DEM integrated intelligent modeling system based on graph convolutional neural network

Abstract

The invention discloses a DEM integrated intelligent modeling system based on a graph convolutional neural network, and the system comprises a data input and preprocessing module which is used for constructing an initial grid topological structure; the multi-relation graph construction module is used for constructing a multi-relation adjacency tensor and a relation characteristic tensor; the relation subset division module is used for dividing a hydrological relation subset and a landform relation subset; the branching graph convolution modeling module is used for setting a hydrological branch and a landform branch and executing directional modulation and anisotropic modulation; the cross gating fusion module is used for constructing an improved R-GCN model; the DEM embedding generation and mapping module is used for generating a DEM integrated embedding representation vector set; the terrain derived quantity calculation module is used for generating a slope grid, a slope direction grid and a hydrological network result; and the DEM integrated modeling output module is used for generating a DEM integrated modeling result. According to the invention, the elevation precision and hydrological coherence of a complex terrain area are improved.

Description

Technical Field

[0001] This invention relates to the field of GIS digital elevation modeling, and in particular to an integrated intelligent modeling system for DEM based on graph convolutional neural networks. Background Technology

[0002] Digital elevation models (DEMs) play a fundamental role in terrain reconstruction, geomorphological analysis, and hydrological calculations. Existing technologies generally rely on regular grid structures for spatial interpolation, slope analysis, and flow direction inference. However, the inconsistent resolution and complex noise distribution among multi-source DEM data make it difficult for traditional interpolation algorithms to simultaneously maintain terrain continuity and geomorphological structure characteristics, easily leading to problems such as elevation distortion, slope breaks, and water system interruptions.

[0003] In complex terrain conditions, relying solely on local windows or fixed neighborhood structures is insufficient to represent the multi-type, multi-scale relationships between topography, hydrology, and geomorphology. Existing deep learning-based terrain modeling methods mostly employ convolutional structures and rely on fixed windows for feature extraction, making it difficult to characterize spatial relationships such as water flow directionality, slope anisotropy, and geomorphic zoning boundaries. This results in insufficient prediction accuracy of the models in areas of abrupt slope changes, fault structures, and river catchment areas.

[0004] While traditional graph neural networks can handle irregular spatial structures, most models only support single relationships and do not differentiate and model multiple relationship features such as hydrological flow direction, geomorphic structure and spatial adjacency. They cannot express the different impacts of different relationships on elevation, slope and hydrological connectivity, and lack directional constraint mechanisms and relationship selection capabilities under multiple relationship conditions, resulting in insufficient fusion of multi-source spatial information.

[0005] Therefore, how to provide an integrated intelligent modeling system for DEM based on graph convolutional neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an integrated intelligent modeling system for geospatial data models (DEMs) based on graph convolutional neural networks (GCNs). This invention utilizes mechanisms such as multi-relational adjacency tensors, directional edge features, relation subset partitioning, branched graph convolution, and cross-gated fusion to distinguish and model spatial adjacency relationships, hydrological relationships, and geomorphic relationships, thereby achieving joint representation of multi-source geospatial data. By constructing hydrological and geomorphic branches in the R-GCN model and introducing a gated fusion structure, this invention significantly improves the continuity, directional consistency, and geomorphic structure representation capability of topographic element prediction, possessing advantages such as high accuracy, strong structural understanding ability, and adaptability to complex terrain.

[0007] According to an embodiment of the present invention, a DEM integrated intelligent modeling system based on graph convolutional neural networks includes: The data input and preprocessing module is used to collect geospatial basic data sets, perform data preprocessing, calculate and generate a set of basic node features, and construct an initial raster topology structure based on raster units. The multi-relationship graph construction module is used to determine spatial adjacency relationships, hydrological relationships, and geomorphological relationships based on the initial grid topology, construct a multi-relationship adjacency tensor, calculate directional edge features for each edge, and aggregate them into a relationship feature tensor; The relation subset partitioning module is used to partition hydrological relation subsets and geomorphological relation subsets; The branched graph convolution modeling module is used to set up hydrological and geomorphological branches in the R-GCN model and perform directional modulation and anisotropic modulation respectively. The cross-gating fusion module is used to generate hydrological gating coefficients and geomorphological gating coefficients, and to perform weighted fusion based on the gating coefficients, stacking multiple relational graph convolutional units to construct an improved R-GCN model; The DEM embedding generation and mapping module is used to input the node basic feature set into the improved R-GCN model, perform layer-by-layer message passing and feature update, and generate a set of DEM integrated embedding representation vectors and an elevation prediction result raster. The topographic derivation calculation module is used to calculate slope and aspect based on the elevation prediction results raster, generate slope raster and aspect raster, and generate hydrological network results based on hydrological relationships; The DEM integrated modeling output module is used to output elevation prediction raster, slope raster, aspect raster and hydrological network results as DEM integrated modeling results.

[0008] Optionally, modules can be integrated using the following methods: Collect a set of basic geospatial data, perform data preprocessing, calculate elevation, slope, aspect and curvature, combine to generate a set of basic node features, and establish an initial raster topology based on raster cells; Based on the initial grid topology, spatial adjacency, hydrological relationship and geomorphological relationship are determined according to the geometric adjacency, hydrological flow direction and geomorphological zoning relationship between grid cells respectively. Spatial adjacency, hydrological relationship and geomorphological relationship are used as edges connecting grid nodes and a multi-relationship adjacency tensor is constructed. Based on the multi-relational adjacency tensor, directional edge features are calculated for each edge, and the directional edge features are aggregated according to relation type to generate a relational feature tensor; The relation types in the multi-relation adjacency tensor and relation feature tensor are divided into hydrological relation subsets and geomorphological relation subsets; The message passing framework based on the R-GCN model sets a hydrological branch for the hydrological relation subset and a geomorphic branch for the geomorphic relation subset within each layer of relation graph convolutional unit. The two branches perform directional modulation and anisotropic modulation, respectively. A cross-gating unit is added to each layer of graph convolutional unit, gating coefficients are calculated, weighted fusion is performed on the outputs of the two branches, and multiple graph convolutional units are stacked to form an improved R-GCN model; The node basic feature set is input into the improved R-GCN model, and layer-by-layer message passing and feature updates are performed to calculate the slope raster, aspect raster and hydrological network results as the integrated DEM modeling results.

[0009] Optionally, the generation of the node basic feature set and the initial grid topology includes: Collect multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data and land cover data, perform format parsing and spatial range verification, and write the verified data into the geospatial basic data set. Data preprocessing is performed on the geospatial basic dataset, including converting the spatial reference information in the geospatial basic dataset into a unified coordinate reference system, performing resolution resampling according to the target spatial resolution, filling missing values ​​at locations with missing measurement markers, and generating a preprocessed geospatial basic dataset. Based on the DEM data in the preprocessed geospatial basic data set, the elevation, slope, aspect and curvature are calculated grid by grid. The land use type attributes and lithological attributes are combined with the elevation, slope, aspect and curvature to generate a set of node basic features. Based on the DEM data in the preprocessed geospatial basic data set, the spatial position of the raster unit in the unified coordinate reference system is determined. Each raster unit is recorded as a raster node in the node set, adjacency relationships are established, and the initial raster topology is constructed by combining them.

[0010] Optionally, the generation of the multi-relational adjacency tensor includes: Based on the existing set of grid nodes and the spatial location of grid nodes in the initial grid topology, a node index is generated for each grid node; Based on the row and column indices of the raster nodes, establish spatial adjacency relationships and record all raster node pairs with spatial adjacency relationships as a spatial adjacency relationship set. Based on hydrological element data, determine the flow direction of each grid node, establish hydrological relationships, and record all grid node pairs with hydrological relationships as a set of hydrological relationships; Based on topographic and geomorphic interpretation data and geological structure data, determine the geomorphic zoning range and structural boundary range, establish geomorphic relationships, and record all raster node pairs with geomorphic relationships as a geomorphic relationship set; The grid node pairs in the spatial adjacency relationship set are recorded as spatial adjacency relationship edges, the grid node pairs in the hydrological relationship set are recorded as hydrological relationship edges, and the grid node pairs in the geomorphological relationship set are recorded as geomorphological relationship edges, which together form an edge set; Based on the edge set and node index, spatial adjacency relationship channels, hydrological relationship channels and geomorphological relationship channels are constructed respectively, and combined into a multi-relational adjacency tensor stored by relationship type.

[0011] Optionally, the generation of the relation feature tensor includes: Based on the multi-relational adjacency tensor, each edge of the spatial adjacency relationship, hydrological relationship, and geomorphological relationship is read one by one, and the two grid nodes connected by each edge are read. The node feature vectors corresponding to the two grid nodes used for directional edge feature calculation are extracted from the node basic feature set. For hydrological relationship edges, the water flow direction vector is determined based on the water flow direction recorded in the geospatial basic data set. The elevation difference is calculated based on the elevation values ​​of the two grid nodes connected by each hydrological relationship edge, and the hydrological direction features are generated by combining them. For spatial adjacency edges and geomorphic relationship edges, the direction input is determined based on the slope direction and slope aspect recorded in the node basic feature set. The slope difference is calculated according to the slope value of the two grid nodes connected by each spatial adjacency edge and geomorphic relationship edge, and the geomorphic direction features are generated by combining them. All hydrological directional features are aggregated according to the storage order of hydrological relationship channels, and all geomorphic directional features are aggregated according to the storage order of geomorphic relationship channels. Relationship feature tensors are generated based on the aggregation results.

[0012] Optionally, the partitioning of the hydrological relation subset and the geomorphological relation subset includes: In the multi-relational adjacency tensor, spatial adjacency channels, hydrological channels, and geomorphological channels are identified by channel, and in the relational feature tensor, the corresponding relational feature channels are identified by the same channel order. Hydrological relationship channels are combined with their corresponding relationship feature channels to form a hydrological relationship subset, and geomorphic relationship channels are combined with spatial adjacency relationship channels and their corresponding relationship feature channels to form a geomorphic relationship subset.

[0013] Optionally, the generation of the hydrological and geomorphological branches includes: In each layer of the relation graph convolutional unit, the hydrological relation channel and the corresponding relation feature channel are input into the hydrological branch of the R-GCN model, and the geomorphological relation channel, the spatial adjacency relation channel and the corresponding relation feature channel are input into the geomorphological branch of the R-GCN model, so that the hydrological branch and the geomorphological branch form a parallel structure in the same relation graph convolutional unit. In each layer of the graph convolutional unit, a common node transformation matrix is ​​set for the hydrological branch and the geomorphological branch. A linear transformation is performed on the node feature vectors in the node basic feature set input to the hydrological branch and the geomorphological branch to obtain the transformed node feature vectors. In the hydrological branch, the hydrological direction features are combined with the transformation node feature vectors used by the hydrological branch, and the direction modulation is performed on the messages from the hydrological relation subset based on the hydrological direction features to produce directional adjustment; In the geomorphic branch, the geomorphic orientation features are combined with the transformation node feature vectors used by the geomorphic branch, and anisotropic modulation is performed on the messages from the geomorphic relation subset based on the geomorphic orientation features to generate directional differences.

[0014] Optionally, the construction of the improved R-GCN model includes: A cross-gating unit is set in each layer of the relation graph convolutional unit. The hydrological branch output and the geomorphological branch output are input into the cross-gating unit and spliced ​​in the feature dimension to obtain the spliced ​​features. In the cross-gated unit, linear transformation and nonlinear mapping are performed with splicing features as input to generate hydrological gating coefficients and geomorphological gating coefficients, forming a set of gating coefficients. The hydrological branch output and the geomorphological branch output are weighted and fused according to the gating coefficient set to obtain the fused output. The fused output is used as the node feature update result of the current layer relation graph convolutional unit, and the fused output is used as the input node feature vector of the next layer relation graph convolutional unit. Multiple relational graph convolutional units, each with a pre-defined number of layers and an output fusion function, are stacked sequentially to form an improved R-GCN model composed of multiple layers of relational graph convolutional units connected in series.

[0015] Optionally, the generation of the integrated DEM modeling results includes: The set of basic node features is arranged according to the index order of the grid nodes in the initial grid topology to form the set of input node feature vectors. This set is then input into the improved R-GCN model, and message passing and feature updates are performed sequentially. The set of node feature vectors output by the last layer relation graph convolutional unit of the improved R-GCN model is then used as the set of DEM integrated embedding representation vectors. Based on the correspondence between grid nodes and grid cells in the initial grid topology, each vector in the DEM integrated embedded representation vector set is mapped to the corresponding grid cell position. The elevation prediction component in the corresponding DEM integrated embedded representation vector is extracted at each grid cell to generate an elevation prediction result grid. Slope and aspect are calculated based on the elevation prediction values ​​of adjacent grid cells in the elevation prediction result grid, and slope and aspect grids are generated. Hydrological network extraction is performed based on the water flow direction in the hydrological relationship and the elevation prediction result grid to generate hydrological network results. The elevation prediction result grid, slope grid, aspect grid and hydrological network results are combined into DEM integrated modeling results.

[0016] The beneficial effects of this invention are: First, this invention constructs a multi-relational adjacency tensor that simultaneously includes spatial adjacency, hydrological relations, and geomorphological relations, and uses directional edge features to characterize water flow direction, slope changes, and geomorphological structure. This enables the model to accurately express the differential constraints between multi-source geospatial data, achieving continuity and reliability in elevation prediction under complex terrain conditions, and avoiding the problems of water system interruption, slope inconsistency, and blurred geomorphological boundaries in traditional methods.

[0017] Secondly, this invention sets up hydrological and geomorphological branches in the R-GCN model and combines directional modulation and anisotropic modulation, enabling the model to capture the unidirectional structural features under the control of hydrological flow direction and the structural consistency features within the geomorphological unit, respectively. This effectively improves the joint expression capability of topographic, hydrological and geomorphological information, thereby obtaining more physically consistent spatial prediction results in a multi-relational environment.

[0018] Furthermore, this invention introduces a cross-gating fusion mechanism to achieve adaptive weight control of the outputs of the hydrological and geomorphological branches, enabling the model to dynamically select more suitable relational information based on the topographic attributes of different regions. This improves the modeling accuracy in complex landforms, areas with abrupt slope changes, and fault structures, and significantly enhances the robustness and generalization ability of integrated DEM modeling. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an integrated intelligent modeling method for DEM based on graph convolutional neural networks proposed in this invention. Figure 2 This is a schematic diagram of the multi-relational adjacency tensor construction and directional edge feature generation structure in this invention; Figure 3 This is a schematic diagram of the improved R-GCN model structure of the present invention, which includes hydrological branches, geomorphological branches, and cross-gated units. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-3 A DEM integrated intelligent modeling system based on graph convolutional neural networks includes: The data input and preprocessing module is used to collect multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data and land cover data, perform coordinate unification, resolution resampling and missing value filling, generate geospatial basic dataset, merge and calculate elevation, slope, aspect and curvature, combine land use type attributes and lithological attributes with elevation, slope, aspect and curvature to generate node basic feature set, and construct initial raster topology structure based on resampled raster cells; The multi-relationship graph construction module is used to determine spatial adjacency relationships, hydrological relationships, and geomorphological relationships based on the initial grid topology, forming spatial adjacency relationship edges, hydrological relationship edges, and geomorphological relationship edges. The spatial adjacency relationship edges, hydrological relationship edges, and geomorphological relationship edges are written into the corresponding relationship channels to construct a multi-relationship adjacency tensor. Directional edge features are generated based on water flow direction, elevation difference, slope direction, slope aspect, and slope difference, and aggregated into relationship feature tensors according to relationship type. The relation subset partitioning module is used to identify spatial adjacency relation channels, hydrological relation channels, and geomorphological relation channels based on the multi-relation adjacency tensor. It combines the hydrological relation channels and their corresponding relation feature channels into hydrological relation subsets, and combines the geomorphological relation channels and spatial adjacency channels and their corresponding feature channels into geomorphological relation subsets, thus forming the relation input structure of the R-GCN model. The branched graph convolutional modeling module is used to set up hydrological and geomorphological branches in each layer of the relational graph convolutional unit of the R-GCN model. The two branches share the node transformation matrix. In the hydrological branch, directional modulation is performed based on hydrological directional features, and in the geomorphological branch, anisotropic modulation is performed based on geomorphological directional features. The cross-gated fusion module is used to input the hydrological branch output and the geomorphological branch output into the cross-gated unit, perform feature splicing, linear transformation and nonlinear mapping to generate hydrological gating coefficients and geomorphological gating coefficients, perform weighted fusion on the hydrological branch output and the geomorphological branch output based on the gating coefficients, and use the fused output as the update result of the current relation graph convolutional unit and stack them to form an improved R-GCN model; The DEM embedding generation and mapping module is used to input the node basic feature set into the improved R-GCN model to perform layer-by-layer message passing and feature update to generate a DEM integrated embedding representation vector set, and to map the DEM integrated embedding representation vector set into an elevation prediction result raster based on the initial raster topology. The topographic derivation calculation module is used to calculate the slope and aspect based on the elevation prediction result grid to generate slope grid and aspect grid, and to generate hydrological network results based on the water flow direction corresponding to the hydrological relationship and the elevation prediction result grid. The DEM integrated modeling output module is used to output elevation prediction raster, slope raster, aspect raster and hydrological network results as DEM integrated modeling results and provide unified DEM integrated modeling data to external application interfaces.

[0022] In this embodiment, the modules are interconnected using the following method: Multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data, and land cover data are collected and unified into a geospatial basic data set. Data preprocessing is performed, including coordinate unification, resolution resampling, and missing value imputation. Elevation, slope, aspect, and curvature are calculated. The attribute content in the geospatial basic data set is combined with elevation, slope, aspect, and curvature to generate a set of basic node features. An initial raster topology structure is established based on raster cells. Based on the initial grid topology, spatial adjacency, hydrological relations, and geomorphological relations are determined according to the geometric adjacency, hydrological flow direction, and geomorphological zoning relationship between grid cells. Spatial adjacency is used to connect adjacent grid nodes, hydrological relations are used to connect upstream and downstream nodes along the water flow path, and geomorphological relations are used to connect nodes within the same geomorphological unit and nodes on both sides of the construction boundary. Spatial adjacency, hydrological relations, and geomorphological relations are used as edges connecting grid nodes. Based on spatial adjacency, hydrological relations, and geomorphological relations, a multi-relationship adjacency tensor is constructed and stored in channels according to relation type. Based on the multi-relational adjacency tensor, directional edge features are calculated for each edge. Hydrological relations generate hydrological directional features based on the direction of water flow and the elevation difference between nodes. Spatial adjacency relations and geomorphological relations generate geomorphological directional features based on the slope direction, slope aspect and the slope difference between nodes. The directional edge features are aggregated according to the relation type to generate a relation feature tensor. The relation types in the multi-relation adjacency tensor and relation feature tensor are divided into hydrological relation subsets and geomorphological relation subsets. The hydrological relation subset is composed of hydrological relations, and the geomorphological relation subset is composed of geomorphological relations and spatial adjacency relations. The multi-relation adjacency tensor channels and relation feature tensor channels corresponding to the hydrological relation subset and the geomorphological relation subset are used as relation input data for constructing the R-GCN model. Based on the message passing framework of the R-GCN model, the model structure is improved by branching. In each layer of the relation graph convolutional unit, a hydrological branch is set for the hydrological relation subset and a geomorphological branch is set for the geomorphological relation subset. The two branches share the node transformation matrix. The hydrological branch performs directional modulation based on the hydrological directional features, and the geomorphological branch performs anisotropic modulation based on the geomorphological directional features. Cross-gating units are added to each layer of graph convolutional units. Gating coefficients are calculated based on the hydrological branch output and the geomorphological branch output. Weighted fusion is performed on the two branch outputs. Multiple graph convolutional units that have been improved by branching and gating are stacked to form an improved R-GCN model. The node basic feature set is input into the improved R-GCN model, and layer-by-layer message passing and feature update are performed to generate a DEM integrated embedding representation vector set. Based on the initial grid topology, the DEM integrated embedding representation vector set is mapped to the elevation prediction result grid. The elevation prediction result grid and the slope grid, aspect grid and hydrological network results calculated based on the elevation prediction result grid are output as the DEM integrated modeling result.

[0023] In this embodiment, the generation of the node basic feature set and the initial grid topology includes: Collect multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data, and land cover data. Perform format parsing and spatial extent verification on the multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data, and land cover data. Write the verified multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data, and land cover data into the geospatial basic data set. Data preprocessing is performed on the geospatial basic dataset, including converting the spatial reference information in the geospatial basic dataset into a unified coordinate reference system, performing resolution resampling on the DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data and land cover data in the geospatial basic dataset according to the target spatial resolution, and filling the missing values ​​in the resampled geospatial basic dataset with missing measurement markers according to the neighborhood interpolation rules and the same attribute constraint rules, and generating the preprocessed geospatial basic dataset; Based on the DEM data in the preprocessed geospatial basic data set, the elevation, slope, aspect and curvature are calculated grid by grid. The land use type attribute and lithology attribute corresponding to each grid cell in the preprocessed geospatial basic data set are combined with the elevation, slope, aspect and curvature. A node basic feature vector containing elevation, slope, aspect, curvature, land use type attribute and lithology attribute is generated at the location of each grid cell. The node basic feature vectors corresponding to all grid cells are used to form a node basic feature set. Based on the DEM data in the preprocessed geospatial basic data set, the spatial position of the raster cell in the unified coordinate reference system is determined. Each raster cell is recorded as a raster node in the node set. Adjacency relationships are established between raster nodes according to the adjacency relationship of the raster cells in the row and column directions. The node set and adjacency relationships are combined to construct the initial raster topology.

[0024] In this embodiment, the generation of the multi-relational adjacency tensor includes: Based on the existing set of grid nodes and the spatial location of grid nodes in the initial grid topology, a node index is generated for each grid node, and the node index is stored in a one-to-one correspondence with the spatial location of the grid node in the initial grid topology. Based on the row direction index and column direction index of the raster node, establish spatial adjacency between raster nodes with a row direction index difference of 1 and a column direction index difference of 0, and establish spatial adjacency between raster nodes with a row direction index difference of 0 and a column direction index difference of 1. Record all raster node pairs with spatial adjacency as a spatial adjacency set. Based on the hydrological element data in the geospatial basic data set, the water flow direction of each grid node is determined. Grid nodes located upstream of the water flow direction are paired with grid nodes located downstream of the water flow direction. Hydrological relationships are established between each pair of upstream and downstream grid nodes. All grid node pairs with hydrological relationships are recorded as a set of hydrological relationships. Based on the topographic and geomorphic interpretation data and geological structure data in the geospatial basic data set, the geomorphic zoning range and structural boundary range are determined. Geomorphic relationships are established between raster nodes located within the same geomorphic zoning range, and geomorphic relationships are established between raster nodes located on both sides of the structural boundary and spatially adjacent. All raster node pairs with geomorphic relationships are recorded as a geomorphic relationship set. Record the grid node pairs in the spatial adjacency relationship set as spatial adjacency relationship edges, record the grid node pairs in the hydrological relationship set as hydrological relationship edges, record the grid node pairs in the geomorphological relationship set as geomorphological relationship edges, and combine the spatial adjacency relationship edges, hydrological relationship edges, and geomorphological relationship edges to form an edge set. Based on the edge set and node index, spatial adjacency relationship channels, hydrological relationship channels, and geomorphological relationship channels are constructed respectively. In the spatial adjacency relationship channel, the starting raster node and the target raster node of each spatial adjacency relationship edge are recorded according to the node index. In the hydrological relationship channel, the upstream raster node and the downstream raster node of each hydrological relationship edge are recorded according to the node index. In the geomorphological relationship channel, the two raster nodes of each geomorphological relationship edge are recorded according to the node index. The spatial adjacency relationship channel, the hydrological relationship channel, and the geomorphological relationship channel are combined into a multi-relationship adjacency tensor stored by relationship type.

[0025] In this embodiment, the generation of the relation feature tensor includes: Based on the multi-relational adjacency tensor, each edge of the spatial adjacency relationship, hydrological relationship, and geomorphological relationship is read one by one, and the two grid nodes connected by each edge are read. The node feature vectors corresponding to the two grid nodes used for directional edge feature calculation are extracted from the node basic feature set. For hydrological relationship edges, the water flow direction vector is determined based on the water flow direction recorded in the geospatial basic data set. The elevation difference is calculated based on the elevation values ​​of the two grid nodes connected by each hydrological relationship edge in the node basic feature set. The water flow direction vector and the elevation difference are combined to generate hydrological direction features, and the hydrological direction features are recorded in the hydrological relationship channel corresponding to the hydrological relationship edge. For spatial adjacency edges and geomorphic relationship edges, the direction input is determined based on the slope direction and slope aspect recorded in the node basic feature set. The slope difference is calculated based on the slope values ​​of the two grid nodes connected by each spatial adjacency edge and geomorphic relationship edge in the node basic feature set. The slope direction, slope aspect and slope difference are combined to generate geomorphic direction features, and the geomorphic direction features are recorded in the geomorphic relationship channel corresponding to the spatial adjacency edge and geomorphic relationship edge. All hydrological directional features recorded in the hydrological relationship channel are aggregated according to the storage order of the hydrological relationship channel, and all geomorphic directional features recorded in the geomorphic relationship channel are aggregated according to the storage order of the geomorphic relationship channel. A relationship feature tensor is generated based on the aggregation results of the hydrological relationship channel and the geomorphic relationship channel.

[0026] In this embodiment, the division of the hydrological relation subset and the geomorphological relation subset includes: In the multi-relational adjacency tensor, spatial adjacency channels, hydrological channels, and geomorphological channels are identified by channel, and in the relational feature tensor, the corresponding relational feature channels are identified by the same channel order. Hydrological relationship channels and their corresponding relationship feature channels are combined into hydrological relationship subsets, and geomorphic relationship channels, spatial adjacency relationship channels, and their corresponding relationship feature channels are combined into geomorphic relationship subsets. The multi-relation adjacency tensor and relation feature tensor in the hydrological relation subset are used as the relation input data for the hydrological branch in the R-GCN model, and the multi-relation adjacency tensor and relation feature tensor in the geomorphological relation subset are used as the relation input data for the geomorphological branch in the R-GCN model. The two relation input data together constitute the relation input structure of the R-GCN model.

[0027] In this embodiment, the generation of the hydrological branch and the geomorphological branch includes: In each layer of the relation graph convolutional unit, the hydrological relation channels in the multi-relation adjacency tensor contained in the hydrological relation subset and the relation feature channels corresponding to the hydrological relation channels in the relation feature tensor are input into the hydrological branch of the R-GCN model. The geomorphological relation channels in the multi-relation adjacency tensor contained in the geomorphological relation subset and the spatial adjacency relation channels in the multi-relation adjacency tensor, as well as the relation feature channels corresponding to the geomorphological relation channels and spatial adjacency channels in the relation feature tensor, are input into the geomorphological branch of the R-GCN model, so that the hydrological branch and the geomorphological branch form a parallel structure within the same relation graph convolutional unit. In each layer of the graph convolutional unit, a common node transformation matrix is ​​set for the hydrological branch and the geomorphological branch. The node transformation matrix is ​​used to perform a linear transformation on the node feature vectors in the node basic feature set input to the hydrological branch and the geomorphological branch, so as to obtain the transformed node feature vectors used by the hydrological branch and the transformed node feature vectors used by the geomorphological branch. In the hydrological branch, the hydrological direction features are combined with the transformation node feature vector used by the hydrological branch. Based on the hydrological direction features, the message from the hydrological relation subset is subjected to direction modulation, so that the transmission amplitude and transmission path of the message in the direction of water flow are directionally adjusted according to the hydrological direction features. In the geomorphic branch, the geomorphic orientation features are combined with the transformation node feature vectors used by the geomorphic branch. Anisotropic modulation is performed on messages from the geomorphic relation subset based on the geomorphic orientation features, so that the transmission weight of messages in different geomorphic directions varies according to the geomorphic orientation features. The hydrological branch output with completed directional modulation and the geomorphological branch output with completed anisotropic modulation are provided to the cross-gating unit for use.

[0028] In this embodiment, the construction of the improved R-GCN model includes: A cross-gating unit is set in each layer of the relation graph convolutional unit. The hydrological branch output and the geomorphological branch output are input into the cross-gating unit. The hydrological branch output and the geomorphological branch output are concatenated in the feature dimension in the cross-gating unit to obtain the concatenated features used for gating calculation. In the cross-gating unit, linear transformation and nonlinear mapping are performed with splicing features as input to generate hydrological gating coefficients and geomorphological gating coefficients. The hydrological gating coefficients are used to control the weights of the hydrological branch outputs, and the geomorphological gating coefficients are used to control the weights of the geomorphological branch outputs. The hydrological gating coefficients and geomorphological gating coefficients are combined into a gating coefficient set. The hydrological branch output and the geomorphological branch output are weighted and fused according to the gating coefficient set to obtain the fused output. The fused output is used as the node feature update result of the current layer relation graph convolutional unit, and the fused output is used as the input node feature vector of the next layer relation graph convolutional unit. Multiple relational graph convolutional units, each with a pre-defined number of layers and an output fusion function, are stacked sequentially to form an improved R-GCN model composed of multiple layers of relational graph convolutional units connected in series.

[0029] In this embodiment, the generation of the integrated DEM modeling result includes: The set of basic node features is arranged in the index order of the grid nodes in the initial grid topology to form the set of input node feature vectors. The set of input node feature vectors is then input into the improved R-GCN model. Message passing and feature updates are performed sequentially in the graph convolutional units of each layer of the improved R-GCN model to obtain the set of node feature vectors output by the last graph convolutional unit of the improved R-GCN model. This set of node feature vectors is then used as the DEM integrated embedding representation vector set. Based on the correspondence between grid nodes and grid cells in the initial grid topology, each vector in the DEM integrated embedded representation vector set is mapped to the corresponding grid cell position. The elevation prediction component in the corresponding DEM integrated embedded representation vector is extracted at each grid cell to generate an elevation prediction result grid with elevation prediction value distribution in the grid space. Slope and aspect are calculated based on the elevation prediction values ​​of adjacent grid cells in the elevation prediction result grid, and slope and aspect grids are generated. Hydrological network extraction is performed based on the water flow direction in the hydrological relationship and the elevation prediction result grid to generate hydrological network results. The elevation prediction result grid, slope grid, aspect grid and hydrological network results are combined into DEM integrated modeling results. Example

[0030] To verify the feasibility of this invention in practice, it was applied to a mixed geographical area with complex topographic relief, dense river systems, and distinct geomorphic zoning. The multi-source DEM data in this area suffers from inconsistent resolution, high local noise, and discontinuous hydrological structures. It also includes slope structure information, geomorphic unit boundaries, and multiple flow paths extracted from topographic interpretation results. Existing modeling methods in such areas commonly exhibit abrupt slope changes, localized elevation prediction shifts, and hydrological network interruptions, making it difficult to guarantee the continuity of topography and consistency of hydrological relationships. To evaluate the practical effectiveness of this invention, its system was applied to a complete workflow of multi-source data integration, elevation prediction, and topographic derivation generation.

[0031] In this region, a unified geospatial foundation data set is constructed, including multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data, and land cover data. A consistent raster structure is obtained through coordinate unification and resolution resampling, and elevation, slope, aspect, and curvature are calculated to form a set of basic node features. Subsequently, an initial raster topology is constructed based on the raster cells, and spatial adjacency relationships, hydrological relationships, and geomorphological relationships are derived. These relationships are written into a multi-relational adjacency tensor. Directional edge features are generated by reading topographic derived features and hydrological features, and aggregated into a relational feature tensor according to relation type.

[0032] This region encompasses multiple slope sections, with slope variations ranging from 8 to 32 degrees and elevation variations ranging from 35 to 260 meters. Some areas exhibit significant elevation abrupt changes due to geomorphic zoning and fault structures, and multiple hydrological paths extend from upstream to downstream. Traditional methods in these areas are prone to producing abrupt slope changes, and due to the lack of explicit constraints on hydrological direction, unreasonable reverse elevation increases occur along the flow path. This invention, when constructing the model, inputs hydrological relationships as an independent subset into the hydrological branch and modulates the direction based on hydrological directional characteristics, making the elevation changes along the flow path more consistent with actual flow trends. Geomorphic relationships and spatial adjacency relationships constitute a subset of geomorphic relationships input into the geomorphic branch. Anisotropic modulation strengthens the topographic continuity within the same geomorphic unit, making the prediction results more stable in complex geomorphic regions.

[0033] In the actual modeling process, the improved R-GCN model continuously performs relational graph convolution and gated fusion on the features of each grid node, enabling the model to automatically adjust weights based on relational subsets of different spatial regions. For example, in slope transition sections, the model enhances the influence of geomorphic relationships and reduces interference from noise areas, thus lowering elevation prediction errors. In areas near water systems, the model increases the importance of hydrological relationships, maintaining the overall coherence of the hydrological network. In the final output elevation prediction results, the slope grid accurately reflects the slope change trend, the aspect grid correctly indicates the slope direction, and the hydrological network results maintain a high degree of consistency with the actual water system topology.

[0034] To verify the performance difference between this invention and traditional methods, the model of this invention was tested against the traditional bilinear interpolation method, the conventional convolutional neural network model, and the standard GCN model with undifferentiated relationships. Evaluation metrics included mean square error of elevation, mean absolute error of slope, aspect consistency ratio, and hydrological network coherence ratio. Experimental comparisons are as follows: Table 1. Comparison of DEM modeling performance between the method of this invention and traditional methods. As shown in Table 1, the method of this invention outperforms traditional methods in all four key indicators. In elevation prediction, the mean square error of this invention is 3.9, which is more accurate than bilinear interpolation (6.8), conventional CNN (5.1), and standard GCN (4.6). In slope calculation, the mean absolute error of this invention is 2.7, improving the stability and accuracy of slope in abrupt change regions. In terms of aspect consistency, this invention achieves 92%, thanks to the introduction of directional edge features, which enables the model to correctly capture slope trends. In terms of hydrological network coherence, this invention achieves 90%, significantly higher than traditional methods. The directional modulation mechanism based on hydrological relationships in the hydrological branches makes the elevation changes along the flow path more consistent with actual patterns. This invention achieves performance improvements in elevation accuracy, slope continuity, and hydrological network recognition. This is due to the spatial expression method constructed by the multi-relationship adjacency tensor and directional edge features, as well as the synergistic effect of the hydrological branch, geomorphological branch, and cross-gating fusion mechanism. This allows the model to adaptively strengthen effective relationships for different terrain features, thereby improving the overall performance of integrated DEM modeling.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A DEM integrated intelligent modeling system based on graph convolutional neural networks, characterized in that, include: The data input and preprocessing module is used to collect geospatial basic data sets, perform data preprocessing, calculate and generate a set of basic node features, and construct an initial raster topology structure based on raster units. The multi-relationship graph construction module is used to determine spatial adjacency relationships, hydrological relationships, and geomorphological relationships based on the initial grid topology, construct a multi-relationship adjacency tensor, calculate directional edge features for each edge, and aggregate them into a relationship feature tensor; The relation subset partitioning module is used to partition hydrological relation subsets and geomorphological relation subsets; The branched graph convolution modeling module is used to set up hydrological and geomorphological branches in the R-GCN model and perform directional modulation and anisotropic modulation respectively. The cross-gating fusion module is used to generate hydrological gating coefficients and geomorphological gating coefficients, and to perform weighted fusion based on the gating coefficients, stacking multiple relational graph convolutional units to construct an improved R-GCN model; The DEM embedding generation and mapping module is used to input the node basic feature set into the improved R-GCN model, perform layer-by-layer message passing and feature update, and generate a set of DEM integrated embedding representation vectors and an elevation prediction result raster. The topographic derivation calculation module is used to calculate slope and aspect based on the elevation prediction results raster, generate slope raster and aspect raster, and generate hydrological network results based on hydrological relationships; The DEM integrated modeling output module is used to output elevation prediction raster, slope raster, aspect raster and hydrological network results as DEM integrated modeling results.

2. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 1, characterized in that, The modules are connected in the following way: Collect a set of basic geospatial data, perform data preprocessing, calculate elevation, slope, aspect and curvature, combine to generate a set of basic node features, and establish an initial raster topology based on raster cells; Based on the initial grid topology, spatial adjacency, hydrological relationship and geomorphological relationship are determined according to the geometric adjacency, hydrological flow direction and geomorphological zoning relationship between grid cells respectively. Spatial adjacency, hydrological relationship and geomorphological relationship are used as edges connecting grid nodes and a multi-relationship adjacency tensor is constructed. Based on the multi-relational adjacency tensor, directional edge features are calculated for each edge, and the directional edge features are aggregated according to relation type to generate a relational feature tensor; The relation types in the multi-relation adjacency tensor and relation feature tensor are divided into hydrological relation subsets and geomorphological relation subsets; The message passing framework based on the R-GCN model sets a hydrological branch for the hydrological relation subset and a geomorphic branch for the geomorphic relation subset within each layer of relation graph convolutional unit. The two branches perform directional modulation and anisotropic modulation, respectively. A cross-gating unit is added to each layer of graph convolutional unit, gating coefficients are calculated, weighted fusion is performed on the outputs of the two branches, and multiple graph convolutional units are stacked to form an improved R-GCN model; The node basic feature set is input into the improved R-GCN model, and layer-by-layer message passing and feature updates are performed to calculate the slope raster, aspect raster and hydrological network results as the integrated DEM modeling results.

3. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 2, characterized in that, The generation of the node basic feature set and the initial grid topology includes: Collect multi-source DEM data, topographic and geomorphological interpretation data, hydrological element data, geological structure data and land cover data, perform format parsing and spatial range verification, and write the verified data into the geospatial basic data set. Data preprocessing is performed on the geospatial basic dataset, including converting the spatial reference information in the geospatial basic dataset into a unified coordinate reference system, performing resolution resampling according to the target spatial resolution, filling missing values ​​at locations with missing measurement markers, and generating a preprocessed geospatial basic dataset. Based on the DEM data in the preprocessed geospatial basic data set, the elevation, slope, aspect and curvature are calculated grid by grid. The land use type attributes and lithological attributes are combined with the elevation, slope, aspect and curvature to generate a set of node basic features. Based on the DEM data in the preprocessed geospatial basic data set, the spatial position of the raster unit in the unified coordinate reference system is determined. Each raster unit is recorded as a raster node in the node set, adjacency relationships are established, and the initial raster topology is constructed by combining them.

4. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 2, characterized in that, The generation of the multi-relation adjacency tensor includes: Based on the existing set of grid nodes and the spatial location of grid nodes in the initial grid topology, a node index is generated for each grid node; Based on the row and column indices of the raster nodes, establish spatial adjacency relationships and record all raster node pairs with spatial adjacency relationships as a spatial adjacency relationship set. Based on hydrological element data, determine the flow direction of each grid node, establish hydrological relationships, and record all grid node pairs with hydrological relationships as a set of hydrological relationships; Based on topographic and geomorphic interpretation data and geological structure data, determine the geomorphic zoning range and structural boundary range, establish geomorphic relationships, and record all raster node pairs with geomorphic relationships as a geomorphic relationship set; The grid node pairs in the spatial adjacency relationship set are recorded as spatial adjacency relationship edges, the grid node pairs in the hydrological relationship set are recorded as hydrological relationship edges, and the grid node pairs in the geomorphological relationship set are recorded as geomorphological relationship edges, which together form an edge set; Based on the edge set and node index, spatial adjacency relationship channels, hydrological relationship channels and geomorphological relationship channels are constructed respectively, and combined into a multi-relational adjacency tensor stored by relationship type.

5. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 2, characterized in that, The generation of the relation feature tensor includes: Based on the multi-relational adjacency tensor, each edge of the spatial adjacency relationship, hydrological relationship, and geomorphological relationship is read one by one, and the two grid nodes connected by each edge are read. The node feature vectors corresponding to the two grid nodes used for directional edge feature calculation are extracted from the node basic feature set. For hydrological relationship edges, the water flow direction vector is determined based on the water flow direction recorded in the geospatial basic data set. The elevation difference is calculated based on the elevation values ​​of the two grid nodes connected by each hydrological relationship edge, and the hydrological direction features are generated by combining them. For spatial adjacency edges and geomorphic relationship edges, the direction input is determined based on the slope direction and slope aspect recorded in the node basic feature set. The slope difference is calculated according to the slope value of the two grid nodes connected by each spatial adjacency edge and geomorphic relationship edge, and the geomorphic direction features are generated by combining them. All hydrological directional features are aggregated according to the storage order of hydrological relationship channels, and all geomorphic directional features are aggregated according to the storage order of geomorphic relationship channels. Relationship feature tensors are generated based on the aggregation results.

6. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 2, characterized in that, The division of the hydrological relation subset and the geomorphological relation subset includes: In the multi-relational adjacency tensor, spatial adjacency channels, hydrological channels, and geomorphological channels are identified by channel, and in the relational feature tensor, the corresponding relational feature channels are identified by the same channel order. Hydrological relationship channels are combined with their corresponding relationship feature channels to form a hydrological relationship subset, and geomorphic relationship channels are combined with spatial adjacency relationship channels and their corresponding relationship feature channels to form a geomorphic relationship subset.

7. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 2, characterized in that, The generation of the hydrological and geomorphological branches includes: In each layer of the relation graph convolutional unit, the hydrological relation channel and the corresponding relation feature channel are input into the hydrological branch of the R-GCN model, and the geomorphological relation channel, the spatial adjacency relation channel and the corresponding relation feature channel are input into the geomorphological branch of the R-GCN model, so that the hydrological branch and the geomorphological branch form a parallel structure in the same relation graph convolutional unit. In each layer of the graph convolutional unit, a common node transformation matrix is ​​set for the hydrological branch and the geomorphological branch. A linear transformation is performed on the node feature vectors in the node basic feature set input to the hydrological branch and the geomorphological branch to obtain the transformed node feature vectors. In the hydrological branch, the hydrological direction features are combined with the transformation node feature vectors used by the hydrological branch, and the direction modulation is performed on the messages from the hydrological relation subset based on the hydrological direction features to produce directional adjustment; In the geomorphic branch, the geomorphic orientation features are combined with the transformation node feature vectors used by the geomorphic branch, and anisotropic modulation is performed on the messages from the geomorphic relation subset based on the geomorphic orientation features to generate directional differences.

8. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 2, characterized in that, The construction of the improved R-GCN model includes: A cross-gating unit is set in each layer of the relation graph convolutional unit. The hydrological branch output and the geomorphological branch output are input into the cross-gating unit and spliced ​​in the feature dimension to obtain the spliced ​​features. In the cross-gated unit, linear transformation and nonlinear mapping are performed with splicing features as input to generate hydrological gating coefficients and geomorphological gating coefficients, forming a set of gating coefficients. The hydrological branch output and the geomorphological branch output are weighted and fused according to the gating coefficient set to obtain the fused output. The fused output is used as the node feature update result of the current layer relation graph convolutional unit, and the fused output is used as the input node feature vector of the next layer relation graph convolutional unit. Multiple relational graph convolutional units, each with a pre-defined number of layers and an output fusion function, are stacked sequentially to form an improved R-GCN model composed of multiple layers of relational graph convolutional units connected in series.

9. The integrated intelligent modeling system for DEM based on graph convolutional neural networks according to claim 2, characterized in that, The generation of the integrated DEM modeling results includes: The set of basic node features is arranged according to the index order of the grid nodes in the initial grid topology to form the set of input node feature vectors. This set is then input into the improved R-GCN model, and message passing and feature updates are performed sequentially. The set of node feature vectors output by the last layer relation graph convolutional unit of the improved R-GCN model is then used as the set of DEM integrated embedding representation vectors. Based on the correspondence between grid nodes and grid cells in the initial grid topology, each vector in the DEM integrated embedded representation vector set is mapped to the corresponding grid cell position. The elevation prediction component in the corresponding DEM integrated embedded representation vector is extracted at each grid cell to generate an elevation prediction result grid. Slope and aspect are calculated based on the elevation prediction values ​​of adjacent grid cells in the elevation prediction result grid, and slope and aspect grids are generated. Hydrological network extraction is performed based on the water flow direction in the hydrological relationship and the elevation prediction result grid to generate hydrological network results. The elevation prediction result grid, slope grid, aspect grid and hydrological network results are combined into DEM integrated modeling results.

Images