images based on the final boundary decoding features.

[0120] The modules in the aforementioned forest fire severity inversion device based on dual-temporal feature fusion and boundary constraints can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0121] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores dual-temporal remote sensing image data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for inverting forest fire severity using dual-temporal feature fusion and boundary constraints.

[0122] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0123] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method steps.

[0124] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method steps.

[0125] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method steps.

[0126] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0127] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0128] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for inverting forest fire severity based on dual-temporal feature fusion and boundary constraints, characterized in that, The method includes: Acquire dual-temporal remote sensing images of the target forest area before and after the fire, and develop a forest fire severity prediction model; The dual-temporal remote sensing images include pre-disaster remote sensing images and post-disaster remote sensing images; the forest fire severity prediction model sequentially includes two encoding branches, a boundary-aware decoding branch, a severity regression decoding branch, a cross-attention change perception module, a cascaded knowledge guidance module, and an uncertainty perception and confidence module; the encoding branch includes a basic convolution module and multiple spatial-frequency joint visual state space modules; The pre-disaster remote sensing images and the post-disaster remote sensing images are respectively input into each of the coding branches, and the first coding features are output through the basic convolutional modules in each of the coding branches; Each of the first coding features is input into the space-frequency joint vision state space module, and according to the downsampling rounds of the coding process, the second coding feature, the third coding feature, and the fourth coding feature of each coding branch are output sequentially. Each of the fourth encoded features is input into the cross-attention change perception module, and the target difference features are output. Candidate difference features are determined based on the differences of each of the first coding features, the differences of each of the second coding features, and the differences of each of the third coding features. According to the upsampling rounds of the decoding process, the target difference features and each of the candidate difference features are sequentially input into the boundary-aware decoding branch. The severity regression decoding branch is guided layer by layer by the cascaded knowledge guidance module, and the final regression decoding features and the final boundary decoding features are output. The final regression decoding features are input into the uncertainty perception and confidence module, which outputs the differential normalized combustion rate continuous numerical map and uncertainty confidence map corresponding to the dual-temporal remote sensing image, and determines the fire spot edge map corresponding to the dual-temporal remote sensing image based on the final boundary decoding features.

2. The method according to claim 1, characterized in that, The spatial-frequency joint vision state space module includes a Fourier frequency domain branch and a spatial state gating branch; The first encoded feature is input into the spatial-frequency joint visual state space module, and the second encoded feature is output, including: The first encoded feature is batch normalized to obtain the normalized feature; Within the Fourier frequency domain branch, the normalized feature two-dimensional real number fast Fourier transform and the frequency domain feature are re-splitned to output the frequency domain branch feature. Within the spatial state gated branch, the normalized features are subjected to separable convolution and activation processing to obtain spatial branch features; The frequency domain branch feature and the spatial branch feature are concatenated by channels to obtain the second coding feature.

3. The method according to claim 1, characterized in that, The step of sequentially inputting the target difference feature and each of the candidate difference features into the boundary-aware decoding branch, and guiding the severity regression decoding branch layer by layer through the cascaded knowledge guidance module, and outputting the final regression decoding feature and the final boundary decoding feature, includes: The target difference feature and the candidate difference feature are concatenated through channels and then input into the boundary-aware decoding branch to output the final boundary decoding feature. The final boundary decoding features are input into the cascaded knowledge guidance module, and the cascaded output features are output. The cascaded output features, the target difference features, and the candidate difference features are input into the severity regression decoding branch, and the final regression decoding features are output.

4. The method according to claim 1, characterized in that, The training process for the forest fire severity prediction model includes: Acquire multiple remote sensing biphasic images of the forest before and after the fire; Target band images with a center wavelength of 865 nm in the near-infrared narrowband band and with center wavelengths of 1610 nm and 2190 nm in the short-wave infrared bands are selected from each of the aforementioned remote sensing dual-phase images. Select target occlusion rate images that meet the cloud occlusion rate selection criteria from the images of each target band; By unifying the resolution of the target occlusion rate images corresponding to different target bands, dual-temporal images of forest fires are obtained. Based on the dual-temporal images of the forest fire, the continuous value regression labels corresponding to the dual-temporal images of the forest fire, and the boundary-aware mask, the training model for forest fire severity is trained. When the loss function satisfies the model iteration termination condition, the forest fire severity prediction model is output.

Images