Mining area vegetation recovery capability prediction and evaluation method based on remote sensing image

Abstract

The invention relates to a mining area vegetation recovery capability prediction and evaluation method based on a remote sensing image, and the method is based on a deep neural network technology, and comprises the steps: multi-dimensional feature fusion, data normalization and null value processing, and dense embedding representation of mining area specific factors. The mining area specificity regulation and control coefficient is innovatively and dynamically injected into the branch network multi-head attention mechanism of the trunk model based on the Transform structure, and the migration and adaptive ability of the model is improved by constructing a dynamic parameter space. In addition, a multi-task learning strategy is adopted to introduce mining area factor importance weighting, and joint optimization of prediction performance among multiple mining areas is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis of remote sensing images and prediction of ecological restoration, and in particular to a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing images. Background Technology

[0002] With the promotion of ecological reconstruction and the concept of green mining, the refined assessment and prediction of vegetation restoration capacity in mining areas has become a crucial technical aspect of ecological environment management and sustainable operation of mining areas. Currently, vegetation restoration analysis in mining areas based on remote sensing imagery is being conducted in the industry. However, existing vegetation restoration prediction models generally face the following technical bottlenecks when deployed in actual mining areas. First, model parameters are highly dependent on the distribution characteristics of local historical data in the mining area. Pre-trained models are often highly "fixed" for local features such as topography, mineral type, climate, and soil in specific mining areas. Once migrated to mining areas with different specific factors, their prediction accuracy and generalization ability decrease significantly. Second, there are significant differences between different mining areas in engineering and natural elements such as mining methods, reclamation paths, and vegetation types, directly resulting in extremely limited applicability of the same model parameters to different mining areas. Therefore, when current mainstream models are migrated to new mining areas, a significant amount of time and computing power is required for data re-collection, parameter fine-tuning, and even model retraining, making it difficult to form a general prediction tool that can be deployed at scale. Summary of the Invention

[0003] This application provides a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing images, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0004] This application provides a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing imagery, specifically including:

[0005] S1: Collect the dynamic sequence of vegetation restoration in the target mining area as the target variable, and simultaneously acquire mining area-specific factors, mining methods, reclamation measures, and multidimensional characteristic data of topography, climate, and soil.

[0006] S2: Perform normalization and null value imputation preprocessing on the vegetation restoration dynamic sequence to generate a standardized vegetation index time series dataset.

[0007] S3: Construct an embedding layer based on the discrete encoding of mining area specific factors, and generate mining area factor embedding vectors through a dimensionality compression algorithm. The embedding vectors have a dimension of 8-16 and are used as mining area feature representations after L2 regularization.

[0008] S4: Construct the backbone model branch network, and use the encoder-decoder structure of the Transformer architecture to extract spatiotemporal features from the standardized vegetation index time series to generate a general evolutionary feature matrix for vegetation restoration.

[0009] S5: Establish a mining area factor branch network, and perform nonlinear transformation on the mining area factor embedding vector through a fully connected layer to generate a parameter mapping function containing mining area-specific regulation coefficients. The regulation coefficients are matched with the dimension of the attention weight matrix in the backbone model.

[0010] S6: Dynamically inject the mining area-specific regulation coefficients into the multi-head attention mechanism of the backbone model, generate adaptive attention weights through weighted fusion operations, and construct the dynamic parameter space of the vegetation restoration prediction model.

[0011] S7: A multi-task learning strategy is adopted to jointly train the vegetation restoration prediction model. The importance coefficient of mining area factors is introduced into the loss function. The importance coefficient is obtained by gradient backpropagation and is used to quantify the influence weight of different mining area-specific factors on the prediction results.

[0012] S8: When deploying in a new mining area, the specific factors of the target mining area are encoded and input into the trained model. The model parameter configuration adapted to the mining area is automatically generated through the parameter mapping function, and the vegetation restoration capacity prediction and evaluation are performed.

[0013] This application provides a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing imagery, which has the following beneficial effects:

[0014] (1) Existing methods typically embed mining area features implicitly into the training data, leading to overfitting of the model to specific mining areas. Once applied to a new mining area, a large number of samples need to be collected again and time-consuming optimization is required, which is difficult to meet the needs of rapid response in actual engineering. This solution introduces key specific factors such as mineral type, mining method, and reclamation measures as explicit inputs into the model structure and constructs a learnable mapping path from factors to model parameters, realizing the adaptive adjustment of the model to different mining area environments. This mechanism allows the backbone network to focus on learning the general spatiotemporal evolution law of vegetation restoration, while the control weights dynamically generated by the mining area factor branches are responsible for capturing individual differences, thereby completing multi-mining area adaptation while maintaining the uniformity of the model structure. Compared with the traditional "one model per area" mode, this method does not require repeating the modeling process for each new mining area. Only basic mining area description information is required to achieve plug-and-play deployment, which greatly reduces operation and maintenance costs and time overhead. It effectively overcomes the technical bottleneck of frequent retraining and low parameter tuning efficiency caused by the differences in specific factors of mining areas when applying traditional vegetation restoration prediction models across mining areas, and significantly improves the model's generalization ability and deployment flexibility.

[0015] (2) By designing a dual-branch neural network architecture and parameter generation mechanism, the collaborative optimization of common knowledge extraction and individual feature fusion was achieved, significantly enhancing the prediction accuracy and stability of the model in complex and heterogeneous mining area scenarios. Based on deep neural network technology, the main branch adopts LSTM or Transformer structure to model the time series evolution trend of vegetation index, fully exploring its inherent dynamic laws; at the same time, the mining area factor branch transforms discrete mining area attributes into continuous semantic vectors through embedding layers and fully connected networks, and further maps them to the adjustment coefficients of key components in the main model, forming a hybrid modeling paradigm of "data-driven + factor-guided". This structural design not only improves the model's sensitivity to input changes and response accuracy, but also promotes the automatic separation of common patterns and individual influences through multi-task joint training strategy, avoiding prediction bias caused by feature confusion. In addition, the introduction of sparse constraints or factor importance evaluation modules in the parameter mapping path further enhances the interpretability of the model, supports the identification and analysis of key influencing factors (such as the promoting effect of a certain type of reclamation measures on the recovery speed), and provides a scientific basis for ecological restoration decision-making.

[0016] (3) A vegetation restoration prediction framework with adaptive capabilities and high scalability has been constructed, promoting the transformation and upgrading of mine ecological governance from "experience-driven" to "model-driven," with significant system-level advantages and broad application prospects. This scheme breaks through the dependence of traditional black-box models on fixed training domains, realizing the technological leap of "one-time modeling, applicable to multiple areas." It is not only applicable to mining areas with various mineral types such as coal and metals, but can also flexibly access data from newly added mining areas without reconstructing the entire model system, demonstrating excellent horizontal scalability. At the same time, since the internal parameters of the model can be dynamically adjusted according to input factors, it can still maintain stable output when facing combinations of unseen mining areas or extreme geographical and climatic conditions, effectively improving prediction robustness. This feature is particularly suitable for the needs of ecological system restoration planning and dynamic monitoring in mining areas with wide distribution, diverse types, and complex ecological backgrounds in my country. It can be widely used in major application scenarios such as the national land space ecological restoration supervision platform and the green mine construction assessment system, providing intelligent technical support for the sustainable development of mining areas. Attached Figure Description

[0017] Figure 1 This is the main flowchart of a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing imagery.

[0018] Figure 2 This is a sub-flowchart of a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing imagery.

[0019] Figure 3 This is another sub-flowchart of a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing imagery. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0022] like Figure 1 As shown, this application provides a method for predicting and assessing the vegetation restoration capacity of mining areas based on remote sensing imagery, specifically including:

[0023] S1: Collect the dynamic sequence of vegetation restoration in the target mining area as the target variable, and simultaneously acquire mining area-specific factors, mining methods, reclamation measures, and multidimensional characteristic data of topography, climate, and soil.

[0024] S2: Perform normalization and null value imputation preprocessing on the vegetation restoration dynamic sequence to generate a standardized vegetation index time series dataset.

[0025] S3: Construct an embedding layer based on the discrete encoding of mining area specific factors, and generate mining area factor embedding vectors through a dimensionality compression algorithm. The embedding vectors have a dimension of 8-16 and are used as mining area feature representations after L2 regularization.

[0026] S4: Construct the backbone model branch network, and use the encoder-decoder structure of the Transformer architecture to extract spatiotemporal features from the standardized vegetation index time series to generate a general evolutionary feature matrix for vegetation restoration.

[0027] S5: Establish a mining area factor branch network, and perform nonlinear transformation on the mining area factor embedding vector through a fully connected layer to generate a parameter mapping function containing mining area-specific regulation coefficients. The regulation coefficients are matched with the dimension of the attention weight matrix in the backbone model.

[0028] S6: Dynamically inject the mining area-specific regulation coefficients into the multi-head attention mechanism of the backbone model, generate adaptive attention weights through weighted fusion operations, and construct the dynamic parameter space of the vegetation restoration prediction model.

[0029] S7: A multi-task learning strategy is adopted to jointly train the vegetation restoration prediction model. The importance coefficient of mining area factors is introduced into the loss function. The importance coefficient is obtained by gradient backpropagation and is used to quantify the influence weight of different mining area-specific factors on the prediction results.

[0030] S8: When deploying in a new mining area, the specific factors of the target mining area are encoded and input into the trained model. The model parameter configuration adapted to the mining area is automatically generated through the parameter mapping function, and the vegetation restoration capacity prediction and evaluation are performed.

[0031] Step S1: Collect the dynamic sequence of vegetation restoration in the target mining area as the target variable, and simultaneously acquire mining area-specific factors, including mineral type, mining method, reclamation measures, and multidimensional characteristic data of topography, climate, and soil. Specifically, this includes:

[0032] S1.1: Based on the remote sensing image dataset, the vegetation coverage of the target mining area is inverted using the time series NDVI (Normalized Difference Vegetation Index) synthesis algorithm to generate dynamic sequence data of vegetation restoration for consecutive years.

[0033] For the remote sensing image dataset of the target mining area, a multi-source data fusion method is adopted (the data source includes multispectral images such as Landsat and Sentinel-2 from previous years, with a spatial resolution of 10-30 meters and a temporal resolution of no less than monthly) to construct a multi-temporal image set of continuous time series.

[0034] Cloud masking and shadow detection algorithms are used to automatically identify and remove cloud-covered and shadowed areas in image data, ensuring that inversion calculations are based on effective pixels.

[0035] The Normalized Difference Vegetation Index (NDVI) was synthesized using the following algorithm (formula: ,in For near-infrared reflectivity, (Reflectance in the red light band) to achieve quantitative inversion of vegetation coverage at each time phase.

[0036] Furthermore, by using a time series synthesis method (parameters: target time step is annual, synthesis strategy is maximum value synthesis to reduce seasonal bare land noise), a cross-year NDVI value sequence is generated, resulting in a continuous annual vegetation index sequence covering the entire observation period.

[0037] Furthermore, a spatial registration algorithm (parameters: matching accuracy of 1 pixel, reference coordinate system of WGS84 / UTM partition) is used to achieve unified spatial reference and geometric correction for image sequences from different sources, ensuring that the spatial position of the time series remains consistent.

[0038] Through the above synthesis and correction processing, the multi-temporal remote sensing image results of the previous step are transformed into continuous annual vegetation restoration dynamic sequence data of the target mining area, realizing accurate capture of spatiotemporal change trends and providing a stable data foundation for subsequent smoothing filtering and feature extraction.

[0039] S1.2: Perform sliding window smoothing on the generated vegetation restoration dynamic sequence to eliminate abnormal fluctuations caused by cloud cover or atmospheric interference in the remote sensing image, and obtain smoothed vegetation index time series data.

[0040] S1.3: Based on the basic geographic information database of the mining area, extract the topographic factor data of the target mining area, including elevation, slope and aspect, and use spatial interpolation algorithm to generate continuous topographic feature raster data to support the construction of multidimensional features of the model input.

[0041] Based on the vector boundary data of the mining area's basic geographic information database, the geospatial overlay analysis method (parameter: the projected coordinate system is WGS 84 / UTM Zone corresponding to the mining area zone number) is used to achieve spatial clipping and positioning of the target mining area.

[0042] Furthermore, based on digital elevation model (DEM) data, a slope calculation algorithm (parameters: 3×3 convolution kernel, unit: degrees) is used to extract the slope value of each pixel, and the continuous slope values ​​are mapped to a preset slope level range through a slope grading function to obtain a slope feature raster.

[0043] Furthermore, using a slope aspect analysis algorithm (parameters: radius of 1 pixel, calculation rule: counterclockwise 0° to 360°), the slope aspect value of each pixel is calculated, and the results are encoded as an eight-directional classification feature to characterize the solar radiation exposure of the mining area.

[0044] Furthermore, spatial interpolation algorithms (parameters: inverse distance weighted IDW, power parameter p is 2, sampling radius is 5 pixels) are executed on the elevation, slope, and aspect raster data respectively to achieve continuous feature filling between discrete sampling points and obtain continuous terrain feature raster data that fully covers the target mining area.

[0045] Through the spatial analysis and interpolation methods described above, the boundary clipping results from the previous step are transformed into standardized raster data containing three types of continuous terrain features: elevation, slope, and aspect. This achieves the expected technical effect of constructing multidimensional terrain features to support model input.

[0046] For example, in an open-pit coal mine application scenario, the basic geographic information database provides polygonal data of the mining area boundary vector in the UTM Zone 50N coordinate system, and the digital elevation model has a resolution of 30m. The slope calculation algorithm uses a 3×3 window convolution kernel, and the obtained slope values ​​range from 0° to 40°, which are divided into 5 slope levels through a grading function. The aspect analysis results adopt an eight-directional classification encoding, with directions such as east-north, due east, and south-east each identified by corresponding integer codes. The spatial interpolation algorithm uses the inverse distance weighting method, with the power parameter p set to 2 and the sampling radius of 5 pixels. The generated continuous rasters of elevation, slope, and aspect are stored as GeoTIFF format files. The interpolation results maintain a smooth transition in the raster edge area, effectively avoiding the disruption of the continuity of model training data caused by DEM sampling holes. The final output terrain feature raster serves as part of the input to the vegetation restoration prediction model, significantly improving the model's sensitivity and adaptability to terrain changes in the subsequent feature fusion stage.

[0047] S1.4: Obtain climate factor data such as annual average temperature, annual precipitation, and annual evaporation of the target mining area based on the meteorological data interface, and convert the climate factors into climate feature vectors with uniform dimensions through standardization and normalization methods.

[0048] Based on the meteorological data interface call mechanism (parameters: API endpoint address, authentication key, time range, spatial range), it is possible to acquire meteorological observation data of the target mining area in batches within a specified period, including core climate factors such as annual average temperature, annual precipitation, and annual evaporation.

[0049] Furthermore, by using a data integrity verification algorithm (parameters: missing data threshold 5%, outlier detection range ±3σ), the acquired climate factor data is screened for validity, eliminating observational anomalies and missing data records that occur within the time span, while retaining data samples that meet the accuracy requirements.

[0050] Furthermore, a time series annual aggregation algorithm (parameters: aggregation period of 1 year, statistical function mean / sum) is adopted to generate annual average temperature vector, annual total precipitation vector and annual total evaporation vector based on the original daily or monthly meteorological observation records, so as to ensure that the climate characteristics are comparable across mining areas.

[0051] Furthermore, based on the standardization and normalization method (parameters: mean μ and standard deviation σ are taken from the entire mining area sample set), the dimensional transformation of the climate factor vectors for each year is performed, and the standardized value Z is calculated using the following formula:

[0052]

[0053] in, These are the raw values ​​of the climate factors. The sample mean. This represents the sample standard deviation.

[0054] Furthermore, a normalization algorithm (parameters: maximum value max and minimum value min are taken from the entire mining area sample set) is used to perform interval mapping on the standardized climate factor vector.

[0055] For example, during meteorological data acquisition and feature conversion in a mining area, the API endpoint address was configured as the climate center interface, the authentication key was a 32-bit hash value, the time range was set to 2010 to 2020, and the spatial range was limited to the polygonal boundary of the mining area. A total of 4015 daily temperature records, the same number of precipitation records, and 3900 evaporation records were obtained. After a 5% missing data threshold check, the valid records were 4000, 4000, and 3800 respectively. Using an annual aggregation algorithm, an annual average temperature vector of length 11 was generated, and precipitation and evaporation vectors were of the same length. In the standardization phase, the average of the entire mining area sample set was used. ℃, standard deviation Using ℃ as a parameter, the temperature value of a certain year ℃ converted to standardized value In the normalization phase, the maximum value of the sample is used. Minimum value For parameters, the standardized values Mapped to normalized values Ultimately, a climate feature vector input model with a length of 11 was formed, which significantly improved the feasibility of comparing climate conditions between different mining areas and the stability of the input data of the prediction model.

[0056] S1.5: Based on the soil type layer and soil physicochemical property database, soil factor data of the target mining area are extracted, including organic matter content, pH value and water holding capacity. Principal component analysis algorithm is used to reduce the dimensionality of soil factors and generate low-dimensional soil feature vectors.

[0057] Based on structured data input from soil type layers and soil physicochemical property databases, a spatial overlay analysis method (parameters: maintaining consistent projection coordinate system, resolution not less than 30m×30m) is used to achieve precise matching between soil type units and physicochemical index records within the target mining area.

[0058] Furthermore, the soil type category and corresponding physicochemical indicators are merged through the attribute association retrieval method (parameter: unique plot number as association key value), including indicators such as organic matter content, pH value and water holding capacity, and the original feature matrix of soil factors is generated, with the rows of the matrix representing plots and the columns representing each physicochemical indicator.

[0059] Furthermore, a missing value processing algorithm (parameters: mean filling for random missing values, interpolation filling for spatially continuous missing values) is used to complete the original feature matrix of soil factors, remove invalid or abnormal records, and obtain a quality-controlled feature matrix to ensure the reliability of subsequent calculations.

[0060] Furthermore, based on the normalization method (parameter: Z-score standardization), the physicochemical indicators are uniformly mapped to a dimensional space with zero mean and unit variance, generating a standardized soil factor matrix that removes scale differences.

[0061] S1.6: Based on the mining area management ledger and remote sensing interpretation results, obtain mining area-specific factors such as mineral type, mining method, and reclamation measures of the target mining area, and use One-Hot coding method to quantify discrete mining area factors to generate mining area factor coding vectors.

[0062] Based on the mining area management ledger and remote sensing interpretation data files, discrete classification attribute fields such as mineral type, mining method, and reclamation measures of the target mining area are analyzed to construct the original mining area specific factor set.

[0063] The attribute value standardization test method (parameter: category enumeration list) is adopted to achieve consistency matching between the mining area factor field values ​​and the standard category label library, ensuring that the semantics and value range of the input features conform to the factor category system preset by the model.

[0064] Furthermore, by using a category index mapping algorithm (parameter: a table of category label sequences and index numbers), the discrete features of mining area factors are mapped to integer indices, resulting in index-type feature vectors that can be directly used for encoding matrix construction.

[0065] Furthermore, the One-Hot encoding method (parameters: indexed feature vector, total number of categories) is adopted to convert each discrete mining area factor into a sparse binary vector with a length equal to the total number of categories, where the position belonging to the current category is set to 1 and the rest are set to 0, and a mining area factor encoding matrix is ​​generated.

[0066] Furthermore, through matrix concatenation operations (parameter: factor coding vectors of each mining area), the column-wise merging of multiple discrete mining area factor codes is achieved, resulting in a structured mining area factor coding vector.

[0067] By using the One-Hot encoding algorithm and matrix concatenation, the discrete attributes of the original mining area-specific factors from the previous step are transformed into numerical mining area factor encoding vectors, thus achieving the unified numerical format required for subsequent embedding layer inputs.

[0068] For example, in an embodiment of an open-pit coal mine, the mineral type field is "coal", corresponding to a total of 4 mineral categories (coal, iron, copper, gold), with a coal category index number of 0; the mining method field is "open-pit", corresponding to a total of 3 mining method categories (open-pit, underground, seabed), with an open-pit category index number of 0; the reclamation measures field is "natural restoration", corresponding to a total of 5 reclamation measures categories (natural restoration, artificial vegetation, land leveling, irrigation improvement, mixed measures), with a natural restoration category index number of 0. The above indexed feature vector is represented as [0,0,0] using a category index mapping algorithm, and then sparse binary vectors are generated by One-Hot encoding: mineral type is encoded as [1,0,0,0], mining method as [1,0,0], and reclamation measures as [1,0,0,0,0]. These are then merged using matrix concatenation operations into a mining area factor encoding vector of length 12: [1,0,0,0,1,0,0,1,0,0,0,0]. The encoded vector is input into the embedding layer, which can generate low-dimensional dense features required for subsequent parameter mapping. In actual model training, it is verified that it can significantly improve the model's adaptability and prediction stability under different mining conditions.

[0069] S1.7: The vegetation restoration dynamic sequence, terrain feature vector, climate feature vector, soil feature vector and mining area factor encoding vector are concatenated and fused to generate a structured dataset containing vegetation restoration target variables and multi-dimensional input features, which serves as the basic data input for subsequent model training and feature embedding.

[0070] Step S2: Perform normalization and null value imputation preprocessing on the vegetation restoration dynamic sequence to generate a standardized vegetation index time series dataset. Specifically, this includes:

[0071] S2.1: Perform max-min normalization on the NDVI / EVI time series data in the original vegetation restoration dynamic sequence to map the vegetation index to the [0,1] interval, obtain normalized vegetation index time series data, and eliminate the problem of dimension inconsistency caused by sensor differences or atmospheric correction in remote sensing images of different mining areas or different years.

[0072] S2.2: The outliers in the normalized vegetation index time series are identified based on the sliding window correlation analysis method, and the outliers are replaced by the time series reconstruction algorithm based on neighborhood similarity to obtain the vegetation index time series data after preliminary cleaning, so as to improve the data quality and reduce the interference of outliers on model training.

[0073] S2.3: The missing values ​​in the vegetation index time series are filled in by a null value imputation method that combines linear interpolation and seasonal decomposition. The imputed values ​​are generated based on the periodic trend of the time series to obtain a continuous and complete vegetation index time series dataset, ensuring the integrity and continuity of the model input data.

[0074] For the normalized and outlier-replaced vegetation index time series data, a linear interpolation method (parameters: time index and index value of adjacent valid observation points) is used to initially fill in missing points. Further, a seasonal decomposition algorithm (parameters: time series length, seasonal cycle length, trend decomposition order) is used to split the vegetation index series into three parts: long-term trend, seasonal component, and random residuals, obtaining the numerical expression of the seasonal component. Further, a periodic trend analysis method (parameters: mean of seasonal component, periodic phase shift) is used to calculate the seasonal filling values ​​for missing locations, generating interpolated data with periodic consistency. Further, an interpolation fusion algorithm (parameters: weight of linear interpolation results, weight of seasonal trend results) is used to weightedly combine the interpolation results, forming a complete time series filling output. A continuity detection algorithm (parameters: time step threshold, sequence stationarity index) is used to transform the filling sequence from the previous step into a continuous and complete vegetation index time series dataset, ensuring the integrity and continuity of the model input data.

[0075] S2.4: Perform Z-score standardization on the infilled vegetation index time series data to eliminate the influence of differences in the initial state of vegetation restoration in different mining areas, obtain standardized vegetation index time series data, and improve the model's ability to generalize and recognize dynamic patterns of vegetation restoration.

[0076] S2.5: The time series-based dynamic window segmentation mechanism segments and classifies the standardized vegetation index time series, divides the training samples and validation samples according to the vegetation recovery cycle characteristics, and generates a structured vegetation recovery time series dataset, providing a unified data input format for subsequent model training and validation.

[0077] like Figure 2 As shown, step S3 involves constructing an embedding layer based on the discrete encoding of mining area-specific factors, generating mining area factor embedding vectors using a dimensionality compression algorithm. These embedding vectors have a dimension of 8-16 and are then processed using L2 regularization to represent the mining area features. Specifically, this includes:

[0078] S3.1: Perform category feature encoding processing on the original discrete codes of mining area-specific factors to obtain a discrete factor encoding sequence with a unified format.

[0079] S3.2: Based on the discrete factor coding sequence, construct an embedding matrix. The number of rows in the embedding matrix corresponds to the total number of categories of factors in the mining area, and the number of columns is the initial embedding dimension, which is 128-256 dimensions.

[0080] Based on the input discrete mining area factor encoding sequence, an embedding matrix initialization method (parameters: total number of categories, initial embedding dimension 128-256) is used to map the factor index of each category to a dense vector representation in a high-dimensional continuous space.

[0081] Furthermore, by using a random normal distribution weight initialization method (parameters: mean 0, standard deviation 0.01), the initial values ​​of each element of the embedding matrix are obtained, resulting in a high-dimensional embedding parameter matrix that conforms to the optimization characteristics of gradient descent.

[0082] Furthermore, a matrix dimension setting strategy is adopted (parameters: number of rows = total number of factor categories, number of columns = 128-256) to realize the physical configuration of the embedded matrix row and column structure, so that each row strictly corresponds to a single mining area factor category, and the number of matrix columns corresponds to the initial feature dimension of the embedding space.

[0083] Furthermore, by utilizing the index mapping lookup algorithm (parameter: factor category ID index), the high-dimensional vector representation of the corresponding row in the embedding matrix is ​​quickly retrieved from the discrete factor encoding sequence, and a complete high-dimensional mining area factor vector set is generated.

[0084] The matrix construction method described above transforms the discrete factor encoding sequence from the previous step into an initial dense embedding matrix with dimensions of 128-256, thus enabling input preparation and semantic preservation for subsequent PCA dimensionality reduction and feature compression steps.

[0085] For example, in a mining area's specific factor set containing 8 types of minerals, 5 types of mining methods, and 6 types of reclamation measures, the total number of categories is 19, and the initial embedding dimension is set to 192. First, an embedding matrix with 19 rows and 192 columns is constructed, and each element is randomly initialized using a Gaussian distribution with a mean of 0 and a standard deviation of 0.01. For a specific factor in the encoded sequence, such as mineral type coded as 3, the 192-dimensional floating-point vector in the 3rd row of the matrix is ​​directly extracted as the dense representation of that factor using an index mapping algorithm. This operation is performed on all discrete encoded fields, ultimately obtaining a complete high-dimensional factor vector set. This embedding matrix can preserve the semantic information of the input discrete factors during subsequent PCA processing and compress them to 8-16 dimensions to reduce the computational cost of parameter mapping. In model validation, this high-dimensional embedding matrix shows significantly improved specific factor representation ability and stability compared to datasets without embedded feature vectors.

[0086] S3.3: Perform dimensionality reduction on the embedding matrix using the principal component analysis (PCA) algorithm to generate a low-dimensional dense embedding vector with a dimension of 8-16.

[0087] S3.4: Perform L2 regularization on the low-dimensional dense embedding vector to obtain the mining area factor embedding vector under the unit norm constraint. The embedding vector serves as a standardized representation of the mining area-specific characteristics.

[0088] The low-dimensional dense embedding vector obtained after dimensionality reduction based on the PCA algorithm is used as input data, with a dimension range of 8-16 and the values ​​being real-number floating-point vectors.

[0089] The L2 regularization method (parameter: norm p=2) is used to implement the unit norm constraint on the low-dimensional dense embedding vector, so as to eliminate the dimensional differences in the vector length of the factor embedding vectors of different mining areas.

[0090] Furthermore, by calculating the L2 norm of the embedding vector This allows you to obtain the length of the vector, where For the i-th component of the embedding vector, For vector dimensions.

[0091] Furthermore, through vector division operations This process normalizes each component and yields a normalized embedding vector with a norm of 1.

[0092] Furthermore, a vector scaling correction algorithm (parameter: error tolerance ε=1e-6) is adopted to achieve the stability of numerical calculation and generate a standardized representation of mining area factor embedding that can be used for subsequent network input.

[0093] By using L2 regularization, the embedding vectors after PCA dimensionality reduction in the previous step are transformed into mining area feature vectors with a consistent norm and eliminating scale bias. This enables the model to align input features under multiple mining area conditions and improve the robustness of the embedding representation.

[0094] For example, in the embedding generation process of a copper mine, the PCA dimensionality reduction output embedding vector has a dimension of 10, with component values ​​such as: [0.12, -0.33, 0.45, 0.27, -0.56, 0.18, 0.08, -0.41, 0.29, 0.36]. Using L2 regularization, the L2 norm is first calculated: = ≈ Then, a division operation is performed on each component to obtain a standardized vector of approximately [0.112, -0.308, 0.42, 0.252, -0.522, 0.168, 0.075, -0.383, 0.271, 0.336]. This vector has a norm of 1, satisfying the unity norm constraint. After inputting this mining area factor embedding vector into the mining area factor branch network, the model can effectively adjust the parameters specific to the mining area when fusing this vector with the vegetation restoration time-series features, significantly improving the stability of cross-regional prediction, and demonstrating a more stable restoration capacity prediction effect in the validation set.

[0095] S3.5: Input the L2-regularized mining area factor embedding vector into the mining area factor branch network to generate the input feature vector of the subsequent parameter mapping function.

[0096] like Figure 3 As shown, step S4 involves constructing a backbone model branch network and using a Transformer-based encoder-decoder structure to extract spatiotemporal features from the standardized vegetation index time series, generating a general evolutionary feature matrix for vegetation restoration. Specifically, this includes:

[0097] S4.1: Based on standardized vegetation index time series data, a time window sliding mechanism is constructed to segment the vegetation index series with a preset time step to generate spatiotemporal sequence samples for model input.

[0098] Based on standardized vegetation index time series data, a time window sliding mechanism (parameters: window length L, sliding step size P) is used to segment the time series, thereby dividing the long time series data into several subsequence samples containing local spatiotemporal patterns.

[0099] Furthermore, by using a window length selection algorithm (set according to the annual cycle characteristics of vegetation restoration), the range of values ​​for L is determined so that each sequence segment covers at least one complete growing season cycle, thereby ensuring the sufficiency of capturing spatiotemporal features and obtaining a list of window start and end indices.

[0100] Furthermore, a sliding step size optimization method (parameter: value of P) is adopted to control the start and end index interval to no more than half a growing season cycle, thereby improving the time coverage between samples and the model's sensitivity to the transition phase.

[0101] Furthermore, through index slicing operations, the original standardized vegetation index time series is segmented according to the window index list to generate a set of multiple time series subsamples of length L, and the timestamp range of each subsample is recorded for location encoding.

[0102] Furthermore, a sequence boundary completion algorithm (parameter: linear interpolation) is used to complete subsamples that are shorter than L due to the window sliding to the end of the sequence, ensuring that all input samples have the same length and forming a spatiotemporal sequence tensor of uniform dimension.

[0103] By using a time window sliding mechanism, the standardized sequence data from the previous step is transformed into a structured spatiotemporal sequence sample set, achieving the effect of local pattern decomposition of the model input data in the time dimension.

[0104] S4.2: Perform location encoding embedding processing on the vegetation index sequence within the time window, and construct location vectors using sine and cosine functions to introduce the temporal prior information of the time series and generate spatiotemporal feature vectors with location information enhancement.

[0105] S4.3: Based on the enhanced spatiotemporal feature vectors, construct the encoder module of the Transformer architecture, and use a multi-head self-attention mechanism to perform context modeling on the input sequence to generate high-level semantic feature representations with context dependencies.

[0106] Based on the spatiotemporal feature vector input with positional encoding enhancement, a Transformer encoder module is constructed using a multi-head self-attention mechanism to achieve multi-scale dependency modeling of internal contextual information of the sequence.

[0107] Layer normalization (parameter: Pick The method normalizes the input spatiotemporal feature vector, eliminates the numerical scale difference of features at different time steps, and obtains a normalized feature matrix with zero mean and one variance.

[0108] Furthermore, through a linear mapping algorithm (parameter: mapping dimension d) k =64, d v =64) Map the normalized feature matrix to the query, key, and value spaces respectively, generating the corresponding... , , A matrix is ​​used as the basic input for calculating attention weights.

[0109] Furthermore, using the scaled dot product attention algorithm, the attention distribution is calculated according to the following formula:

[0110]

[0111] in, For querying the matrix, Let T be the key matrix, and T represent the transpose of the matrix. For value matrices, This is a scaling factor used to prevent the dot product magnitude from becoming too large in high-dimensional spaces.

[0112] Furthermore, a multi-head attention mechanism is constructed (parameter: h=8 heads) to receive the input. , , The matrix is ​​divided into 8 subspaces to perform the above attention calculations, obtain the weighted output of each head, and perform a concatenation operation on the head dimension to generate a comprehensive attention output matrix.

[0113] Furthermore, the residual connection mechanism is used to fuse the integrated attention output matrix with the encoder input element-wise by addition, and layer normalization is performed again to ensure the stability of the gradient flow and the consistency of feature representation.

[0114] Furthermore, a nonlinear transformation is performed on the attention fusion result through a feedforward network (parameters: two fully connected layers, 256 hidden units, and ReLU activation function) to enhance the high-order feature abstraction capability and generate the final output high-dimensional semantic feature representation matrix of the encoder module.

[0115] The encoder module, composed of the multi-head self-attention and feedforward network, transforms the spatiotemporal features with positional encoding enhancement into high-level semantic feature representations with contextual dependencies, enabling the model to deeply capture the dynamic patterns of vegetation restoration across time windows.

[0116] S4.4: Based on the high-level semantic feature representation of the encoder output, construct the decoder module of the Transformer architecture, and introduce a cross-attention mechanism to interactively model the encoder output and the historical vegetation restoration state in order to generate the temporal evolution feature sequence of the vegetation restoration process.

[0117] Based on the high-level semantic feature representation of the encoder output, a decoder module of the Transformer architecture is constructed to combine existing general regular features with the historical recovery state of the target task, forming a temporal evolution feature sequence that can be used for subsequent parameter mapping and prediction.

[0118] A multi-head self-attention mechanism is used to model the internal correlation of historical vegetation restoration state sequences, thereby capturing internal patterns and enhancing features of historical states.

[0119] Furthermore, a cross-attention mechanism (parameters: query matrix from decoder self-attention output, key / value matrix from encoder output) is used to interactively calculate the high-level semantic features of the encoder and the current time step state of the decoder, and obtain the cross-sequence context relevance matrix to fuse historical state information and general regular features.

[0120] Furthermore, by utilizing residual connections and layer normalization methods, the cross-attention output and the decoder input features are summed element-wise and normalized to generate a stable semantic fusion sequence, thereby reducing gradient vanishing and enhancing cross-domain consistency of features.

[0121] Furthermore, a feedforward fully connected network is used to perform a nonlinear transformation on the fused sequence, mapping the high-order interaction features to a temporal pattern space that is adapted to subsequent predictions, and generating a temporal evolution feature sequence.

[0122] Through the above-mentioned decoder module processing method, the high-level semantic features output by the encoder are fused with the historical vegetation restoration status and transformed into a temporal evolution feature sequence with the ability to perceive general laws and specific conditions across time series and mining areas, thereby realizing a refined driving effect on the subsequent regulation of mining area factors.

[0123] S4.5: Perform global average pooling on the temporal evolution feature sequence output by the decoder to compress the temporal dimension and generate a fixed-dimensional vegetation restoration general evolution feature matrix for subsequent parameter mapping.

[0124] Step S5: Establish a mining area factor branch network, and perform a nonlinear transformation on the mining area factor embedding vector through a fully connected layer to generate a parameter mapping function containing mining area-specific regulation coefficients. These regulation coefficients match the dimension of the attention weight matrix in the backbone model. Specifically, this includes:

[0125] S5.1: Based on the embedding vector of mining area factors, a nonlinear mapping process is performed using a fully connected network. The ReLU activation function is used to construct a deep feature transformation path to generate high-order feature representations of mining area factors.

[0126] Based on the input condition of the embedded vector of mining area factors, a fully connected network structure (parameters: 3 layers, 512 neurons per layer, and weight initialization method of He Normal distribution) is adopted to realize the nonlinear mapping processing of mining area factor features.

[0127] Furthermore, by introducing the ReLU activation function (parameter: negative slope coefficient 0), a deep feature transformation path is constructed to improve the model's ability to fit the complex correlations of factors in different mining areas, and the output feature matrix of the hidden layer in the first stage is obtained.

[0128] Furthermore, the batch normalization algorithm (parameters: momentum 0.9, ε=1e-5) is used to standardize the distribution of the hidden layer output feature matrix, thereby stabilizing the feature distribution and reducing the vanishing or exploding gradient phenomenon.

[0129] Furthermore, a residual connection mechanism (parameter: connection span of 2 layers) is adopted to perform element-wise addition of the current layer output with the previous layer features to maintain the traceability of the initial mining area factor embedding features and enhance the training convergence of the high-dimensional nonlinear transformation.

[0130] The feature matrix from the previous step is sparsified by using the Dropout random deactivation algorithm (parameter: deactivation rate 0.3), which improves the network's robustness to noise and reduces the risk of overfitting. This generates the final high-order feature representation matrix of the mining area factors, achieving the technical effects of expanding feature representation capacity and compressing high-dimensional information.

[0131] S5.2: Perform dimension alignment operation on the high-order feature representation of mining area factors output by the fully connected network. Use affine transformation to map the feature dimension to be consistent with the key vector dimension in the multi-head attention mechanism in the backbone model, so as to generate the control vector for the adaptation parameter injection.

[0132] S5.3: Design parameter mapping function based on control vector, and use Softmax normalization algorithm to generate control coefficients of mining area factors. The control coefficients are used to characterize the influence of mining area specific factors on the distribution of attention weights in the backbone model.

[0133] S5.4: The mining area factor regulation coefficients are weighted and fused with the original attention weight matrix in the backbone model. An adaptive attention weight matrix is ​​generated based on the weighted summation algorithm to realize the dynamic adjustment of the mining area factors on the internal decision-making mechanism of the model.

[0134] S5.5: Perform a differentiable parameter freezing operation on the generated adaptive attention weight matrix and use the gradient masking mechanism to retain the update capability of learnable parameters in the mining area factor branch, so as to support the adaptive adjustment of parameters when the model is deployed in a new mining area.

[0135] Step S6: Dynamically inject the mining area-specific regulation coefficients into the multi-head attention mechanism of the backbone model, generate adaptive attention weights through weighted fusion operations, and construct the dynamic parameter space of the vegetation restoration prediction model. Specifically, this includes:

[0136] S6.1: Based on the original attention weight matrix of the multi-head attention mechanism in the backbone model, the original attention weight is weighted and fused using the regulation coefficient vector output by the mining area factor branch to generate an adaptive attention weight matrix that is adapted to the specificity of the current mining area.

[0137] S6.2: Perform dimension alignment on the mining area-specific control coefficient vector to match the number of heads and the dimension of each head in the multi-head attention mechanism in the backbone model, so as to ensure that the control coefficients can be accurately mapped to the corresponding attention heads, thereby achieving fine-grained attention mechanism regulation.

[0138] S6.3: The aligned mining area-specific control coefficient vector is applied to the query, key, and value matrices of each attention head in the multi-head attention mechanism. The similarity calculation process in the attention mechanism is dynamically adjusted through element-wise multiplication to enhance the model's responsiveness to mining area-specific factors.

[0139] Based on the mining area-specific control coefficient vector after dimension alignment, a matrix element-wise multiplication algorithm (parameter: the size of the coefficient matrix is ​​consistent with the size of the Query, Key, and Value matrices of the corresponding attention head) is used to achieve weighted adjustment of the query matrix of each attention head in the multi-head attention mechanism, and obtain the Query matrix adjustment result with mining area-specific factors.

[0140] Furthermore, by using a matrix element-wise multiplication algorithm (parameter: the same control coefficient vector as the Query matrix), a weighted correction is performed on the key matrix of each attention head in the multi-head attention mechanism, injecting the regulatory information of mining area factors into the similarity calculation process, so that the key matrix strengthens the semantic association of mining area conditions during the spatial feature matching process.

[0141] Furthermore, by using a matrix element-wise multiplication algorithm (parameter: control coefficient vector consistent with the above adjustment), a weighted update is performed on the value matrix of each attention head in the multi-head attention mechanism, embedding the influence of mining area-specific factors on the final feature aggregation result into the weight allocation of the value matrix output.

[0142] Furthermore, an attention similarity calculation method based on matrix dot product is adopted, which multiplies the adjusted Query matrix and the adjusted Key matrix, and performs scaling processing on the product result (the scaling factor is the reciprocal of the square root of the key dimension). ,in This is the adjusted query matrix. The adjusted key matrix, This indicates the transpose operation.

[0143] By introducing an element-wise multiplication algorithm for the mining area-specific control coefficient vector, the original Query, Key, and Value matrices are transformed into weighted matrices containing mining area-specific information, thereby enhancing the dynamic sensitivity of the attention calculation within the backbone model to mining area conditions.

[0144] For example, in a Transformer encoder environment with 6 attention heads, the dimension of the mine-specific modulation coefficient vector is... The column dimension of the Query, Key, and Value matrices for each attention head is 64. For the Query matrix of the first attention head, the control coefficient vector is a floating-point column vector of length 64. After element-wise multiplication, the new Query matrix shows a significant increase in principal component vector energy under the conditions of mining area A. In the key matrix processing, the same 64-length control coefficient vector is applied, and the main diagonal value of the matrix dot product results increases overall, indicating that the similarity calculation has an enhanced ability to match typical features of mining area A. In the value matrix processing, the proportion of the cumulative distribution curve of the adjusted matrix in the high-weight interval increases, reflecting a significant enhancement in the response of the output features to key recovery factors under the specific conditions of the mining area. Finally, the Query, Key, and Value matrices processed as described above are used to calculate the adjusted attention weight matrix, and output feature representations that better match the actual conditions of mining area A in vegetation recovery capacity prediction. The contribution of each core feature in the prediction results is highly consistent with the historical recovery pattern of the mining area, achieving accurate model adaptation to the conditions of the mining area.

[0145] S6.4: Based on the dynamically adjusted query, key, and value matrix, the attention weight distribution of each attention head is recalculated and normalized using the softmax function to obtain an attention weight distribution matrix with specific perception capabilities for mining areas.

[0146] Based on the query matrix, key matrix, and value matrix adjusted element-wise by the mining area-specific control coefficient in the multi-head attention mechanism, a multi-head similarity calculation method (parameter: scaling factor equal to the reciprocal of the square root of the key vector dimension) is adopted to generate the relevance score of each attention head.

[0147] Furthermore, the unnormalized weight distribution of each attention head is calculated through matrix multiplication (parameter: multiplying the query matrix by the transpose of the key matrix), resulting in a correlation matrix that includes the specific adjustment effects of the mining area. This calculation can be explicitly expressed as:

[0148]

[0149] in, The adjusted query matrix, The adjusted bond matrix, It is the transpose operator. The dimension of the key vector.

[0150] S6.5: The weighted feature vectors output by each attention head are concatenated and their dimensions are restored through a linear transformation layer to generate a fusion feature representation containing mining area-specific information, thereby constructing the dynamic parameter space of the vegetation restoration prediction model and realizing the model's adaptive migration capability in different mining areas.

[0151] Step S7: A multi-task learning strategy is used to jointly train the vegetation restoration prediction model. An importance coefficient for mining area factors is introduced into the loss function. This importance coefficient is calculated through gradient backpropagation and is used to quantify the influence weights of different mining area-specific factors on the prediction results. Specifically, this includes:

[0152] S7.1: Based on the vegetation restoration prediction task in multiple mining areas, a multi-task learning framework is constructed. The multi-task learning framework includes multiple mining area sub-tasks. Each sub-task corresponds to a dynamic sequence prediction target for vegetation restoration in a mining area, so as to realize the joint modeling of common laws and individual characteristics.

[0153] Based on sample data from vegetation restoration prediction tasks across multiple mining areas, a task partitioning module is used to associate the input standardized vegetation index time series with the mining area-specific factor encoding vectors, generating independent sub-task datasets corresponding to each mining area. A task instantiation algorithm (parameters: number of tasks = number of mining areas M, task type: time series regression) is used to instantiate the model input for each mining area dataset, ensuring that each sub-task can independently execute the time series prediction process. Furthermore, a task parameter sharing mechanism (parameters: shared layer structure = the first N layers of the Transformer encoder) enables the learning of common patterns in the low-level feature space for each sub-task, while retaining high-level task-specific parameters for personalized feature modeling of mining areas. Finally, a task dependency construction method (parameters: dependency matrix is ​​a symmetric positive definite matrix, element values ​​are calculated from the feature similarity of mining areas) maps the feature similarity between sub-tasks to constraints between tasks, guiding the degree of feature sharing or separation in the model during joint training. Through the joint processing of task division, parameter sharing, and dependency construction, the input data is structurally mapped into a multi-task learning framework input that can simultaneously characterize both commonalities and individual characteristics, thereby achieving the ability to model prediction targets for different mining areas in parallel.

[0154] For example, in a prediction task covering five mining areas, the task partitioning module maps the five-year standardized vegetation index sequence (60 time points in total, 12 months per year) of each mining area to the encoding vectors of mineral type, mining method, and reclamation measures, forming five task datasets. The task instantiation algorithm is configured with M=5. Each task calls the same Transformer encoder for the first three layers of parameters to extract low-level features, while the last two layers maintain task-independent parameters to adapt to the individual conditions of each mining area. The mining area feature similarity matrix is ​​calculated by combining the Euclidean distances of topography, climate, and soil features. The matrix has a diagonal of 1 and off-diagonal elements ranging from 0.45 to 0.82. The task dependency construction method assigns higher weights to mining areas with similarity higher than 0.7 and lower weights to those with similarity lower than 0.5. During joint training, by sharing the feature space, the model can significantly improve the prediction accuracy for mining areas with similar climate patterns while maintaining its specific response capability to mining areas with significant climate differences. The final multi-task prediction framework showed a significant reduction in mean absolute error on the five mining area validation sets, and its generalization performance was significantly improved.

[0155] S7.2: The prediction errors of each mining area sub-task are weighted and summarized to construct a multi-task joint loss function. The loss function includes the product of the mean square error term of each sub-task and the importance coefficient of the mining area factor, so as to reflect the influence of mining area-specific factors on model training.

[0156] S7.3: During the backpropagation of the model, the importance coefficient of the mining area factor is calculated based on the gradient information of each mining area sub-task. The importance coefficient of the mining area factor represents the sensitivity of the mining area-specific factor to the prediction error, so as to guide the dynamic adjustment of the model parameters.

[0157] S7.4: Normalize the importance coefficients of mining area factors to eliminate the bias caused by the difference in sample size and prediction difficulty among different mining areas, and obtain standardized importance weights of mining area factors as the basis for the allocation of task weights for each mining area in the loss function.

[0158] S7.5: Based on the standardized importance weights of mining area factors, the task weight configuration in the multi-task joint loss function is optimized. The gradient descent algorithm is used to jointly update the parameters of the backbone model and the mining area factor branch network to improve the model's generalization ability and prediction stability in different mining areas.

[0159] Step S8: When deploying in a new mining area, the specific factors of the target mining area are encoded and input into the trained model. A parameter mapping function is used to automatically generate model parameter configurations adapted to the mining area, and vegetation restoration capacity prediction and assessment are performed. Specifically, this includes:

[0160] S8.1: Based on the discrete encoding input of specific factors such as mineral type, mining method, and reclamation measures of the target mining area, construct the mining area factor encoding vector as the input condition of the parameter mapping function.

[0161] S8.2: Input the mining area factor encoding vector into the pre-trained embedding layer, generate 8-16 dimensional mining area factor embedding vectors through a dimensionality compression algorithm, and perform L2 regularization to obtain standardized mining area feature representation vectors.

[0162] S8.3: Based on the feature representation vector of the mining area, a nonlinear transformation operation is performed through the fully connected layer in the mining area factor branch network to generate a parameter mapping function output containing mining area-specific control coefficients. The control coefficients are matched with the weight matrix dimension of the multi-head attention mechanism in the backbone model.

[0163] S8.4: The control coefficients output by the parameter mapping function are dynamically injected into the multi-head attention mechanism of the backbone model. An adaptively adjusted attention weight matrix is ​​generated through weighted fusion operations to construct a dynamic parameter space adapted to the target mining area.

[0164] The mining area-specific control coefficient vector output by the parameter mapping function is used as input to perform weighted fusion processing on the original attention weight matrix in the multi-head attention mechanism of the backbone model. An element-wise multiplication and addition weight adjustment algorithm is adopted (the parameter is set so that the length of the coefficient vector is equal to the number of rows corresponding to the weight matrix) to achieve quantitative adjustment of the response intensity of each attention unit.

[0165] Furthermore, by using dimension matching and broadcast expansion algorithms (the dimension expansion mode is from 1×N to H×N, where H is the number of attention heads), the control coefficient vector is aligned in the dimension space of each attention head, ensuring that the coefficients are uniformly mapped at the corresponding position of each head during the parameter injection process, thereby achieving fine-grained multi-head control.

[0166] Furthermore, by employing element-wise multiplication, the regulation coefficient vector is applied to the query matrix, key matrix, and value matrix of the attention mechanism, respectively, so that the regulation effect of mining area-specific factors is reflected in the three stages of similarity calculation, weight allocation, and feature aggregation.

[0167] S8.5: Based on the dynamic parameter space, perform spatiotemporal feature extraction and evolution prediction operations on the time series of standardized vegetation indices to generate prediction results of vegetation restoration capacity of the target mining area, and output a visual assessment report to support ecological restoration decision-making.

[0168] Based on a standardized vegetation index time series with dynamic parameter spatial input generated according to the target mining area, a Transformer encoder-decoder structure (parameters: number of heads = 8, embedding dimension = 512, positional encoding period constant set based on time step) is adopted to achieve multi-scale extraction of spatiotemporal features. Adaptive attention weights are injected into the key-value matrix through dynamic queries, and a multi-head self-attention mechanism is used to enhance the responsiveness to mining area-specific factors, resulting in a high-dimensional evolutionary feature matrix containing multiple time periods and spatial scales. Further, based on the evolutionary feature matrix, a time-series prediction decoding operation is performed, using a cross-attention mechanism to associate historical recovery states with current features, generating a vegetation index prediction sequence for multiple future time steps. Further, through a linear regression layer and an activation function output layer, the prediction sequence is transformed into a quantitative vegetation recovery capacity scoring index, quantifying the two core indicators: recovery speed and final coverage. Further, residual analysis is used to calculate the prediction error distribution, and residual variance is used to measure prediction stability, generating model confidence interval information. The visualization generation module (parameters: coordinate axis range automatically matches the predicted value range, color scheme is based on the recovery capacity level stratification) constructs an assessment report in the form of curves, bar charts and spatial heat maps to provide intuitive support for decision-making on ecological restoration of mining areas by generating recovery capacity scores, time series change trends and confidence interval ranges.

[0169] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0170] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0171] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery, characterized in that... Specifically, it includes: S1: Based on the remote sensing image dataset, obtain the dynamic sequence of vegetation restoration in the target mining area as the target variable, and simultaneously obtain multidimensional feature data of the mining area; S2: Perform normalization and null value imputation preprocessing on the vegetation restoration dynamic sequence to generate a standardized vegetation index time series dataset; S3: An embedding layer is constructed based on the discrete encoding of mining area specific factors. The embedding vector of mining area factors is generated by the dimensionality compression algorithm and then used as the feature representation of the mining area after L2 regularization. S4: Construct the backbone model branch network, and use the encoder-decoder structure of the Transformer architecture to extract spatiotemporal features from the standardized vegetation index time series dataset to generate a general evolutionary feature matrix for vegetation restoration. S5: Establish a mining area factor branch network, perform nonlinear transformation on the mining area factor embedding vector, and generate a parameter mapping function containing mining area-specific control coefficients; S6: Dynamically inject the mining area-specific regulation coefficients into the multi-head attention mechanism of the main model branch network, and generate adaptive attention weights through weighted fusion operations to construct a vegetation restoration prediction model. S7: A multi-task learning strategy is used to jointly train the vegetation restoration prediction model; S8: When deploying in a new mining area, the specific factors of the target mining area are encoded and input into the trained vegetation restoration prediction model. The model parameter configuration adapted to the mining area is automatically generated through the parameter mapping function, and the vegetation restoration capacity prediction and evaluation are performed.

2. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 1, characterized in that, The multidimensional feature data mentioned in step S1 includes mining area-specific factors such as mineral type, mining method, reclamation measures, and topographic, climatic, and soil data.

3. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 2, characterized in that, The embedding vector in step S1 has a dimension of 8-16.

4. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 1, characterized in that, The mining area-specific control coefficient mentioned in step S5 is matched with the dimension of the attention weight matrix in the branch network of the backbone model.

5. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 2, characterized in that, Topographic data includes elevation, slope, and aspect.

6. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 2, characterized in that, Climate data includes average annual temperature, annual precipitation, and annual evaporation.

7. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 2, characterized in that, Soil data includes organic matter content, pH value, and water holding capacity.

8. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 1, characterized in that, Step S4 specifically includes: Based on standardized vegetation index time series data, a time window sliding mechanism is constructed to segment the vegetation index series with a preset time step to generate spatiotemporal sequence samples for model input. Location encoding embedding is performed on the vegetation index sequence within the time window, and location vectors are constructed using sine and cosine functions to generate spatiotemporal feature vectors with enhanced location information. Based on the enhanced spatiotemporal feature vectors, an encoder module of the Transformer architecture is constructed, and a multi-head self-attention mechanism is used to perform context modeling on the input sequence to generate high-level semantic feature representations with context dependencies. Based on the high-level semantic feature representation of the encoder output, a decoder module of the Transformer architecture is constructed. A cross-attention mechanism is introduced to model the interaction between the encoder output and the historical vegetation restoration state, so as to generate a temporal evolution feature sequence of the vegetation restoration process. A global average pooling operation is performed on the temporal evolution feature sequence output by the decoder to generate a fixed-dimensional vegetation restoration general evolution feature matrix for subsequent parameter mapping.

9. The method for predicting and assessing vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 1, characterized in that, Step S6 specifically includes: Based on the original attention weight matrix of the multi-head attention mechanism in the backbone model branch network, the original attention weight is weighted and fused by the output regulation coefficient vector of the mining area factor branch network to generate an adaptive attention weight matrix that is adapted to the specificity of the current mining area. Perform dimension alignment on the mining area-specific control coefficient vector; apply the aligned mining area-specific control coefficient vector to the query, key and value matrices of each attention head in the multi-head attention mechanism, and dynamically adjust them through element-wise multiplication. Based on the dynamically adjusted query, key and value matrix, the attention weight distribution of each attention head is recalculated and normalized to obtain an attention weight distribution matrix with mining area-specific perception capabilities. The weighted feature vectors output by each attention head are concatenated and their dimensions are reduced by a linear transformation layer to generate a fusion feature representation containing mining area-specific information, in order to construct a vegetation restoration prediction model.

10. The method for predicting and evaluating vegetation restoration capacity in mining areas based on remote sensing imagery according to claim 9, characterized in that, The attention weight distribution of each attention head is recalculated and normalized using the softmax function.

Images