Satellite-based burned area intelligent identification and atmospheric pollution evaluation method and device

Abstract

The embodiment of the invention provides a method and device for intelligent recognition and air pollution evaluation of burned areas based on satellites, and relates to the technical field of image recognition, and the method comprises the following steps: constructing a comprehensive feature vector of each pixel; inputting the comprehensive feature vector into a pre-trained adaptive feature fusion network, outputting a dynamic combustion index of each pixel by the adaptive feature fusion network through a dynamic weight generation mechanism, and generating a dynamic combustion index graph; based on the dynamic combustion index diagram, adopting a self-adaptive threshold segmentation method to perform preliminary identification, and performing logic verification and optimization on a preliminary identification result; based on the burned area product and the dynamic combustion index corresponding to the burned area product, constructing an effective combustion intensity field; and inputting the effective combustion intensity field into a Bayesian time sequence estimation model, and carrying out inversion to obtain an aerosol optical thickness spatial distribution diagram. According to the scheme, physical significance and data driving are integrated, and the adaptivity and robustness of burned area identification are enhanced.

Description

A method and apparatus for intelligent identification of burned areas and assessment of air pollution based on satellite. Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and apparatus for intelligent identification of burned areas and assessment of air pollution based on satellite. Background Technology

[0002] Open burning of biomass releases large amounts of aerosols, pollutants, and greenhouse gases, significantly impacting regional air quality, public health, and climate change. It is a key area of ​​concern for global and Chinese atmospheric environmental governance and ecological protection. Developing biomass burning monitoring technologies is of significant practical importance for quantifying emission inventories, assessing ecological losses, and supporting environmental regulation and policy formulation. Satellite remote sensing-based monitoring technology is a crucial means for large-scale, routine monitoring. However, existing technologies still face significant challenges in terms of identification accuracy and assessment reliability.

[0003] At the identification level, current mainstream methods essentially rely on predefined, globally fixed spectral indices and discrimination thresholds. Because the spectral signals of biomass combustion are complexly modulated by multiple factors such as vegetation moisture, combustion intensity, and soil background, this static feature utilization paradigm struggles to adapt to the spatiotemporal heterogeneity of real-world scenarios. This results in insufficient generalization ability across different regions, vegetation types, and combustion states, leading to high rates of missed and false positives, and susceptibility to interference from bare soil and water bodies. These limitations in the identification process further exacerbate the uncertainty in subsequent assessment stages. Furthermore, while some researchers have proposed machine learning methods in recent years, these primarily involve inputting all variables and having the machine learning model perform opaque operations, resulting in weak interpretation of physical meaning. At the assessment level, existing common methods obtain binary burning patches and areas or acquire the thermal radiation value of burning ignition points, multiplying these by emission factors to estimate pollution emissions. This pollution assessment method mainly focuses on estimating emissions and does not consider meteorological factors or the real-time, continuous spatial impact on surrounding areas. In summary, how to achieve a high-precision method for identifying burned areas that is both data-driven and based on physical logic in an adaptive scenario, and on this basis, to realize a technology for spatially continuous real-time and near-real-time pollution and uncertainty assessment, has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a satellite-based intelligent identification method for burned areas and atmospheric pollution assessment, to solve the technical problem that existing burned area identification methods cannot simultaneously consider data-driven and physical logic. The method includes:

[0005] Multi-temporal satellite imagery data of the target area within a preset time range is acquired and preprocessed. Based on the preprocessed satellite imagery data, a comprehensive feature vector is constructed for each pixel. This comprehensive feature vector includes surface reflectance across multiple bands, various spectral indices, and the difference in spectral indices before and after burning. The comprehensive feature vector is input into a pre-trained adaptive feature fusion network. This network outputs a dynamic combustion index for each pixel through a dynamic weight generation mechanism, generating a dynamic combustion index map. The dynamic combustion index is used to quantitatively characterize the intensity of biomass combustion and provide source strength for pollution emission estimation. Based on the dynamic combustion index map, an adaptive threshold segmentation method is used to initially identify burned areas. The system identifies a region and uses a multidimensional knowledge base to perform logical verification and optimization on the preliminary identification results, generating high-confidence burned area products. Based on the wind speed and direction data from meteorological data, and using the burned area products and their corresponding dynamic combustion indices, it calculates effective combustion intensity values ​​and constructs an effective combustion intensity field. This field characterizes the cumulative influence intensity of upwind source pixels on each pixel through the effective combustion intensity value of each pixel. The effective combustion intensity field and aerosol optical thickness data obtained through satellite observation are input into a Bayesian time-series estimation model. The model then inverts the spatial distribution map of aerosol optical thickness contributed by biomass incineration, using this spatial distribution map as the result of air pollution assessment.

[0006] This invention also provides a satellite-based intelligent identification and air pollution assessment device for burned areas, addressing the technical problem that existing burned area identification methods cannot simultaneously consider data-driven and physical logic. The device includes: a feature vector extraction module, used to acquire multi-temporal satellite image data of a target area within a preset time range, preprocess the satellite image data, and construct a comprehensive feature vector for each pixel based on the preprocessed satellite image data. The comprehensive feature vector includes surface reflectance in multiple bands, multiple spectral indices, and the difference in spectral indices before and after burning; a dynamic combustion index calculation module, used to input the comprehensive feature vector into a pre-trained adaptive feature fusion network. The adaptive feature fusion network outputs a dynamic combustion index for each pixel through a dynamic weight generation mechanism and generates a dynamic combustion index map. The dynamic combustion index is used to quantitatively characterize the intensity of biomass combustion and provide source strength basis for pollution emission estimation; and a burned area product generation module, used to initially generate a product based on the dynamic combustion index map using an adaptive threshold segmentation method. The study identifies burned areas and uses a multi-dimensional knowledge base to logically verify and optimize the preliminary identification results, generating high-confidence burned area products. An effective combustion intensity field module is constructed to calculate the effective combustion intensity value based on the burned area products and their corresponding dynamic combustion indices using wind speed and direction data from meteorological data. This effective combustion intensity field characterizes the cumulative influence intensity of upwind source pixels on each pixel through the effective combustion intensity value. An aerosol thickness distribution map module is also constructed. This module inputs the effective combustion intensity field and aerosol optical thickness data obtained through satellite observation into a Bayesian time-series estimation model. The model then inverts the spatial distribution map of aerosol optical thickness contributed by biomass incineration, using this map as the result of air pollution assessment.

[0007] Compared with the prior art, the beneficial effects that can be achieved by the above-mentioned at least one technical solution adopted in the embodiments of this specification include at least the following: the satellite-based intelligent identification of burned areas and atmospheric pollution assessment method of the present invention integrates physical meaning and data-driven approach, thereby enhancing the adaptability and robustness of burned area identification. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 is a flowchart of a satellite-based intelligent identification and air pollution assessment method for burned areas provided by an embodiment of the present invention; Figure 2 is a schematic diagram of the loss curve based on the AFFNet training results; Figure 3 is a schematic diagram of the DBI probability distribution based on the AFFNet training results; Figure 4 is a schematic diagram of the performance index based on the AFFNet training results; Figure 5 is a schematic diagram of the burned area range; Figure 6 is a schematic diagram of the AOD of the incineration contribution calculated using the method of an embodiment of the present invention; Figure 7 is an example of the uncertainty distribution corresponding to the AOD in Figure 6; Figure 8 is a structural block diagram of a satellite-based intelligent identification and air pollution assessment device provided by an embodiment of the present invention. Detailed Implementation

[0010] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0011] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] In this embodiment of the invention, a method for intelligent identification of burned areas and assessment of air pollution based on satellite is provided, as shown in Figure 1. The method includes: Step S101: acquiring multi-temporal satellite image data of the target area within a preset time range, and preprocessing the satellite image data. Based on the preprocessed satellite image data, a comprehensive feature vector is constructed for each pixel, wherein the comprehensive feature vector includes surface reflectance of multiple bands, multiple spectral indices, and the difference between spectral indices before and after burning; Step S102: inputting the comprehensive feature vector into a pre-trained adaptive feature fusion network, wherein the adaptive feature fusion network outputs a dynamic combustion index for each pixel through a dynamic weight generation mechanism, and generates a dynamic combustion index map, wherein the dynamic combustion index is used to quantitatively characterize the intensity of biomass combustion and provide source strength basis for pollution emission estimation; Step S103: based on the... A dynamic combustion index map is generated, and an adaptive threshold segmentation method is used to initially identify burned areas. A multi-dimensional knowledge base is then used to logically verify and optimize the initial identification results, generating high-confidence burned area products. Step S104: Based on the burned area products and their corresponding dynamic combustion indices, and using wind speed and direction data from meteorological data, an effective combustion intensity value is calculated, and an effective combustion intensity field is constructed. This field characterizes the cumulative influence intensity of upwind source pixels on each pixel through the effective combustion intensity value of each pixel. Step S105: The effective combustion intensity field and aerosol optical thickness data obtained through satellite observation are input into a Bayesian time-series estimation model. The Bayesian time-series estimation model is used to invert and obtain a spatial distribution map of aerosol optical thickness contributed by biomass incineration. This spatial distribution map is then used as the result of air pollution assessment.

[0013] In specific implementation, the following steps are used to construct a comprehensive feature vector for each pixel based on the preprocessed satellite image data: obtaining vegetation cover, normalized building index (NDI), and normalized water index (NDI) from the preprocessed satellite image data; removing areas from the preprocessed satellite image data where the vegetation cover is greater than or equal to a first threshold, the NDI is greater than or equal to a second threshold, and the NDI is greater than or equal to a third threshold, to obtain the image data to be analyzed; and extracting the reflectance of each pixel in multiple key bands from the image data to be analyzed. In this process, the key bands include the shortwave infrared band, the near-infrared band, and the visible light band. Based on the reflectance of multiple bands, an improved combustion index NBR' is calculated, and the improved combustion index NBR' is used as one of multiple spectral indices. For the multiple spectral indices, the difference between the multiple spectral indices of the same pixel before and after incineration is calculated as the difference before and after incineration. Based on the reflectance of the multiple bands, the multiple spectral indices, and the difference before and after incineration of the spectral indices, a comprehensive feature vector is generated to characterize the spectral state and dynamic changes of each pixel.

[0014] In specific implementation, the improved combustion index NBR' is calculated based on the reflectivity of multiple key wavebands through the following steps: NBR' = (ρ SWIR2 ρ NIR ) / ((ρ SWIR2 +ρ NIR )×(1+NDVI pre )), where ρ SWIR2 ρ is the reflectivity in the shortwave infrared band. NIR For near-infrared reflectance, NDVI pre The normalized vegetation index is the value before the fire.

[0015] In specific implementation, the comprehensive feature vector is input into a pre-trained adaptive feature fusion network through the following steps. The adaptive feature fusion network outputs a dynamic burning index for each pixel through a dynamic weight generation mechanism and generates a dynamic burning index map: Constructing a function that minimizes the target loss function. An adaptive feature fusion network is trained by minimizing the objective loss function, wherein, For loss, To measure the consistency between the predicted dynamic combustion index and the actual label, the binary cross-entropy loss is used. For weight difference loss, For losses with a clear direction, and All are balance coefficients; each original feature value in the comprehensive feature vector is standardized by using the mean and standard deviation calculated from the global training data to generate the corresponding standardized feature value. ,in, , For the first Standardized feature values ​​of each feature For the first 1 eigenvalue, For the first The mean of each feature, For the first The standard deviation of each feature; based on the standardized feature value of each feature. Parameters of the corresponding learnable direction Construct a feature transformation function based on direction parameters, and apply the feature transformation function to the standardized eigenvalues. A nonlinear transformation is performed to generate transformed eigenvalues, where the eigenvalue transformation function is based on the direction parameter. , It is an exponential function. For the first The orientation parameters of each feature For the first Standardized feature values ​​of each feature A preset threshold is set, and C is an optional constant; the original molecular weight is set for each transformed feature value. and the original denominator weights The original numerator weights and the original denominator weights are then normalized by absolute value to obtain the normalized numerator weights. and normalized denominator weights ,in, , For the total number of features, Assign feature numbers; based on all transformed feature values ​​and the normalized molecule weights and the normalized denominator weights A dynamic combustion index is generated for each pixel; based on the dynamic combustion index of all pixels in the target area and the spatial location of all pixels, a dynamic combustion index map is generated.

[0016] In specific implementation, the following steps are used to achieve the following: based on all transformed eigenvalues ​​and the normalized numerator weights. and the normalized denominator weights Generate the dynamic burning index for each pixel: Dynamic Burning Index ,in, For learnable bias terms, To prevent small constants from being divided by zero.

[0017] In specific implementation, the following steps are used to initially identify burned areas based on the dynamic combustion index map using an adaptive threshold segmentation method, and to logically verify and optimize the preliminary identification results using a multidimensional knowledge base to generate a high-confidence burned area product: Adaptive threshold segmentation is performed on the dynamic combustion index map to obtain the binarized range of the burned area; logical verification and optimization are performed on the binarized range of the burned area based on the multidimensional knowledge base to generate logically verified patches; edge smoothing and hole filling operations are performed on the logically verified patches, and the evolution process of each patch is tracked in the multi-temporal satellite image data; the temporal result with the highest dynamic combustion index value for each patch during the evolution process is selected as the final high-confidence burned area product.

[0018] In specific implementation, the effective combustion intensity value is calculated based on the wind speed and direction in the meteorological data, the burned area product, and the corresponding dynamic combustion index. This is achieved by repeating the following steps until all target pixels within the target area are traversed, generating a spatially continuous effective combustion intensity field: The upwind region of the target pixel is determined based on the wind speed and direction in the meteorological data; the diffusion scale parameter is determined based on the wind speed and atmospheric stability parameters; the upwind source pixel of the refined burned area located within the upwind region and the corresponding dynamic combustion index are obtained; the dynamic combustion index is standardized and scaled to generate a standardized dynamic combustion index; and the effective combustion intensity value of the target pixel is calculated based on the upwind source pixel, the standardized dynamic combustion index of the upwind source pixel, and the diffusion scale parameter.

[0019] In specific implementation, the effective combustion intensity value of the target pixel is calculated based on the upwind source pixel, the dynamic combustion index of the upwind source pixel, and the diffusion scale parameter through the following steps: calculating the Euclidean distance between the upwind source pixel and the target pixel. ,in, , For the target pixel, For upwind source pixels; obtain diffusion scale parameters using diffusion parameter tables or empirical formulas based on wind speed and atmospheric stability categories. ; through the upwind source pixel, the dynamic combustion index of the upwind source pixel, and the Euclidean distance and the diffusion scale parameter The effective combustion intensity value of the target pixel is calculated, wherein the effective combustion intensity value of the target pixel is... ,in, Upwind source pixels Standardized dynamic combustion index, Let p be the set of source pixels located upwind of the target pixel p.

[0020] In specific implementation, the following steps are used to obtain the spatial distribution map of aerosol optical thickness contributed by biomass incineration through the Bayesian time-series estimation model: obtaining the effective combustion intensity field and corresponding satellite-observed aerosol optical thickness of the target region on continuous time slices; constructing a Bayesian linear regression model for each time slice t, which is used to invert the aerosol optical thickness contributed by incineration, wherein... , For satellite-monitored aerosol optical thickness, Emission efficiency coefficient The intercept term for the optical thickness of background aerosols. To conform to a mean of zero and a variance of The observation error of the normal distribution is calculated; a prior distribution is applied to the parameters of the Bayesian linear regression model, and the posterior distribution is obtained based on the Bayesian linear regression model and the prior distribution to generate the emission efficiency coefficient. Intercept term of background aerosol optical thickness The posterior distribution; based on the posterior distribution and the effective combustion intensity field, the aerosol optical thickness value contributed by biomass incineration to each pixel is calculated, and an aerosol optical thickness spatial distribution map is output, wherein the aerosol optical thickness spatial distribution map includes the aerosol optical thickness value of each pixel and the corresponding confidence interval.

[0021] In one embodiment of the present invention, a method for intelligent identification of burned areas and assessment of air pollution based on satellite is provided, including the following steps: Step 1, satellite data preprocessing and feature vector preparation.

[0022] Acquire medium- to high-resolution satellite time-series images of the target area before and after the peak burning period, and perform standard preprocessing such as radiometric calibration, atmospheric correction, cloud and cloud shadow masking. Construct sample sets of burned areas (positive samples) and other types (negative samples). Develop a high-dimensional comprehensive feature vector for each sample pixel, including surface reflectance in key bands sensitive to burning (such as visible light, near-infrared, and shortwave infrared), as well as various spectral indices that can indicate vegetation status, burning traces, and background interference (such as the improved burn index NBR', enhanced vegetation index EVI, bare land index, etc.), and the difference in spectral indices before and after burning, to capture the dynamic spectral changes caused by burning.

[0023] Acquire medium- to high-resolution satellite time-series imagery (such as Sentinel-2 and Landsat-8 / 9) of the target area before and after peak biomass burning periods, and perform preprocessing including radiometric calibration, atmospheric correction, and cloud and cloud shadow masking. Acquire contemporaneous auxiliary data, including land surface type data, precipitation, wind direction, wind speed, and other meteorological parameters, as well as aerosol optical depth (AOD) products.

[0024] Step 1.1: Preliminary screening of potential incineration areas.

[0025] Image mask extraction is performed by combining high-precision forest, grassland, and farmland classification data. Furthermore, based on the vegetation cover (VFC), normalized building index (NDBI), and normalized water index (NDWI) calculated from the preprocessed images, areas with VFC ≥ θ1, NDWI ≥ θ2, or NDBI ≥ θ3 (θ1 (first threshold), θ2 (second threshold), and θ3 (third threshold) are empirical thresholds; typical values ​​for θ1 are 0.75-0.85, for θ2 are 0.2-0.4, and for θ3 are 0.1-0.3) are excluded. Pixels within the masked area that are densely vegetated and clearly do not show signs of burning, such as high-confidence buildings and water bodies, are also excluded.

[0026] Step 1.2: Training sample preparation.

[0027] Within the pre-screened area, by fusing high-confidence active fire point products from the same period with visual interpretation results, the confirmed burned areas are manually or semi-automatically delineated as positive samples, while negative sample polygons are delineated in other vegetated areas, bare soil areas, etc. Pixels are randomly and uniformly sampled from the polygons, and their feature vectors and corresponding labels (burning, background) constitute the training and validation sets of the model.

[0028] Based on the preprocessed image, a comprehensive feature vector is constructed for each pixel. D represents the dimension of the input features. R is the set of real numbers. This vector contains: surface reflectance of multiple bands (i.e., key bands, with center wavelengths of approximately 490nm, 560nm, 665nm, 842nm, 1610nm, and 2190nm), combined spectral indices (such as Enhanced Vegetation Index (EVI), Improved Combustion Index (NBR'), Bare Land Index, Water Index, etc.), and temporal (before and after burning) spectral index differences. Among these, this invention proposes an improved combustion index NBR' = (ρ SWIR2 ρ NIR ) / ((ρ SWIR2 +ρ NIR )×(1+NDVI pre )), where ρ SWIR2 ρ is the reflectance in the short-wave infrared band (2190nm). NIRFor reflectance in the near-infrared band (842nm), NDVI pre This is the normalized vegetation index before the fire. The goal is to incorporate pre-fire vegetation background values ​​so that the index can more accurately reflect the relative burning intensity.

[0029] Step 2: Construction and training of a dynamic feature fusion network for adaptive enhancement of incineration features.

[0030] A dynamic weight generation mechanism is introduced to construct a fully connected neural network. The core idea is to dynamically generate a set of optimal feature fusion weights for each pixel through a data-driven approach, thereby adaptively enhancing the combustion spectral signal that varies due to differences in land surface type, vegetation moisture content, vegetation underlying surface, and combustion state.

[0031] An adaptive feature fusion network is constructed, trained by inputting the high-dimensional feature vector constructed in step 1, to generate the Dynamic Burning Index (DBI). This includes the following steps.

[0032] Step 2.1: Configure learnable directional parameters and corresponding statistical parameters for each input feature. First, standardize the features to obtain... This addresses the issue of preprocessing consistency in real-time deployment.

[0033] ,in, For the first 1 eigenvalue, and The mean and standard deviation of this feature are calculated from the global training data and stored as fixed parameters.

[0034] Set the direction parameters that can be learned. It automatically learns the feature directionality using the following formula. .in, For the first The orientation parameters of each feature; is the preset threshold; C is an optional constant representing a high background reference point in the standardized space. This is a feature transformation function based on direction parameters.

[0035] Step 2.2, Feature Weight Learning. For each feature... Set the numerator weights separately and denominator weights , ,in and These are the original network parameters. The total number of features is used to ensure that the weights are non-negative and normalized through the above calculations.

[0036] Step 2.3: Calculate the flammability index (DBI). .in, The total number of features; The numerator weight represents the degree to which this feature enhances the combustion signal; The weight in the denominator represents the importance of this feature in the background benchmark. For learnable bias terms; To prevent small constants from being divided by zero, the DBI is compared with a learnable global threshold θ during decision-making. When DBI > θ, the incineration feature signal is stronger than the background benchmark and is judged as an incineration area; when DBI < θ, the incineration feature signal is weaker than the background benchmark and is judged as a background area. The larger the DBI value, the more significant the incineration feature.

[0037] Step 2.4: Using an end-to-end training approach, design a loss function to optimize network parameters:

[0038] in, The cross-entropy loss is used for binary classification to ensure the consistency between the DBI value and the true label; To mitigate the weighting difference loss, greater weighting is encouraged for features important for incineration detection within the molecule. , To define the directionality of the loss, the direction parameter is encouraged to be far from zero to clarify the characteristic directionality. ; This is the balance coefficient.

[0039] Step 3: Identification and optimization of burned areas.

[0040] The trained network is applied to the imagery of the entire study area to obtain the DBI map. Then, an adaptive threshold segmentation algorithm, such as the Otsu algorithm, is used to initially extract the extent of burned areas. The following are optional steps: To further improve product accuracy, a multi-dimensional knowledge base containing information on land use, crop phenology, and meteorological conditions can be introduced to logically verify and optimize the preliminary results (e.g., removing patches with extremely unreasonable shapes, and combining precipitation data to eliminate interference from wet soil). The "appearance-clarity-fading" process of patches is tracked on the time-series imagery, and the clearest patch with the highest DBI is selected as the final high-confidence refined burned area vector product.

[0041] The trained adaptive feature fusion network was applied to the target area imagery to obtain a dynamic burning index map. An adaptive threshold segmentation algorithm was then used to obtain the preliminary binary fire area range. To improve recognition accuracy, a multi-dimensional knowledge base was constructed to serve the optimization process, and the preliminary results were logically verified and optimized. Morphological rules were applied to remove linear features with aspect ratios greater than a certain proportion (such as roads and riverbanks) or point noise with an area less than n pixels (n can be 4). Temporal rules were applied to exclude suspected patches appearing within crop growth windows by combining key phenological calendars for different vegetation types, but patches appearing during harvest, withering, or other historically possible burning periods were retained. Rainfall interference was compared: if suspected patches appeared after rainfall, they should be marked, and their presence in the preceding and following images should be used to determine whether they were wetland surfaces. The patches that pass the verification are subjected to edge smoothing, small hole filling and other operations, and their evolution process of "appearance-clarity-fading" is tracked in the time series images. The time phase result with the highest DBI value and the clearest morphology is selected as the final RBS product of the burned area.

[0042] Step 4: Pollution assessment and uncertainty quantification based on physical constraints.

[0043] Using the DBI spatial distribution output from step 2 or step 3 as input, the downwind air pollution impact is assessed. First, a pre-trained global model and global feature statistical benchmark are used, combined with local feature statistics of the target area, to standardize and scale the original DBI, generating a globally comparable combustion detection index DBI_s (Standardized Dynamic Combustion Index). An effective combustion intensity field is constructed, considering the spatial diffusion effect of pollutants. Using wind speed and direction data, the DBI_s values ​​at discrete points are transformed into a continuous spatial influence field—the Effective Combustion Intensity (EBI) field—through a diffusion model. The EBI quantitatively characterizes the cumulative influence intensity of all upwind combustion sources experienced by any downstream point.

[0044] Based on the spatial distribution of the incineration sites and their corresponding dynamic combustion indices obtained in step 2 or step 3, and taking any point p downwind as the target, calculate the cumulative impact of all incineration sources upwind on it, i.e., the effective combustion intensity EBIp. The calculation formula is as follows:

[0045] in, Represents a point Source pixels within the upwind refined burn site (RBS) (upwind source pixels).

[0046] Ω upwind(p) Represents all pairs of points The set of contributing upwind source pixels. It is the source pixel The comparable dynamic combustion index characterizes the relative combustion intensity at that point. For pixels With fire point pixels The Euclidean distance between them. The diffusion scale parameter L can be obtained by querying a predefined diffusion parameter table based on wind speed and atmospheric stability category (such as the Pasquale-Trnal stability classification), or by using the empirical formula L=a The calculation is performed using U / sqrt(N), where U is the wind speed, N is the stability level coefficient, and a is an empirical constant. This step transforms discrete incineration patches into a continuous spatial influence field.

[0047] Using EBI as the independent variable and AOD monitored by satellite as the dependent variable, a physically constrained Bayesian model is established to invert the AOD contributed by incineration. For each time slice... , assuming .in, Emission efficiency coefficient For the background AOD intercept term, To conform to a mean of zero and a variance of The observation error follows a normal distribution. The following prior distribution is applied to the key parameters of the above model: where, Obey the scale parameter is seminormal prior , For pre-defined scale parameters, the model is forced to learn a positive physical relationship: "the greater the combustion intensity, the higher the AOD contribution." Simultaneously, it is assumed that the emission efficiency coefficients of adjacent time slices are correlated. Characterizing the temporal continuity of the combustion emission process, among which, For a pre-defined random walk prior The standard deviation parameter in the figure, where N represents the normal distribution. Controlling the adjacent The range of change between The smaller the value, the stronger the temporal smoothness. Methods such as Markov chain Monte Carlo or variational inference are used to solve the problem, ultimately outputting a spatial distribution map of AOD contribution caused by biomass incineration, along with confidence intervals for the estimated value of each pixel.

[0048] The model is solved by Markov chain Monte Carlo or variational inference methods, and the spatial distribution of AOD contributed by incineration and the posterior probability distribution or confidence interval of the estimated value of each pixel are finally output, so as to achieve the unification of pollution assessment and uncertainty measurement.

[0049] To verify the effectiveness of the method of this invention, a dataset containing 2000 samples was constructed based on Sentinel-2 imagery, divided into training, validation, and independent test sets in a 7:2:1 ratio. Comparisons were made with the method of this invention, the traditional Normalized Burning Index (NBR) thresholding method, and the standard neural network method. The AFF-Net method achieved an overall accuracy of 95%, a recall rate of 96% for incineration samples, and a sample precision of 92%, showing significant improvement compared to the traditional Normalized Burning Index (NBR) thresholding method and also demonstrating improvement compared to the standard neural network method. (See Figures 2 to 4.)

[0050] Based on this model, a straw burning event in farmland in late June 2025 was monitored. Combined with meteorological data (northwest wind, level 2, neutral stability), the spatial distribution of AOD (Average Displacement of Occurrence) contributed by burning was obtained through physical constraint Bayesian model inversion. As shown in Figures 5 to 7, the inversion results show that in the surrounding area of ​​the main burning source, the AOD contributed by burning ranged from 0.2 to 1.0, with a 95% confidence interval width of 0.03 to 0.15.

[0051] Based on the same inventive concept, this invention also provides a satellite-based intelligent identification and air pollution assessment device for burned areas, as described in the following embodiments. Since the principle of the satellite-based intelligent identification and air pollution assessment device is similar to that of the satellite-based intelligent identification and air pollution assessment method, the implementation of the satellite-based intelligent identification and air pollution assessment device can refer to the implementation of the satellite-based intelligent identification and air pollution assessment method, and will not be repeated. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0052] Figure 8 is a structural block diagram of a satellite-based intelligent identification and atmospheric pollution assessment device according to an embodiment of the present invention. As shown in Figure 8, it includes: a feature vector extraction module 801, a dynamic combustion index calculation module 802, a burned area product generation module 803, an effective combustion intensity field construction module 804, and an aerosol thickness distribution map construction module 805. The structure will be described below.

[0053] The feature vector extraction module 801 is used to acquire multi-temporal satellite image data of the target area within a preset time range, and preprocess the satellite image data. Based on the preprocessed satellite image data, a comprehensive feature vector is constructed for each pixel. The comprehensive feature vector includes surface reflectance of multiple bands, multiple spectral indices, and the difference between spectral indices before and after burning. The dynamic combustion index calculation module 802 is used to input the comprehensive feature vector into a pre-trained adaptive feature fusion network. The adaptive feature fusion network outputs the dynamic combustion index of each pixel through a dynamic weight generation mechanism and generates a dynamic combustion index map. The dynamic combustion index is used to quantitatively characterize the intensity of biomass combustion and provide source strength basis for pollution emission estimation. The burned area product generation module 803 is used to preliminarily identify the burned area based on the dynamic combustion index map using an adaptive threshold segmentation method. The burned areas are identified, and the preliminary identification results are logically verified and optimized using a multi-dimensional knowledge base to generate high-confidence burned area products. An effective combustion intensity field module 804 is constructed to calculate the effective combustion intensity value based on the burned area products and the corresponding dynamic combustion index using wind speed and direction data from meteorological data, and to construct an effective combustion intensity field. The effective combustion intensity field is used to characterize the cumulative influence intensity of the upwind source pixels on each pixel through the effective combustion intensity value of the pixels. An aerosol thickness distribution map module 805 is constructed to input the effective combustion intensity field and aerosol optical thickness data obtained through satellite observation into a Bayesian time series estimation model. The Bayesian time series estimation model is used to invert and obtain the spatial distribution map of aerosol optical thickness contributed by biomass incineration, and the spatial distribution map of aerosol optical thickness is used as the result of air pollution assessment.

[0054] In one embodiment, the feature vector extraction module includes: a preprocessing unit for obtaining vegetation cover, normalized building index, and normalized water index from the preprocessed satellite image data; a data removal unit for removing areas from the preprocessed satellite image data where the vegetation cover is greater than or equal to a first threshold, the normalized water index is greater than or equal to a second threshold, and the normalized building index is greater than or equal to a third threshold, to obtain image data to be analyzed; and a reflectance extraction unit for extracting the reflectance of each pixel in multiple bands from the image data to be analyzed, wherein the multiple bands include shortwave infrared band, near-wave infrared band, and near-wave infrared band. The system includes an infrared band and a visible light band; an improved combustion index calculation unit, used to calculate the improved combustion index NBR' based on the reflectance of multiple bands, and to use the improved combustion index NBR' as one of multiple spectral indices; a difference calculation unit, used to calculate the difference between the multiple spectral indices of the same pixel before and after incineration, as the difference before and after incineration; and a comprehensive feature vector construction unit, used to generate a comprehensive feature vector characterizing the spectral state and dynamic changes of each pixel based on the reflectance of the multiple bands, the multiple spectral indices, and the difference before and after incineration of the spectral indices.

[0055] In one embodiment, the improved combustion index calculation unit is also used for NBR'= (ρ SWIR2 ρ NIR ) / ((ρ SWIR2 +ρ NIR )×(1+NDVI pre )), where ρ SWIR2 ρ is the reflectivity in the shortwave infrared band. NIR For near-infrared reflectance, NDVI pre The normalized vegetation index is the value before the fire.

[0056] In one embodiment, the dynamic combustion index calculation module includes: a loss function construction unit for constructing a target loss function that minimizes the target loss function. An adaptive feature fusion network is trained by minimizing the objective loss function, wherein, For loss, To measure the consistency between the predicted dynamic combustion index and the actual label, the binary cross-entropy loss is used. For weight difference loss, For losses with a clear direction, and All are balance coefficients; the standardization processing unit is used to standardize each original feature value in the comprehensive feature vector, generating the corresponding standardized feature value using the mean and standard deviation calculated from the global training data. ,in, , For the first Standardized feature values ​​of each feature For the first 1 eigenvalue, For the first The mean of each feature, For the first The standard deviation of each feature; an eigenvalue transformation unit, used for transforming each of the standardized eigenvalues. Parameters of the corresponding learnable direction Construct a feature transformation function based on direction parameters, and apply the feature transformation function to the standardized eigenvalues. A nonlinear transformation is performed to generate transformed eigenvalues, where the eigenvalue transformation function is based on the direction parameter. , It is an exponential function. For the first The orientation parameters of each feature For the first Standardized feature values ​​of each feature The preset threshold is C, which is an optional constant; the normalization unit is used to set the original molecular weights for each transformed feature value. and the original denominator weights The original numerator weights and the original denominator weights are then normalized by absolute value to obtain the normalized numerator weights. and normalized denominator weights ,in, , For the total number of features, Feature numbering; Pixel dynamic combustion index calculation unit, used to calculate based on all transformed feature values ​​and the normalized molecular weights. and the normalized denominator weights The dynamic combustion index of each pixel is generated; the index map generation unit is used to generate a dynamic combustion index map based on the dynamic combustion index of all pixels in the target area and the spatial position of all pixels.

[0057] In one embodiment, the pixel dynamic combustion index calculation unit is also used for dynamic combustion index calculation. ,in, For learnable bias terms, To prevent small constants from being divided by zero.

[0058] In one embodiment, a burned area product generation module includes: a binarization range acquisition unit, used to perform adaptive threshold segmentation on the dynamic combustion index map to obtain a binarized range of the burned area; a logically verified patch generation unit, used to perform logical verification and optimization on the binarized range of the burned area based on a multi-dimensional knowledge base to generate logically verified patches; an evolution process acquisition unit, used to perform edge smoothing and hole filling operations on the logically verified patches, and track the evolution process of each patch in the multi-temporal satellite image data; and a burned area product generation unit, used to select the temporal result with the highest dynamic combustion index value of each patch during the evolution process as the final high-confidence burned area product.

[0059] In one embodiment, an effective combustion intensity field construction module includes: a looping unit, used to repeat the following steps until all target pixels within the target area are traversed to generate a spatially continuous effective combustion intensity field; an upwind region determination unit, used to determine the upwind region of the target pixel based on wind speed and wind direction in meteorological data; a diffusion scale parameter determination unit, used to determine the diffusion scale parameter based on the wind speed and atmospheric stability parameter; a dynamic combustion index acquisition unit, used to acquire the upwind source pixel of the refined burning site located within the upwind region and the dynamic combustion index corresponding to the upwind source pixel, and to standardize and scale the dynamic combustion index to generate a standardized dynamic combustion index; and an intensity value calculation unit, used to calculate the effective combustion intensity value of the target pixel based on the upwind source pixel, the standardized dynamic combustion index of the upwind source pixel, and the diffusion scale parameter.

[0060] In one embodiment, the intensity value calculation unit is further configured to calculate the Euclidean distance between the upwind source pixel and the target pixel. ,in, , For the target pixel, For upwind source pixels; obtain diffusion scale parameters using diffusion parameter tables or empirical formulas based on wind speed and atmospheric stability categories. ; through the upwind source pixel, the dynamic combustion index of the upwind source pixel, and the Euclidean distance and the diffusion scale parameter The effective combustion intensity value of the target pixel is calculated, wherein the effective combustion intensity value of the target pixel is... ,in, Upwind source pixels Standardized dynamic combustion index, Let p be the set of source pixels located upwind of the target pixel p.

[0061] In one embodiment, the aerosol thickness distribution map construction module includes: a data acquisition unit for acquiring the effective combustion intensity field and corresponding satellite-observed aerosol optical thickness of the target region on continuous time slices; and a Bayesian linear regression model construction unit for constructing a Bayesian linear regression model for each time slice t, wherein the Bayesian linear regression model is used to invert the aerosol optical thickness contributed by combustion. , For satellite-monitored aerosol optical thickness, Emission efficiency coefficient The intercept term for the optical thickness of background aerosols. To conform to a mean of zero and a variance of The observation error of the normal distribution is calculated; the posterior distribution unit is solved to apply the prior distribution to the parameters of the Bayesian linear regression model; based on the Bayesian linear regression model and the prior distribution, the posterior distribution is obtained, and the emission efficiency coefficient is generated. Intercept term of background aerosol optical thickness The posterior distribution is obtained; a spatial distribution map unit is constructed to calculate the aerosol optical thickness value contributed by biomass incineration to each pixel based on the posterior distribution and the effective combustion intensity field, and output an aerosol optical thickness spatial distribution map, wherein the aerosol optical thickness spatial distribution map includes the aerosol optical thickness value of each pixel and the corresponding confidence interval.

[0062] The embodiments of this invention achieve the following technical effects: The neural network-based identification method of this invention is no longer a purely opaque operation, but rather integrates its physical meaning and data-driven approach, enhancing the adaptability and robustness of burned area identification, and enabling near real-time, spatially continuous assessment of atmospheric pollution impacts. Furthermore, its high degree of automation significantly reduces labor costs, providing an efficient and reliable technical tool for large-scale, operational biomass incineration monitoring and assessment. The embodiments of this invention, by constructing an adaptive feature fusion network, configure learnable directional parameters and weights for each spectral feature, dynamically generating a combustion index (DBI) for the current scenario, overcoming the limitations of traditional methods that rely on fixed spectral indices and global thresholds. This network can adaptively enhance burning feature signals based on factors such as land surface type, vegetation moisture, and burning intensity, significantly reducing misjudgments and missed judgments caused by interference from bare soil, water bodies, and cloud shadows. It maintains high recall and high precision even in complex geographical environments. Unlike the "black box" operation of traditional machine learning models, the dynamic burning index of this invention adopts a weighted ratio structure of numerator and denominator. The numerator enhances the burning signal, the denominator represents the background benchmark, the directional parameter clearly defines the direction of feature increase or decrease, and the loss function explicitly encourages important features to enter the numerator and the directional parameter to move away from zero. This design gives the parameters learned by the network clear physical meaning, and the model behavior is traceable and interpretable, facilitating engineering deployment and business applications. This invention introduces a multi-dimensional knowledge base, including land use, crop phenology, and precipitation time series, into the post-processing workflow of burned areas. It performs multi-dimensional logical verification on the preliminary identification results, including morphological rules, time windows, and precipitation interference, and combines time series images to track the evolution process of patches from "appearance-clarity-fading," ultimately outputting the temporal result with the highest dynamic burning index and the clearest morphology. This optimization strategy effectively eliminates false burn patches along roads, wetland surfaces, and during non-burning periods, generating high-confidence, high-spatial-accuracy burn patch vector products. Based on meteorological wind fields and Gaussian diffusion models, this invention transforms discrete burn patch structures and their dynamic combustion indices into a spatially continuous effective combustion intensity field (EBI). This quantitatively characterizes the cumulative impact intensity of all upwind combustion sources at any downstream location, extending pollution assessment from total emission estimation to spatially continuous impact mapping. This provides high-resolution, near-real-time technical support for regional air quality monitoring and population exposure assessment. Furthermore, this invention constructs a Bayesian hierarchical model embedding semi-normal and time-series smoothing priors, using the effective combustion intensity field as the independent variable and satellite aerosol optical thickness as the dependent variable to invert the spatial distribution of air odor caused by biomass burning.The semi-normal prior forces the model to learn the positive physical relationship that "the stronger the combustion, the greater the contribution," and the temporal smoothing prior uses multi-temporal information to stabilize parameter estimation. For the first time, it achieves direct separation of combustion source contribution and background aerosols from satellite observations without the support of emission factor inventory. In the Bayesian inversion stage, the embodiment of this invention outputs the posterior probability distribution and confidence interval of each pixel estimate, so that the pollution assessment results not only include mean estimates but also come with clear confidence levels. The uncertainty measurement information provides risk boundaries for environmental regulatory decisions, significantly improving the scientificity and usability of the assessment results and filling the gap in the lack of uncertainty quantification in existing biomass incineration pollution assessment technologies. The method of this embodiment of the invention does not depend on specific satellite sensors, specific regions, or specific vegetation types. The trained adaptive feature fusion network and Bayesian model can be directly transferred to satellite data of different geographical regions and different resolutions, and have good cross-scenario generalization ability. It can be widely used in environmental monitoring operations such as farmland straw burning, forest and grassland fires, and open biomass burning.

[0063] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of burned areas and assessment of air pollution based on satellite, characterized in that, include: Multi-temporal satellite imagery data of the target area within a preset time range is acquired and preprocessed. Based on the preprocessed satellite imagery data, a comprehensive feature vector is constructed for each pixel. This comprehensive feature vector includes surface reflectance across multiple bands, various spectral indices, and the difference in spectral indices before and after burning. The comprehensive feature vector is input into a pre-trained adaptive feature fusion network. This network outputs a dynamic combustion index for each pixel through a dynamic weight generation mechanism, generating a dynamic combustion index map. The dynamic combustion index is used to quantitatively characterize the intensity of biomass combustion and provide source strength for pollution emission estimation. Based on the dynamic combustion index map, an adaptive threshold segmentation method is used to initially identify burned areas. The system identifies a region and uses a multidimensional knowledge base to perform logical verification and optimization on the preliminary identification results, generating high-confidence burned area products. Based on the wind speed and direction data from meteorological data, and using the burned area products and their corresponding dynamic combustion indices, it calculates effective combustion intensity values ​​and constructs an effective combustion intensity field. This field characterizes the cumulative influence intensity of upwind source pixels on each pixel through the effective combustion intensity value of each pixel. The effective combustion intensity field and aerosol optical thickness data obtained through satellite observation are input into a Bayesian time-series estimation model. The model then inverts the spatial distribution map of aerosol optical thickness contributed by biomass incineration, using this spatial distribution map as the result of air pollution assessment.

2. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 1, characterized in that, Based on the preprocessed satellite image data, a comprehensive feature vector is constructed for each pixel, including: obtaining vegetation cover, normalized building index, and normalized water index from the preprocessed satellite image data; removing areas from the preprocessed satellite image data where the vegetation cover is greater than or equal to a first threshold, the normalized water index is greater than or equal to a second threshold, and the normalized building index is greater than or equal to a third threshold, to obtain image data to be analyzed; and extracting the reflectance of each pixel in multiple bands from the image data to be analyzed, wherein multiple bands... The spectrum includes shortwave infrared, near-infrared, and visible light bands. Based on the reflectance of multiple bands, an improved combustion index NBR' is calculated, and the improved combustion index NBR' is used as one of multiple spectral indices. For the multiple spectral indices, the difference between the multiple spectral indices of the same pixel before and after incineration is calculated as the difference before and after incineration. Based on the reflectance of the multiple bands, the multiple spectral indices, and the difference before and after incineration of the spectral indices, a comprehensive feature vector is generated to characterize the spectral state and dynamic changes of each pixel.

3. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 2, characterized in that, Based on the reflectivity of multiple wavebands, the improved combustion index NBR' is calculated, including: NBR' = (ρ SWIR2 ρ NIR ) / ((ρ SWIR2 +ρ NIR )×(1+NDVI pre )), where ρ SWIR2 ρ is the reflectivity in the shortwave infrared band. NIR For near-infrared reflectance, NDVI pre The normalized vegetation index is the value before the fire.

4. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 1, characterized in that, The integrated feature vector is input into a pre-trained adaptive feature fusion network. The adaptive feature fusion network outputs a dynamic burning index for each pixel through a dynamic weight generation mechanism and generates a dynamic burning index map, including: constructing a function that minimizes the target loss function. An adaptive feature fusion network is trained by minimizing the objective loss function, wherein, For loss, To measure the consistency between the predicted dynamic combustion index and the actual label, the binary cross-entropy loss is used. For weight difference loss, For losses with a clear direction, and All are balance coefficients; each original feature value in the comprehensive feature vector is standardized by using the mean and standard deviation calculated from the global training data to generate the corresponding standardized feature value. ,in, ， For the first Standardized feature values ​​of each feature For the first 1 eigenvalue, For the first The mean of each feature, For the first The standard deviation of each feature; based on the standardized feature value of each feature. Parameters of the corresponding learnable direction Construct a feature transformation function based on direction parameters, and apply the feature transformation function to the standardized eigenvalues. A nonlinear transformation is performed to generate transformed eigenvalues, where the eigenvalue transformation function is based on the direction parameter. ， It is an exponential function. For the first The orientation parameters of each feature For the first Standardized feature values ​​of each feature A preset threshold is set, and C is an optional constant; the original molecular weight is set for each transformed feature value. and the original denominator weights The original numerator weights and the original denominator weights are then normalized by absolute value to obtain the normalized numerator weights. and normalized denominator weights ,in, ， For the total number of features, Assign feature numbers; based on all transformed feature values ​​and the normalized molecule weights and the normalized denominator weights A dynamic combustion index is generated for each pixel; based on the dynamic combustion index of all pixels in the target area and the spatial location of all pixels, a dynamic combustion index map is generated.

5. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 4, characterized in that, Based on all transformed eigenvalues ​​and the normalized molecule weights and the normalized denominator weights Generate a dynamic burning index for each pixel, including: dynamic burning index ,in, For learnable bias terms, To prevent small constants from being divided by zero.

6. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 1, characterized in that, Based on the dynamic combustion index map, an adaptive threshold segmentation method is used to initially identify burned areas. A multidimensional knowledge base is then used to logically verify and optimize the initial identification results, generating a high-confidence burned area product. This includes: performing adaptive threshold segmentation on the dynamic combustion index map to obtain a binarized range of burned areas; performing logical verification and optimization on the binarized range of burned areas based on the multidimensional knowledge base to generate logically verified patches; performing edge smoothing and hole filling operations on the logically verified patches, and tracking the evolution process of each patch in multi-temporal satellite imagery data; selecting the temporal result with the highest dynamic combustion index value for each patch during the evolution process as the final high-confidence burned area product.

7. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 1, characterized in that, Based on the wind speed and direction in the meteorological data, and the dynamic combustion index corresponding to the burned area product, the effective combustion intensity value is calculated. This includes: repeating the following steps until all target pixels within the target area are traversed to generate a spatially continuous effective combustion intensity field: determining the upwind region of the target pixel based on the wind speed and direction in the meteorological data; determining the diffusion scale parameter based on the wind speed and atmospheric stability parameter; obtaining the upwind source pixel of the refined burned area located within the upwind region and the dynamic combustion index corresponding to the upwind source pixel, standardizing and scaling the dynamic combustion index to generate a standardized dynamic combustion index; and calculating the effective combustion intensity value of the target pixel based on the upwind source pixel, the standardized dynamic combustion index of the upwind source pixel, and the diffusion scale parameter.

8. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 7, characterized in that, Based on the upwind source pixel, the dynamic combustion index of the upwind source pixel, and the diffusion scale parameter, the effective combustion intensity value of the target pixel is calculated, including: calculating the Euclidean distance between the upwind source pixel and the target pixel. ,in, ， For the target pixel, For upwind source pixels; obtain diffusion scale parameters using diffusion parameter tables or empirical formulas based on wind speed and atmospheric stability categories. ; through the upwind source pixel, the dynamic combustion index of the upwind source pixel, and the Euclidean distance and the diffusion scale parameter The effective combustion intensity value of the target pixel is calculated, wherein the effective combustion intensity value of the target pixel is... ,in, Upwind source pixels Standardized dynamic combustion index, Let p be the set of source pixels located upwind of the target pixel p.

9. The method for intelligent identification of burned areas and assessment of air pollution based on satellite as described in claim 1, characterized in that, The spatial distribution map of aerosol optical thickness contributed by biomass incineration is obtained by inverting the Bayesian time-series estimation model, including: acquiring the effective combustion intensity field and corresponding satellite-observed aerosol optical thickness of the target region on continuous time slices; for each time slice t, constructing a Bayesian linear regression model, which is used to invert the aerosol optical thickness contributed by incineration, wherein... ， For satellite-monitored aerosol optical thickness, Emission efficiency coefficient The intercept term for the optical thickness of background aerosols. To conform to a mean of zero and a variance of The observation error of the normal distribution is calculated; a prior distribution is applied to the parameters of the Bayesian linear regression model, and the posterior distribution is obtained based on the Bayesian linear regression model and the prior distribution to generate the emission efficiency coefficient. Intercept term of background aerosol optical thickness The posterior distribution; based on the posterior distribution and the effective combustion intensity field, the aerosol optical thickness value contributed by biomass incineration to each pixel is calculated, and an aerosol optical thickness spatial distribution map is output, wherein the aerosol optical thickness spatial distribution map includes the aerosol optical thickness value of each pixel and the corresponding confidence interval.

10. A device for intelligent identification of burned areas and assessment of air pollution based on satellite, characterized in that, include: A feature vector extraction module is used to acquire multi-temporal satellite imagery data of the target area within a preset time range, and preprocess the satellite imagery data. Based on the preprocessed satellite imagery data, a comprehensive feature vector is constructed for each pixel. The comprehensive feature vector includes surface reflectance of multiple bands, multiple spectral indices, and the difference in spectral indices before and after burning. A dynamic combustion index calculation module is used to input the comprehensive feature vector into a pre-trained adaptive feature fusion network. The adaptive feature fusion network outputs a dynamic combustion index for each pixel through a dynamic weight generation mechanism and generates a dynamic combustion index map. The dynamic combustion index is used to quantitatively characterize the intensity of biomass combustion and provide source strength basis for pollution emission estimation. A burned area product generation module is used to preliminarily identify fires based on the dynamic combustion index map using an adaptive threshold segmentation method. The burned areas are identified, and the preliminary identification results are logically verified and optimized using a multi-dimensional knowledge base to generate high-confidence burned area products. An effective combustion intensity field module is constructed to calculate the effective combustion intensity value based on the burned area products and their corresponding dynamic combustion indices using wind speed and direction data from meteorological data. The effective combustion intensity field is used to characterize the cumulative influence intensity of the upwind source pixels on each pixel through the effective combustion intensity value of each pixel. An aerosol thickness distribution map module is constructed to input the effective combustion intensity field and aerosol optical thickness data obtained through satellite observation into a Bayesian time series estimation model. The Bayesian time series estimation model is used to invert and obtain the spatial distribution map of aerosol optical thickness contributed by biomass incineration. The spatial distribution map of aerosol optical thickness is used as the result of air pollution assessment.

Images