CN-122023697-A - DEM integrated intelligent modeling system based on graph convolution neural network

CN122023697ACN 122023697 ACN122023697 ACN 122023697ACN-122023697-A

Abstract

The invention discloses a Digital Elevation Model (DEM) integrated intelligent modeling system based on a graph convolution neural network, which comprises a data input and preprocessing module, a multi-relation graph construction module, a relation subset division module, a branching graph convolution modeling module, a cross gating fusion module, a DEM embedding generation and mapping module, a terrain derivative calculation module and a DEM integrated modeling output module, wherein the data input and preprocessing module is used for constructing an initial grid topological structure, the multi-relation graph construction module is used for constructing a multi-relation adjacent tensor and a relation characteristic tensor, the relation subset division module is used for dividing a hydrologic relation subset and a geomorphic relation subset, the branching graph convolution modeling module is used for setting hydrologic branches and landform branches and executing direction 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 expression vector set, and the terrain derivative calculation module is used for generating gradient grids, slope grids and hydrologic network results. The invention improves the elevation precision and hydrologic consistency of the complex terrain area.

Inventors

NI JIANJIANG
MAO QIFENG
YU ZHEKE
CHEN XUDONG
Shen Shangxiang
WANG CHEN

Assignees

余姚市规划测绘设计有限公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (9)

1. The utility model provides a Digital Elevation Model (DEM) integration intelligent modeling system based on graph convolution neural network which characterized in that includes: the data input and preprocessing module is used for acquiring a geospatial basic data set, executing data preprocessing, calculating to generate a node basic feature set, and constructing an initial grid topological structure according to the grid units; The multi-relation graph construction module is used for determining a space adjacent relation, a hydrologic relation and a geomorphic relation based on the initial grid topological structure, constructing a multi-relation adjacent tensor, calculating directional edge characteristics for each edge, and aggregating the directional edge characteristics into a relation characteristic tensor; the relation subset dividing module is used for dividing the hydrologic relation subset and the geomorphic relation subset; The branched graph convolution modeling module is used for setting hydrologic branches and landform branches in the R-GCN model and respectively executing direction modulation and anisotropic modulation; The cross gating fusion module is used for generating hydrologic gating coefficients and landform gating coefficients, performing weighted fusion according to the gating coefficients, and stacking a plurality of relation graph convolution units to construct an improved R-GCN model; The DEM embedding generation and mapping module is used for inputting the node basic feature set into the improved R-GCN model, performing layer-by-layer message transfer and feature update, and generating a DEM integrated embedding representation vector set and an elevation prediction result grid; The terrain derivative calculation module is used for calculating gradient and slope direction based on the elevation prediction result grid, generating a gradient grid and a slope direction grid, and generating a hydrologic network result according to the hydrologic relation; and the DEM integrated modeling output module is used for outputting elevation prediction result grids, gradient grids, slope grids and hydrologic network results as DEM integrated modeling results.
2. The intelligent modeling system for integrating DEM based on graph roll-up neural network as claimed in claim 1, wherein the modules are realized by the following method: collecting a geospatial basic data set, performing data preprocessing, calculating elevation, gradient, slope direction and curvature, combining to generate a node basic feature set, and establishing an initial grid topological structure based on grid units; Based on the initial grid topological structure, respectively determining a space adjacent relation, a hydrologic relation and a landform relation according to geometric adjacency, hydrologic flow direction and landform partition relation among grid units, taking the space adjacent relation, the hydrologic relation and the landform relation as edges for connecting grid nodes, and constructing a multi-relation adjacency tensor; Calculating directional edge characteristics for each edge based on the multi-relation adjacent tensor, and aggregating the directional edge characteristics according to relation types to generate a relation characteristic tensor; Dividing relationship types in the multi-relationship adjacent tensor and the relationship characteristic tensor into a hydrologic relationship subset and a landform relationship subset; Setting hydrologic branches for the hydrologic relation subsets in each layer of relation graph rolling units based on a message transmission framework of the R-GCN model, setting landform branches for the landform relation subsets, and respectively executing direction modulation and anisotropic modulation by the two branches; adding a cross gating unit in each layer of relation graph rolling unit, calculating gating coefficients, performing weighted fusion on two branch outputs, and stacking a plurality of relation graph rolling units to form an improved R-GCN model; inputting the node basic feature set into an improved R-GCN model, executing layer-by-layer message transfer and feature update, and calculating gradient grids, slope grids and hydrologic network results as DEM integrated modeling results.
3. The intelligent modeling system of claim 2, wherein the generating of the node base feature set and the initial grid topology comprises: collecting multisource DEM data, landform interpretation data, hydrologic element data, geological structure data and land coverage data, executing format analysis and space range verification, and uniformly writing the verified data into a geospatial basic data set; Performing data preprocessing on the geospatial base data set, wherein the data preprocessing comprises the steps of converting spatial reference information in the geospatial base data set into a unified coordinate reference system, performing resolution resampling according to target spatial resolution, filling missing values in positions with missing measurement marks, and generating a preprocessed geospatial base data set; Calculating elevation, gradient, slope direction and curvature grid by grid based on DEM data in the preprocessed geospatial base data set, and combining land utilization type attribute and lithology attribute with the elevation, gradient, slope direction and curvature to generate a node base feature set; and determining the spatial position of the grid units in the unified coordinate reference system based on the DEM data in the preprocessed geospatial basic data set, recording each grid unit as a grid node in the node set, establishing an adjacent relation, and combining to construct an initial grid topological structure.
4. The intelligent modeling system for DEM integration based on a graph roll-up neural network as claimed in claim 2, wherein said generating of said multi-relational-adjacency tensor includes: Generating a node index for each grid node based on the existing grid node set and the spatial positions of the grid nodes in the initial grid topology; establishing a space adjacent relation according to the row direction index and the column direction index of the grid nodes, and recording all grid node pairs with the space adjacent relation as a space adjacent relation set; determining the water flow direction of each grid node according to the hydrologic element data, establishing a hydrologic relation, and recording all grid node pairs with hydrologic relations as a hydrologic relation set; determining a geomorphic partition range and a construction boundary range according to the geomorphic interpretation data and the geological construction data, establishing a geomorphic relation, and recording all grid node pairs with the geomorphic relation as a geomorphic relation set; Recording grid node pairs in the spatial adjacent relation set as spatial adjacent relation edges, recording grid node pairs in the hydrologic relation set as hydrologic relation edges, recording grid node pairs in the geomorphic relation set as geomorphic relation edges, and forming an edge set together; And respectively constructing a spatial adjacent relation channel, a hydrologic relation channel and a geomorphic relation channel according to the edge set and the node index, and combining the spatial adjacent relation channel, the hydrologic relation channel and the geomorphic relation channel into a multi-relation adjacent tensor stored according to the relation type sub-channel.
5. The intelligent modeling system for DEM integration based on a graph roll-up neural network as claimed in claim 2, wherein said generating of the relational-feature tensor includes: Based on the multi-relation adjacent tensor, reading two grid nodes connected with each side one by one on the space adjacent relation side, the hydrologic relation side and the geomorphic relation side, and extracting node feature vectors corresponding to the two grid nodes for directional side feature calculation from a node basic feature set; Aiming at the hydrologic relation edges, determining a water flow direction vector according to the water flow direction recorded in the geospatial basic data set, calculating elevation differences according to elevation values of two grid nodes connected with each hydrologic relation edge, and combining to generate hydrologic direction features; determining a direction input quantity based on the gradient direction and the slope direction recorded in the node basic feature set aiming at the space adjacent relation side and the landform relation side, calculating gradient differences according to gradient values of two grid nodes connected with each space adjacent relation side and each landform relation side, and combining to generate a landform direction feature; all hydrologic direction features are aggregated according to the storage sequence of the hydrologic relation channels, all geomorphic direction features are aggregated according to the storage sequence of the geomorphic relation channels, and a relation feature tensor is generated according to an aggregation result.
6. The intelligent modeling system of claim 2, wherein the division of the hydrologic and geomorphic relationship subsets comprises: Identifying a spatial adjacent relationship channel, a hydrologic relationship channel and a geomorphic relationship channel according to channels in the multi-relationship adjacent tensor, and identifying corresponding relationship feature channels according to the same channel sequence in the relationship feature tensor; Combining the hydrologic relation channel and the corresponding relation characteristic channel into a hydrologic relation subset, and combining the geomorphic relation channel, the spatial adjacent relation channel and the corresponding relation characteristic channel into a geomorphic relation subset.
7. The intelligent modeling system of claim 2, wherein the generating of the hydrologic and geomorphic branches comprises: in each layer of relation graph rolling unit, inputting a hydrologic relation channel and a corresponding relation characteristic channel into a hydrologic branch of an R-GCN model, and inputting a landform relation channel, a space adjacent relation channel and a corresponding relation characteristic channel into a landform branch of the R-GCN model, so that the hydrologic branch and the landform branch form a parallel structure in the same relation graph rolling unit; setting a common node transformation matrix for the hydrologic branch and the geomorphic branch in each layer of relation graph rolling units, and performing linear transformation on node feature vectors in node basic feature sets input to the hydrologic branch and the geomorphic branch to obtain transformed node feature vectors; In the hydrologic branch, combining hydrologic direction characteristics with transformation node characteristic vectors used by the hydrologic branch, and executing direction modulation on messages from the hydrologic relation subset according to the hydrologic direction characteristics to generate directivity adjustment; in the landform branch, combining the landform direction feature with a transformation node feature vector used by the landform branch, performing anisotropic modulation on a message from the landform relation subset according to the landform direction feature, and generating a direction difference.
8. The intelligent modeling system for DEM integration based on a graph roll-up neural network according to claim 2, wherein the construction of the improved R-GCN model includes: a cross gating unit is arranged in each layer of relation graph rolling unit, hydrologic branch output and landform branch output are input into the cross gating unit, and splicing is carried out on characteristic dimensions to obtain splicing characteristics; In the cross gating unit, linear transformation and nonlinear mapping are carried out by taking the splicing characteristic as input, and a hydrologic gating coefficient and a landform gating coefficient are generated to form a gating coefficient set; Performing weighted fusion on hydrological branch output and landform branch output according to the gating coefficient set to obtain fusion output, taking the fusion output as a node characteristic updating result of a convolution unit of the relation graph of the layer, and taking the fusion output as an input node characteristic vector of a convolution unit of the relation graph of the next layer; And stacking a plurality of relation graph convolution units which are internally provided with the cross gating units and output fusion output in sequence according to a preset layer number to form an improved R-GCN model formed by connecting the relation graph convolution units in series.
9. The intelligent modeling system for DEM integration based on the graph roll-up neural network according to claim 2, wherein the generating of the modeling result for DEM integration includes: Arranging the node basic feature set into an input node feature vector set according to the index sequence of grid nodes in an initial grid topological structure, inputting an improved R-GCN model, and sequentially executing message transmission and feature updating to obtain a node feature vector set output by a final layer relationship graph convolution unit of the improved R-GCN model as a DEM integrated embedded representation vector set; Mapping each vector in the DEM integrated embedded representation vector set to a corresponding grid unit position based on a corresponding relation between grid nodes and grid units in the initial grid topological structure, extracting an elevation prediction component in the corresponding DEM integrated embedded representation vector at each grid unit, and generating an elevation prediction result grid; And calculating gradient and slope direction based on the elevation prediction values of the adjacent grid units in the elevation prediction result grid, generating a gradient grid and a slope direction grid, executing hydrologic network extraction according to the water flow direction and the elevation prediction result grid in the hydrologic relationship, generating a hydrologic network result, and combining the elevation prediction result grid, the gradient grid, the slope direction grid and the hydrologic network result into a DEM integrated modeling result.

Description

DEM integrated intelligent modeling system based on graph convolution neural network Technical Field The invention relates to the field of GIS digital elevation modeling, in particular to a Digital Elevation Model (DEM) integrated intelligent modeling system based on a graph convolution neural network. Background The digital elevation model has fundamental roles in terrain reconstruction, landform analysis and hydrologic calculation, and the prior art generally relies on a regular grid structure to carry out spatial interpolation, slope analysis and flow direction inference. However, the resolution ratio of the multisource DEM data is inconsistent, the noise distribution is complex, so that the traditional interpolation algorithm can not simultaneously maintain the terrain continuity and the landform structural characteristics, and the problems of elevation distortion, gradient fracture, water system interruption and the like are easy to generate. Under the condition of complex terrains, the multi-type and multi-scale association relationship among terrains, hydrology and landforms is difficult to express only by means of local windows or fixed neighborhood structures. The existing terrain modeling method based on deep learning mostly adopts a convolution structure, relies on a fixed window to perform feature extraction, and is difficult to describe spatial relationships such as water flow directivity, slope anisotropy and landform partition boundaries, so that the prediction precision of the model in a slope mutation area, a fracture construction area and a water system water collecting area is insufficient. Although the traditional graph neural network can process an irregular space structure, most models only support a single relation, do not carry out differentiated modeling on multi-relation features such as hydrologic flow direction, geomorphic structure, space adjacency and the like, cannot express the difference influence of different relations on elevation, gradient and hydrologic connectivity, and lack a direction constraint mechanism and relation selection capability under the multi-relation condition, so that multi-source space information fusion is insufficient. Therefore, how to provide a DEM integrated intelligent modeling system based on a graph convolution neural network is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a DEM integrated intelligent modeling system based on a graph convolution neural network, which utilizes mechanisms such as multi-relation adjacency tensor, directional edge characteristics, relation subset division, branching graph convolution and cross gating fusion and the like to carry out differential modeling and direction constraint on space adjacent relation, hydrologic relation and geomorphic relation so as to realize joint expression of multi-source geospatial data. According to the invention, the hydrologic branch and the landform branch are constructed in the R-GCN model, and the gating fusion structure is introduced, so that the continuity, the direction consistency and the landform structure expression capability of the landform element prediction are obviously improved, and the method has the advantages of high precision, strong structure understanding capability and adaptation to complex landforms. According to an embodiment of the invention, a DEM integrated intelligent modeling system based on a graph convolution neural network comprises: the data input and preprocessing module is used for acquiring a geospatial basic data set, executing data preprocessing, calculating to generate a node basic feature set, and constructing an initial grid topological structure according to the grid units; The multi-relation graph construction module is used for determining a space adjacent relation, a hydrologic relation and a geomorphic relation based on the initial grid topological structure, constructing a multi-relation adjacent tensor, calculating directional edge characteristics for each edge, and aggregating the directional edge characteristics into a relation characteristic tensor; the relation subset dividing module is used for dividing the hydrologic relation subset and the geomorphic relation subset; The branched graph convolution modeling module is used for setting hydrologic branches and landform branches in the R-GCN model and respectively executing direction modulation and anisotropic modulation; The cross gating fusion module is used for generating hydrologic gating coefficients and landform gating coefficients, performing weighted fusion according to the gating coefficients, and stacking a plurality of relation graph convolution units to construct an improved R-GCN model; The DEM embedding generation and mapping module is used for inputting the node basic feature set into the improved R-GCN model, performing layer-by-layer message transfer and feature update, and generating a DEM integrat