A vegetation coverage inversion method based on multi-source data fusion and LR-attention mechanism

Abstract

The application discloses a kind of vegetation coverage inversion methods based on multi-source data fusion and LR-Attention mechanism, method includes: to multi-source data set is aligned and feature extraction obtains multi-dimensional heterogeneous feature image data;After manifold algebra synthesis operation, it is obtained pure vegetation manifold matrix component by eliminating non-vegetation background interference processing;Topological gradient measure matrix is obtained using pure vegetation manifold matrix component, spatial heterogeneity perception weight and is cooperated with subgroup left coset projection dimension reduction processing and obtains simplified core representation matrix;Through manifold topological forest model, obtain pure vegetation manifold evolution abundance preliminary survey value;Weighted aggregation is obtained vegetation coverage prediction value in topological near neighbor pixel set.The application improves the inversion accuracy and reliability of vegetation coverage under complex disturbance scene such as mining area, with very high physical fidelity and robustness, provides accurate decision support for mining area ecological monitoring.

Description

Technical Field

[0001] This invention relates to the field of vegetation cover extraction and inversion based on multi-source data fusion, and particularly to a vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism. Background Technology

[0002] Vegetation cover refers to the percentage of the vertically projected area of ​​vegetation on the ground to the total area of ​​the statistical area. It is a core indicator for measuring the energy exchange, carbon sequestration capacity, and ecological health of an ecosystem. Achieving rapid and accurate estimation of vegetation cover is of great significance for the scientific regulation of natural resource utilization, maintaining regional ecological balance, and conducting ecological restoration assessments. Traditional methods for obtaining vegetation cover mainly rely on ground-based field surveys, such as needle sampling and quadrat sampling. While these methods offer high accuracy at single points, they suffer from inherent drawbacks such as high labor intensity, long statistical cycles, and limited spatial coverage, making it difficult to meet the real-time requirements of large-scale, long-term ecological monitoring. In recent years, satellite remote sensing technology has provided a powerful tool for macroscopic and dynamic monitoring of surface vegetation. Researchers often utilize remote sensing images to extract spectral features such as the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EDI) to construct linear or nonlinear inversion models. However, in complex scenarios such as open-pit mines where the ecology is extremely fragile and human disturbance is severe, traditional inversion methods face serious challenges: First, the surface background composition of mining areas is complex, with high-reflectivity slag, coal dust cover, and bare soil patches creating severe background noise interference; second, the phenomenon of "different objects with the same spectrum" is serious, resulting in a high degree of overlap between vegetation signals and non-vegetation background signals; third, the vegetation distribution in mining areas is highly fragmented, and traditional linear mixture models are prone to physical distortion in low-coverage areas, resulting in a large amount of salt-and-pepper noise in the inversion results.

[0003] Currently, although some studies have attempted to fuse multi-source remote sensing data to enrich feature representation, the following shortcomings remain: First, feature fusion is mostly shallow splicing, lacking organic coupling of heterogeneous features at the deep geometric structure, making it difficult to characterize the complex nonlinear ecological relocation of mining areas; second, insufficient utilization of spatial heterogeneity and topological constraints makes it difficult to maintain the integrity of vegetation manifold structure under background noise interference; third, conventional attention mechanisms are not combined with geometric priors such as geodesic distance, making them susceptible to spectral aliasing and focusing on the wrong background. Therefore, how to deeply fuse multi-source heterogeneous features, construct a nonlinear representation manifold capable of characterizing the complex ecological relocation of mining areas, and introduce geometric and topological constraints to enhance feature robustness and inversion accuracy has become a key problem urgently needing to be solved in the field of remote sensing monitoring of mining environment. Summary of the Invention

[0004] The purpose of this invention is to provide a vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism. Through the collaborative optimization of multi-source data fusion, manifold mechanism modeling, topological geometric constraints and neighborhood collaborative correction, the accuracy and reliability of vegetation cover estimation in complex disturbance scenarios such as mining areas are improved from multiple dimensions such as data representation, feature dimensionality reduction and model inversion.

[0005] The objective of this invention is achieved through the following technical solution: A vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism, the method includes: S1. Acquire multi-source datasets including UAV image data and satellite image data, and perform alignment processing and feature extraction on the multi-source datasets to obtain multi-dimensional heterogeneous feature image data. S2. Construct a high-order correlation potential field and use the correlation kernel function to obtain the similarity between any two pixels. Use the local evolution entropy function to obtain the feature adaptive channel weight vector and high-dimensional manifold matrix representation of each pixel. After manifold algebra synthesis operation and removal of non-vegetation background interference, obtain the pure vegetation manifold matrix components. S3. Spatial heterogeneity perception weights are constructed within a local spatial window using pure vegetation manifold matrix components. Topological gradient measure matrix is ​​obtained using pure vegetation manifold matrix components and spatial heterogeneity perception weights. Dimensionality reduction is achieved by synergistic subgroup left coset projection of pure vegetation manifold matrix components and topological gradient measure matrix to obtain a simplified core representation matrix. S4. Construct an LR-Attention hybrid model. The LR-Attention hybrid model obtains self-attention weights and topology-enhanced feature fields. The trained manifold topology forest model uses the topology-enhanced feature fields and feature adaptive channel weight vectors to obtain the initial measurement value of the pure vegetation manifold evolution abundance. S5. Obtain the topological nearest neighbor set of the target pixel, and perform weighted aggregation of the initial measured values ​​of the pure vegetation manifold evolution abundance of the nearest neighbor pixels within the topological nearest neighbor set to obtain the predicted value of vegetation cover of the target pixel.

[0006] To better implement this invention, in method S5, the distance to the target pixel is filtered. The topological nearest K cells constitute the topological nearest neighbor set. Target image The expression for the predicted vegetation cover value is as follows: ,in For pixels Topological nearest neighbor cells Topological gradient measure between For manifold scale parameters, Topological nearest neighbor pixels Preliminary values ​​of pure vegetation manifold evolution abundance. The normalization factor is... The expression is as follows: ; For affinity weights, topological nearest neighbor pixels With pixels The greater the topological distance, the lower the affinity weight.

[0007] Preferably, in method S1, the UAV image data in the multi-source dataset undergoes the following preprocessing: UAV image data and corresponding high-precision positioning data are acquired; feature point matching and aerial triangulation are performed using the Scale Invariant Feature Transform (SIFT) algorithm to obtain an orthophoto DOM and a corresponding digital surface model (DSM); the high-precision positioning data is the spatial coordinates at the time of UAV capture; radiometric correction is performed using the whiteboard correction method; the satellite image data in the multi-source dataset is Sentinel-2 image data, which undergoes the following preprocessing: cloud and fog noise removal is performed on each band of the satellite image data; resampling and resolution unification are performed; spatiotemporal alignment is performed on the UAV image data and satellite image data in the multi-source dataset; texture features are extracted from the UAV image data and spectral features are extracted from the satellite image data and fused to construct multidimensional heterogeneous feature image data. .

[0008] Preferably, the multidimensional heterogeneous feature image data Including texture features and spectral features, the higher-order correlation potential field is formed by texture features and spectral features, and the correlation kernel function expression is as follows: , For pixels With pixels similarity, , Multidimensional heterogeneous feature image data medium pixel With pixels The multidimensional features, , Each pixel With pixels spectral characteristics, , Each pixel With pixels Texture features, The manifold kernel bandwidth parameter is used to control the rate of similarity decay. As a spatial co-evolutionary adjustment factor, Represents a cell With pixels The square of the Euclidean distance of the texture features.

[0009] Preferably, in method S2, a local evolutionary entropy (LEE) feedback mechanism is introduced to quantify environmental uncertainty. The expression for the local evolutionary entropy function is as follows: ,in The total feature dimension of a pixel's multidimensional features. For pixels Feature Dimension The probability distribution, This is the disturbance intensity correction factor. For adjustment coefficients, Represents a cell The index is affected by environmental disturbances; the feature adaptive channel weight vector is dynamically generated on the feature axis pixel by pixel. Suppression and downsampling are applied to spectral bands affected by environmental pollution, and gain compensation is applied to high-reliability microtexture high-order fingerprints to obtain high-dimensional manifold pre-characterization elements after environmental weight correction and physical potential field reshaping.

[0010] Preferably, in method S2, the feature local autocorrelation symmetric positive definite matrix corresponding to each pixel is reconstructed, as shown in the following expression: , Let the square matrix be the result of the outer product of vectors. High-dimensional manifold matrix representations are obtained by mapping in the HEAM manifold or Lie group space. The algebraic composition operation of manifolds is performed by applying the mapping between the logarithm and exponent of the Lie algebra matrix, as shown in the following expression: ,in , Each pixel Pixel High-dimensional manifold matrix representation, For pixels Pixel The result after completing the algebraic composition operation of manifolds; express and Lie algebra field operations or Lie bracket operations; after removing non-vegetation background interference at the algebraic manifold level, the pure vegetation manifold matrix components are obtained. .

[0011] Preferably, in method S3, surface patches are obtained by acquiring the patchy fragmentation features of the mining area from multidimensional heterogeneous feature image data, and the spatial heterogeneity perception weight expression is as follows: ,in Represents a cell The spatial heterogeneity perception weights corresponding to feature dimension k; This represents the Gini index. Indicates in pixels Surface patches within a central local spatial window The probability of the area occupied. This represents the total number of surface patches within a local spatial window. Represented by pixels The set of the sliding spatial neighborhood within the local spatial window centered on the center, where N represents the set of the sliding spatial neighborhood. The total number of neighboring pixels contained within; Indicates neighboring pixels , components of the pure vegetation manifold matrix with feature dimension k; Let represent the mean of the pure vegetation manifold matrix components of all cell feature dimension k in the slip space neighborhood set. Let Frobenius norm squared.

[0012] Preferably, in method S3, the pixels in the topological gradient measure matrix and Topological gradient measure between The expression is as follows: ,in The physical fingerprint sensitivity adjustment parameter represents the feature dimension k. This is the inverse matrix operation of the global feature covariance matrix. Represents a cell Pixel The difference between the components of the pure vegetation manifold matrix in feature dimension k; the topological gradient measure matrix is ​​constructed by obtaining the topological gradient measure of any two pixels according to the above method.

[0013] Preferably, in method S4, the LR-Attention hybrid model uses the topological gradient measure matrix. With a simplified core representation matrix Construct a joint attention scoring function to calculate the pixel. Pixel Self-attention weights Utilizing self-attention weights For the simplified core representation matrix Topology enhancement is performed to obtain the topology-enhanced feature field. The manifold topological forest model consists of multiple decision trees and utilizes topology-enhanced feature fields. The feature-adaptive channel weight vector was used to obtain the initial measurement of the abundance of pure vegetation manifold evolution.

[0014] Preferably, the manifold topological forest model uses a modified Riemannian Gini index adapted to nonlinear manifold surfaces to construct the splitting criterion. The expression for the modified Riemannian Gini index is as follows: ,in To obtain the weight parameters of sample n from the feature adaptive channel weight vector, This represents the sum of the weights of all samples in the split node of the decision tree; This is the penalty intensity adjustment coefficient. For the local feature manifold structure corresponding to all samples within the split node of the decision tree Custom manifold compactness factor; ,in Represents the matrix logarithm operation. This represents the covariance matrix corresponding to the node samples. Represents the Frobenius matrix norm; when a decision tree split node is mixed with a non-vegetation background and If so, the decision tree is forced to abandon the corresponding interfering split path.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention constructs a multi-source dataset containing multispectral remote sensing images and UAV high-resolution images. By performing cross-scale spatiotemporal alignment and collaborative fusion, it effectively solves the heterogeneity problem of different sensors in terms of spatial resolution and spectral representation. The high-order texture fingerprint extracted from UAV images and the spectral features extracted from satellite images complement each other, which not only eliminates the limitation that a single image is difficult to distinguish between vegetation and mining land in detail, but also provides a rich physical representation basis for subsequent manifold modeling, significantly enhancing the ability to completely characterize the surface information of complex mining areas from the data source level. This invention achieves the following by constructing a high-order correlated potential field and introducing a feedback mechanism of niche correlation entropy and local evolution entropy: The invention achieves deep nonlinear coupling between spectral and texture features; it can dynamically generate adaptive channel weight vectors for features pixel by pixel, suppress and reduce contaminated spectral bands, and perform gain compensation for high-reliability texture fingerprints, effectively solving the problem that feature fusion in traditional methods only involves shallow splicing and has poor noise resistance; moreover, by mapping weighted features to the heterogeneous evolutionary algebraic manifold HEAM space, and utilizing the algebraic properties of Lie groups and inverse operations, the invention can accurately remove the perturbation shift of non-vegetation backgrounds such as slag and coal dust at the geometric level, and correct the benchmark shift caused by environmental interference from a physical mechanism, significantly improving the purity of weak vegetation signal extraction in scenes with foreign objects and mixed pixels.

[0016] (2) This invention utilizes pure vegetation manifold matrix components to construct spatial heterogeneity perception weights and topological gradient measurement matrices, which can accurately measure the geodesic distance of pixels on the path of surface vegetation evolution manifold, reducing complex multidimensional heterogeneous features to a measurement space with clear physical semantics; by constructing environmental disturbance invariant subgroups and performing dimensionality reduction processing of cooperative subgroup left coset projection, redundant noise dimensions affected by background conjugate space interference are eliminated in a targeted manner, realizing feature refinement without degrading quality, and significantly enhancing the feature expression stability and spatial consistency of the model in the context of patchy and fragmented open-pit mining areas.

[0017] (3) This invention embeds the topological gradient measure as a geometric prior constraint into the LR-Attention hybrid model. By using the manifold topological constraint attention mechanism, it forces the focus to the sample with the highest affinity on the manifold path, thereby eliminating the interference of misfocusing of objects with the same spectrum but different physical properties caused by similar spectra. This generates a topologically enhanced feature field with a high signal-to-noise ratio. The manifold topological forest model uses the modified Riemann Gini index adapted to the nonlinear manifold surface as the splitting standard, and combines the compactness factor to force the abandonment of the interference splitting path mixed with non-vegetation background. It can accurately distinguish between high reflectance soil noise and low-coverage vegetation boundary. Through the nearest neighbor aggregation correction based on topological geodesic distance weighting, it automatically repairs the isolated misjudged points and salt-and-pepper noise caused by the abrupt change of spatial fragmentation zone, so that the vegetation cover prediction value is significantly improved in terms of accuracy, spatial consistency and robustness.

[0018] (4) This invention improves the accuracy and reliability of vegetation cover estimation in complex disturbance scenarios such as mining areas through multi-source data fusion, manifold mechanism modeling, topological geometric constraints and neighborhood collaborative correction. It improves the accuracy and reliability of vegetation cover estimation in complex disturbance scenarios such as mining areas from multiple dimensions such as data representation, feature dimensionality reduction and model inversion. It not only significantly improves the inversion accuracy of sparse vegetation areas and high coverage saturation areas, but also has extremely high physical fidelity and robustness, providing accurate decision support for ecological monitoring, resource management and environmental assessment in mining areas. Attached Figure Description

[0019] Figure 1 This is a flowchart of the vegetation cover inversion method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of removing non-vegetation background in the embodiment; Figure 3 This is a schematic diagram illustrating the principle of obtaining the topologically enhanced feature field using the LR-Attention hybrid model in the embodiment. Figure 4 This is a schematic diagram illustrating the principle of obtaining the initial measurement of the pure vegetation manifold evolution abundance using the manifold topological forest model in the embodiment. Figure 5 This is a schematic diagram illustrating the principle of using weighted aggregation inversion of topological nearest neighbor pixel sets to predict vegetation cover in the embodiment. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to embodiments: Example like Figure 1 As shown, a vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism is proposed, the method comprising: S1. Acquire multi-source datasets, including UAV image data and satellite image data, and perform alignment processing and feature extraction on the multi-source datasets to obtain multi-dimensional heterogeneous feature image data.

[0021] In some embodiments, UAV imagery data and satellite imagery data in a multi-source dataset undergo preprocessing including atmospheric correction, radiometric correction, and geometric fine correction to achieve precise spatiotemporal alignment of data with different payloads and resolutions, thereby constructing a multi-source heterogeneous dataset (i.e., a multi-source dataset). This embodiment uses the Shendong mining area (110°52′-111°31′E, 39°23′-39°54′N) as the research area, acquiring UAV and satellite imagery data for the research area. The UAV imagery data (preferably high-resolution UAV imagery) in the multi-source dataset undergoes the following preprocessing: UAV imagery data and corresponding high-precision positioning data are acquired; feature point matching and aerial triangulation are performed using the Scale Invariant Feature Transform (SIFT) algorithm to obtain the orthophoto DOM and the corresponding digital surface model (DSM); the high-precision positioning data is the spatial coordinates at the time of UAV capture. Radiometric correction is performed using the whiteboard correction method. The satellite imagery data (preferably high-resolution satellite imagery) in the multi-source dataset is Sentinel-2 imagery data. The satellite imagery data undergoes the following preprocessing: cloud and fog noise removal is performed on each band of the satellite imagery data, and resampling is performed to unify the resolution. In the case study of the Shendong mining area, a DJI Phantom 4 RTK drone equipped with a 20-megapixel RGB visible light camera was used to conduct on-site image acquisition during the peak vegetation season (mid-July). The flight altitude was set at 120m, the overlap rate was 80%, and high spatial resolution images with a ground sampling interval (GSD) of 0.035m were acquired. Simultaneously, Sentinel-2 L2A level satellite imagery (e.g., downloaded from the European Space Agency (Copernicus Open Access Hub)) with a spatial resolution of 10m was acquired to ensure consistency between the satellite and drone data on the phenological baseline.

[0022] When preprocessing UAV imagery data in the Shendong mining area as a case study, the following technical solution can be adopted: First, the acquired raw images and corresponding RTK high-precision positioning data are imported into Pix4Dmapper, and the SIFT (Scale Invariant Feature Transform) algorithm is used for feature point matching and aerial triangulation to generate an orthophoto image (DOM) and digital surface model (DSM) with a spatial resolution of 0.05m. Precise cropping is then performed based on the vector boundaries of the mining area. Subsequently, using field-deployed calibration whiteboards, a linear regression model between grayscale values ​​(DN) and measured reflectance (Ref) is constructed using the whiteboard calibration method to eliminate radiation distortion caused by differences in illumination at different times. Finally, using the GLCM tool in ENVI software, with a step size of 1 in a 3×3 pixel window, eight types of texture feature variables, including second-order angular moment, entropy, contrast, and correlation, are calculated as physical fingerprints characterizing the microstructure of vegetation.

[0023] When using Sentinel-2 L2A imagery (i.e., satellite imagery) in the Shendong mining area as a case study, the following technical approach can be adopted: First, the images of the study area are screened using the quality assessment report of SNAP software, and invalid phases with cloud cover exceeding 10% are removed. For low cloud cover images, thin cloud pixels are identified based on the NDVI (Normalized Difference Vegetation Index) threshold determination method (with a threshold set at 0.12), and spatial replacement and restoration are performed using corresponding pixels from the previous ten-day cloudless image. Bilinear interpolation is used to uniformly resample the 20m resolution red-edge bands (B5, B6, B7, etc.) and shortwave infrared bands to 10m resolution. Finally, based on the resampled red band (B4) and near-infrared band (B8), an NDVI time-series dataset is calculated to support the dynamic characterization of vegetation growth and evolution.

[0024] Spatiotemporal alignment processing is performed on UAV imagery data and satellite imagery data from multiple sources. Texture features are extracted from the UAV imagery data and spectral features are extracted from the satellite imagery data, and then fused to construct multidimensional heterogeneous feature image data. In the Shendong mining area as a case study, the following technical solution can be adopted: Texture features are extracted from UAV imagery data, and spectral features are extracted from satellite imagery data. Within a 10m×10m satellite pixel grid, all UAV pixels falling within this grid are considered as a non-Euclidean statistical whole. The joint probability density distribution function of the texture features within the grid is calculated pixel by pixel, and high-order statistical combinations of these texture features, including local skewness, local kurtosis, and joint covariance matrix, are extracted and concatenated into a multidimensional vector as the internal texture feature (or microstructural heterogeneity texture feature). Furthermore, a super-resolution manifold mapping operator is used to spatially overlap and align the satellite spectral information (i.e., spectral features) with the UAV microstructural features (i.e., texture features). Finally, the network uses Savitzky-Golay (SG) filtering to smooth and denoise the satellite spectral trajectory, ultimately outputting encapsulated multidimensional heterogeneous feature image data. The total number of pixels in the image is denoted as The total dimension of the encapsulated features is denoted as The multidimensional feature corresponding to pixel i is denoted as .

[0025] S2. Construct a high-order correlation potential field and use the correlation kernel function to obtain the similarity between any two pixels; based on the preprocessed multidimensional heterogeneous feature image data. This invention constructs a heterogeneous ecologicaltope algebraic manifold (HEAM) adapted to the complex ecological succession process in mining areas. The HEAM manifold, through nonlinear algebraic mapping, preserves the algebraic symmetry of multi-source data while achieving deep mechanistic fusion of spectral and textural aspects, providing a framework for accurate extraction of vegetation signals in complex backgrounds. This invention preferably proposes a HEAM manifold modeling framework adapted to the complex ecological succession process in mining areas. It introduces niche association entropy (NCE) and local evolutionary entropy (LEE) to achieve adaptive weighting and collaborative mapping of multi-source features, and constructs a manifold synthesis mechanism based on Lie group algebraic structures. Combined with inverse operations, it achieves algebraic stripping of background disturbances such as slag and buildings, maintaining consistency between data geometry and algebraic structure while achieving accurate decoupling and enhancement of vegetation signals. In some embodiments, multidimensional heterogeneous feature image data... Including texture features and spectral features, a higher-order correlation potential field (constructed through the niche correlation entropy operator) is formed by texture features and spectral features. The correlation kernel function expression is as follows: , For pixels With pixels similarity, , Multidimensional heterogeneous feature image data medium pixel With pixels The multidimensional features, , Each pixel With pixels The spectral features (also known as spectral sub-feature vectors). , Each pixel With pixels Texture features (also known as texture sub-feature vectors). Example of a manifold kernel bandwidth parameter used to control the rate of similarity decay: The value is 0.45. This is a spatial co-evolution adjustment factor (constraining feature evolution to focus on the spatial topological dimension, allowing feature clustering to refer to the geographic spatial layout). Represents a cell With pixels The square of the Euclidean distance of the texture features. Represents a cell With pixels The square of the Euclidean distance of the spectral characteristics , The differences in texture dimension and spectral dimension between two pixels are measured separately. This means that if either the texture dimension or the spectral dimension differs significantly, the similarity between two pixels will decrease rapidly.

[0026] The adaptive channel weight vector and high-dimensional manifold matrix representation of each pixel are obtained using the local evoked entropy function. After manifold algebraic synthesis and removal of non-vegetation background interference, pure vegetation manifold matrix components are obtained. In some embodiments, a local evoked entropy (LEE) feedback mechanism is introduced to quantify environmental uncertainty. The expression for the local evoked entropy function is as follows: ,in The total feature dimension of a pixel's multidimensional features. For pixels Feature Dimension The probability distribution, The disturbance intensity correction coefficient is specifically adapted to mining scenarios (in this example, the value is set to 1.15 to additionally compensate for environmental fluctuations caused by mining disturbances). For adjustment coefficients, Represents a cell The index affected by environmental disturbances (also known as the pixel-specific entropy); dynamically generating adaptive channel weight vectors for each pixel along the feature axis. , For pixels Feature Dimension The channel weight vector is obtained; suppression and downsampling are performed on spectral bands affected by environmental pollution, and gain compensation is performed on high-reliability micro-texture high-order fingerprints, resulting in high-dimensional manifold pre-representation elements after environmental weight correction and physical potential field reshaping. The features are further upgraded to the Lie group matrix manifold space to achieve advanced operations for removing background noise. The feature outer product operation is used to reconstruct the feature local autocorrelation symmetric positive definite matrix corresponding to each pixel, as expressed below: , Let the square matrix be the result of the outer product of vectors. High-dimensional manifold matrix representations are obtained by mapping in the HEAM manifold or Lie group space. The algebraic composition operation of manifolds is performed by applying the mapping between the logarithm and exponent of the Lie algebra matrix, as shown in the following expression: ,in , Each pixel Pixel High-dimensional manifold matrix representation, For pixels Pixel The result after completing the algebraic composition operation of manifolds; Indicates to and Perform Lie algebra field operations or Lie bracket operations; after removing non-vegetation background interference at the algebraic manifold level, obtain the pure vegetation manifold matrix components. .

[0027] To accurately remove solid waste and high-reflectivity background noise from complex mixed pixels (including real vegetation features and background disturbances such as slag / solid waste piles / exposed rock and soil), this embodiment adopts the following method for removing non-vegetation background interference: Selecting a pristine, purely vegetated area with minimal disturbance within the mining area as the manifold unit cell. (Its purpose is to select the characteristics of natural, pure vegetation plots with extremely low disturbance within the mining area as a benchmark, representing an ideal vegetation reference benchmark without background pollution.) Environmental disturbance operators are obtained by sampling adjacent pure waste piles. (Quantifying background factors for the original Introducing a systematic offset), background stripping correction is performed using the inverse operation property of Lie groups: , To perform the high-dimensional manifold matrix representation after background stripping correction, also known as the pure vegetation manifold matrix component; For pixels The high-dimensional manifold matrix representation of vegetation is superimposed with the offset perturbations caused by backgrounds such as solid waste and bright bare soil; For environmental disturbance operators Inverse operation under the Lie group framework. Ultimately, the pure vegetation manifold matrix components, stripped of non-vegetation background interference, are obtained. This method ensures the purity of vegetation signal feature extraction from a physical mechanism perspective. In this embodiment, background offset is removed in a high-dimensional manifold geometric space, so the subtle growth characteristics of the vegetation itself can also be preserved, resulting in a more stable purification effect.

[0028] S3. Spatial heterogeneity perception weights are constructed within a local spatial window (a 7×7 local sliding spatial window is used in this embodiment) using the pure vegetation manifold matrix components. In some embodiments, the patchy fragmentation features of the mining area are obtained from multidimensional heterogeneous feature image data to obtain surface patches (surface vegetation, bare soil, and waste slag are interspersed, and spatial fragmentation is very prominent). The expression for the spatial heterogeneity perception weights is as follows: ,in Represents a pixel The spatial heterogeneity perception weights corresponding to the feature dimension (or feature channel) k. , The total dimension of the multidimensional features of a pixel; This represents the Gini index. The larger the Gini index, the more fragmented the local surface patch types and the stronger the spatial heterogeneity. Indicates in pixels Surface patches within a central local spatial window The probability of the area occupied. Index labels indicating surface patch types , This represents the total number of surface patches within a local spatial window. Represented by pixels The set of the sliding spatial neighborhood within the local spatial window centered on the center, where N represents the set of the sliding spatial neighborhood. The total number of neighboring pixels contained within; Indicates neighboring pixels The pure vegetation manifold matrix components with feature dimension (or feature channel) k; Let $k$ be the mean of the pure vegetation manifold matrix components for all pixels in the slip space neighborhood set, representing the feature dimension (or feature channel) $k$. Let Frobenius norm squared.

[0029] Using pure vegetation manifold matrix components The topological gradient measure matrix is ​​obtained by sensing spatial heterogeneity weights. In some embodiments, the pixels in the topological gradient measure matrix... and Topological gradient measure between (Also known as topological gradient distance, see [link]) Figure 2The topological gradient measure is the geodesic length of the HEAM curved manifold space along the path of surface vegetation evolution. The topological gradient measure accurately reflects the topological path of a pixel's evolution from completely bare land to fully vegetated land. By reducing multidimensional heterogeneous features to the measure space, the spectral distribution shift caused by noise pollution can be significantly reduced. The expression is as follows: ,in The physical fingerprint sensitivity adjustment parameter represents the feature dimension k. The global feature covariance matrix Inverse matrix operations; Represents a pixel Pixel The difference between the components of the pure vegetation manifold matrix in feature dimension k; the topological gradient measure matrix is ​​constructed by obtaining the topological gradient measure of any two pixels according to the above method. .

[0030] For pure vegetation manifold matrix components Topological gradient measure matrix A simplified core representation matrix is ​​obtained by performing dimensionality reduction through left coset projection of the cooperative subgroup. In the case study of the Shendong mining area, the specific method is as follows: an environmental disturbance invariant subgroup is established based on the metric space, consisting mainly of non-vegetation components such as waste and strongly reflective soil background. By calculating the features of the high-dimensional manifold in the left coset space The Riemann projection on the matrix converges the high-dimensional heterogeneous features to the core coverage component. While ensuring the retention of over 95.5% of the variance contribution rate, redundant noise dimensions affected by subgroup conjugate space interference are selectively removed, achieving feature extraction without dimensionality reduction. The final output is a concise core representation matrix that is spatiotemporally continuous and physically semantically clear. This lays a data foundation with an extremely high signal-to-noise ratio for the final inversion.

[0031] S4. Construct an LR-Attention hybrid model. The LR-Attention hybrid model obtains self-attention weights and a topology-enhanced feature field. The trained manifold topology forest model uses the topology-enhanced feature field and adaptive channel weight vectors to obtain the initial prediction value of the pure vegetation manifold evolution abundance. This invention's manifold topology forest model uses a sample set of topology-enhanced feature fields containing ground truth labels (ground truth labels for pure vegetation manifold evolution abundance) and weight vector data for prediction training of the initial prediction value of the pure vegetation manifold evolution abundance. Alternatively, a multi-source data sample set containing ground truth labels (ground truth labels for pure vegetation manifold evolution abundance) can be used to obtain the topology-enhanced feature field and adaptive channel weights according to methods S1 to S4, and then input into the manifold topology forest model for prediction training of the initial prediction value of the pure vegetation manifold evolution abundance. In some embodiments, such as... Figure 3 As shown, the LR-Attention hybrid model uses the topological gradient measure matrix. With a simplified core representation matrix Construct a joint attention scoring function (corresponding to LR-Attention and topologically constrained attention) to calculate the pixel. Pixel Self-attention weights This invention integrates a manifold topological constraint-based LR-Attention feature enhancement mechanism. It innovatively proposes an LR-Attention mechanism combining Lie group structure and Riemannian geometric measure, introducing topological gradient distance constraints into traditional attention calculations. Through an exponential decay function, it achieves synergistic modeling of spectral similarity and manifold geometric affinity, enabling the model to effectively distinguish between similar and dissimilar objects, significantly suppressing nonlinear interference in complex mining environments, and improving the discriminative power and physical consistency of feature representations. Specifically, it simplifies the core representation matrix... Generate a query vector from the input through a linear transformation matrix mapping. and key vector (The dimension of the weight matrix is ​​set to...) Design the geometric association layer to retrieve the topological gradient distance. Self-attention weights The expression is as follows: ,in The inner product scaling factor for stabilizing gradients. The manifold gain coefficient is set to 0.5 in this embodiment. The topological attenuation factor is set to 0.85 in this embodiment; This indicates the processing of the Softmax normalization function. For pixels Pixel Unnormalized raw attention scores between them , Representing pixels Pixel Transpose of the key vector For pixels The query vector, For pixels With pixels The topological gradient measure between them.

[0032] Using self-attention weights For the simplified core representation matrix Topology enhancement is performed to obtain the topology-enhanced feature field. This invention uses geometric correction of the attention weights by an exponential term to force the model to prioritize focusing on samples with the highest affinity along the manifold path. This eliminates interference from spectrally similar but physically dissimilar objects from a physical mechanism perspective, ultimately outputting a high signal-to-noise ratio topology-enhanced feature field that eliminates spectral aliasing and environmental background noise. .like Figure 4 As shown, the manifold topological forest model consists of multiple decision trees, and it utilizes topology-enhanced feature fields. The feature-adaptive channel weight vector was used to obtain the initial measurement of the abundance of pure vegetation manifold evolution.

[0033] In some embodiments, the manifold topological forest model uses a modified Riemannian Gini index adapted to nonlinear manifold surfaces to construct a splitting criterion. The expression for the modified Riemannian Gini index is as follows: ,in To obtain the weight parameters of sample n from the feature adaptive channel weight vector, This represents the sum of the weights of all samples in the split node of the decision tree; This is the penalty intensity adjustment coefficient. For the local feature manifold structure corresponding to all samples within the split node of the decision tree Custom manifold compactness factor; ,in Represents the matrix logarithm operation. This represents the covariance matrix corresponding to the node samples. Represents the Frobenius matrix norm; when a decision tree split node is mixed with a non-vegetation background and In this embodiment, the threshold is set to 1.45, which forces the decision tree to abandon the corresponding interfering splitting path; the manifold topological forest model is guided to strictly optimize the splitting along the pure topological path, and finally outputs the preliminary value of the pure vegetation manifold evolution abundance at the micro-dimensional level of each pixel. This invention accurately distinguishes between high-reflectivity soil noise and low-coverage vegetation boundaries. Based on the Riemannian Gini index-based manifold topological forest (MTF) splitting mechanism, it introduces a manifold compactness factor to adaptively regulate the node splitting process. When the sample distribution distortion is large, it automatically increases the penalty weight and abandons unreasonable splitting paths; when the distribution is compact, it guides the model to optimize the splitting direction along the manifold structure, thereby significantly improving the model's ability to fit the nonlinear and complex features of mining areas.

[0034] S5. Obtain the set of topological nearest neighbors of the target pixel, such as Figure 5 As shown, the vegetation cover prediction value of the target pixel is obtained by weighted aggregation of the initial measured values ​​of pure vegetation manifold evolution abundance of neighboring pixels within the topological nearest neighbor pixel set, i.e., the neighborhood correction mechanism is adopted. In some embodiments, the neighborhood correction mechanism method is as follows: filtering pixels from the target pixel. The topological nearest K cells constitute the topological nearest neighbor set. Target image The expression for the predicted vegetation cover value is as follows: ,in For pixels Topological nearest neighbor cells Topological gradient measure between For manifold scale parameters, Topological nearest neighbor pixels Preliminary values ​​of pure vegetation manifold evolution abundance. Normalization factor The expression is as follows: ; For affinity weights, topological nearest neighbor pixels With pixels The greater the topological distance, the lower the affinity weight. In this embodiment, the neighborhood correction mechanism uses geometric proximity to adaptively repair isolated salt-and-pepper noise and misjudged points caused by abrupt changes in spatially fragmented zones (i.e., weighted aggregation of initial measurement results from nearby reliable pixels to smooth out isolated misjudged points and salt-and-pepper noise generated by the fragmented surface). The final output is a predicted vertical projection vegetation cover value that combines high spatial consistency and physical fidelity. .

[0035] This invention first utilizes high-resolution UAV imagery to extract texture features, including micro-texture and spatial positional characteristics, and combines this with multi-temporal satellite remote sensing imagery to extract spectral features, including spectral information. A heterogeneous representation algebraic manifold (HEAM) adapted to the complex ecological positional characteristics of mining areas is constructed. By introducing the niche association entropy (NCE) manifold evolution operator, the co-evolutionary laws among multi-source features are captured. Furthermore, Lie algebraic exponential mapping and inverse operations are used to achieve algebraic-level removal of background noise. Finally, differential geometric properties are used to construct a topological gradient measure (TGM) to perceive inter-pixel relationships within the manifold tangent space. The biophysical distance is used to simplify and enhance the feature dimension through Riemann projection of the left coset space. Subsequently, the enhanced features under manifold topological constraints are obtained through the LR-Attention mechanism of the LR-Attention hybrid model, identifying and suppressing interference from homospectral heterogeneous objects. A manifold topological forest (MTF) model is constructed to improve the traditional splitting criterion to the Riemann Gini index adapted to the nonlinear surface, achieving preliminary discrimination of vegetation cover. Finally, through local collaborative correction of manifold weight evolution, spatial consistency enhancement is driven by the geometric proximity of topological nearest neighbors. This invention not only significantly improves the inversion accuracy of sparse vegetation areas and high-coverage saturated areas, but also has extremely high physical fidelity and robustness, providing accurate decision support for ecological monitoring, resource management, and environmental assessment in mining areas.

[0036] To quantitatively evaluate the accuracy and stability of the vegetation cover inversion method of this invention, this embodiment constructs a multi-index evaluation system and compares and analyzes the model output results with the measured ground values. The specific steps are as follows: To quantitatively evaluate the stability of the vegetation cover estimation method based on HEAM manifold and LR-Attention proposed in this invention, 120 representative 10m×10m quadrats were selected in the Shendong mining area as a case study. Measured vegetation cover data were obtained through visual interpretation of high-resolution UAV images combined with field needle acupuncture, and these data were used as the reference ground truth. (i.e., the true value of vegetation cover); simultaneously extract the corresponding estimated pixel value output by this invention. (i.e., predicted vegetation cover), construct a multi-index evaluation system based on the coefficient of determination R² and root mean square error RMSE, as shown in the following expression: , ;in, These are measured values ​​(i.e., the true value of vegetation coverage). These are the model predictions (i.e., vegetation cover predictions). This is the average of the measured values ​​(i.e., the average of the true values ​​of vegetation coverage). The sample is used; R² is used to measure the model's ability to fit the trend of real data changes, and its value range is [0, 1]. RMSE is used to measure the level of absolute error between the predicted value and the real value. Through comparative analysis, the mean R² of the method of the present invention in the Shendong mining area is stable at a high level, and the RMSE remains in a low range, which fully verifies the robustness of the method in complex backgrounds.

[0037] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism, characterized in that: The methods include: S1. Acquire multi-source datasets including UAV image data and satellite image data, and perform alignment processing and feature extraction on the multi-source datasets to obtain multi-dimensional heterogeneous feature image data. S2. Construct a high-order correlation potential field and use the correlation kernel function to obtain the similarity between any two pixels. Use the local evolution entropy function to obtain the feature adaptive channel weight vector and high-dimensional manifold matrix representation of each pixel. After manifold algebra synthesis operation and removal of non-vegetation background interference, obtain the pure vegetation manifold matrix components. S3. Spatial heterogeneity perception weights are constructed within a local spatial window using pure vegetation manifold matrix components. Topological gradient measure matrix is ​​obtained using pure vegetation manifold matrix components and spatial heterogeneity perception weights. Dimensionality reduction is achieved by synergistic subgroup left coset projection of pure vegetation manifold matrix components and topological gradient measure matrix to obtain a simplified core representation matrix. S4. Construct an LR-Attention hybrid model. The LR-Attention hybrid model obtains self-attention weights and topology-enhanced feature fields. The trained manifold topology forest model uses the topology-enhanced feature fields and feature adaptive channel weight vectors to obtain the initial measurement value of the pure vegetation manifold evolution abundance. S5. Obtain the topological nearest neighbor set of the target pixel, and perform weighted aggregation of the initial measured values ​​of the pure vegetation manifold evolution abundance of the nearest neighbor pixels within the topological nearest neighbor set to obtain the predicted value of vegetation cover of the target pixel.

2. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1, characterized in that: In method S5, the distance to the target pixel is filtered. The topological nearest K cells constitute the topological nearest neighbor set. Target image The expression for the predicted vegetation cover value is as follows: ,in For pixels Topological nearest neighbor cells Topological gradient measure between For manifold scale parameters, Topological nearest neighbor pixels Preliminary values ​​of pure vegetation manifold evolution abundance. The normalization factor is... The expression is as follows: ; For affinity weights, topological nearest neighbor pixels With pixels The greater the topological distance, the lower the affinity weight.

3. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1, characterized in that: In method S1, the UAV image data in the multi-source dataset undergoes the following preprocessing: UAV image data and corresponding high-precision positioning data are acquired; feature point matching and aerial triangulation are performed using the Scale Invariant Feature Transform (SIFT) algorithm to obtain an orthophoto DOM and a corresponding digital surface model (DSM); the high-precision positioning data is the spatial coordinates taken by the UAV; radiometric correction is performed using the whiteboard correction method. The satellite image data in the multi-source dataset is Sentinel-2 image data, which undergoes the following preprocessing: cloud and fog noise removal is performed on each band of the satellite image data; resampling and resolution unification are performed; spatiotemporal alignment is performed on the UAV image data and satellite image data in the multi-source dataset; texture features are extracted from the UAV image data and spectral features are extracted from the satellite image data and fused to construct multidimensional heterogeneous feature image data. .

4. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1, characterized in that: The multidimensional heterogeneous feature image data Including texture features and spectral features, the higher-order correlation potential field is formed by texture features and spectral features, and the correlation kernel function expression is as follows: , For pixels With pixels similarity, , Multidimensional heterogeneous feature image data medium pixel With pixels The multidimensional features, , Each pixel With pixels spectral characteristics, , Each pixel With pixels Texture features, The manifold kernel bandwidth parameter is used to control the rate of similarity decay. As a spatial co-evolutionary adjustment factor, Represents a pixel With pixels The square of the Euclidean distance of the texture feature.

5. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1, characterized in that: In method S2, a local evolutionary entropy (LEE) feedback mechanism is introduced to quantify environmental uncertainty. The expression for the local evolutionary entropy function is as follows: ,in The total feature dimension of a pixel's multidimensional features. For pixels Feature Dimension The probability distribution, This is the disturbance intensity correction factor. For adjustment coefficients, Represents a pixel The index is affected by environmental disturbances; the feature adaptive channel weight vector is dynamically generated pixel by pixel on the feature axis. Suppression and downsampling are applied to spectral bands affected by environmental pollution, and gain compensation is applied to high-reliability microtexture high-order fingerprints to obtain high-dimensional manifold pre-characterization elements after environmental weight correction and physical potential field reshaping.

6. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1, characterized in that: In method S2, the feature local autocorrelation symmetric positive definite matrix corresponding to each pixel is reconstructed, as shown in the following expression: , Let the square matrix be the result of the outer product of vectors. High-dimensional manifold matrix representations are obtained by mapping in the HEAM manifold or Lie group space. The algebraic composition operation of manifolds is performed by applying the mapping between the logarithm and exponent of the Lie algebra matrix, as shown in the following expression: ,in , Each pixel Pixel High-dimensional manifold matrix representation, For pixels Pixel The result after completing the algebraic composition operation of manifolds; express and Lie algebra field operations or Lie bracket operations; after removing non-vegetation background interference at the algebraic manifold level, the pure vegetation manifold matrix components are obtained. .

7. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1, characterized in that: In method S3, the patchy fragmentation features of the mining area are obtained from multidimensional heterogeneous feature image data to obtain surface patches. The spatial heterogeneity perception weight expression is as follows: ,in Represents a pixel The spatial heterogeneity perception weights corresponding to feature dimension k; This represents the Gini index. Indicates in pixels Surface patches within a central local spatial window The probability of the area occupied. This represents the total number of surface patches within a local spatial window. Represented by pixels The set of the sliding spatial neighborhood within the local spatial window centered on the center, where N represents the set of the sliding spatial neighborhood. The total number of neighboring pixels contained within; Indicates neighboring pixels , components of the pure vegetation manifold matrix with feature dimension k; Let represent the mean of the pure vegetation manifold matrix components of all cell feature dimension k in the slip space neighborhood set. Let Frobenius norm squared.

8. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 7, characterized in that: In method S3, the pixels in the topological gradient measure matrix and Topological gradient measure between The expression is as follows: ,in The physical fingerprint sensitivity adjustment parameter represents the feature dimension k. This is the inverse matrix operation of the global feature covariance matrix. Represents a pixel Pixel The difference between the components of the pure vegetation manifold matrix in feature dimension k; the topological gradient measure matrix is ​​constructed by obtaining the topological gradient measure of any two pixels according to the above method.

9. The vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1, characterized in that: In method S4, the LR-Attention hybrid model uses the topological gradient measure matrix. With a simplified core representation matrix Construct a joint attention scoring function to calculate the pixel. Pixel Self-attention weights Utilizing self-attention weights For the simplified core representation matrix Topology enhancement is performed to obtain the topology-enhanced feature field. ; The manifold topological forest model consists of multiple decision trees and utilizes topology-enhanced feature fields. The feature-adaptive channel weight vector was used to obtain the initial measurement of the abundance of pure vegetation manifold evolution.

10. A vegetation cover inversion method based on multi-source data fusion and LR-Attention mechanism according to claim 1 or 9, characterized in that: The manifold topological forest model uses a modified Riemannian Gini index adapted to nonlinear manifold surfaces to construct a splitting criterion. The expression for the modified Riemannian Gini index is as follows: ,in To obtain the weight parameters of sample n from the feature adaptive channel weight vector, This represents the sum of the weights of all samples in the split node of the decision tree; This is the penalty intensity adjustment coefficient. For the local feature manifold structure corresponding to all samples within the split node of the decision tree Custom manifold compactness factor; ,in Represents the matrix logarithm operation. This represents the covariance matrix corresponding to the node samples. Represents the Frobenius matrix norm; when a decision tree split node is mixed with a non-vegetation background and If so, the decision tree is forced to abandon the corresponding interfering split path.

Images