A multi-source remote sensing driven object-level wild fire spread prediction method and system

Abstract

The application discloses a multi-source remote sensing driven object-level wild fire spread prediction method and system. The method comprises the following steps: acquiring multi-source remote sensing data in a target region and performing spatio-temporal unified processing; aggregating active fire observation data into a burning mask according to days, performing connected domain marking on the daily burning mask, and constructing an object-level wild fire event sequence according to the spatial overlap relationship or the adjacent relationship between the burning patches of adjacent dates; generating a wild fire state variable based on the object-level wild fire time sequence, the vegetation index data and the historical burning state; inputting the wild fire state variable into a convolutional long short-term memory network to obtain a predicted continuous wild fire spread response field at a future time; and performing thresholding processing on the continuous wild fire spread response field at the future time to generate a wild fire spread prediction map at the future time. The application can reduce object sample mixed noise, enhance the expression of fire line continuous advance and non-adjacent diffusion relationship, and improve the reliability of regional scale wild fire spread prediction and early warning.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-source remote sensing-driven method, system, storage medium, and electronic device for predicting object-level wildfire spread. Background Technology

[0002] Wildfire spread is influenced by a combination of factors, including historical combustion conditions, combustible continuity, near-surface wind fields, topographic conditions, and human intervention, exhibiting significant spatiotemporal nonstationarity and scale dependence. Traditional physical or semi-empirical fire spread models typically require relatively complete data on fuel type, moisture content, topography, and high-frequency meteorological boundary conditions. While they offer clear mechanistic explanations, they often face challenges in obtaining parameters, high regional calibration costs, and significant computational overhead in large-scale, multi-event, and rapidly updated remote sensing operational scenarios.

[0003] In recent years, deep learning-based wildfire spread prediction methods have been able to directly learn the fire line expansion patterns from multi-source remote sensing sequences. However, existing methods still have three main shortcomings: First, many methods use fixed raster frames as samples, which can easily mix different fire events or discontinuous fire spots as the same training object, weakening the physical consistency of the samples. Second, they often simply concatenate data such as NDVI (Normalized Difference Vegetation Index), fire points, and wind fields as input channels, lacking an explicit expression of the dynamic relationship between "unburned combustible areas, previously burned areas, and newly burned areas." Third, standard convolutional recurrent networks tend to focus on local neighborhood propagation, making it difficult to simultaneously characterize the advancement of continuous fire lines and non-neighborly jump spread, while directly using global attention may lead to high computational costs. Summary of the Invention

[0004] This invention provides a multi-source remote sensing-driven object-level wildfire spread prediction method, system, storage medium, and electronic device, which can overcome the shortcomings of existing technologies such as inconsistent sample organization, insufficient expression of state variables, difficulty in taking into account both local propagation and non-local diffusion, and lack of flexibility in early warning thresholds.

[0005] This invention provides a multi-source remote sensing-driven object-level wildfire spread prediction method, comprising: Acquire multi-source remote sensing data within the target area, wherein the multi-source remote sensing data includes at least active fire observation data, vegetation index data, and near-surface wind field data; The multi-source remote sensing data are subjected to spatiotemporal unified processing; The active fire observation data is aggregated into a burning mask on a daily basis. The burning masks of each day are labeled with connected components, and an object-level wildfire event sequence is constructed based on the spatial overlap or adjacency relationship between burning patches on adjacent dates. Based on the object-level wildfire time series, the vegetation index data, and historical combustion status, wildfire state variables are generated. The wildfire state variables are input into a convolutional long short-term memory network to obtain the predicted continuous wildfire spread response field at future moments; Thresholding is applied to the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map for the future time.

[0006] Furthermore, according to the above-mentioned multi-source remote sensing-driven object-level wildfire spread prediction method, the spatiotemporal unified processing of the multi-source remote sensing data includes: Transform the spatial projection coordinate system of all data layers into a pre-defined target projection coordinate system; Based on the target projection coordinate system, a standard regular rectangular analysis grid is established, and the standard regular rectangular analysis grid is used as a unified spatial reference benchmark to define the cell spatial layout of each data layer. A spatial resampling algorithm is used to register all data layers to the standard rule analysis grid. Each data layer after resampling is cropped according to a preset common spatial range boundary to ensure that the number of pixels in the row and column directions of all data layers is completely consistent, and that each pixel corresponds in spatial position.

[0007] Furthermore, according to the above-mentioned multi-source remote sensing-driven object-level wildfire spread prediction method, the active fire observation data is aggregated daily into a combustion mask, the daily combustion mask is labeled with connected components, and an object-level wildfire event sequence is constructed based on the spatial overlap or adjacency relationship between combustion patches on adjacent dates, including: A daily-scale binary combustion state indicator factor is constructed based on the active fire observation data acquired daily. The daily-scale binary combustion state indicator factor is assigned a value based on a preset assignment rule to obtain a combustion mask of the active fire observation data. The daily combustion masks were spatially clustered using a connected component labeling algorithm to obtain several independent combustion patches; Spatial overlap matching is performed on the fire patches extracted from adjacent dates. If the fire patch of the following day and the fire patch of the previous day meet the preset spatial continuity judgment rule, then the two are determined to belong to the same object-level wildfire event, and the latter is assigned to the object-level wildfire event sequence of the former.

[0008] Furthermore, according to the above-mentioned multi-source remote sensing-driven object-level wildfire spread prediction method, the preset assignment rules include pixel center fall rules and buffer rasterization rules. The pixel center falling rules include: if the geographic latitude and longitude coordinates of any VIIRS active fire point fall within the spatial coverage of any pixel in the target spatial grid, then the pixel is assigned a value of 1; if it does not fall within the spatial coverage of any pixel in the target spatial grid, then the pixel is assigned a value of 0. The buffer rasterization rule includes: taking each active fire point location as the center, setting a buffer radius that matches the spatial resolution of the VIIRS sensor; if the center point of the target pixel is within the buffer range, or if the spatial area covered by the target pixel has geometric intersection with the buffer, then the target pixel is assigned a value of 1; if the center point of the target pixel is not within the buffer range, and the spatial area covered by the target pixel does not have geometric intersection with the buffer, then the target pixel is assigned a value of 0.

[0009] Furthermore, according to the above-mentioned multi-source remote sensing-driven object-level wildfire spread prediction method, the convolutional long short-term memory network includes an input layer, a convolutional gated recursive unit, a non-local attention enhancement module, and an output mapping layer. The processing procedure of the convolutional long short-term memory network includes: The wildfire state variables are stacked along the time dimension to obtain a multi-channel spatiotemporal input tensor. The multi-channel spatiotemporal input tensor is then input into the input layer to obtain the input slice corresponding to the current moment. The current input slice is concatenated with the hidden state from the previous time step and then input into the convolutional gated recursive unit to generate the hidden state at the current time step. The current hidden state is input into the nonlocal spatial attention enhancement module to obtain the enhanced nonlocal spatial feature tensor. The enhanced nonlocal spatial feature tensor is added to the current hidden state to obtain the enhanced hidden state that integrates global spatial correlation information. The steps of generating the hidden state at the current time step and obtaining the enhanced hidden state are executed iteratively in chronological order until all input slices at all time steps have been processed. The current hidden state and the enhanced hidden state are fused and input into the output mapping layer to obtain the continuous wildfire spread response field for future time steps.

[0010] Furthermore, according to the above-mentioned multi-source remote sensing-driven object-level wildfire spread prediction method, the process of inputting the current hidden state into the non-local spatial attention enhancement module to obtain the enhanced non-local spatial feature tensor includes: The hidden state at the current moment is mapped to query vector Q, key vector K, and value vector V through convolutional blocks, and the spatial dimensions of query vector Q, key vector K, and value vector V are expanded into a sequence of H×W position vectors; where H is the number of pixels in the target spatial grid in the row direction, and W is the number of pixels in the target spatial grid in the row direction. Calculate the matrix product between the query vector Q and the transpose of the key vector K to obtain the pairwise relevance matrix, and use an activation function to normalize each row to obtain the spatial attention weight map. The spatial attention weight map is weighted and aggregated with the value feature V to generate an enhanced nonlocal spatial feature tensor.

[0011] Furthermore, according to the above-mentioned multi-source remote sensing-driven object-level wildfire spread prediction method, the thresholding process is performed on the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map at the future time, including: A response threshold parameter is preset. For any cell in the target spatial grid, if the response value of the cell in the continuous wildfire spread response field is greater than or equal to the response threshold parameter, the cell is determined to be a predicted spread cell. If the response value of the cell in the continuous wildfire spread response field is less than the response threshold parameter, it is determined to be a predicted non-spread cell, so as to generate a wildfire spread prediction map.

[0012] This invention also provides a multi-source remote sensing-driven object-level wildfire spread prediction system, comprising: The data acquisition module is used to acquire multi-source remote sensing data within the target area. The multi-source remote sensing data includes at least active fire observation data, vegetation index data, and near-surface wind field data. The spatiotemporal registration module is used to perform spatiotemporal unified processing on the multi-source remote sensing data; The object-level event construction module is used to aggregate the active fire observation data into a burning mask on a daily basis, mark the connected components of the daily burning mask, and construct an object-level wildfire event sequence based on the spatial overlap or adjacency relationship between burning patches on adjacent dates. The state variable generation module is used to generate wildfire state variables based on the object-level wildfire time series, the vegetation index data, and historical burning status. The spatiotemporal prediction module is used to input the wildfire state variables into a convolutional long short-term memory network to obtain the predicted continuous wildfire spread response field at future moments; The thresholding output module is used to threshold the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map at the future time.

[0013] The present invention also provides a computer-readable storage medium storing a plurality of instructions adapted to be loaded by a processor to execute any of the above-described multi-source remote sensing-driven object-level wildfire spread prediction methods.

[0014] The present invention also provides an electronic device, including a processor and a memory, wherein the processor is electrically connected to the memory, the memory is used to store instructions and data, and the processor is used in the steps of the multi-source remote sensing driven object-level wildfire spread prediction method described in any of the preceding claims.

[0015] The multi-source remote sensing-driven object-level wildfire spread prediction method, system, storage medium, and electronic device provided by this invention have the following beneficial effects: (1) By constructing object-level wildfire event sequences, sample noise is effectively suppressed, and the purity of model training data is improved. This invention rasterizes the daily VIIRS active fire points into a binary combustion mask, and then performs connected component labeling and spatiotemporal overlap and adjacency matching of cross-day patches, merging patches belonging to the same physical combustion process into a continuous object-level event sequence. This processing method avoids mixing the combustion pixels of multiple independent fire events in the same training sample, thereby significantly reducing feature-label mismatch noise caused by event mixing, and providing a supervision signal with clear physical meaning and strong temporal consistency for subsequent deep learning models.

[0016] (2) By constructing wildfire state variables that include vegetation combustibility, historical combustion memory, and new fire line expansion, the key driving factors are explicitly coupled, enhancing the correspondence between input features and the physical process of wildfire spread. This invention combines vegetation index data after spatiotemporal unified processing, historical combustion mask, and current combustion state into multi-channel input variables that characterize wildfire state. This allows the model input layer to directly include combustible conditions, historical combustion memory, and dynamic fire line advancement information, avoiding the physical semantic ambiguity problem caused by relying solely on indirect indicators such as original remote sensing brightness temperature or fire point density in traditional methods. This improves the model's ability to perceive complex behaviors such as fire line stagnation, acceleration, and turning.

[0017] (3) Introducing pairwise self-attention or local patch attention enhancement modules into the convolutional long short-term memory network, taking into account both the modeling of local fireline advancement and global non-neighbor spatial relationships, improves prediction accuracy. This invention sets a non-local spatial attention mechanism at the output of the convolutional gated recursive unit, enabling the network to explicitly model the dependencies of non-neighbor spatial locations such as skip-burning, multi-point fireline merging, and long-distance wind direction influence while maintaining the efficient capture of local expansion patterns in the fireline neighborhood by ConvLSTM; when using local patch attention as an alternative, the continuity and detail representation of the fireline boundary are enhanced while reducing computational overhead. This design breaks through the bottleneck of the limited receptive field of traditional convolutional recursive networks and effectively improves the prediction fidelity for complex spread patterns.

[0018] (4) The continuous response field replaces the direct binary classification output, and the adjustable threshold enables flexible configuration of prediction decisions, taking into account the application needs of multiple scenarios such as early warning and accuracy balance. This invention designs the network output layer as a single-channel continuous grid response field, instead of directly outputting binary class labels, so that the model prediction results retain the confidence difference information contained in the continuous values; in the subsequent thresholding process, the threshold can be set to a low value (such as 0.1) to reduce the false alarm rate and meet the sensitivity requirements of early fire warning; or the threshold can be optimized to the range of 0.2~0.3 based on the validation set F1 or ROC curve to balance the risks of false alarms and false alarms. This adjustable threshold strategy enables the same trained model to adapt to different business needs, significantly enhancing the practicality and robustness of the method. Attached Figure Description

[0019] The technical solution and other beneficial effects of the present invention will become apparent from the following detailed description of specific embodiments of the invention, in conjunction with the accompanying drawings.

[0020] Figure 1 A flowchart of a multi-source remote sensing-driven object-level wildfire spread prediction method provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a convolutional long short-term memory network provided in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the nonlocal spatial attention enhancement module provided in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram comparing the target combustion mask, model predicted response field, and residual of a wildfire event provided in an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram comparing the target state, model prediction, and residual changes of continuously added combustion zones over historical time steps T1 to T10, as provided in an embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram of the structure of a multi-source remote sensing-driven object-level wildfire spread prediction system provided in an embodiment of the present invention.

[0026] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0027] 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.

[0028] This invention provides a multi-source remote sensing-driven object-level wildfire spread prediction method, system, storage medium, and electronic device. The multi-source remote sensing-driven object-level wildfire spread prediction system provided by this invention can be integrated into an electronic device, such as a terminal or server. The terminal can include a tablet computer, laptop computer, personal computer (PC), microprocessor box, or other devices.

[0029] Please see Figure 1 , Figure 1 The flowchart illustrates a multi-source remote sensing-driven object-level wildfire spread prediction method provided in this embodiment of the invention. This method, applied in electronic devices, includes the following steps: S1. Acquire multi-source remote sensing data within the target area. The multi-source remote sensing data includes at least active fire observation data, vegetation index data, and near-surface wind field data.

[0030] The active fire observation data uses the Visible Infrared Imaging Radiometer Suite (VIIRS) active fire product, with a nominal spatial resolution of 375 m, to provide daily active fire location and fire radiation power information. The vegetation index data uses the Moderate Resolution Imaging Spectroradiometer (MODIS) Normalized Difference Vegetation Index (NDVI) product, with a nominal spatial resolution of 250 m and a nominal temporal resolution of 16 days, to characterize surface vegetation greenness and combustible load. The near-surface wind field data uses the 10 m height zonal wind component u10 and meridional wind component v10 from the European Centre for Medium-Range Weather Forecasts' Generation 5 Reanalysis product (ERA5-Land), with a nominal spatial resolution of approximately 0.1° (approximately 11 km) and a nominal temporal resolution of hourly scale, to characterize the meteorological and dynamic conditions affecting the direction and rate of wildfire spread.

[0031] Preferably, the spatial extent of the target area is set according to the boundary of the study area or the boundary of the administrative division to ensure that the acquired multi-source remote sensing data have a consistent overlapping area in the horizontal spatial coverage.

[0032] S2 performs spatiotemporal unified processing on multi-source remote sensing data.

[0033] In one embodiment, step S2 includes the following steps: S21, convert the spatial projection coordinate system of all data layers to the pre-defined target projection coordinate system.

[0034] S22. Based on the target projection coordinate system, a standard regular rectangular analysis grid is established, and the standard regular rectangular analysis grid is used as a unified spatial reference benchmark to define the cell spatial layout of each data layer.

[0035] S23 employs a spatial resampling algorithm to register all data layers to a standard rule-based analysis grid.

[0036] S24. After resampling, each data layer is cropped according to a preset common spatial range boundary to ensure that the number of pixels in the row and column directions of all data layers is completely consistent and that each pixel corresponds in spatial position.

[0037] Specifically, VIIRS active fire observation data were synthesized on a daily scale; MODIS vegetation index data were matched to a daily scale sequence according to the corresponding date to characterize vegetation combustibility conditions; and the u10 and v10 components in ERA5-Land near-surface wind field data were aggregated from hourly scales into daily scale statistics. Subsequently, the active fire observation data, vegetation index data, and near-surface wind field data were unified to the same projection coordinate system and registered to the same target spatial grid, thus forming a multi-source remote sensing daily scale input sequence with consistent spatial extent and time step.

[0038] Preferably, the target grid resolution is set to 375 m, and each event sample is cropped into a 100×100 pixel window, corresponding to a spatial range of approximately 37.5 km×37.5 km. Nearest neighbor resampling is used for the combustion mask, bilinear or nearest-time synthesis is used to complete continuous positive variables such as NDVI, and nearest neighbor registration is used for the u10 and v10 wind components; this wind field processing is only used for spatial alignment and does not constitute physical downscaling.

[0039] S3 aggregates active fire observation data into a burning mask on a daily basis, marks the connected components of the daily burning mask, and constructs an object-level wildfire event sequence based on the spatial overlap or adjacency relationship between burning patches on adjacent dates.

[0040] In one embodiment, step S3 includes the following steps: S31. Construct a daily-scale binary combustion state indicator factor based on the daily active fire observation data. Assign values ​​to the daily-scale binary combustion state indicator factor based on the preset assignment rules to obtain the combustion mask of the active fire observation data.

[0041] Among them, the preset assignment rules include the pixel center falling rule and the buffer rasterization rule; The cell center falling rules include: if the geographic latitude and longitude coordinates of any VIIRS active fire point fall within the spatial coverage of any cell in the target spatial grid, then the cell is assigned a value of 1; if it does not fall within the spatial coverage of any cell in the target spatial grid, then the cell is assigned a value of 0. The buffer rasterization rules include: setting a buffer radius that matches the spatial resolution of the VIIRS sensor, with each active fire point as the center; if the center point of the target pixel is within the buffer range, or if the spatial area covered by the target pixel has geometric intersection with the buffer, then the target pixel is assigned a value of 1; if the center point of the target pixel is not within the buffer range, and the spatial area covered by the target pixel does not have geometric intersection with the buffer, then the target pixel is assigned a value of 0.

[0042] Specifically, under the pixel center fall rule, if the latitude and longitude coordinates of a certain VIIRS active fire point fall within the range of a certain pixel in the target grid, then that pixel is assigned a value of 1, indicating that the pixel... Combustion observations are present on day t; pixels not covered by fire points are assigned a value of 0, indicating no combustion observations. Under the buffer rasterization rule, a buffer radius matching the VIIRS spatial resolution is set with the active fire point location as the center, preferably 375 m or a range of half a pixel to one pixel; if the center of the target pixel is within this buffer range, or if the pixel spatially intersects with the buffer, it is assigned a value of 1; otherwise, it is assigned a value of 0. This yields the binary combustion mask for day t.

[0043] S32 uses a connected component labeling algorithm to perform spatial clustering on the daily combustion masks, resulting in several independent combustion patches.

[0044] Specifically, the generated daily binary burning masks are spatially clustered using a connected component labeling algorithm (preferably the eight-neighbor connectivity criterion), and adjacent pixels with a value of 1 are aggregated into several independent burning patches.

[0045] S33, perform spatial overlap matching on the fire patches extracted from adjacent dates. If the fire patch of the following day and the fire patch of the previous day meet the preset spatial continuity judgment rule, then determine that the two belong to the same object-level wildfire event, and classify the latter into the object-level wildfire event sequence of the former.

[0046] The preset spatial continuity judgment rule is as follows: (1) The two burning patches have a direct overlapping area in space; (2) The two burning patches satisfy the eight-adjacency condition, that is, there is at least one pixel in the burning patch of the next day, and its eight-neighbor range contains at least one pixel in the burning patch of the previous day. (3) After performing a single-pixel morphological dilation operation on the previous day's burning patch, there is a spatial intersection with the subsequent day's burning patch.

[0047] If any of the preset spatial continuity judgment rules are met, the two burning patches are determined to belong to the same object-level wildfire event; otherwise, they are initialized as a new event.

[0048] For a wildfire event being tracked, if there are consecutive daily-scale observation gaps in the time series, and the duration of the gaps does not exceed a preset discontinuity tolerance threshold (preferably not exceeding one daily-scale time step, i.e., consecutive gaps do not exceed one day), and the burning patches corresponding to the event before and after the gaps satisfy the spatial continuity criterion, then the gap is considered an observation gap within the same event, the event identity remains unchanged, and continuity is maintained by spatial interpolation or morphological constraints of the burning patches before and after the gaps; if the duration for which no new burning patches are matched in the continuous time series for an event exceeds the preset discontinuity tolerance threshold, then the wildfire event is determined to be terminated, the tracking and recording of the event ends, and complete event attribute information including the event's spatiotemporal range, burning pixel set, and duration is output.

[0049] S4 generates wildfire state variables based on object-level wildfire time series, vegetation index data, and historical burning status.

[0050] Specifically, for each object-level wildfire event, at each time step t, a cumulative burning history mask is first constructed based on VIIRS fire point observations up to the current time. Based on this, three types of state variables are generated: the unburned combustible region U, the previously burned region B, and the newly added burned region N. The unburned combustible region is defined as... This means that only positive NDVI values ​​of pixels that have not yet been detected as burning in the event history are retained; once a pixel is detected as burning, its U value remains 0 in subsequent time steps. Previously burning areas are... Time is defined as ,exist Time is defined as This is used to represent the accumulated area that has been detected as burning before the current moment. Newly added burning areas are... Time is defined as ,exist Time is defined as This is used to characterize newly detected burning pixels at the current moment relative to the previous cumulative burning history. This construction method ensures that the fuel state after pixel burning will not be restored due to NDVI product updates, thus maintaining the irreversibility of the fuel consumption process in the event sequence.

[0051] S5 inputs the wildfire state variables into a convolutional long short-term memory network to obtain the predicted continuous wildfire spread response field for future moments.

[0052] Figure 2 This is a schematic diagram of the structure of a convolutional long short-term memory network provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the Convolutional Long Short-Term Memory (LSM) network includes an input layer, convolutionally gated recurrent units, a non-local attention enhancement module, and an output mapping layer. The processing steps of the LSM network include: S51: Stack the wildfire state variables along the time dimension to obtain a multi-channel spatiotemporal input tensor. Input the multi-channel spatiotemporal input tensor into the input layer to obtain the input slice corresponding to the current moment.

[0053] Specifically, the five types of wildfire state variables, namely unburned and combustible areas, are considered over T consecutive historical time points. U、 Previous burning area B、 New combustion zone N The zonal wind component u10 and the meridional wind component v10 are stacked along the time dimension to construct a multi-channel spatiotemporal input tensor of size T×H×W×C; where T represents the historical time step, H and W represent the number of rows and columns of the target spatial grid, respectively, and C represents the number of channels, corresponding to the total number of categories of the above variables. Preferably, T=10, H=W=100, and C=5. Then, the multi-channel spatiotemporal input tensor is input into the input layer to obtain the input slice corresponding to the current time.

[0054] S52 concatenates the current input slice with the hidden state from the previous time step and inputs it into the convolutional gated recursive unit to generate the hidden state at the current time step.

[0055] Specifically, the convolutional gated recursive unit sequentially reads the multi-channel raster slice corresponding to each moment in the multi-channel spatiotemporal input tensor along the time axis. Through the input gate, forget gate, output gate and memory cell update mechanism driven by two-dimensional convolution operation, it iteratively updates the hidden state and cell state at each moment, thereby capturing the spatiotemporal dynamic evolution characteristics of wildfire spread while maintaining the integrity of the two-dimensional spatial topology.

[0056] It should be noted that the convolutional long short-term memory network contains at least one layer of convolutionally gated recursive units (preferably ConvLSTM units), in which two-dimensional convolutional kernels are used to replace the matrix multiplication transformation in the traditional fully connected long short-term memory network; Preferably, the convolutional long short-term memory network adopts a two-layer stacked structure, with the number of hidden channels set to 32 and 64 respectively, and the kernel size of each layer is uniformly set to 3×3 pixels.

[0057] S53, input the current hidden state into the nonlocal spatial attention enhancement module to obtain the enhanced nonlocal spatial feature tensor, add the enhanced nonlocal spatial feature tensor to the current hidden state to obtain the enhanced hidden state that integrates global spatial correlation information.

[0058] After the product-gated recursive unit outputs the hidden state feature map at the current moment, a non-local spatial attention enhancement module is further introduced to explicitly model the long-range dependencies between the local spatial context of the fire line's adjacent area and the global spatial location, as well as the spatial associations between non-adjacent combustion areas, so as to enhance the network's ability to express continuous fire line advancement, skip-spreading, and spatially discontinuous combustion response modes.

[0059] Figure 3 This is a schematic diagram of the structure of the nonlocal spatial attention enhancement module provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the nonlocal spatial attention enhancement module preferably adopts a pairwise self-attention mechanism, and its specific calculation process is as follows: S531, the hidden state at the current time step is mapped to query vector Q, key vector K and value vector V through convolutional blocks, and the spatial dimensions of query vector Q, key vector K and value vector V are expanded into a sequence of H×W position vectors; where H is the number of pixels in the target space grid in the row direction and W is the number of pixels in the target space grid in the row direction.

[0060] S532, calculate the matrix product between the query vector Q and the transpose of the key vector K to obtain the pairwise relevance matrix, and use the activation function to normalize each row to obtain the spatial attention weight map.

[0061] S533 weighted aggregates the spatial attention weight map with the value feature V to generate an enhanced nonlocal spatial feature tensor.

[0062] Specifically, the hidden state tensor output by the ConvLSTM unit at the current time t As the input feature map for the nonlocal spatial attention enhancement module, the hidden state tensor has integrated multi-dimensional spatiotemporal features including historical combustion states, current combustible area distribution, newly added combustion sites, and wind field dynamic conditions; the hidden state The query feature map Q, key feature map K, and value feature map V are generated by 1×1 convolution mapping or linear projection transformation, respectively. Q, K, and V are expanded into a feature vector sequence containing H×W spatial locations in the spatial dimension. Then, the correlation weight between any two spatial locations is calculated, and after normalization by the Softmax function, the value feature map V is weighted and aggregated to obtain the enhanced non-local spatial feature representation. As an alternative implementation of the aforementioned attention mechanism, the non-local spatial attention enhancement module can also be configured as a local patch attention module, that is, only focusing on the area around the target pixel. Spatial attention aggregation is applied to the hidden state features within the neighborhood to enhance the network's ability to represent continuous fireline boundary details while reducing the computational overhead of global attention; preferably, the neighborhood window size... Set to 7.

[0063] S54, iteratively execute the steps to generate the hidden state at the current time step and the steps to obtain the enhanced hidden state in chronological order, until all input slices at all time steps have been processed.

[0064] S55 merges the current hidden state with the enhanced hidden state and inputs it into the output mapping layer to obtain the continuous wildfire spread response field for future time steps.

[0065] Specifically, the enhanced nonlocal spatial features are linked with the original hidden state using residual connections. Perform element-wise fusion and feed the fused features back to the subsequent time state update step or output mapping layer.

[0066] The pre-trained convolutional long short-term memory network used above, during its training process, uses real-time label burning. conduct Transformation, network output and using mean square error Training is performed. In one embodiment, the optimizer is Adam, with an initial learning rate of 0.001, 100 training epochs, and a batch size of 1 to 4 depending on the available memory. Training, validation, and testing are divided by event or year to ensure that the same fire event does not appear simultaneously in the training and test sets.

[0067] S6 performs thresholding on the continuous wildfire spread response field at future time points to generate a wildfire spread prediction map for future time points.

[0068] A response threshold parameter is preset. For any cell in the target spatial grid, if the response value of the cell in the continuous wildfire spread response field is greater than or equal to the response threshold parameter, the cell is determined to be a predicted spread cell. If the response value of the cell in the continuous wildfire spread response field is less than the response threshold parameter, it is determined to be a predicted non-spread cell, so as to generate a wildfire spread prediction map.

[0069] Specifically, a threshold parameter is applied to the continuous response field. This yields a binary spread prediction map for the next time step. The continuous response field refers to the consistent single-output wildfire spread prediction result for the next time step, specifically represented as a single-channel continuous raster image with the same grid size as the target space. The value of each pixel in this image represents the model's response intensity to the location at the next time step, which is not directly equivalent to the calibrated probability of occurrence. This is achieved by setting threshold parameters. This allows for the identification of pixels with response values ​​greater than or equal to a threshold as spreading pixels, and pixels with response values ​​below the threshold as non-spreading pixels, thereby generating a binary wildfire spread prediction map. In early warning scenarios, A threshold of 0.1 can be used to reduce the risk of false negatives. In scenarios where accuracy and recall are balanced, a threshold between 0.2 and 0.3 can be selected based on the validation set F1 value or ROC curve. If fire point data for a particular day is missing, the event state can be maintained using the most recent valid combustion mask and marked as a low-confidence input. If wind field data is missing, interpolation between adjacent time points or reanalysis of the most recent valid value of the product can be used. If information such as manual extinguishing, firebreaks, or manual removal of combustibles exists, it can be used as an external constraint mask to suppress and correct the predicted response field.

[0070] The following experiments verify the effectiveness of the wildfire spread prediction method proposed in this invention: This invention uses wildfire events in California from 2012 to 2021 as the verification object. VIIRS active fire products are used as the source for fire monitoring, MODIS NDVI products are used to assist in characterizing vegetation combustibility and fuel continuity, and the 10 m near-surface wind fields u10 and v10 from the ERA5-Land reanalysis product are used as meteorological driving variables. After unified projection, spatial registration, and diurnal-scale aggregation, all data sources are organized into a target-level wildfire event sequence.

[0071] In one specific implementation, the spatial resolution of the VIIRS active fire observation is 375 m, the spatial resolution of the MODIS NDVI is 250 m and the temporal resolution is a composite of 16 days, and the raw spatial resolution of the ERA5-Land wind field is approximately 0.1° and the temporal resolution is on an hourly scale. Each wildfire event is represented by a 100×100 pixel analysis window, corresponding to a spatial range of approximately 37.5 km × 37.5 km, and the wildfire spread response field for the next day is predicted using 10 historical time steps.

[0072] The model input channels include the unburned combustible zone U, the previously burning zone B, the newly added burning zone N, and the u10 and v10 wind components. U is determined by the NDVI and the current burning state, B by the burning state at the previous moment, and N by the difference in burning states between adjacent moments. The model output is a continuous wildfire spread response field transformed by log(1+Y), and can generate a binary spread prediction map for the next moment based on a threshold.

[0073] Figure 4 Target combustion mask for wildfire events provided in embodiments of the present invention ( Figure 4 Target in the model, predicted response field ( Figure 4 Prediction and residuals in Figure 4 A comparison diagram of Prediction Errors in [the context of the diagram]. Figure 4 This paper compares the standard combustion mask, model prediction response field, and residual map of models such as CNN-LSTM, Vision Transformer, ConvLSTM, Pairwise-ConvLSTM, Patchwise-ConvLSTM, and ST-GNN. It can be seen that CNN-LSTM and Vision Transformer have wider residual bands at the fire line boundary, and some prediction results exhibit smoothing or blockiness. ConvLSTM and its self-attention enhancement variant can better maintain the continuity of the fire line contour. Pairwise-ConvLSTM performs better in non-neighbor region association and background false alarm suppression. Patchwise-ConvLSTM shows more stability in local boundary continuity and precision-recall balance.

[0074] Figure 5 This is a schematic diagram comparing the target state, model prediction, and residual changes of continuously added combustion zones over historical time steps T1 to T10, as provided in an embodiment of the present invention. Figure 5 It is evident that the main errors are concentrated in the high-gradient boundary region of the fire line advance, while the residuals are smaller in the already burned interior and stable unburned background far from the fire line. Compared with CNN-LSTM and VisionTransformer, ConvLSTM, Pairwise-ConvLSTM, and Patchwise-ConvLSTM can maintain temporal continuity more stably and produce narrower residual bands at the fire line edge, indicating that attention-enhanced convolutional long short-term memory networks are beneficial for improving the prediction ability of wildfire spread boundaries for the next day.

[0075] To verify the effectiveness of the technical solution of this invention, the prediction performance of different models on the cumulative wildfire dataset and the newly added burning area dataset was compared. Tables 1 and 2 below list examples of the prediction performance of the main baseline model, the pairwise self-attention model, and the local patch attention model on the cumulative wildfire dataset. Evaluation metrics include Overall RMSE, Fire RMSE, Non-fire RMSE, Precision, Recall, and F1-Score. A lower RMSE indicates a smaller prediction response field error, while higher Precision, Recall, and F1-Score indicate better identification of the spread area after thresholding.

[0076] Table 1. Examples of prediction performance of different models on the cumulative wildfire dataset.

[0077] As shown in Table 1, Pairwise-ConvLSTM achieved the lowest Overall RMSE (0.0495), indicating that the pairwise self-attention mechanism helps reduce the overall prediction error. Patchwise-ConvLSTM achieved a combined result of Precision=0.962, Recall=0.949, and F1-Score=0.955 at Ri=7, indicating that local patch attention has a better balance between fire continuity and background false alarm control.

[0078] Table 2. Examples of prediction performance of different models on the newly added combustion zone dataset.

[0079] As shown in Table 2, in the newly added burning region, which better reflects the difficulty of advancing the fire line, the Overall RMSE of Pairwise-ConvLSTM is 0.0484, while Patchwise-ConvLSTM achieves Precision=0.916, Recall=0.888, and F1-Score=0.902 at Ri=7. Compared to CNN-LSTM and Vision Transformer, the self-attention-enhanced convolutional long short-term memory network has a stronger ability to recognize the boundaries of the newly added fire line, and can reduce over-prediction and under-prediction near the boundaries.

[0080] The experimental results above demonstrate that the self-attention-enhanced convolutional long short-term memory network provided in this invention exhibits more stable fireline boundary representation capabilities compared to CNN-LSTM and Vision Transformer. Pairwise-ConvLSTM shows superior overall error performance, achieving an Overall RMSE of 0.0495; Patchwise-ConvLSTM... R i When R_i=7, the overall performance of Precision=0.962, Recall=0.949, and F1=0.955 is achieved, indicating that local patch attention can maintain good fire line continuity while keeping background false alarms low. In the evaluation of newly added combustion areas, the Overall RMSE of Pairwise-ConvLSTM is 0.0484, and the Precision / Recall / F1 of Patchwise-ConvLSTM (R_i=7) are 0.916 / 0.888 / 0.902, further demonstrating the technical effectiveness of the state variable construction and attention-enhanced spatiotemporal prediction scheme for identifying newly added fire lines. Threshold sensitivity results show that τ=0.1 is suitable for recall-oriented early warning, and τ=0.2 to 0.3 generally achieves a more balanced performance in Precision, Recall, and F1.

[0081] It should be noted that the above embodiments use data from the California region as the verification object and do not imply that the present invention is limited to this region. When the target area has similar remote sensing observation conditions such as active fire, vegetation index, and wind field, those skilled in the art can implement the present invention according to the above registration, event construction, state variable calculation, and model training process. For areas with complex terrain, slope, aspect, or altitude can be used as extended covariates input into the network or used to construct adjacency weights; for scenarios with sufficient information on human-induced fire suppression, the response field can be corrected through external constraint masks. The conventional data download, coordinate transformation, tensor storage, and model deployment steps not described in detail in this specification can all be implemented using existing technologies in the field.

[0082] Based on the method described in the above embodiments, this embodiment will further describe the object-level wildfire spread prediction system from the perspective of a multi-source remote sensing driven system. The multi-source remote sensing driven system can be implemented as an independent entity or integrated into an electronic device. The electronic device can be a terminal, server, or other device. The terminal can include a tablet computer, a laptop computer, a personal computer (PC), a microprocessor box, or other devices.

[0083] Please see Figure 6 , Figure 6This invention specifically describes a multi-source remote sensing-driven object-level wildfire spread prediction system, applied in electronic devices. This multi-source remote sensing-driven object-level wildfire spread prediction system may include: The data acquisition module is used to acquire multi-source remote sensing data within the target area. The multi-source remote sensing data includes at least active fire observation data, vegetation index data, and near-surface wind field data. The spatiotemporal registration module is used to perform spatiotemporal unified processing on the multi-source remote sensing data; The object-level event construction module is used to aggregate the active fire observation data into a burning mask on a daily basis, mark the connected components of the daily burning mask, and construct an object-level wildfire event sequence based on the spatial overlap or adjacency relationship between burning patches on adjacent dates. The state variable generation module is used to generate wildfire state variables based on the object-level wildfire time series, the vegetation index data, and historical burning status. The spatiotemporal prediction module is used to input the wildfire state variables into a convolutional long short-term memory network to obtain the predicted continuous wildfire spread response field at future moments; The thresholding output module is used to threshold the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map at the future time.

[0084] In specific implementation, the above modules and / or units can be implemented as independent entities, or they can be arbitrarily combined and implemented as the same or several entities. For the specific implementation of the above modules and / or units, please refer to the previous method embodiments. For the specific beneficial effects that can be achieved, please also refer to the beneficial effects in the previous method embodiments, which will not be repeated here.

[0085] In addition, this embodiment of the invention also provides an electronic device, which may be a computer, tablet computer, or other similar device. This electronic device can implement the steps in any embodiment of the multi-source remote sensing-driven object-level wildfire spread prediction method provided in this embodiment of the invention. Therefore, it can achieve the beneficial effects that any multi-source remote sensing-driven object-level wildfire spread prediction method provided in this embodiment of the invention can achieve, as detailed in the preceding embodiments, and will not be repeated here.

[0086] Figure 7 A specific structural block diagram of an electronic device provided in an embodiment of the present invention is shown. This electronic device can be used to implement the multi-source remote sensing-driven object-level wildfire spread prediction method provided in the above embodiments. The electronic device 500 can be a terminal, server, or other device. The terminal can include a tablet computer, laptop computer, personal computer (PC), microprocessor box, or other devices.

[0087] The memory 520 can be used to store software programs and modules, such as the program instructions / modules corresponding to those in the above embodiments. The processor 580 executes various functional applications and data processing by running the software programs and modules stored in the memory 520. The memory 520 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 520 may further include memory remotely located relative to the processor 580, and these remote memories can be connected to the electronic device 500 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0088] The input unit 530 can be used to receive input numeric or character information, and to generate a keyboard and mouse related to user settings and function control.

[0089] Display unit 540 can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces, which can be composed of graphics, text, icons, video, and any combination thereof. Display unit 540 may include display panel 541, which may optionally be configured in the form of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar forms.

[0090] Electronic device 500, through transmission module 570 (e.g., Wi-Fi module), can help users receive requests, send information, etc., providing users with wireless broadband internet access. Although transmission module 570 is shown in the figure, it is understood that it is not an essential component of electronic device 500 and can be omitted as needed without changing the essence of the invention.

[0091] The processor 580 is the control center of the electronic device 500. It connects to various parts of the phone via various interfaces and lines, and performs various functions and processes data of the electronic device 500 by running or executing software programs and / or modules stored in the memory 520, and by calling data stored in the memory 520, thereby providing overall monitoring of the electronic device. Optionally, the processor 580 may include one or more processing cores; in some embodiments, the processor 580 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 580.

[0092] Electronic device 500 also includes a power supply 590 (such as a battery) that supplies power to various components. In some embodiments, the power supply may be logically connected to processor 580 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 590 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0093] Although not shown, the electronic device 500 also includes cameras (such as front-facing cameras and rear-facing cameras), Bluetooth modules, etc., which will not be described in detail here. Specifically, in this embodiment, the display unit of the electronic device is a touch screen display, and the mobile terminal also includes a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors. One or more programs contain instructions for performing the following operations: Acquire multi-source remote sensing data within the target area, wherein the multi-source remote sensing data includes at least active fire observation data, vegetation index data, and near-surface wind field data; The multi-source remote sensing data are subjected to spatiotemporal unified processing; The active fire observation data is aggregated into a burning mask on a daily basis. The burning masks of each day are labeled with connected components, and an object-level wildfire event sequence is constructed based on the spatial overlap or adjacency relationship between burning patches on adjacent dates. Based on the object-level wildfire time series, the vegetation index data, and historical combustion status, wildfire state variables are generated. The wildfire state variables are input into a convolutional long short-term memory network to obtain the predicted continuous wildfire spread response field at future moments; Thresholding is applied to the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map for the future time.

[0094] In practice, the above modules can be implemented as independent entities or combined in any way to be implemented as the same or several entities. For the specific implementation of the above modules, please refer to the previous method implementation examples, which will not be repeated here.

[0095] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. Therefore, embodiments of the present invention provide a storage medium storing multiple instructions that can be loaded by a processor to execute the steps of any embodiment of the multi-source remote sensing-driven object-level wildfire spread prediction method provided by the present invention.

[0096] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0097] Since the instructions stored in the storage medium can execute the steps in any embodiment of the multi-source remote sensing driven object-level wildfire spread prediction method provided in the embodiments of the present invention, the beneficial effects that any multi-source remote sensing driven object-level wildfire spread prediction method provided in the embodiments of the present invention can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0098] The foregoing has provided a detailed description of a multi-source remote sensing-driven object-level wildfire spread prediction method, system, storage medium, and electronic device provided by embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-source remote sensing-driven object-level wildfire spread prediction method, characterized in that, The method includes: Acquire multi-source remote sensing data within the target area, wherein the multi-source remote sensing data includes at least active fire observation data, vegetation index data, and near-surface wind field data; The multi-source remote sensing data are subjected to spatiotemporal unified processing; The active fire observation data is aggregated into a burning mask on a daily basis. The burning masks of each day are labeled with connected components, and an object-level wildfire event sequence is constructed based on the spatial overlap or adjacency relationship between burning patches on adjacent dates. Based on the object-level wildfire time series, the vegetation index data, and historical combustion status, wildfire state variables are generated. The wildfire state variables are input into a convolutional long short-term memory network to obtain the predicted continuous wildfire spread response field at future moments; Thresholding is applied to the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map for the future time.

2. The multi-source remote sensing-driven object-level wildfire spread prediction method according to claim 1, characterized in that, Spatiotemporal unified processing of the multi-source remote sensing data includes: Transform the spatial projection coordinate system of all data layers into a pre-defined target projection coordinate system; Based on the target projection coordinate system, a standard regular rectangular analysis grid is established, and the standard regular rectangular analysis grid is used as a unified spatial reference benchmark to define the cell spatial layout of each data layer. A spatial resampling algorithm is used to register all data layers to the standard rule analysis grid. Each data layer after resampling is cropped according to a preset common spatial range boundary to ensure that the number of pixels in the row and column directions of all data layers is completely consistent, and that each pixel corresponds in spatial position.

3. The multi-source remote sensing-driven object-level wildfire spread prediction method according to claim 1, characterized in that, The active fire observation data is aggregated into combustion masks on a daily basis. Connectivity is labeled for each daily combustion mask, and an object-level wildfire event sequence is constructed based on the spatial overlap or adjacency relationships between combustion patches on adjacent dates, including: A daily-scale binary combustion state indicator factor is constructed based on the active fire observation data acquired daily. The daily-scale binary combustion state indicator factor is assigned a value based on a preset assignment rule to obtain a combustion mask of the active fire observation data. The daily combustion masks were spatially clustered using a connected component labeling algorithm to obtain several independent combustion patches; Spatial overlap matching is performed on the fire patches extracted from adjacent dates. If the fire patch of the following day and the fire patch of the previous day meet the preset spatial continuity judgment rule, then the two are determined to belong to the same object-level wildfire event, and the latter is assigned to the object-level wildfire event sequence of the former.

4. The multi-source remote sensing-driven object-level wildfire spread prediction method according to claim 3, characterized in that, The preset assignment rules include pixel center falling rules and buffer rasterization rules; The pixel center falling rules include: if the geographic latitude and longitude coordinates of any VIIRS active fire point fall within the spatial coverage of any pixel in the target spatial grid, then the pixel is assigned a value of 1; if it does not fall within the spatial coverage of any pixel in the target spatial grid, then the pixel is assigned a value of 0. The buffer rasterization rule includes: taking each active fire point location as the center, setting a buffer radius that matches the spatial resolution of the VIIRS sensor; if the center point of the target pixel is within the buffer range, or if the spatial area covered by the target pixel has geometric intersection with the buffer, then the target pixel is assigned a value of 1; if the center point of the target pixel is not within the buffer range, and the spatial area covered by the target pixel does not have geometric intersection with the buffer, then the target pixel is assigned a value of 0.

5. The multi-source remote sensing-driven object-level wildfire spread prediction method according to claim 1, characterized in that, The convolutional long short-term memory network includes an input layer, convolutional gated recursive units, a non-local attention enhancement module, and an output mapping layer. The processing procedure of the convolutional long short-term memory network includes: The wildfire state variables are stacked along the time dimension to obtain a multi-channel spatiotemporal input tensor. The multi-channel spatiotemporal input tensor is then input into the input layer to obtain the input slice corresponding to the current moment. The current input slice is concatenated with the hidden state from the previous time step and then input into the convolutional gated recursive unit to generate the hidden state at the current time step. The current hidden state is input into the nonlocal spatial attention enhancement module to obtain the enhanced nonlocal spatial feature tensor. The enhanced nonlocal spatial feature tensor is added to the current hidden state to obtain the enhanced hidden state that integrates global spatial correlation information. The steps of generating the hidden state at the current time step and obtaining the enhanced hidden state are executed iteratively in chronological order until all input slices at all time steps have been processed. The current hidden state and the enhanced hidden state are fused and input into the output mapping layer to obtain the continuous wildfire spread response field for future time steps.

6. The multi-source remote sensing-driven object-level wildfire spread prediction method according to claim 5, characterized in that, The hidden state at the current moment is input into the nonlocal space attention enhancement module to obtain the enhanced nonlocal space feature tensor, including: The hidden state at the current moment is mapped to query vector Q, key vector K, and value vector V through convolutional blocks, and the spatial dimensions of query vector Q, key vector K, and value vector V are expanded into a sequence of H×W position vectors; where H is the number of pixels in the target spatial grid in the row direction, and W is the number of pixels in the target spatial grid in the row direction. Calculate the matrix product between the query vector Q and the transpose of the key vector K to obtain the pairwise relevance matrix, and use an activation function to normalize each row to obtain the spatial attention weight map. The spatial attention weight map is weighted and aggregated with the value feature V to generate an enhanced nonlocal spatial feature tensor.

7. The multi-source remote sensing-driven object-level wildfire spread prediction method according to claim 1, characterized in that, Thresholding is applied to the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map for the future time, including: A response threshold parameter is preset. For any cell in the target spatial grid, if the response value of the cell in the continuous wildfire spread response field is greater than or equal to the response threshold parameter, the cell is determined to be a predicted spread cell. If the response value of the cell in the continuous wildfire spread response field is less than the response threshold parameter, it is determined to be a predicted non-spread cell, so as to generate a wildfire spread prediction map.

8. A multi-source remote sensing-driven object-level wildfire spread prediction system, characterized in that, include: The data acquisition module is used to acquire multi-source remote sensing data within the target area. The multi-source remote sensing data includes at least active fire observation data, vegetation index data, and near-surface wind field data. The spatiotemporal registration module is used to perform spatiotemporal unified processing on the multi-source remote sensing data; The object-level event construction module is used to aggregate the active fire observation data into a burning mask on a daily basis, mark the connected components of the daily burning mask, and construct an object-level wildfire event sequence based on the spatial overlap or adjacency relationship between burning patches on adjacent dates. The state variable generation module is used to generate wildfire state variables based on the object-level wildfire time series, the vegetation index data, and historical burning status. The spatiotemporal prediction module is used to input the wildfire state variables into a convolutional long short-term memory network to obtain the predicted continuous wildfire spread response field at future moments; The thresholding output module is used to threshold the continuous wildfire spread response field at the future time to generate a wildfire spread prediction map at the future time.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted to be loaded by a processor to execute the multi-source remote sensing-driven object-level wildfire spread prediction method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, The method includes a processor and a memory, the processor being electrically connected to the memory, the memory being used to store instructions and data, and the processor being used to execute the steps in the multi-source remote sensing-driven object-level wildfire spread prediction method according to any one of claims 1 to 7.

Images