An ecosystem service function evaluation method based on multi-source geospatial-temporal data

Abstract

The application discloses an ecosystem service function evaluation method based on multi-source geographic spatio-temporal data, which comprises the following steps: S1, collecting multi-source geographic spatio-temporal data to form a geographic spatio-temporal data set; S2, encoding spatial position information and time sequence information in the geographic spatio-temporal data set, and forming a spatio-temporal correlation feature set through the correlation relationship of data in the spatio-temporal dimension; S3, inputting the spatio-temporal correlation feature set into a constructed ecosystem service function evaluation model to obtain a nonlinear mapping relationship between different features and the ecosystem service function, and establishing a quantitative correlation between the features and the ecosystem service function; S4, adopting a training strategy combining reinforcement learning and supervised learning to optimize the model; and S5, inputting spatio-temporal correlation features of a region to be evaluated into the trained ecosystem service function evaluation model to obtain a quantitative evaluation result. The application realizes comprehensive, accurate and dynamic evaluation of the ecosystem service function.

Description

Technical Field

[0001] This invention relates to the field of ecosystem service assessment technology, and specifically to a method for assessing ecosystem service functions based on multi-source geospatial data. Background Technology

[0002] With the increasing severity of global ecological and environmental problems, accurate assessment of ecosystem service functions is crucial for ecological protection, resource management, and sustainable development planning. Good water conservation functions ensure a stable water supply, soil retention functions reduce soil erosion and maintain soil fertility, windbreak and sand fixation functions are of great significance to the ecological security of arid and semi-arid regions, and biodiversity maintenance functions help maintain ecological balance and ecosystem stability. However, an ecosystem is a complex dynamic system, and its service functions are influenced by a variety of geographical and temporal factors, including land use type, ecosystem type, vegetation cover, topography, and meteorological and hydrological conditions. To achieve accurate assessment of ecosystem service functions, a method for evaluating ecosystem service functions is needed.

[0003] Traditional methods for assessing ecosystem service functions primarily rely on data from ground-based monitoring stations. Ecosystem-related indicators are obtained through field measurements and surveys. For example, when assessing soil conservation, soil erosion in a specific plot is measured to infer the soil conservation capacity of the entire region. While this method can obtain relatively accurate data in localized areas, it has significant limitations. Firstly, ground-based monitoring stations are sparsely distributed, failing to cover large areas and resulting in insufficient spatial representativeness of the data, thus failing to reflect the spatial heterogeneity of ecosystem service functions within a region. Secondly, field measurements require substantial human, material, and time resources, making them inefficient and unsuitable for meeting the needs of long-term dynamic monitoring and assessment of ecosystem service functions.

[0004] Existing technologies have begun to utilize satellite remote sensing data for assessment, but often simply use a single band or a combination of a few bands from remote sensing imagery to estimate indicators related to ecosystem service functions. For example, vegetation indexes are used alone to assess vegetation cover and infer biodiversity maintenance functions. While this method improves data acquisition efficiency and spatial coverage to some extent, it neglects the complexity of ecosystems and the interactions of multiple factors, resulting in poor accuracy and reliability of the assessment results. For instance, vegetation indices cannot accurately reflect the impact of topography on ecosystem service functions. In mountainous areas, topographic relief affects conditions such as sunlight and water, thus influencing vegetation growth and the functioning of ecosystem services, and these factors cannot be effectively taken into account.

[0005] At the same time, existing ecosystem service function assessment technologies also have shortcomings in data processing, model building, and computational efficiency, and cannot meet the requirements for a comprehensive, accurate, and dynamic assessment of ecosystem service functions. Summary of the Invention

[0006] Based on the above, this application discloses a method for assessing ecosystem service functions based on multi-source geospatial data, including:

[0007] S1. Collect multi-source geospatial data, and combine the land use type, ecosystem type and vegetation cover obtained after classifying satellite remote sensing image data with the elevation, slope and aspect data extracted from topographic data to form a geospatial dataset through data fusion technology.

[0008] S2. Encode the spatial location information and time series information in the geographic spatiotemporal dataset, obtain the spatiotemporal location encoding vector through the spatiotemporal correlation analysis model, and form a spatiotemporal correlation feature set through the correlation relationship of data in the spatiotemporal dimension;

[0009] S3. Using the spatiotemporal correlation feature set as input, combined with the type of ecosystem service function, input the constructed ecosystem service function assessment model to obtain the nonlinear mapping relationship between different features and ecosystem service functions, and initially establish the quantitative correlation between features and ecosystem service functions.

[0010] S4. A training strategy combining reinforcement learning and supervised learning is adopted for iterative training, so that the ecosystem service function assessment model can accurately predict ecosystem service functions based on spatiotemporal correlation characteristics.

[0011] S5. Input the spatiotemporal correlation characteristics of the area to be evaluated into the trained ecosystem service function assessment model to obtain quantitative assessment results of water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance in each area.

[0012] Preferably, the multi-source geospatial data in S1 includes high-resolution satellite remote sensing image data with spatial geographic information at different times, high-precision topographic data, rainfall, evapotranspiration, wind speed, runoff, soil type, soil moisture, soil physicochemical properties, species list and species distribution data.

[0013] Preferably, the step S1, which involves classifying land use types and ecosystem types based on satellite remote sensing image data, includes: constructing a spatiotemporal fusion feature classification model to classify the satellite remote sensing image data using a fusion architecture of multi-scale convolutional neural networks and spatiotemporal attention mechanisms, as shown in the formula: in, For position The classification results at the location, For position Multispectral image features at the location, For position The spatiotemporal context features of the location For time scale numbers, For spatial attention heads, and These are the attention weights for the time and space dimensions, respectively. For feature fusion operations, weights are adaptively allocated through a spatiotemporal attention mechanism to obtain temporal change features and spatial correlation features in image data at different times, and output classification results including land use types such as cultivated land, forest land, grassland, and water area, as well as ecosystem types such as forest ecosystem, grassland ecosystem, and wetland ecosystem.

[0014] Preferably, in step S1, vegetation cover is obtained by classifying high-resolution satellite remote sensing image data, specifically as follows:

[0015] Vegetation cover is calculated by constructing a spatiotemporal spectral fusion vegetation index model, using the following formula: ,in, For position Vegetation coverage at the location For position The conversion soil-adjusted vegetation index at the location is calculated using the following formula: , For near-infrared reflectivity, Reflectivity in the red light band Soil adjustment coefficient, This is the noise correction factor. and The TSI values ​​are for areas with pure soil and areas with pure vegetation, respectively. For time attention weights, The spatiotemporal attention mechanism is used to obtain the temporal variation characteristics of vegetation growth at different times; the spatiotemporal spectral fusion vegetation index model adaptively enhances the weight of vegetation spectral features through the spatiotemporal attention mechanism to classify vegetation coverage under complex terrain.

[0016] Preferably, in step S1, a geographic spatiotemporal dataset is formed using data fusion technology. By constructing a multi-source data fusion framework with a spatiotemporal feature pyramid network, land use type, ecosystem type, and vegetation cover data obtained from satellite remote sensing image classification are structurally processed with elevation, slope, and aspect data extracted from high-precision topographic data to generate feature vectors under a unified spatiotemporal reference. The weights of different data types are dynamically allocated using an attention mechanism. Through a hierarchical feature extraction and fusion strategy, the spatial topological features of the topographic data are cross-coupled with the land cover attribute features of the remote sensing classification data to form a geographic spatiotemporal dataset containing spatiotemporal attributes, land cover categories, and topographic features.

[0017] Preferably, in step S2, the spatiotemporal correlation analysis model is constructed to obtain the spatiotemporal location encoding vector, specifically as follows:

[0018] A spatiotemporal dual-attention encoding network is constructed to encode spatial location information and time series information in the geographic spatiotemporal dataset. The formula is as follows: ,in, For position At any moment The spatiotemporal location encoding vector, The spatial location encoding function generates spatial location embeddings using two-dimensional sine and cosine functions. The time-position encoding function generates time-position embeddings through a periodic function; Indicates position At any moment Multi-source feature vectors This is a spatial convolution operation used to extract local spatial features. For Long Short-Term Memory networks, capture temporal dependencies; and These are spatial and temporal attention mechanisms, respectively, which generate weight matrices through self-attention computation. Represents element-wise multiplication. For a multilayer perceptron, feature fusion and nonlinear transformation are performed; a spatiotemporal dual attention coding network is used to encode complex spatiotemporal relationships in geographic spatiotemporal data to obtain spatiotemporal location coding vectors.

[0019] Preferably, the ecosystem service function assessment model constructed in S3 is specifically as follows:

[0020] A hierarchical heterogeneous attention fusion network is constructed using the acquired spatiotemporal correlation feature set. A multi-scale spatiotemporal feature decomposition module is used to hierarchically decompose the spatiotemporal correlation features according to spatial scale and temporal granularity. Functional feature enhancement subnetworks are constructed for different ecosystem service functions such as water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance. Each subnetwork adopts a dynamic graph convolution structure, adaptively adjusting the correlation weights between nodes according to functional characteristics. The functional fusion features are calculated through a cross-functional attention fusion mechanism, using the following formula: ,in For the first Characteristics of ecosystem service functions For the first Each graph convolutional layer, For the spatiotemporal features after splitting, Based on the hierarchical heterogeneous network model that comprehensively evaluates multiple ecosystem service functions, an ecosystem service function assessment model is formed, with function-specific attention weights.

[0021] Preferably, in step S4, the ecosystem service function assessment model is optimized using a training strategy that combines reinforcement learning and supervised learning. Through a dual-space parameter optimization framework, actual observation data of known ecosystem service functions are used as supervised learning labels. The mean square error between the model's prediction results and the actual observation data is used as the supervised learning loss function. The reinforcement learning mechanism is used, with environmental state as the deviation between the model's prediction results and the actual situation, and action space as the model parameter adjustment strategy. The accuracy is evaluated through a reward function, enabling the model to continuously adjust parameters in a dynamic environment and optimize the model structure.

[0022] Preferably, the optimization model structure using the dual-space parameter optimization framework is formulated as follows: ,in, For model parameters, Optimize model parameters, The formula for the supervised learning mean squared error loss function is: , Based on actual observation data, The model predicts the result; in the reinforcement learning part, the state... The mean absolute error of the prediction bias, action For parameter adjustment amount, For the reward function, For balance coefficient, As a discount factor, The total number of time steps. The total number of samples.

[0023] Preferably, in step S5, quantitative assessment results of water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance in each region are obtained. Through a geographic information system, using three-dimensional visualization technology, the assessment results are integrated and displayed with a geographic map in the form of a three-dimensional layer, and the functional values ​​are distinguished by different colors, heights, and textures.

[0024] Compared with the prior art, the technical solution of this application has the following technical effects:

[0025] This invention collects multi-source geospatial data, including high-resolution satellite remote sensing images and high-precision topographic data. It uses a spatiotemporal fusion feature classification model to obtain accurate land use types and ecosystem types, calculates vegetation coverage using a spatiotemporal spectral fusion vegetation index model, and fuses multi-source data through a spatiotemporal feature pyramid network. This makes the data processing comprehensive and accurate, enabling in-depth mining of the information behind the data and a complete presentation of the characteristics of the ecosystem in the spatiotemporal dimension.

[0026] This invention constructs a hierarchical heterogeneous attention fusion network in the ecosystem service function assessment model. It splits spatiotemporal correlation features according to spatial scale and temporal granularity, constructs functional feature enhancement subnetworks for different ecosystem service functions, and adaptively adjusts weights using a dynamic graph convolution structure. Through a cross-functional attention fusion mechanism, it comprehensively evaluates multiple functions, effectively capturing the nonlinear mapping relationship between different features and ecosystem service functions. Compared with traditional models, it is more efficient in processing complex ecosystem data, significantly improves the accuracy and adaptability of the assessment model, and accurately quantifies the association between features and ecosystem service functions.

[0027] This invention employs a training strategy that combines reinforcement learning and supervised learning. Actual observation data serves as the label for supervised learning, and the loss function is calculated using mean squared error. In reinforcement learning, prediction bias is used as the environmental state, parameter adjustment as the action space, and improved evaluation accuracy as the reward function. Under this dual-space parameter optimization framework, the model continuously adjusts its parameters in a dynamic environment, achieving adaptive fusion of supervisory signals and environmental feedback. This training strategy optimizes the model structure, enabling it to better fit actual data, enhancing its generalization ability, and ensuring accurate prediction of ecosystem service functions in different scenarios.

[0028] This invention utilizes geographic information systems and 3D visualization technology to integrate quantitative assessment results of water conservation, soil retention, and other functions in various regions into a 3D layer and geographic map. Different colors, heights, and textures are used to distinguish the magnitude and distribution of functional values, making it easier for decision-makers to intuitively understand the spatial distribution characteristics of ecosystem service functions. At the same time, it can analyze the changing trends and evolution patterns based on assessment results at different times, providing intuitive and scientific decision-making basis for ecological protection, resource management, and regional planning, and helping to achieve sustainable development of ecosystems.

[0029] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0030] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0032] Figure 1 This is a flowchart of an ecosystem service function assessment method based on multi-source geospatial data according to the present invention.

[0033] Figure 2 A structural diagram of the multi-source data fusion framework for the spatiotemporal feature pyramid network;

[0034] Figure 3 A schematic diagram illustrating the selection of sampling points for ecosystem service function assessment methods;

[0035] Figure 4 Comparison chart of the accuracy of land use type classification for ecosystem service function assessment methods;

[0036] Figure 5 A graph showing the comparison of vegetation cover at sample point 1 between this application and existing technologies;

[0037] Figure 6 This is a comparison chart of the absolute error between the present application and the prior art for sample point 1;

[0038] Figure 7 This is a comparison chart of the error indices of this application and the prior art for sample point 1. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0040] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0041] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0042] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0043] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0044] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0045] Example 1

[0046] This embodiment mainly describes a method for assessing ecosystem service functions based on multi-source geospatial data, such as... Figure 1 As shown, it includes:

[0047] S1. Collect multi-source geospatial data, and combine the land use type, ecosystem type and vegetation cover obtained after classifying satellite remote sensing image data with the elevation, slope and aspect data extracted from topographic data to form a geospatial dataset through data fusion technology.

[0048] S2. Encode the spatial location information and time series information in the geographic spatiotemporal dataset, obtain the spatiotemporal location encoding vector through the spatiotemporal correlation analysis model, and form a spatiotemporal correlation feature set through the correlation relationship of data in the spatiotemporal dimension;

[0049] S3. Using the spatiotemporal correlation feature set as input, combined with the type of ecosystem service function, input the constructed ecosystem service function assessment model to obtain the nonlinear mapping relationship between different features and ecosystem service functions, and initially establish the quantitative correlation between features and ecosystem service functions.

[0050] S4. A training strategy combining reinforcement learning and supervised learning is adopted for iterative training, so that the ecosystem service function assessment model can accurately predict ecosystem service functions based on spatiotemporal correlation characteristics.

[0051] S5. Input the spatiotemporal correlation characteristics of the area to be evaluated into the trained ecosystem service function assessment model to obtain quantitative assessment results of water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance in each area.

[0052] Furthermore, S1 contains multi-source geospatial data, including high-resolution satellite remote sensing imagery with spatial geographic information at different times, high-precision topographic data, rainfall, evapotranspiration, wind speed, runoff, soil type, soil moisture, soil physicochemical properties, species list and species distribution data.

[0053] Furthermore, in S1, the land use types and ecosystem types classified based on satellite remote sensing image data include: a spatiotemporal fusion feature classification model is constructed using a fusion architecture of multi-scale convolutional neural networks and spatiotemporal attention mechanisms to classify satellite remote sensing image data, with the following formula: in, For position The classification results at the location, For position Multispectral image features at the location, For position The spatiotemporal context features of the location For time scale numbers, For spatial attention heads, and These are the attention weights for the time and space dimensions, respectively. For feature fusion operations, weights are adaptively allocated through a spatiotemporal attention mechanism to obtain temporal change features and spatial correlation features in image data at different times, and output classification results including land use types such as cultivated land, forest land, grassland, and water area, as well as ecosystem types such as forest ecosystem, grassland ecosystem, and wetland ecosystem.

[0054] Furthermore, vegetation cover is obtained in S1 through classification using high-resolution satellite remote sensing image data, specifically as follows:

[0055] Vegetation cover is calculated by constructing a spatiotemporal spectral fusion vegetation index model, using the following formula: ,in, For position Vegetation coverage at the location For position The conversion soil-adjusted vegetation index at the location is calculated using the following formula: , For near-infrared reflectivity, Reflectivity in the red light band Soil adjustment coefficient, This is the noise correction factor. and The TSI values ​​are for areas with pure soil and areas with pure vegetation, respectively. For time attention weights, The spatiotemporal attention mechanism is used to obtain the temporal variation characteristics of vegetation growth at different times; the spatiotemporal spectral fusion vegetation index model adaptively enhances the weight of vegetation spectral features through the spatiotemporal attention mechanism to classify vegetation coverage under complex terrain.

[0056] Furthermore, in S1, data fusion technology is used to form a geographic spatiotemporal dataset. By constructing a multi-source data fusion framework of a spatiotemporal feature pyramid network, land use type, ecosystem type, and vegetation cover data obtained from satellite remote sensing image classification are structured with elevation, slope, and aspect data extracted from high-precision topographic data to generate feature vectors under a unified spatiotemporal benchmark. An attention mechanism is used to dynamically allocate weights for different data types. Through a hierarchical feature extraction and fusion strategy, the spatial topological features of the topographic data and the land cover attribute features of the remote sensing classification data are cross-coupled to form a geographic spatiotemporal dataset containing spatiotemporal attributes, land cover categories, and topographic features.

[0057] Furthermore, in S2, a spatiotemporal correlation analysis model is constructed to obtain the spatiotemporal location encoding vector, specifically as follows:

[0058] A spatiotemporal dual-attention encoding network is constructed to encode spatial location information and time series information in the geographic spatiotemporal dataset. The formula is as follows: ,in, For position At any moment The spatiotemporal location encoding vector, The spatial location encoding function generates spatial location embeddings using two-dimensional sine and cosine functions. The time-position encoding function generates time-position embeddings through a periodic function; Indicates position At any moment Multi-source feature vectors This is a spatial convolution operation used to extract local spatial features. For Long Short-Term Memory networks, capture temporal dependencies; and These are spatial and temporal attention mechanisms, respectively, which generate weight matrices through self-attention computation. Represents element-wise multiplication. For a multilayer perceptron, feature fusion and nonlinear transformation are performed; a spatiotemporal dual attention coding network is used to encode complex spatiotemporal relationships in geographic spatiotemporal data to obtain spatiotemporal location coding vectors.

[0059] Furthermore, an ecosystem service function assessment model is constructed in S3, specifically as follows:

[0060] A hierarchical heterogeneous attention fusion network is constructed using the acquired spatiotemporal correlation feature set. A multi-scale spatiotemporal feature decomposition module is used to hierarchically decompose the spatiotemporal correlation features according to spatial scale and temporal granularity. Functional feature enhancement subnetworks are constructed for different ecosystem service functions such as water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance. Each subnetwork adopts a dynamic graph convolution structure, adaptively adjusting the correlation weights between nodes according to functional characteristics. The functional fusion features are calculated through a cross-functional attention fusion mechanism, using the following formula: ,in For the first Characteristics of ecosystem service functions For the first Each graph convolutional layer, For the spatiotemporal features after splitting, Based on the hierarchical heterogeneous network model that comprehensively evaluates multiple ecosystem service functions, an ecosystem service function assessment model is formed, with function-specific attention weights.

[0061] Furthermore, in S4, the ecosystem service function assessment model adopts a training strategy that combines reinforcement learning and supervised learning to optimize the model structure. Through a dual-space parameter optimization framework, the actual observation data of known ecosystem service functions are used as the supervised learning label. The mean square error between the model prediction results and the actual observation data is calculated as the supervised learning loss function. The reinforcement learning mechanism is used, with the environmental state as the deviation between the model prediction results and the actual situation, the action space as the model parameter adjustment strategy, and the accuracy is evaluated through the reward function. This allows the model to continuously adjust parameters in the dynamic environment and optimize the model structure.

[0062] Furthermore, the model structure is optimized using a dual-space parameter optimization framework, as shown in the following formula: ,in, For model parameters, Optimize model parameters, The formula for the supervised learning mean squared error loss function is: , Based on actual observation data, The model predicts the result; in the reinforcement learning part, the state... The mean absolute error of the prediction bias, action For parameter adjustment amount, For the reward function, For balance coefficient, As a discount factor, The total number of time steps. The total number of samples.

[0063] Furthermore, S5 obtains quantitative assessment results of water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance in various regions. Through the geographic information system, using three-dimensional visualization technology, the assessment results are integrated and displayed with the geographic map in the form of a three-dimensional layer, and the functional values ​​are distinguished by different colors, heights, and textures.

[0064] This embodiment details how multi-source geospatial data is collected, classified, fused, and encoded to construct an assessment model. A strategy combining reinforcement learning and supervised learning is used for training and optimization to achieve accurate assessment. This model can comprehensively and accurately quantify ecosystem service functions, clearly present the status of functions such as water conservation and soil retention in different regions, and provide intuitive and easy-to-understand assessment results through visualization. This facilitates information access for decision-makers and provides a scientific basis for ecological protection and resource management.

[0065] Based on Example 1, this implementation details how this application constructs a multi-source data fusion framework using a spatiotemporal feature pyramid network to fuse data into a geographic spatiotemporal dataset containing spatiotemporal attributes, land cover categories, and terrain features, such as... Figure 2 As shown, specifically:

[0066] A multi-source data fusion framework, the spatiotemporal feature pyramid network, is constructed. Pre-processing is performed on land use type, ecosystem type, and vegetation cover data obtained from satellite remote sensing image classification, as well as elevation, slope, and aspect data extracted from high-precision topographic data. These data are then divided according to their spatiotemporal characteristics, and a hierarchical spatiotemporal feature pyramid network is designed. From the bottom to the top, the spatial resolution of the data gradually decreases, but the level of feature abstraction gradually increases. At the bottom, the original high-resolution data is directly input, and a convolutional neural network (CNN) is used for feature extraction. For land use type data, the CNN can capture the boundary and texture features of different land features; for topographic data, the CNN can extract the undulation features of the terrain. As the network layer increases, pooling operations reduce the data resolution, while skip connections are used to fuse the detailed features of the bottom layer with the abstract features of the higher layer, forming a multi-scale feature representation.

[0067] To achieve effective fusion of different types of data, all data are unified under the same spatiotemporal reference, generating a unified feature vector. For remote sensing classification data such as land use type, ecosystem type and vegetation cover, coordinate transformation and resampling are used to make them spatially consistent with topographic data. The coordinate system of remote sensing images is transformed into the same coordinate system as the topographic data, and the remote sensing data is resampled according to the resolution of the topographic data.

[0068] In the time dimension, for multi-source data with time series, interpolation or aggregation operations are performed according to a uniform time interval. If some data have a high time resolution while others have a low time resolution, high-resolution data can be aggregated using statistical methods such as mean, maximum, or minimum values ​​to align it with low-resolution data in time. The processed data are then combined into feature vectors according to certain rules. For example, the independent codes of land use types, vegetation cover values, elevation values, slope values, and aspect values ​​can be arranged in sequence to form a feature vector containing multiple types of information. In this way, the feature vector of each location can comprehensively reflect the integrated information of that location in the spatiotemporal dimension.

[0069] After data preprocessing and feature vector generation, an attention mechanism is used to dynamically allocate weights for different data types. A hierarchical feature extraction and fusion strategy is employed to cross-couple the spatial topological features of terrain data with the land cover attribute features of remote sensing classification data. The core idea of ​​the attention mechanism is to allow the model to automatically learn the importance of different data features, thereby assigning higher weights to more important features. For example, in mountainous areas, topography has a greater impact on the ecosystem, so the attention mechanism will assign higher weights to topographic data such as elevation, slope, and aspect. In plains areas, land use type and vegetation cover have a more significant impact on ecosystem service functions, and the corresponding data feature weights will be increased.

[0070] Hierarchical feature extraction and fusion strategies process data through multiple levels to progressively mine and fuse features. At lower levels, the focus is on extracting and fusing local features, such as convolving feature vectors from adjacent locations to extract local spatial relationship features. As the levels increase, the scope of feature fusion gradually expands, comprehensively analyzing features from different regions to uncover more macroscopic spatiotemporal correlation features. In this way, the inherent connections between terrain data and remote sensing classification data can be fully explored, achieving deep feature fusion and forming a geographic spatiotemporal dataset that includes spatiotemporal attributes, land cover categories, and terrain features.

[0071] This embodiment describes in detail how a spatiotemporal feature pyramid network is constructed to fuse multi-source data, generate a unified spatiotemporal benchmark feature vector, and utilize attention mechanisms and hierarchical strategies to deeply fuse terrain and remote sensing data features. This enables the precise integration of multi-source geographic spatiotemporal data, the mining of complex spatiotemporal relationships, and the formation of a comprehensive dataset. This provides an accurate and reliable data foundation for the assessment of ecosystem service functions, thereby improving the accuracy and scientific rigor of the assessment.

[0072] Based on Example 1, this embodiment details the accuracy data of land use type and ecosystem type classification and the accuracy data of vegetation cover calculation in this application, specifically as follows:

[0073] This application utilizes a multi-scale convolutional neural network and a spatiotemporal attention mechanism to construct a spatiotemporal fusion feature classification model. A large number of samples were selected from the experimental region, such as... Figure 3 As shown, classification work is carried out on satellite remote sensing image data. Sample 1 is selected, such as... Figure 4 As shown, the land use type classification accuracy is as high as 93.56%. Among the 80 cultivated land samples, 74 were correctly classified; among the 65 forest land samples, 60 were correctly classified. In terms of ecosystem types, among the 50 forest ecosystem samples, 46 were correctly classified, with a classification accuracy of 92%; and among the 45 grassland ecosystem samples, 42 were correctly classified, with a classification accuracy of 93.33%.

[0074] However, when using single satellite remote sensing image data and the maximum likelihood classification method for classification, in the same experimental area, selecting sample 1, the land use type classification accuracy was only 70.32%. For example, forest land and grassland, due to their similar spectral characteristics, are easily misclassified under existing technology, leading to large deviations in classification results. As can be clearly seen from the figure, the accuracy curve of this application's technology is consistently higher than that of existing technologies across different land use types. In cultivated land classification accuracy, this application is approximately 23.24% higher than existing technologies; the difference is even more significant in forest land classification.

[0075] Regarding vegetation cover, such as Figure 5-7 As shown, this application constructs a spatiotemporal spectral fusion vegetation index model to calculate vegetation cover. Using sample 1, the results of field measurements at 20 sample points were compared with the model calculations. The average absolute error was 0.038, the root mean square error was 0.053, and the correlation coefficient was as high as 0.925. For example, the field measured vegetation cover at sample point 1 was 0.583, while the model calculated value was 0.615, with an absolute error of 0.032; the field measured value was 0.502, while the calculated value was 0.534, with an absolute error of 0.032. The overall error was controlled within a very small range, and the correlation coefficient curve of this application highly matched the fitted curve.

[0076] Existing technologies are mostly based on simple vegetation index algorithms, relying only on a few bands of satellite remote sensing imagery to calculate vegetation cover. In comparisons of similar points, their average absolute error reaches 0.156, root mean square error is 0.201, and correlation coefficient is only 0.701. Taking sample point 1 as an example, the vegetation cover calculated by existing technologies deviates significantly from the actual value compared to the technology in this application. In terms of root mean square error, the technology in this application is only about one-quarter of that of existing technologies. This fully demonstrates that the technology in this application is more accurate in calculating vegetation cover and can more accurately reflect the actual vegetation cover, providing more precise data support for the assessment of ecosystem service functions such as biodiversity maintenance and soil conservation.

[0077] This embodiment details how the present application constructs a classification and vegetation cover calculation model by integrating multi-scale convolutional neural networks and spatiotemporal attention mechanisms. The model significantly outperforms existing technologies in terms of accuracy in land use and ecosystem type classification. In vegetation cover calculation, the mean absolute error and root mean square error are smaller, and the correlation coefficient is higher, providing accurate data and strongly supporting the assessment of ecosystem service functions.

[0078] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for assessing ecosystem service functions based on multi-source geospatial data, characterized in that, Includes the following steps: S1. Collect multi-source geospatial data, and combine the land use type, ecosystem type and vegetation cover obtained after classifying satellite remote sensing image data with the elevation, slope and aspect data extracted from topographic data to form a geospatial dataset through data fusion technology. S2. Encode the spatial location information and time series information in the geographic spatiotemporal dataset, obtain the spatiotemporal location encoding vector through the spatiotemporal correlation analysis model, and form a spatiotemporal correlation feature set through the correlation relationship of data in the spatiotemporal dimension; S3. Using the spatiotemporal correlation feature set as input, combined with the type of ecosystem service function, input the constructed ecosystem service function assessment model to obtain the nonlinear mapping relationship between different features and ecosystem service functions, and initially establish the quantitative correlation between features and ecosystem service functions. The construction of an ecosystem service function assessment model includes: A hierarchical heterogeneous attention fusion network is constructed using the acquired spatiotemporal correlation feature set. A multi-scale spatiotemporal feature decomposition module is used to hierarchically decompose the spatiotemporal correlation features according to spatial scale and temporal granularity. Functional feature enhancement subnetworks are constructed for different ecosystem service functions such as water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance. Each subnetwork adopts a dynamic graph convolution structure, adaptively adjusting the correlation weights between nodes according to functional characteristics. The functional fusion features are calculated through a cross-functional attention fusion mechanism, using the following formula: ,in For the first Characteristics of ecosystem service functions For the first Each graph convolutional layer, For the spatiotemporal features after splitting, Based on the hierarchical heterogeneous network model that comprehensively evaluates multiple ecosystem service functions, an ecosystem service function assessment model is formed, with function-specific attention weights. S4. A training strategy combining reinforcement learning and supervised learning is adopted for iterative training, so that the ecosystem service function assessment model can accurately assess the ecosystem service function based on spatiotemporal correlation characteristics. S5. Input the spatiotemporal correlation characteristics of the area to be evaluated into the trained ecosystem service function assessment model to obtain quantitative assessment results of water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance in each area.

2. The method for assessing ecosystem service functions based on multi-source geospatial data according to claim 1, characterized in that, The multi-source geospatial data in S1 includes high-resolution satellite remote sensing imagery with spatial geographic information at different times, high-precision topographic data, rainfall, evapotranspiration, wind speed, runoff, soil type, soil moisture, soil physicochemical properties, species list and species distribution data.

3. The method for assessing ecosystem service functions based on multi-source geospatial data according to claim 1, characterized in that, S1 classifies land use types and ecosystem types based on satellite remote sensing image data as follows: A spatiotemporal fusion feature classification model is constructed using a multi-scale convolutional neural network and spatiotemporal attention mechanism to classify satellite remote sensing image data. The formula is as follows: in, For position The classification results at the location, For position Multispectral image features at the location, For position The spatiotemporal context features of the location For time scale numbers, For spatial attention heads, and These are the attention weights for the time and space dimensions, respectively. For feature fusion operations, weights are adaptively allocated through a spatiotemporal attention mechanism to obtain temporal change features and spatial correlation features in image data at different times, and output classification results including land use types such as cultivated land, forest land, grassland, and water area, as well as ecosystem types such as forest ecosystem, grassland ecosystem, and wetland ecosystem.

4. A method for assessing ecosystem service functions based on multi-source geospatial data according to claim 1 or 3, characterized in that, In S1, vegetation cover is obtained through classification using high-resolution satellite remote sensing image data, specifically: Vegetation cover is calculated by constructing a spatiotemporal spectral fusion vegetation index model, using the following formula: ,in, For position Vegetation coverage at the location For position The conversion soil-adjusted vegetation index at the location is calculated using the following formula: , For near-infrared reflectivity, For red light band reflectivity, Soil adjustment coefficient, This is the noise correction factor. and The TSI values ​​are for areas with pure soil and areas with pure vegetation, respectively. For time attention weights, The spatiotemporal attention mechanism is used to obtain the temporal variation characteristics of vegetation growth at different times; the spatiotemporal spectral fusion vegetation index model adaptively enhances the weight of vegetation spectral features through the spatiotemporal attention mechanism to classify vegetation coverage under complex terrain.

5. The method for assessing ecosystem service functions based on multi-source geospatial data according to claim 1, characterized in that, S1 utilizes data fusion technology to form a geospatial dataset. By constructing a multi-source data fusion framework with a spatiotemporal feature pyramid network, it structurally processes land use type, ecosystem type, and vegetation cover data obtained from satellite remote sensing image classification with elevation, slope, and aspect data extracted from high-precision topographic data to generate feature vectors under a unified spatiotemporal reference. It uses an attention mechanism to dynamically allocate weights for different data types and employs a hierarchical feature extraction and fusion strategy to cross-couple the spatial topological features of topographic data with the land cover attribute features of remote sensing classification data, thus forming a geospatial dataset containing spatiotemporal attributes, land cover categories, and topographic features.

6. The method for assessing ecosystem service functions based on multi-source geospatial data according to claim 1, characterized in that, In S2, a spatiotemporal correlation analysis model is constructed to obtain the spatiotemporal location encoding vector, specifically as follows: A spatiotemporal dual-attention encoding network is constructed to encode spatial location information and time series information in the geographic spatiotemporal dataset. The formula is as follows: ,in, For position At any moment The spatiotemporal location encoding vector, The spatial location encoding function generates spatial location embeddings using two-dimensional sine and cosine functions. The time-position encoding function generates time-position embeddings through a periodic function; Indicates position At any moment Multi-source feature vectors This is a spatial convolution operation used to extract local spatial features. For Long Short-Term Memory networks, capture temporal dependencies; and These are spatial and temporal attention mechanisms, respectively, which generate weight matrices through self-attention computation. Represents element-wise multiplication. For a multilayer perceptron, feature fusion and nonlinear transformation are performed; a spatiotemporal dual attention coding network is used to encode complex spatiotemporal relationships in geographic spatiotemporal data to obtain spatiotemporal location coding vectors.

7. The method for assessing ecosystem service functions based on multi-source geospatial data according to claim 1, characterized in that, In S4, a training strategy combining reinforcement learning and supervised learning is used to optimize the model structure of the ecosystem service function assessment model. Through a dual-space parameter optimization framework, actual observational data of known ecosystem service functions are used as labels for supervised learning. The mean squared error between the model's prediction results and the actual observational data is used as the supervised learning loss function. The reinforcement learning mechanism is used, with environmental state as the deviation between the model's prediction results and the actual situation, and action space as the model parameter adjustment strategy. The accuracy is evaluated through a reward function, enabling the model to continuously adjust parameters in a dynamic environment and optimize the ecosystem service function assessment model structure.

8. The method for assessing ecosystem service functions based on multi-source geospatial data according to claim 7, characterized in that, The structure of the ecosystem service function assessment model is optimized using a dual-space parameter optimization framework, and the formula is as follows: ,in, For model parameters, To optimize model parameters, The formula for the supervised learning mean squared error loss function is: , Based on actual observation data, The model predicts the result; in the reinforcement learning part, the state... The mean absolute error of the prediction bias, action For parameter adjustment amount, For the reward function, For balance coefficient, As a discount factor, The total number of time steps. The total number of samples.

9. The method for assessing ecosystem service functions based on multi-source geospatial data according to claim 1, characterized in that, In S5, quantitative assessment results of water conservation, soil conservation, windbreak and sand fixation, and biodiversity maintenance in various regions are obtained. Through the geographic information system, the assessment results are integrated with the geographic map in the form of a three-dimensional layer using three-dimensional visualization technology. The functional values ​​are distinguished by different colors, heights, and textures.

Images