Grassland fertility dynamic evaluation method and system based on carbon-nitrogen ratio and carbon adhesion ratio

Abstract

The application discloses a grassland fertility dynamic evaluation method and system based on carbon-nitrogen ratio and carbon-clay ratio, which comprises multi-source data acquisition, data preliminary processing, multi-source data fusion, fertility evaluation model construction and grassland fertility dynamic evaluation. The application obtains original data through multi-source data acquisition; adopts a data preliminary processing method of time and space benchmark unification, remote sensing image preprocessing, soil data quality control and data set segmentation; adopts a method of taking carbon-nitrogen ratio and carbon-clay ratio as core indexes, deducing them from discrete sampling points to continuous spatial surface and fusing multi-modal data, so that the generated surface data has spatial continuity and contains rich ecological correlation semantics, providing a reliable and physically meaningful data basis for subsequent evaluation; and adopts a method of constructing a modal interaction graph and using a graph neural network as an evaluation model, so that the evaluation is not only based on the properties of the unit itself, but also fully considers the synergistic influence of the surrounding environment and ecological background.

Description

Technical Field

[0001] This invention relates to the field of dynamic evaluation technology for grassland fertility, specifically to a method and system for dynamic evaluation of grassland fertility based on carbon-nitrogen ratio and carbon-viscosity ratio. Background Technology

[0002] Grassland soil fertility assessment is a process of dynamically monitoring and evaluating the fertility level, nutrient status, and physicochemical properties of grassland soil by combining Geographic Information Systems (GIS) with advanced technologies such as machine learning and data mining. This technology not only helps to accurately assess the current status and potential of grassland resources, but also enables the prediction of grassland degradation, restoration, and sustainable utilization strategies based on changing trends, thus providing a scientific basis for grassland protection and management, ecological restoration, and sustainable agricultural development.

[0003] However, traditional grassland fertility assessment methods rely on intensive field sampling to create attribute spatial maps, which is costly and time-consuming. Even when auxiliary data is introduced, simple linear or statistical interpolation methods are often used, which are difficult to characterize the nonlinear and heterogeneous interaction mechanisms between soil properties and complex environmental factors, resulting in insufficient spatial prediction accuracy and high uncertainty. Traditional grassland fertility assessment methods also suffer from the problem that conventional models usually assume that each assessment unit is statistically independent, ignoring the spatial autocorrelation and ecological dependence phenomena that are common in grassland landscapes. They are unable to model the cascade and interaction effects of soil properties, vegetation status and other fertility elements in space, which makes fertility assessment prone to bias or smoothing distortion. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for dynamic evaluation of grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio. Traditional grassland soil fertility evaluation methods rely on intensive field sampling to create attribute spatial maps, resulting in high costs and poor timeliness. Even when auxiliary data is introduced, simple linear or statistical interpolation methods are often used, failing to characterize the nonlinear and heterogeneous interaction mechanisms between soil properties and complex environmental factors, leading to insufficient spatial prediction accuracy and high uncertainty. This solution creatively employs carbon-nitrogen ratio and carbon-viscosity ratio as core indicators, extrapolating them from discrete sampling points to a continuous spatial surface and fusing multimodal data. This results in spatially continuous areal data with rich ecological semantic associations, providing reliable and physically meaningful data for subsequent evaluation. Data Foundation: Traditional grassland fertility assessment methods often assume statistical independence among assessment units, neglecting the prevalent spatial autocorrelation and ecological dependence in grassland landscapes. They fail to model the cascading and interactive effects of soil properties, vegetation status, and other fertility elements in space, leading to biases or smoothing distortions in fertility assessments. This solution creatively employs a modal interaction graph and graph neural network as the assessment model. By constructing the modal interaction graph, the spatial heterogeneity and ecological correlation of grasslands are accurately expressed in a structured manner. The graph neural network model effectively aggregates and transmits information, simulating the spatial diffusion and interaction of ecological impacts. This allows the assessment to not only be based on the unit's own attributes but also fully consider the synergistic effects of its surrounding environment and ecological background.

[0005] The technical solution adopted in this invention is as follows: The method for dynamic evaluation of grassland fertility based on carbon-nitrogen ratio and carbon-viscosity ratio provided by this invention includes the following steps:

[0006] Step S1: Multi-source data acquisition;

[0007] Step S2: Preliminary data processing;

[0008] Step S3: Multi-source data fusion;

[0009] Step S4: Construction of the soil fertility evaluation model;

[0010] Step S5: Dynamic evaluation of grassland fertility.

[0011] Further, in step S1, the multi-source data acquisition is used to collect multi-source raw data required for dynamic evaluation of grassland fertility. Specifically, through data acquisition, a raw dataset for grassland fertility evaluation is obtained. The raw dataset for grassland fertility evaluation specifically includes a historical evaluation raw dataset and a current evaluation raw dataset. Both the historical evaluation raw dataset and the current evaluation raw dataset contain soil property modal data, remote sensing modal data, topographic modal data, and climate modal data. The historical evaluation raw dataset also contains grassland fertility evaluation level labeling data.

[0012] Further, in step S2, the preliminary data processing is used to perform preliminary processing on the collected multi-source raw data, specifically including the following steps:

[0013] Step S21: Spatiotemporal reference unification is used to eliminate inconsistencies in coordinate system, spatial resolution and time reference of multi-source data. Specifically, all data are unified to the CGCS2000 Gauss-Krüger projection through projection transformation, data at different resolutions are resampled to 10m resolution using bilinear interpolation, and time series data are aligned through temporal synthesis to obtain a spatiotemporally consistent dataset.

[0014] Step S22: Remote sensing image preprocessing, used to eliminate atmospheric interference, cloud cover and terrain shadow effects in remote sensing images. Specifically, radiometric calibration, atmospheric correction, cloud masking and terrain correction are performed on the remote sensing images, vegetation indices are calculated, and Savitzky-Golay filtering is used to smooth the time series to obtain continuous vegetation index data.

[0015] The vegetation index data include, but are not limited to, the normalized differential vegetation index, the enhanced vegetation index, the soil-regulating vegetation index, and the normalized differential water index.

[0016] Step S23: Soil data quality control, used to detect and process outliers and missing values ​​in soil data, and align point data with the raster grid, specifically using 3D... Outliers in the carbon-nitrogen ratio and carbon-viscosity ratio data were removed in principle. Missing data were filled in by Kriging interpolation. The global Moran index was calculated to assess spatial autocorrelation. The soil sampling point data was spatially matched with a 10m grid to obtain the processed soil property data.

[0017] Step S24: Dataset segmentation, used to obtain training data and test data, specifically, to segment the original historical evaluation dataset;

[0018] By unifying the spatiotemporal benchmark, preprocessing the remote sensing images, and controlling the soil data quality, the current evaluation original dataset is pre-processed to obtain a dataset to be fused. By unifying the spatiotemporal benchmark, preprocessing the remote sensing images, controlling the soil data quality, and segmenting the dataset, the historical evaluation original dataset is pre-processed to obtain a fusion training set and a fusion test set.

[0019] Furthermore, in step S3, the multi-source data fusion is used to extrapolate the carbon-nitrogen ratio and carbon-viscosity ratio from discrete sampling points to a continuous spatial surface and fuse multimodal data. Specifically, it involves constructing a random forest model, using the carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables for prediction, and fusing multimodal data to construct a modal interaction graph.

[0020] The multi-source data fusion specifically includes the following steps:

[0021] Step S31: Spatial expansion of soil properties, used to establish a random forest model to expand the carbon-nitrogen ratio and carbon-viscosity ratio from point observations to a continuous spatial surface. Specifically, with the carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables and vegetation index data, soil property modal data, topographic modal data, and climate modal data as independent variables, two random forest regression models are constructed to output the carbon-nitrogen ratio and carbon-viscosity ratio respectively. The model is trained and its performance is verified based on the fused training set and the fused test set. The dataset to be fused is used as input to obtain a 10m resolution spatial distribution map of the carbon-nitrogen ratio and a spatial distribution map of the carbon-viscosity ratio.

[0022] Step S32: Multimodal feature pyramid encoding, used to achieve multi-scale feature abstraction and modality fusion, specifically involves constructing a four-level feature pyramid, including the following steps:

[0023] Step S321: Original feature layer, used to retain all the information of the original features. Specifically, vegetation index data, soil attribute modal data, topographic modal data and climate modal data are respectively constructed into feature maps, which are directly used as the output of the original feature layer to obtain the original features of each modality.

[0024] Step S322: Intramodal fusion layer, used to perform feature fusion within each modality, specifically by performing 3×3 convolution, batch normalization and ReLU activation on each modality to obtain intramodal enhanced features for each modality;

[0025] Step S323: Cross-modal interaction layer, used to realize feature interaction between different modalities. Specifically, it obtains modal fusion features by using the intra-modal enhancement features of soil attribute modalities as queries and the concatenation of intra-modal enhancement features of non-soil attribute modalities as keys and values ​​through a cross-attention mechanism.

[0026] Step S324: High-level semantic layer, used to extract abstract ecological pattern features, specifically by further processing modality fusion features through 3×3 convolution and residual connections to obtain high-level modality fusion features;

[0027] Step S33: Dual-channel attention mechanism, used to enhance key information through spatial and feature dual channels, includes the following steps:

[0028] Step S331: Spatial attention channel, used to enhance important spatial regions in the feature map. Specifically, spatial context is obtained through global average pooling and global max pooling, and spatial attention weights are generated through 7×7 convolution and sigmoid activation to obtain spatial weighted features.

[0029] Step S332: Feature attention channel, used to enhance feature channels related to carbon-nitrogen ratio and carbon-viscosity ratio. Specifically, it calculates feature attention weights and obtains feature-weighted features by using a weighted fusion feature of carbon-nitrogen ratio and carbon-viscosity ratio as the query and high-level modality fusion feature as the key and value.

[0030] Step S333: Dual-channel attention fusion, used to adaptively fuse the output features of spatial and feature attention. Specifically, by learning channel fusion weight parameters, the outputs of spatial attention channel and feature attention channel are weighted and fused to obtain channel attention fusion features.

[0031] Step S334: Feature compression and enhancement, used to adjust feature dimensions and enhance feature expressive power. Specifically, it compresses the number of feature channels of channel attention fusion features through 1×1 convolution to obtain compressed and enhanced features.

[0032] Step S34: Modal interaction graph construction, used to convert raster features into graph structure data. Specifically, each 10m grid is defined as a graph node, compressed and enhanced features are used as node features, three types of edge connections are constructed: spatial adjacency, feature similarity, and ecological type. The adjacency matrix obtained based on the three types of edge connections is weighted and fused, and a comprehensive adjacency matrix is ​​obtained after normalization.

[0033] Spatial adjacency edges connect each node to its eight adjacent nodes (up, down, left, right, and diagonally) using the 8-neighborhood connection rule.

[0034] Feature-similar edges are obtained by calculating the cosine similarity of node features and connecting each node to the top M non-spatial adjacent nodes with the highest similarity.

[0035] Ecological type edges connect nodes of the same type using soil type and land use type information;

[0036] Step S35: Dataset processing, specifically, steps S31 to S34 are performed sequentially to process the dataset to be fused, the fusion training set, and the fusion test set to obtain the dataset to be evaluated, the evaluation training set, and the evaluation test set.

[0037] Furthermore, in step S4, the soil fertility evaluation model is constructed to build the model required to achieve accurate prediction of grassland soil fertility evaluation level, specifically by constructing a graph neural network model as the soil fertility evaluation model;

[0038] The construction of the soil fertility evaluation model specifically includes the following steps:

[0039] Step S41: Multi-head attention calculation, used to calculate the attention weights between nodes. Specifically, the similarity between node features is calculated by using 8 attention heads respectively, and the attention coefficients between node pairs are obtained after LeakyReLU activation and softmax normalization.

[0040] Step S42: Adaptive edge weight generation, used to dynamically adjust the importance of edges based on node features. Specifically, by learning adaptive edge weight parameters, based on the concatenation of adjacent node features and the comprehensive adjacency matrix, the sigmoid activation value is calculated to obtain the adaptive edge weight, which is then multiplied with the attention coefficient between each attention head node pair to obtain the adjusted attention coefficient.

[0041] Step S43: Feature aggregation, used to aggregate neighbor features and fuse multi-head outputs, includes the following steps:

[0042] Step S431: Attention head aggregation, used to aggregate neighbor features, specifically by weighting and aggregating neighbor features based on the adjusted attention coefficients to obtain the output node features of each attention head;

[0043] Step S432: Gated fusion, used to fuse multi-head outputs. Specifically, it fuses the output node features of 8 attention heads by learning gating weights, adds them to the original node features after linear transformation, and then performs layer normalization to obtain multi-head attention-enhanced node features.

[0044] Step S44: Obtain the model output, specifically by integrating the multi-head attention calculation, the adaptive edge weight generation, and the feature aggregation to construct a graph convolutional layer and stack multiple layers. The multi-head attention-enhanced node features output from the last graph convolutional layer are used as the input of an independent multilayer perceptron to obtain the model's predicted grassland fertility evaluation level label.

[0045] Step S45: Model construction and training, specifically, constructing a graph neural network model through multi-head attention calculation, adaptive edge weight generation, feature aggregation, and obtaining model output; training the model based on the evaluation training set and the evaluation test set and verifying its performance to obtain a graph neural network model as a soil fertility evaluation model.

[0046] Further, in step S5, the dynamic evaluation of grassland fertility specifically involves using the dataset to be evaluated as input to the fertility evaluation model to obtain the model-predicted grassland fertility evaluation level label. Based on spatial location information, the model maps back to geographic space to generate 10m resolution raster data with geographic coordinates. A model-predicted grassland fertility distribution map is constructed using GIS technology. Simultaneously, the soil carbon-nitrogen ratio and soil carbon-clay ratio in the current evaluation original dataset are transformed using natural logarithms. Based on preset critical values, a four-quadrant classification is performed. The geographic coordinates of sampling points and the four-quadrant classification information are integrated using GIS technology to draw a two-dimensional index grassland fertility distribution map. The model-predicted grassland fertility distribution map and the two-dimensional index grassland fertility distribution map are overlaid and adjusted to obtain a comprehensive visualization result that simultaneously carries the model prediction accuracy and the objectivity of the two-dimensional index.

[0047] The grassland soil fertility dynamic evaluation system based on carbon-nitrogen ratio and carbon-viscosity ratio provided by this invention includes a multi-source data acquisition module, a preliminary data processing module, a multi-source data fusion module, a soil fertility evaluation model construction module, and a grassland soil fertility dynamic evaluation module.

[0048] The multi-source data acquisition module is used to collect multi-source raw data, obtain the grassland fertility evaluation raw dataset by collecting multi-source raw data, and send the grassland fertility evaluation raw dataset to the data preliminary processing module.

[0049] The preliminary data processing module is used for preliminary data processing. Through preliminary data processing, a dataset to be fused, a fusion training set, and a fusion test set are obtained, and the dataset to be fused, the fusion training set, and the fusion test set are sent to the multi-source data fusion module.

[0050] The multi-source data fusion module is used for multi-source data fusion. By constructing a random forest model, it makes predictions with carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables, and fuses multimodal data to obtain the dataset to be evaluated, the evaluation training set, and the evaluation test set. The dataset to be evaluated is sent to the grassland soil fertility dynamic evaluation module, and the evaluation training set and the evaluation test set are sent to the soil fertility evaluation model construction module.

[0051] The soil fertility evaluation model construction module is used to construct a soil fertility evaluation model. It constructs a graph neural network model as a soil fertility evaluation model and sends the soil fertility evaluation model to the grassland soil fertility dynamic evaluation module.

[0052] The grassland soil fertility dynamic evaluation module is used for dynamic evaluation of grassland soil fertility. It processes the current data using the soil fertility evaluation model to obtain the model-predicted grassland soil fertility evaluation level label, and combines it with two-dimensional indicator information to obtain a comprehensive visualization result based on GIS technology.

[0053] The beneficial effects achieved by the present invention using the above solution are as follows:

[0054] (1) Traditional grassland fertility evaluation methods rely on intensive field sampling to draw attribute space maps, which is costly and time-consuming. Even when auxiliary data is introduced, simple linear or statistical interpolation methods are often used, which are difficult to characterize the nonlinear and heterogeneous interaction mechanism between soil properties and complex environmental factors, resulting in insufficient spatial prediction accuracy and high uncertainty. This scheme creatively adopts the carbon-nitrogen ratio and carbon-viscosity ratio as core indicators, extrapolates them from discrete sampling points to continuous spatial surfaces and integrates multimodal data, so that the generated surface data has spatial continuity and contains rich ecological association semantics, providing a reliable and physically meaningful data foundation for subsequent evaluation.

[0055] (2) In view of the technical problems of traditional grassland soil fertility evaluation methods, traditional models usually assume that each evaluation unit is statistically independent, ignore the spatial autocorrelation and ecological dependence phenomena that are common in grassland landscapes, and cannot model the cascade and interaction effects of soil properties, vegetation status and other soil fertility elements in space, which leads to bias or smoothing distortion in soil fertility evaluation. This scheme creatively adopts the method of constructing modal interaction diagrams and using graph neural networks as evaluation models. By constructing modal interaction diagrams, the spatial heterogeneity and ecological correlation of grassland are accurately expressed in a structured way. The graph neural network model can effectively aggregate and transmit information, simulate the diffusion and interaction process of ecological impact in space, so that the evaluation is not only based on the attributes of the unit itself, but also fully considers the synergistic effects of its surrounding environment and ecological background. Attached Figure Description

[0056] Figure 1 A schematic diagram of the process for the dynamic evaluation method of grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio provided by the present invention;

[0057] Figure 2 A schematic diagram of the modules of the dynamic evaluation system for grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio provided by the present invention;

[0058] Figure 3 This is a flowchart illustrating the preliminary data processing in step S2.

[0059] Figure 4 This is a flowchart illustrating the process of multi-source data fusion in step S3.

[0060] Figure 5 This is a schematic diagram of the process for constructing the geotechnical evaluation model in step S4.

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0064] Example 1, see Figure 1 The present invention provides a method for dynamic evaluation of grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio, the method comprising the following steps:

[0065] Step S1: Multi-source data acquisition;

[0066] Step S2: Preliminary data processing;

[0067] Step S3: Multi-source data fusion;

[0068] Step S4: Construction of the soil fertility evaluation model;

[0069] Step S5: Dynamic evaluation of grassland fertility.

[0070] Example 2, see Figure 1 and Figure 2This embodiment is based on the above embodiment. In step S1, the multi-source data acquisition is used to collect multi-source raw data required for dynamic evaluation of grassland fertility. Specifically, through data acquisition, a raw dataset for grassland fertility evaluation is obtained. The raw dataset for grassland fertility evaluation specifically includes a historical evaluation raw dataset and a current evaluation raw dataset. Both the historical evaluation raw dataset and the current evaluation raw dataset contain soil property modal data, remote sensing modal data, topographic modal data, and climate modal data. The historical evaluation raw dataset also contains grassland fertility evaluation level labeling data.

[0071] The soil property modal data specifically includes soil organic carbon content, soil total nitrogen content, and soil clay content used to calculate soil carbon-nitrogen ratio and carbon-clay ratio, the calculated soil carbon-nitrogen ratio and soil carbon-clay ratio, as well as soil type data, land use type data, soil pH data, and soil moisture content data.

[0072] The remote sensing modal data specifically includes Sentinel-2 data and Landsat-8 data with cloud cover below 20%;

[0073] The terrain modal data specifically includes geographic location data, elevation data, slope data, aspect data, and terrain relief data;

[0074] The climate modal data specifically includes ambient temperature and humidity data, precipitation data, sunshine duration data, and wind speed and direction data;

[0075] The grassland fertility evaluation level labeling data specifically refers to grassland fertility evaluation level labels, including excellent, good, medium, and poor.

[0076] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the preliminary data processing is used to perform preliminary processing on the collected multi-source raw data, specifically including the following steps:

[0077] Step S21: Spatiotemporal reference unification is used to eliminate inconsistencies in coordinate system, spatial resolution and time reference of multi-source data. Specifically, all data are unified to the CGCS2000 Gauss-Krüger projection through projection transformation, data at different resolutions are resampled to 10m resolution using bilinear interpolation, and time series data are aligned through temporal synthesis to obtain a spatiotemporally consistent dataset.

[0078] Step S22: Remote sensing image preprocessing, used to eliminate atmospheric interference, cloud cover and terrain shadow effects in remote sensing images. Specifically, radiometric calibration, atmospheric correction, cloud masking and terrain correction are performed on the remote sensing images, vegetation indices are calculated, and Savitzky-Golay filtering is used to smooth the time series to obtain continuous vegetation index data.

[0079] The vegetation index data include, but are not limited to, the normalized differential vegetation index, the enhanced vegetation index, the soil-regulating vegetation index, and the normalized differential water index.

[0080] Step S23: Soil data quality control, used to detect and process outliers and missing values ​​in soil data, and align point data with the raster grid, specifically using 3D... Outliers in the carbon-nitrogen ratio and carbon-viscosity ratio data were removed in principle. Missing data were filled in by Kriging interpolation. The global Moran index was calculated to assess spatial autocorrelation. The soil sampling point data was spatially matched with a 10m grid to obtain the processed soil property data.

[0081] Step S24: Dataset segmentation, used to obtain training data and test data, specifically, to segment the original historical evaluation dataset;

[0082] By unifying the spatiotemporal benchmark, preprocessing the remote sensing images, and controlling the soil data quality, the current evaluation original dataset is pre-processed to obtain a dataset to be fused. By unifying the spatiotemporal benchmark, preprocessing the remote sensing images, controlling the soil data quality, and segmenting the dataset, the historical evaluation original dataset is pre-processed to obtain a fusion training set and a fusion test set.

[0083] Example 4, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the multi-source data fusion is used to extrapolate the carbon-nitrogen ratio and carbon-viscosity ratio from discrete sampling points to a continuous spatial surface and fuse multimodal data. Specifically, it is to construct a random forest model, use the carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables for prediction, and fuse multimodal data to construct a modal interaction graph.

[0084] The multi-source data fusion specifically includes the following steps:

[0085] Step S31: Spatial expansion of soil properties, used to establish a random forest model to expand the carbon-nitrogen ratio and carbon-viscosity ratio from point observations to a continuous spatial surface. Specifically, with the carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables and vegetation index data, soil property modal data, topographic modal data, and climate modal data as independent variables, two random forest regression models are constructed to output the carbon-nitrogen ratio and carbon-viscosity ratio respectively. The model is trained and its performance is verified based on the fused training set and the fused test set. The dataset to be fused is used as input to obtain a 10m resolution spatial distribution map of the carbon-nitrogen ratio and a spatial distribution map of the carbon-viscosity ratio.

[0086] Step S32: Multimodal feature pyramid encoding, used to achieve multi-scale feature abstraction and modality fusion, specifically involves constructing a four-level feature pyramid, including the following steps:

[0087] Step S321: Original feature layer, used to retain all the information of the original features. Specifically, vegetation index data, soil attribute modal data, topographic modal data and climate modal data are respectively constructed into feature maps, which are directly used as the output of the original feature layer to obtain the original features of each modality.

[0088] Step S322: Intramodal fusion layer, used to perform feature fusion within each modality, specifically by performing 3×3 convolution, batch normalization and ReLU activation on each modality to obtain intramodal enhanced features for each modality;

[0089] Step S323: Cross-modal interaction layer, used to realize feature interaction between different modalities. Specifically, it obtains modal fusion features by using the intra-modal enhancement features of soil attribute modalities as queries and the concatenation of intra-modal enhancement features of non-soil attribute modalities as keys and values ​​through a cross-attention mechanism.

[0090] Step S324: High-level semantic layer, used to extract abstract ecological pattern features, specifically by further processing modality fusion features through 3×3 convolution and residual connections to obtain high-level modality fusion features;

[0091] Step S33: Dual-channel attention mechanism, used to enhance key information through spatial and feature dual channels, includes the following steps:

[0092] Step S331: Spatial attention channel, used to enhance important spatial regions in the feature map. Specifically, spatial context is obtained through global average pooling and global max pooling, and spatial attention weights are generated through 7×7 convolution and sigmoid activation to obtain spatially weighted features. The formula used is as follows:

[0093] ;

[0094] In the formula, Represents spatial attention weights. This represents the sigmoid activation function. This represents a 7×7 convolution function. This represents the global average pooling function. Represents the global max pooling function. Indicates high-level modality fusion features, Represents spatial weighted features. This represents element-wise multiplication.

[0095] Step S332: Feature attention channel, used to enhance feature channels related to carbon-nitrogen ratio and carbon-viscosity ratio. Specifically, it calculates feature attention weights and obtains feature-weighted features by using a weighted fusion feature of carbon-nitrogen ratio and carbon-viscosity ratio as the query and high-level modality fusion feature as the key and value.

[0096] Step S333: Dual-channel attention fusion, used to adaptively fuse the output features of spatial and feature attention. Specifically, by learning channel fusion weight parameters, the outputs of the spatial attention channel and the feature attention channel are weighted and fused to obtain the channel attention fusion features. The formula used is as follows:

[0097] ;

[0098] In the formula, Indicates the channel fusion weight, This represents the learnable channel fusion mapping weights. This represents the learnable channel fusion mapping bias term. This represents the channel attention fusion feature. This represents a feature-weighted feature, and T represents the transpose operation;

[0099] Step S334: Feature compression and enhancement, used to adjust feature dimensions and enhance feature expressive power. Specifically, it compresses the number of feature channels of channel attention fusion features through 1×1 convolution to obtain compressed and enhanced features.

[0100] Step S34: Modal interaction graph construction, used to convert raster features into graph structure data. Specifically, each 10m grid is defined as a graph node, compressed and enhanced features are used as node features, three types of edge connections are constructed: spatial adjacency, feature similarity, and ecological type. The adjacency matrix obtained based on the three types of edge connections is weighted and fused, and a comprehensive adjacency matrix is ​​obtained after normalization.

[0101] Spatial adjacency edges connect each node to its eight adjacent nodes (up, down, left, right, and diagonally) using the 8-neighborhood connection rule.

[0102] Feature-similar edges are obtained by calculating the cosine similarity of node features and connecting each node to the top M non-spatial adjacent nodes with the highest similarity.

[0103] Ecological type edges connect nodes of the same type using soil type and land use type information;

[0104] Step S35: Dataset processing, specifically, steps S31 to S34 are performed sequentially to process the dataset to be fused, the fusion training set, and the fusion test set to obtain the dataset to be evaluated, the evaluation training set, and the evaluation test set.

[0105] By performing the above operations, this solution addresses the technical problems of traditional grassland fertility evaluation methods, which rely on intensive field sampling to draw attribute spatial maps, resulting in high costs and poor timeliness. Even when auxiliary data is introduced, simple linear or statistical interpolation methods are often used, making it difficult to characterize the nonlinear and heterogeneous interaction mechanisms between soil properties and complex environmental factors, leading to insufficient spatial prediction accuracy and high uncertainty. This solution creatively adopts a method that uses carbon-nitrogen ratio and carbon-viscosity ratio as core indicators, extrapolates them from discrete sampling points to continuous spatial surfaces, and integrates multimodal data. This results in the generated areal data having spatial continuity and containing rich ecological association semantics, providing a reliable and physically meaningful data foundation for subsequent evaluation.

[0106] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the soil fertility evaluation model is constructed to build the model required to achieve accurate prediction of grassland soil fertility evaluation level. Specifically, a graph neural network model is constructed as the soil fertility evaluation model.

[0107] The construction of the soil fertility evaluation model specifically includes the following steps:

[0108] Step S41: Multi-head attention calculation, used to calculate the attention weights between nodes. Specifically, the similarity between node features is calculated using 8 attention heads, followed by LeakyReLU activation and softmax normalization to obtain the attention coefficients between node pairs for each attention head. The formula used is as follows:

[0109] ;

[0110] In the formula, This represents the attention coefficient between the h-th attention head node a and node b. This represents the LeakyReLU activation function. This represents the attention vector of the h-th attention head. This represents the node feature mapping weight of the h-th attention head. This represents the node characteristics of node a. Represents the node characteristics of node b. Let a represent the set of neighbors of node a. Represents the node characteristics of node c;

[0111] Step S42: Adaptive edge weight generation, used to dynamically adjust the importance of edges based on node features. Specifically, by learning adaptive edge weight parameters, calculating sigmoid activation values ​​based on the concatenation of adjacent node features and the comprehensive adjacency matrix, the adaptive edge weights are obtained. These weights are then multiplied by the attention coefficients between each pair of attention head nodes to obtain the adjusted attention coefficients. The formula used is as follows:

[0112] ;

[0113] In the formula, This represents the adaptive edge weight between node a and node b. This indicates learnable adaptive mapping weights. This represents the influence strength coefficient of the combined adjacency matrix. This represents a learnable adaptive mapping bias term. This represents the adjusted attention coefficient between the h-th attention head node a and node b;

[0114] Step S43: Feature aggregation, used to aggregate neighbor features and fuse multi-head outputs, includes the following steps:

[0115] Step S431: Attention head aggregation, used to aggregate neighbor features. Specifically, it aggregates neighbor features based on the adjusted attention coefficients to obtain the output node features of each attention head. The formula used is as follows:

[0116] ;

[0117] In the formula, This represents the output node feature of node a in the h-th attention head. This represents the adjusted attention coefficient between the h-th attention head node a and node c. Let h represent the learnable aggregate weight matrix of the h-th attention head;

[0118] Step S432: Gated fusion, used to fuse multi-head outputs. Specifically, it fuses the output node features of 8 attention heads by learning gating weights, adds them to the original node features after linear transformation, and then performs layer normalization to obtain the multi-head attention-enhanced node features. The formula used is as follows:

[0119] ;

[0120] In the formula, This represents the gating weight of node a in the h-th attention head. This represents the learnable gating mapping weights. This represents the learnable gating mapping bias term. Indicates the first The output node features of node a in the attention head, This indicates the multi-head attention enhancement node feature of node a. The layer normalization function is represented. This means concatenating the gated weighted results of all attention heads for node a. Represents the weights of the original node feature mapping;

[0121] Step S44: Obtain the model output, specifically by integrating the multi-head attention calculation, the adaptive edge weight generation, and the feature aggregation to construct a graph convolutional layer and stack multiple layers. The multi-head attention-enhanced node features output from the last graph convolutional layer are used as the input of an independent multilayer perceptron to obtain the model's predicted grassland fertility evaluation level label.

[0122] Step S45: Model construction and training, specifically, constructing a graph neural network model through multi-head attention calculation, adaptive edge weight generation, feature aggregation, and obtaining model output; training the model based on the evaluation training set and the evaluation test set and verifying its performance to obtain a graph neural network model as a soil fertility evaluation model.

[0123] By performing the above operations, this solution addresses the technical problems of traditional grassland fertility evaluation methods. Traditional models typically assume that each evaluation unit is statistically independent, neglecting the spatial autocorrelation and ecological dependence phenomena prevalent in grassland landscapes. They also fail to model the cascading and interactive effects of fertility elements such as soil properties and vegetation status in space, leading to biases or smoothing distortions in fertility evaluations. This solution creatively adopts a method of constructing modal interaction diagrams and using graph neural networks as the evaluation model. By constructing modal interaction diagrams, the spatial heterogeneity and ecological correlation of grasslands are accurately expressed in a structured way. The graph neural network model can effectively aggregate and transmit information, simulating the spatial diffusion and interaction process of ecological impacts. This allows the evaluation to not only be based on the unit's own attributes but also to fully consider the synergistic effects of its surrounding environment and ecological background.

[0124] Example 6, see Figure 1 and Figure 2This embodiment is based on the above embodiment. In step S5, the dynamic evaluation of grassland fertility specifically involves using the dataset to be evaluated as the input of the fertility evaluation model to obtain the model-predicted grassland fertility evaluation level label. Based on the spatial location information, it is mapped back to the geographic space to generate 10m resolution raster data with geographic coordinates. A model-predicted grassland fertility distribution map is constructed using GIS technology. At the same time, the soil carbon-nitrogen ratio and soil carbon-clay ratio in the current evaluation original dataset are transformed by natural logarithm. Based on the preset critical value, four-quadrant classification is performed. The geographic coordinates of the sampling points and the four-quadrant classification information are integrated using GIS technology to draw a two-dimensional index grassland fertility distribution map. The model-predicted grassland fertility distribution map and the two-dimensional index grassland fertility distribution map are overlaid and adjusted to obtain a comprehensive visualization result that simultaneously carries the model prediction accuracy and the objectivity of the two-dimensional index.

[0125] The preset critical values ​​are determined comprehensively based on the statistical analysis of measured soil data in the area to be evaluated, combined with the experience and demonstration of experts in the field. Specifically, they include the critical values ​​for the natural logarithmic transformation of soil carbon-nitrogen ratio and the critical values ​​for the natural logarithmic transformation of soil carbon-viscosity ratio.

[0126] The four-quadrant classification based on preset critical values ​​is specifically defined as follows: when the natural logarithmic transformation value of soil carbon-nitrogen ratio is less than or equal to the natural logarithmic transformation critical value of soil carbon-nitrogen ratio and the natural logarithmic transformation value of soil carbon-viscosity ratio is greater than or equal to the natural logarithmic transformation critical value of soil carbon-viscosity ratio, it is in the first quadrant, and the corresponding grassland fertility evaluation level label is excellent.

[0127] When the natural logarithmic transformation value of soil carbon-nitrogen ratio is greater than the critical value of natural logarithmic transformation of soil carbon-nitrogen ratio and the natural logarithmic transformation value of soil carbon-clay ratio is greater than or equal to the critical value of natural logarithmic transformation of soil carbon-clay ratio, it is in the second quadrant, and the corresponding grassland fertility evaluation grade label is good.

[0128] When the natural logarithmic transformation value of soil carbon-nitrogen ratio is less than or equal to the critical value of natural logarithmic transformation of soil carbon-nitrogen ratio and the natural logarithmic transformation value of soil carbon-clay ratio is less than the critical value of natural logarithmic transformation of soil carbon-clay ratio, it is in the third quadrant, and the corresponding grassland fertility evaluation level label is medium.

[0129] When the natural logarithmic transformation value of soil carbon-nitrogen ratio is greater than the critical value of the natural logarithmic transformation value of soil carbon-nitrogen ratio and the natural logarithmic transformation value of soil carbon-clay ratio is less than the critical value of the natural logarithmic transformation value of soil carbon-clay ratio, it is in the fourth quadrant, and the corresponding grassland fertility evaluation grade label is poor.

[0130] Example 7, see Figure 1 and Figure 2Based on the above embodiments, the grassland soil fertility dynamic evaluation system based on carbon-nitrogen ratio and carbon-viscosity ratio provided by the present invention includes a multi-source data acquisition module, a preliminary data processing module, a multi-source data fusion module, a soil fertility evaluation model construction module, and a grassland soil fertility dynamic evaluation module.

[0131] The multi-source data acquisition module is used to collect multi-source raw data, obtain the grassland fertility evaluation raw dataset by collecting multi-source raw data, and send the grassland fertility evaluation raw dataset to the data preliminary processing module.

[0132] The preliminary data processing module is used for preliminary data processing. Through preliminary data processing, a dataset to be fused, a fusion training set, and a fusion test set are obtained, and the dataset to be fused, the fusion training set, and the fusion test set are sent to the multi-source data fusion module.

[0133] The multi-source data fusion module is used for multi-source data fusion. By constructing a random forest model, it makes predictions with carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables, and fuses multimodal data to obtain the dataset to be evaluated, the evaluation training set, and the evaluation test set. The dataset to be evaluated is sent to the grassland soil fertility dynamic evaluation module, and the evaluation training set and the evaluation test set are sent to the soil fertility evaluation model construction module.

[0134] The soil fertility evaluation model construction module is used to construct a soil fertility evaluation model. It constructs a graph neural network model as a soil fertility evaluation model and sends the soil fertility evaluation model to the grassland soil fertility dynamic evaluation module.

[0135] The grassland soil fertility dynamic evaluation module is used for dynamic evaluation of grassland soil fertility. It processes the current data using the soil fertility evaluation model to obtain the model-predicted grassland soil fertility evaluation level label, and combines it with two-dimensional indicator information to obtain a comprehensive visualization result based on GIS technology.

[0136] It should 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, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0137] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0138] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A dynamic evaluation method for grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio, characterized in that: The method includes the following steps: S1: Multi-source data acquisition. Through data acquisition, the original dataset for grassland fertility evaluation is obtained. Specifically, the original dataset for grassland fertility evaluation includes historical evaluation datasets and current evaluation datasets. S2: Preliminary data processing: The collected multi-source raw data is preliminarily processed to obtain the dataset to be fused, the fusion training set, and the fusion test set; S3: Multi-source data fusion is used to extrapolate the carbon-nitrogen ratio and carbon-viscosity ratio from discrete sampling points to a continuous spatial surface and fuse multi-modal data. Specifically, it involves constructing a random forest model, using the carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables for prediction, fusing multi-modal data, constructing a modal interaction graph, and obtaining the dataset to be evaluated, the evaluation training set, and the evaluation test set. S4: Construction of the soil fertility evaluation model, which is used to build the model required to accurately predict the evaluation level of grassland soil fertility. Specifically, it is to build a graph neural network model as the soil fertility evaluation model. S5: Dynamic evaluation of grassland fertility, specifically, using the dataset to be evaluated as input to the fertility evaluation model to obtain the model's predicted grassland fertility evaluation level label, mapping it back to geographic space based on spatial location information to generate 10m resolution raster data with geographic coordinates, constructing a model-predicted grassland fertility distribution map using GIS technology, simultaneously performing a natural logarithmic transformation on the soil carbon-nitrogen ratio and soil carbon-clay ratio in the current evaluation original dataset, performing four-quadrant classification based on preset critical values, integrating the geographic coordinates of sampling points with the four-quadrant classification information using GIS technology, drawing a two-dimensional index grassland fertility distribution map, and overlaying and adjusting the model-predicted grassland fertility distribution map and the two-dimensional index grassland fertility distribution map to obtain a comprehensive visualization result that simultaneously carries the model's prediction accuracy and the objectivity of the two-dimensional index.

2. The method for dynamic evaluation of grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio according to claim 1, characterized in that: The multi-source data fusion specifically includes the following steps: Step S31: Spatial expansion of soil properties, used to establish a random forest model to expand the carbon-nitrogen ratio and carbon-viscosity ratio from point observations to a continuous spatial surface. Specifically, with the carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables and vegetation index data, soil property modal data, topographic modal data, and climate modal data as independent variables, two random forest regression models are constructed to output the carbon-nitrogen ratio and carbon-viscosity ratio respectively. The model is trained and its performance is verified based on the fused training set and the fused test set. The dataset to be fused is used as input to obtain a 10m resolution spatial distribution map of the carbon-nitrogen ratio and a spatial distribution map of the carbon-viscosity ratio. Step S32: Multimodal feature pyramid encoding, used to achieve multi-scale feature abstraction and modality fusion, specifically involves constructing a four-level feature pyramid, including the following steps: Step S321: Original feature layer, used to retain all the information of the original features. Specifically, vegetation index data, soil attribute modal data, topographic modal data and climate modal data are respectively constructed into feature maps, which are directly used as the output of the original feature layer to obtain the original features of each modality. Step S322: Intramodal fusion layer, used to perform feature fusion within each modality, specifically by performing 3×3 convolution, batch normalization and ReLU activation on each modality to obtain intramodal enhanced features for each modality; Step S323: Cross-modal interaction layer, used to realize feature interaction between different modalities. Specifically, it obtains modal fusion features by using the intra-modal enhancement features of soil attribute modalities as queries and the concatenation of intra-modal enhancement features of non-soil attribute modalities as keys and values ​​through a cross-attention mechanism. Step S324: High-level semantic layer, used to extract abstract ecological pattern features, specifically by further processing modality fusion features through 3×3 convolution and residual connections to obtain high-level modality fusion features; Step S33: Dual-channel attention mechanism, used to enhance key information through spatial and feature dual channels, includes the following steps: Step S331: Spatial attention channel, used to enhance important spatial regions in the feature map. Specifically, spatial context is obtained through global average pooling and global max pooling, and spatial attention weights are generated through 7×7 convolution and sigmoid activation to obtain spatial weighted features. Step S332: Feature attention channel, used to enhance feature channels related to carbon-nitrogen ratio and carbon-viscosity ratio. Specifically, it calculates feature attention weights and obtains feature-weighted features by using a weighted fusion feature of carbon-nitrogen ratio and carbon-viscosity ratio as the query and high-level modality fusion feature as the key and value. Step S333: Dual-channel attention fusion, used to adaptively fuse the output features of spatial and feature attention. Specifically, by learning channel fusion weight parameters, the outputs of spatial attention channel and feature attention channel are weighted and fused to obtain channel attention fusion features. Step S334: Feature compression and enhancement, used to adjust feature dimensions and enhance feature expressive power. Specifically, it compresses the number of feature channels of channel attention fusion features through 1×1 convolution to obtain compressed and enhanced features. Step S34: Modal interaction graph construction, used to convert raster features into graph structure data. Specifically, each 10m grid is defined as a graph node, compressed and enhanced features are used as node features, three types of edge connections are constructed: spatial adjacency, feature similarity, and ecological type. The adjacency matrix obtained based on the three types of edge connections is weighted and fused, and a comprehensive adjacency matrix is ​​obtained after normalization. Spatial adjacency edges connect each node to its eight adjacent nodes (up, down, left, right, and diagonally) using the 8-neighborhood connection rule. Feature-similar edges are obtained by calculating the cosine similarity of node features and connecting each node to the top M non-spatial adjacent nodes with the highest similarity. Ecological type edges connect nodes of the same type using soil type and land use type information; Step S35: Dataset processing, specifically, steps S31 to S34 are performed sequentially to process the dataset to be fused, the fusion training set, and the fusion test set to obtain the dataset to be evaluated, the evaluation training set, and the evaluation test set.

3. The method for dynamic evaluation of grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio according to claim 1, characterized in that: The construction of the soil fertility evaluation model specifically includes the following steps: Step S41: Multi-head attention calculation, used to calculate the attention weights between nodes. Specifically, the similarity between node features is calculated by using 8 attention heads respectively, and the attention coefficients between node pairs are obtained after LeakyReLU activation and softmax normalization. Step S42: Adaptive edge weight generation, used to dynamically adjust the importance of edges based on node features. Specifically, by learning adaptive edge weight parameters, based on the concatenation of adjacent node features and the comprehensive adjacency matrix, the sigmoid activation value is calculated to obtain the adaptive edge weight, which is then multiplied with the attention coefficient between each attention head node pair to obtain the adjusted attention coefficient. Step S43: Feature aggregation, used to aggregate neighbor features and fuse multi-head outputs, includes the following steps: Step S431: Attention head aggregation, used to aggregate neighbor features, specifically by weighting and aggregating neighbor features based on the adjusted attention coefficients to obtain the output node features of each attention head; Step S432: Gated fusion, used to fuse multi-head outputs. Specifically, it fuses the output node features of 8 attention heads by learning gating weights, adds them to the original node features after linear transformation, and then performs layer normalization to obtain multi-head attention-enhanced node features. Step S44: Obtain the model output, specifically by integrating the multi-head attention calculation, the adaptive edge weight generation, and the feature aggregation to construct a graph convolutional layer and stack multiple layers. The multi-head attention-enhanced node features output from the last graph convolutional layer are used as the input of an independent multilayer perceptron to obtain the model's predicted grassland fertility evaluation level label. Step S45: Model construction and training, specifically, constructing a graph neural network model through multi-head attention calculation, adaptive edge weight generation, feature aggregation, and obtaining model output; training the model based on the evaluation training set and the evaluation test set and verifying its performance to obtain a graph neural network model as a soil fertility evaluation model.

4. The method for dynamic evaluation of grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio according to claim 1, characterized in that: Both the historical evaluation original dataset and the current evaluation original dataset contain soil attribute modal data, remote sensing modal data, topographic modal data, and climate modal data. The historical evaluation original dataset also contains grassland fertility evaluation level labeling data.

5. The method for dynamic evaluation of grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio according to claim 1, characterized in that: The preliminary data processing specifically includes the following steps: Step S21: Spatiotemporal reference unification is used to eliminate inconsistencies in coordinate system, spatial resolution and time reference of multi-source data. Specifically, all data are unified to the CGCS2000 Gauss-Krüger projection through projection transformation, data at different resolutions are resampled to 10m resolution using bilinear interpolation, and time series data are aligned through temporal synthesis to obtain a spatiotemporally consistent dataset. Step S22: Remote sensing image preprocessing, used to eliminate atmospheric interference, cloud cover and terrain shadow effects in remote sensing images. Specifically, radiometric calibration, atmospheric correction, cloud masking and terrain correction are performed on the remote sensing images, vegetation indices are calculated, and Savitzky-Golay filtering is used to smooth the time series to obtain continuous vegetation index data. The vegetation index data include, but are not limited to, the normalized differential vegetation index, the enhanced vegetation index, the soil-regulating vegetation index, and the normalized differential water index. Step S23: Soil data quality control, used to detect and process outliers and missing values ​​in soil data, and align point data with the raster grid, specifically using 3D... Outliers in the carbon-nitrogen ratio and carbon-viscosity ratio data were removed in principle. Missing data were filled in by Kriging interpolation. The global Moran index was calculated to assess spatial autocorrelation. The soil sampling point data was spatially matched with a 10m grid to obtain the processed soil property data. Step S24: Dataset segmentation, used to obtain training data and test data, specifically, to segment the original historical evaluation dataset; By unifying the spatiotemporal benchmark, preprocessing the remote sensing images, and controlling the soil data quality, the current evaluation original dataset is pre-processed to obtain a dataset to be fused. By unifying the spatiotemporal benchmark, preprocessing the remote sensing images, controlling the soil data quality, and segmenting the dataset, the historical evaluation original dataset is pre-processed to obtain a fusion training set and a fusion test set.

6. A dynamic evaluation system for grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio, used to implement the dynamic evaluation method for grassland soil fertility based on carbon-nitrogen ratio and carbon-viscosity ratio as described in any one of claims 1-5, characterized in that: It includes a multi-source data acquisition module, a preliminary data processing module, a multi-source data fusion module, a soil fertility evaluation model construction module, and a grassland soil fertility dynamic evaluation module.

7. The grassland soil fertility dynamic evaluation system based on carbon-nitrogen ratio and carbon-viscosity ratio according to claim 6, characterized in that: The multi-source data acquisition module is used to collect multi-source raw data, obtain the grassland fertility evaluation raw dataset by collecting multi-source raw data, and send the grassland fertility evaluation raw dataset to the data preliminary processing module. The preliminary data processing module is used for preliminary data processing. Through preliminary data processing, a dataset to be fused, a fusion training set, and a fusion test set are obtained, and the dataset to be fused, the fusion training set, and the fusion test set are sent to the multi-source data fusion module. The multi-source data fusion module is used for multi-source data fusion. By constructing a random forest model, it makes predictions with carbon-nitrogen ratio and carbon-viscosity ratio as dependent variables, and fuses multimodal data to obtain the dataset to be evaluated, the evaluation training set, and the evaluation test set. The dataset to be evaluated is sent to the grassland soil fertility dynamic evaluation module, and the evaluation training set and the evaluation test set are sent to the soil fertility evaluation model construction module. The soil fertility evaluation model construction module is used to construct a soil fertility evaluation model. It constructs a graph neural network model as a soil fertility evaluation model and sends the soil fertility evaluation model to the grassland soil fertility dynamic evaluation module. The grassland soil fertility dynamic evaluation module is used for dynamic evaluation of grassland soil fertility. It processes the current data using the soil fertility evaluation model to obtain the model-predicted grassland soil fertility evaluation level label, and combines it with two-dimensional indicator information to obtain a comprehensive visualization result based on GIS technology.

Images