CN116992962A - A method of constructing geomorphological knowledge graph based on self-supervised deep learning - Google Patents

Abstract

The present invention relates to the technical field of geomorphological knowledge graph research, specifically a geomorphic knowledge graph construction method based on self-supervised deep learning, including self-supervised pre-training model construction, self-supervised pre-training model evaluation and semantic analysis and knowledge graph construction. , compared with the existing technology, this invention learns multi-scale semantics and spatial characteristics of landform types on global multi-resolution DEM data, forms a landform feature vector representation for the purpose of machine calculation, and builds a set of computer-understandable and reproducible Computational, inferable landscape knowledge graphs.

Description

A method for constructing geomorphological knowledge graph based on self-supervised deep learning

Technical field

The present invention relates to the technical field of geomorphological knowledge graph research, specifically a geomorphic knowledge graph construction method based on self-supervised deep learning.

Background technique

Landform is one of the most important physical geographical elements, affecting the surface climate, ecological environment and spatial distribution of natural resources. Landform research is an important branch of geography, which plays an important role in understanding the shape, structure, distribution changes and laws of the earth's surface, and provides scientific basis for human survival and development. With the development of science and technology, the study of geomorphology has become increasingly information-based and intelligent. In the late 1990s, with the introduction of the concept of geological information maps, scholars in the field of geomorphology made breakthroughs in key technologies such as geomorphological information extraction and classification, and constructed information maps such as landform morphological characteristics and landform development to realize a series of geomorphological knowledge. summary and condensation.

Currently, knowledge graph (KnowledgeGraph) has become a research hotspot in the information field. Knowledge graph represents knowledge in the form of a structured directed graph, using nodes and edges to represent entities and relationships between entities. The semantics of entities and relationships are mapped to a low-dimensional continuous vector space through knowledge representation learning (KnowledgeGraph Embedding). , so that knowledge reasoning and prediction can be realized through vector or matrix operations. Therefore, knowledge graphs have powerful semantic computing capabilities and are widely used in semantic retrieval, intelligent question answering, and personalized recommendations. Domain knowledge graphs have also been developed in many fields of geoscience research. Construct.

Specific to the field of geomorphology, due to the problems of blurred boundaries, difficult to determine types, and multi-scale problems in landforms, it is difficult to classify landforms. There are various landform classification systems, and it is difficult to unify different systems, and it is difficult to construct a landform knowledge ontology. Moreover, the boundaries of the landforms themselves are fuzzy and difficult to determine. The current classification of landform types is based on human cognition and is not a system oriented toward machine calculations and centered on machine understanding. It is difficult to achieve quantitative expressions and calculations that can be understood and utilized by computers.

Domestic and foreign research in the existing technology requires expert knowledge, and the ontology layer is constructed from top to bottom. The degree of automation is not high, and the existing landform classification systems are diverse, making it difficult to construct a unified knowledge ontology. The triples extracted from the text are based on the ontology. It is difficult to provide sufficient entities of various landform types and their characteristics.

Therefore, in order to solve the above problems, this application proposes a method of constructing a geomorphological knowledge graph based on self-supervised deep learning. Self-supervised deep learning has developed rapidly in the field of deep learning. The pre-training model based on the self-supervised learning strategy is trained on a large amount of data, and can automatically extract common features in the data without domain expert knowledge, forming a self-supervised deep learning model and training strategy suitable for DEM data, so as to fully achieve the goal. The purpose is to learn the semantics and spatial characteristics of landforms, explore the landform classification system from the perspective of artificial intelligence, and realize the bottom-up automated construction of a computer-understandable and computable landform knowledge map.

Contents of the invention

The purpose of the present invention is to fill the gaps in the existing technology and provide a method for constructing a geomorphological knowledge map based on self-supervised deep learning. Self-supervised deep learning has developed rapidly through the field of deep learning. The pre-training model based on the self-supervised learning strategy is trained on a large amount of data, and can automatically extract common features in the data without domain expert knowledge, forming a self-supervised deep learning model and training strategy suitable for DEM data, so as to fully achieve the goal. The purpose is to learn the semantics and spatial characteristics of landforms, explore the landform classification system from the perspective of artificial intelligence, and realize the bottom-up automated construction of a computer-understandable and computable landform knowledge map.

In order to achieve the above objectives, the present invention provides a method for constructing a geomorphological knowledge graph based on self-supervised deep learning, including self-supervised pre-training model construction, self-supervised pre-training model evaluation and semantic analysis and knowledge graph construction;

Self-supervised pre-training model construction: improve the self-supervised learning strategy, including building training data sets, designing model structures, exploring learning strategies, designing loss functions, and conducting model training;

Evaluation of self-supervised pre-training models: Apply the pre-training model to downstream tasks, evaluate the learning performance and transfer performance of the model, and feed back problems to the pre-training model to adjust model parameters;

Semantic analysis and knowledge graph construction: Semantic analysis is performed on the pre-trained landscape vector representation to construct a landscape knowledge graph;

The construction of self-supervised pre-training model specifically includes:

S1, pre-training data set:

Utilize existing global DEM products of different resolutions to build data sources for large-scale pre-training data sets, randomly select some areas to build pre-training data sets, cut the data into raster images of uniform size, and build a total data scale In the pre-training data set of 900,000 to 1,000,000 images, a smaller proportion of the data is divided into the verification set, and the rest is used as the training set;

S2, model design:

Construct a self-supervised DEM deep learning model, design the structure of the encoder and decoder, select the mask structure of the MAE model and the basic architecture based on the ViT model, and compare the performance of different ViT model structures, as well as the performance of other convolutional neural network models. Performance, through multiple sets of comparative experiments, the best model is selected to solve key coding technologies;

Change the encoder data input method and adopt the method of synchronous input of multi-resolution data in the same area, so that the model can learn the features of different resolutions synchronously; because the pixels of the same image at different resolutions are inconsistent, change the pixel position encoding method of MAE and learn from Scale -MAE's absolute distance encoding method, based on the correlation of terrain between different frames, designs an appropriate position encoding method to keep the position information of different resolution data consistent, which not only avoids the information leakage problem of absolute position encoding, but also It reflects the periodic patterns of terrain features to a certain extent, and improves the model’s versatility for data of different scales by adjusting the decoder structure;

S3, self-supervised learning strategy:

Improve the self-supervised learning strategy and explore the settings of different hyperparameters. Self-supervised learning uses image masking to randomly cover a certain proportion of the DEM image, and uses the uncovered part as the input of the pre-training model, through the encoder After decoding the recovered image, calculate the loss compared with the original image, and optimize the model; try different data covering methods or cover the data with randomly generated arbitrary shapes, try different covering ratios, and explore the best ratio;

S4, loss function:

Design a reasonable loss function. In the self-supervised learning process, the most direct indicator is the numerical difference between the restored value and the real value. Introduce terrain factors to control the training effect of the model, and add the loss values of these factors to the loss function. , so that the model considers terrain factors when optimizing;

Semantic analysis and knowledge graph construction specifically include:

S10, the supervised learning model represents the landform features as vectors. By calculating the similarity of the vectors, plots with similar characteristics and spatial proximity are merged into landform units. A landform unit is a landform type entity;

S20, obtain the hierarchical structure of landform types through hierarchical clustering;

S30, through the attention mechanism of self-supervised learning, the attention scores between each two plots are obtained, an attention score matrix is constructed, the degree of interdependence between the plots is calculated, and visual analysis is used to explore the relationship between different landform entities. study the interdependence between landform types, study the spatial pattern of landform type distribution, and construct a landform space syntax tree;

S40: Form a geomorphic knowledge graph including the hierarchical structure of geomorphic types, semantic representation of geomorphic entities, and spatial relationships of geomorphic types.

Compared with the existing technology, the present invention learns multi-scale semantics and spatial features of landform types on global multi-resolution DEM data to form a landform feature vector representation for the purpose of machine calculation.

Through exploration, a self-supervised deep learning model and training strategy suitable for DEM data are formed to achieve the purpose of fully learning the semantics and spatial characteristics of landforms; through the analysis and calculation of landform feature vector representation, the landform classification system from the perspective of artificial intelligence is explored to achieve Automatically build a computer-understandable and computable geomorphological knowledge graph from the bottom up.

Description of the drawings

Figure 1 is a schematic diagram of the overall research framework and technical route of the present invention project.

Figure 2 is a schematic structural diagram of the MAE model of the present invention.

Figure 3 is a schematic diagram of the technical route for constructing a knowledge graph according to the present invention.

Detailed ways

The present invention will now be further described with reference to the accompanying drawings.

Please refer to Figures 1 to 3, a method of constructing a geomorphological knowledge graph based on self-supervised deep learning, including self-supervised pre-training model construction, evaluation of the self-supervised pre-training model, semantic analysis and knowledge graph construction;

Self-supervised pre-training model construction: improve the self-supervised learning strategy, including building training data sets, designing model structures, exploring learning strategies, designing loss functions, and conducting model training;

Evaluation of self-supervised pre-training models: Apply the pre-training model to downstream tasks, evaluate the learning performance and transfer performance of the model, and feed back problems to the pre-training model to adjust model parameters;

Semantic analysis and knowledge graph construction: Semantic analysis is performed on the pre-trained landscape vector representation to construct a landscape knowledge graph;

The construction of self-supervised pre-training model specifically includes:

S1, pre-training data set:

Utilize existing global DEM products of different resolutions to build data sources for large-scale pre-training data sets, randomly select some areas to build pre-training data sets, cut the data into raster images of uniform size, and build a total data scale In the pre-training data set of 900,000 to 1,000,000 images, a smaller proportion of the data is divided into the verification set, and the rest is used as the training set;

S2, model design:

Construct a self-supervised DEM deep learning model, design the structure of the encoder and decoder, select the mask structure of the MAE model and the basic architecture based on the ViT model, and compare the performance of different ViT model structures, as well as the performance of other convolutional neural network models. Performance, through multiple sets of comparative experiments, the best model is selected to solve key coding technologies;

Change the encoder data input method and adopt the method of synchronous input of multi-resolution data in the same area, so that the model can learn the features of different resolutions synchronously; because the pixels of the same image at different resolutions are inconsistent, change the pixel position encoding method of MAE and learn from Scale -MAE's absolute distance encoding method, based on the correlation of terrain between different frames, designs an appropriate position encoding method to keep the position information of different resolution data consistent, which not only avoids the information leakage problem of absolute position encoding, but also It reflects the periodic patterns of terrain features to a certain extent, and improves the model’s versatility for data of different scales by adjusting the decoder structure;

S3, self-supervised learning strategy:

Improve the self-supervised learning strategy and explore the settings of different hyperparameters. Self-supervised learning uses image masking to randomly cover a certain proportion of the DEM image, and uses the uncovered part as the input of the pre-training model, through the encoder After decoding the recovered image, calculate the loss compared with the original image, and optimize the model; try different data covering methods or cover the data with randomly generated arbitrary shapes, try different covering ratios, and explore the best ratio;

S4, loss function:

Design a reasonable loss function. In the self-supervised learning process, the most direct indicator is the numerical difference between the restored value and the real value. Introduce terrain factors to control the training effect of the model, and add the loss values of these factors to the loss function. , so that the model considers terrain factors when optimizing;

Semantic analysis and knowledge graph construction specifically include:

S10, the supervised learning model represents the landform features as vectors. By calculating the similarity of the vectors, plots with similar characteristics and spatial proximity are merged into landform units. A landform unit is a landform type entity;

S20, obtain the hierarchical structure of landform types through hierarchical clustering;

S30, through the attention mechanism of self-supervised learning, the attention scores between each two plots are obtained, an attention score matrix is constructed, the degree of interdependence between the plots is calculated, and visual analysis is used to explore the relationship between different landform entities. study the interdependence between landform types, study the spatial pattern of landform type distribution, and construct a landform space syntax tree;

S40: Form a geomorphic knowledge graph including the hierarchical structure of geomorphic types, semantic representation of geomorphic entities, and spatial relationships of geomorphic types.

The above are only the preferred embodiments of the present invention, and are only used to help understand the method and its core idea of the present application. The protection scope of the present invention is not limited to the above-mentioned embodiments. All technical solutions that fall under the ideas of the present invention belong to the present invention. scope of protection. It should be pointed out that for those of ordinary skill in the art, several improvements and modifications may be made without departing from the principles of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention.

The present invention comprehensively solves the problem that when building a knowledge map in the field of geomorphology in the prior art, the landform classification systems are diverse, it is difficult to unify different systems, the degree of dependence is high and the degree of automation is not high, and the existing landform classification systems are diverse and difficult to construct. Unified knowledge ontology and other issues, through self-supervised deep learning models and training strategies suitable for DEM data, can fully learn the semantics and spatial characteristics of landforms. Through the analysis and calculation of landform feature vector representation, the landform classification system from the perspective of artificial intelligence is explored, and a bottom-up automated construction of a computer-understandable and computable landform knowledge map is realized.

Claims (1)

1. A method of constructing a geomorphological knowledge graph based on self-supervised deep learning, which is characterized by including the construction of a self-supervised pre-training model, the evaluation and semantic analysis of the self-supervised pre-training model, and the construction of a knowledge graph; The construction of the self-supervised pre-training model: improving the self-supervised learning strategy, including constructing training data sets, designing model structures, exploring learning strategies and designing loss functions, and conducting model training; the evaluation of the self-supervised pre-training model: The pre-trained model is applied to downstream tasks, the learning performance and transfer performance of the model are evaluated, and problems found are fed back to the pre-trained model to adjust model parameters; The semantic analysis and knowledge map construction: perform semantic analysis on the landform vector representation obtained by pre-training, and construct a landform knowledge map; The construction of the self-supervised pre-training model specifically includes: S1, pre-training data set: Utilize existing global DEM products of different resolutions to build data sources for large-scale pre-training data sets, randomly select some areas to build pre-training data sets, cut the data into raster images of uniform size, and build a total data scale In the pre-training data set of 900,000 to 1,000,000 images, a smaller proportion of the data is divided into the verification set, and the rest is used as the training set; S2, model design: Construct a self-supervised DEM deep learning model, design the structure of the encoder and decoder, select the mask structure of the MAE model and the basic architecture based on the ViT model, and compare the performance of different ViT model structures, as well as the performance of other convolutional neural network models. Performance, through multiple sets of comparative experiments, the best model is selected to solve key coding technologies; Change the encoder data input method and adopt the method of synchronous input of multi-resolution data in the same area, so that the model can learn the features of different resolutions synchronously; because the pixels of the same image at different resolutions are inconsistent, change the pixel position encoding method of MAE and learn from Scale -MAE's absolute distance encoding method, based on the correlation of terrain between different frames, designs an appropriate position encoding method to keep the position information of different resolution data consistent, which not only avoids the information leakage problem of absolute position encoding, but also To a certain extent, it reflects the periodic patterns of terrain features, and by adjusting the decoder structure, it improves the model's versatility for data of different scales; S3, self-supervised learning strategy: Improve the self-supervised learning strategy and explore the settings of different hyperparameters. Self-supervised learning uses image masking to randomly cover a certain proportion of the DEM image, and uses the uncovered part as the input of the pre-training model, through the encoder After decoding the restored image, calculate the loss compared with the original image, and optimize the model; try different data covering methods or cover the data with randomly generated arbitrary shapes, try different covering ratios, and explore the best ratio; S4, Loss function: Design a reasonable loss function. In the self-supervised learning process, the most direct indicator is the numerical difference between the restored value and the real value. Introduce terrain factors to control the training effect of the model, and add the loss values of these factors to the loss function. , so that the model considers terrain factors when optimizing; the semantic analysis and knowledge graph construction specifically include: S10, the supervised learning model represents the landform features as vectors. By calculating the similarity of the vectors, plots with similar characteristics and spatial proximity are merged into landform units. A landform unit is a landform type entity; S20, obtain the hierarchical structure of landform types through hierarchical clustering; S30, through the attention mechanism of self-supervised learning, the attention scores between each two plots are obtained, an attention score matrix is constructed, the degree of interdependence between the plots is calculated, and visual analysis is used to explore the relationship between different landform entities. study the interdependence between landform types, study the spatial pattern of landform type distribution, and construct a landform space syntax tree; S40: Form a geomorphic knowledge graph including the hierarchical structure of geomorphic types, semantic representation of geomorphic entities, and spatial relationships of geomorphic types.