Natural resource satellite remote sensing knowledge base construction method for large model application

Abstract

The invention provides a natural resource satellite remote sensing knowledge base construction method for large model application, and belongs to the technical field of knowledge base construction, and the method comprises the following steps: obtaining satellite remote sensing image data, and preprocessing the remote sensing image data; performing feature extraction and data annotation on the remote sensing image; the method comprises the following steps: carrying out semantic recognition expansion on extracted remote sensing image features and annotation data, carrying out feature recognition through a remote sensing image, and then carrying out feature semantic expansion and annotation data expansion on the remote sensing image, so that the use places of the remote sensing image can be directly increased from the semantic level; according to the method, remote sensing image data which does not appear in practice can be generated, data sources and data fields of remote sensing images are increased, then more comprehensive remote sensing image data explanation is obtained by using corresponding model identification, mapping and an upper and lower directory relation are established according to the process of constructing a knowledge graph, and the mapping efficiency is improved. And an accurate knowledge base is provided for subsequent natural resource remote sensing satellite identification.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge base construction, and in particular to a method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications. Background Art

[0002] Remote sensing image target classification and recognition is a crucial component of information extraction and processing in high-resolution Earth observation systems and automatic target recognition systems. It is a research hotspot and a challenging area in remote sensing, playing a crucial role in applications such as intelligent transportation, smart cities, and dynamic target monitoring and positioning. Remote sensing image target classification and recognition aims to improve the accuracy, intelligence, real-time performance, and processing efficiency of algorithm processing. With the in-depth study of remote sensing image training datasets, the accuracy and efficiency of image target recognition have been significantly improved.

[0003] However, a semantic gap exists between remote sensing image target recognition and remote sensing image cognition. Remote sensing image target recognition and classification essentially utilize machine learning, deep learning, and other methods to achieve perception of remote sensing images. For example, a deep learning model trained on annotated remote sensing image datasets can quickly identify target categories such as "aircraft" and "ship" in target images. However, due to the lack of relevant target semantic information, further "cognition" of remote sensing image targets is very difficult.

[0004] Remote sensing image knowledge primarily addresses the semantic gaps in low-level image information. Researchers in related fields have conducted extensive research on the concept, classification, and application of remote sensing image knowledge. Different applications offer different understandings of remote sensing image knowledge. For example, Li Sheng (2018. Urban Land Cover Change Detection Method Combining Domain Knowledge and Deep Learning. Wuhan: Wuhan University) categorizes remote sensing image knowledge into image knowledge, geographic knowledge, and prior knowledge of change patterns. Yan Pengfei (2018. Research on Knowledge-Constrained High-Resolution Remote Sensing Image Segmentation Method. Beijing: China University of Geosciences) categorizes knowledge constraints in remote sensing image segmentation into internal and external knowledge constraints. Sun Jiabo (2014. Research on Knowledge-Based Automatic Extraction of Cultivated Land from High-Resolution Remote Sensing Images. Beijing: China Agricultural University) comprehensively considers remote sensing imagery at low, medium, and high spatial resolutions and categorizes relevant knowledge into ground feature spectrum knowledge, ground feature texture knowledge, and ground feature geometry knowledge. Gu Haiyan et al. (Gu Haiyan. 2015. Object Classification Technology Driven by Geographic Ontology Modeling for Remote Sensing Images. Wuhan: Wuhan University) categorize geographic entity knowledge into four categories: geographic knowledge, remote sensing image features, image object features, and expert knowledge. Research and application of knowledge semantics in remote sensing can be broadly divided into three categories: knowledge-based remote sensing image segmentation, knowledge-based remote sensing image target recognition, and cognitive computing using remote sensing information knowledge graphs.

[0005] Knowledge graphs are a rapidly developing technology in the field of artificial intelligence. Their core purpose is to build a large-scale semantic web, bridging the "semantic gap" between human perception and cognition. They are beginning to demonstrate their potential in areas such as natural language question answering, machine translation, product recommendations, and knowledge mining. Knowledge graph technologies offer a valuable insight into bridging the semantic gap in remote sensing imagery cognition. Evolving from the original semantic net, knowledge graphs have become a crucial foundational technology in the field of artificial intelligence. Essentially, a knowledge graph is a semantic web, where nodes represent entities or concepts, while edges represent the semantic relationships between entities / concepts. By linking vast amounts of information within a knowledge graph, a series of information can be quickly retrieved based on keywords and adjacency relationships, significantly improving the efficiency of manual searches. At its core, knowledge graphs are structured semantic knowledge bases. Knowledge graphs describe concepts, entities, and their relationships in the physical world in a structured manner, representing internet information in a form closer to human cognition and providing a means to better organize, manage, and understand the vast amount of information on the internet. Current remote sensing data does not have a knowledge base specifically for natural resources, and cannot achieve faster and more accurate semantic descriptions in large-scale model applications. Therefore, it is necessary to design a method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications. SUMMARY

[0006] The purpose of the present application is to provide a natural resource satellite remote sensing knowledge base construction method for large model applications, which solves the technical problem that existing satellite remote sensing knowledge bases do not have knowledge bases for natural resources, resulting in inaccurate semantic description of remote sensing satellite data in the direction of natural resources.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] The natural resource satellite remote sensing knowledge base construction method for large model applications comprises the following steps:

[0009] Step 1: Obtain satellite remote sensing image data and pre-process the remote sensing image data;

[0010] Step 2: Use a deep learning method to extract remote sensing image features and data labeling;

[0011] Step 3: Perform semantic recognition expansion on the extracted remote sensing image features and labeled data to obtain expanded feature data and expanded labeled data;

[0012] Step 4: Perform reverse drawing of remote sensing image on the expanded labeled data and expanded feature data to obtain expanded remote sensing image data;

[0013] Step 5: Perform minimum unit class recognition on the expanded remote sensing image data, real satellite remote sensing image data to obtain a class data set of remote sensing image;

[0014] Step 6: Randomly combine any two or more minimum unit classes to obtain a remote sensing image block, and randomly combine any two or more remote sensing image blocks to obtain a combined remote sensing image;

[0015] Step 7: Perform model training on the expanded remote sensing image data, combined remote sensing image data and real satellite remote sensing image data, and optimize the model parameters;

[0016] Step 8: Use the trained model to recognize remote sensing satellite images, and construct a knowledge graph according to the recognition data.

[0017] Further, the specific process of step 1 is as follows: select Landsat, Sentinel series satellite remote sensing data, perform radiation correction, atmospheric correction or geometric correction on the remote sensing data to eliminate interference factors, and ensure the consistency and availability of the data.

[0018] Further, the specific process of step 2 is: using the U-Net model to extract the texture, shape features and semantic data of the remote sensing image, and automatically labeling the remote sensing image polygons through the LabelMe tool to generate a JSON file. During labeling, the image objects are framed according to their categories, and then the framed graphics are semantically annotated. Each target can be segmented, and point annotation is also performed.

[0019] Further, the specific process of step 3 is: performing semantic recognition on the remote sensing image features and label data, then querying the Internet for similar or identical words, summarizing the similar or identical words queried, and classifying according to semantics. Each feature semantic corresponds to the corresponding label data, and a mapping table is constructed. One feature corresponds to one or more label data, and the semantic meaning of the label data used according to the scene is represented.

[0020] Further, the specific process of step 4 is: according to the meaning of each feature, describing the content of the remote sensing image in reverse, generating a remote sensing image from the semantic meaning, realizing a reverse image generation. In reverse depiction, the smallest image unit of the remote sensing satellite is directly called, and then the image units are combined to obtain the remote sensing image, realizing the expansion of the remote sensing image. After combining the image units, coloring is performed according to the scene used.

[0021] Further, the specific process of step 5 is: using a remote sensing highlighting recognition classification model to recognize all remote sensing images. During recognition, different things on each remote sensing image are separated, and the same things are classified into the same category. However, according to the shape and color of the graphics of the real object, the differences between the different categories of things are distinguished, and then semantic annotation is used to reflect the differences. All category data is summarized to obtain the category data set of the remote sensing image.

[0022] Further, the specific process of step 6 is: combining any two or more smallest unit real object images in the remote sensing image category data set. During combination, the contact position of the two is random, and different contact edges obtain different combined graphics. Then store the image blocks obtained by each combination, and randomly select two image blocks from the stored image blocks to combine. The position of the combination and the direction of the image block are randomly rotated to obtain a single combined remote sensing image. All combined remote sensing images are summarized to obtain a combined remote sensing image data set.

[0023] Furthermore, the specific process of step 7 is: first use the convolutional neural network to classify all remote sensing image data, determine the different scenarios used according to different classifications, and find the basic data of the corresponding usage scenarios when performing semantic interpretation. When using, first train, and then adjust the model parameters according to the training structure, and then use the trained Transformer model to identify remote sensing image sequences and semantic data, and then build a mapping table with the labeled data to improve the prediction accuracy of the model.

[0024] Furthermore, the specific process of step 8 is: defining the ontology model of the natural resources field, clarifying the concepts, entities and the relationships between them in the field, automatically or semi-automatically extracting knowledge from the remote sensing data analysis results, including land feature types, distribution patterns and change trends, integrating data and knowledge from different sources, avoiding data redundancy, and ensuring the integrity and consistency of the knowledge base.

[0025] The present invention has the following beneficial effects due to the adoption of the above technical solution:

[0026] The present invention performs feature recognition on remote sensing images, and then expands the feature semantics and annotation data of the remote sensing images, which can directly increase the use of remote sensing images from the semantic level. When reversely depicting the remote sensing images in the later stage, remote sensing image data that has never appeared in reality can be generated, and the data source and data field of the remote sensing images are increased. Then, the corresponding model recognition is used to obtain a more comprehensive interpretation of the remote sensing image data, and then mapping and upper and lower directory relationships are established in the process of constructing the knowledge graph, providing an accurate knowledge base for subsequent natural resource remote sensing satellite identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and by way of preferred embodiments. However, it should be noted that many of the details listed in this specification are merely provided to help the reader gain a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be practiced even without these specific details.

[0029] like Figure 1 As shown, a method for constructing a natural resources satellite remote sensing knowledge base for large-scale model applications includes the following steps:

[0030] Step 1: Acquire satellite remote sensing image data and preprocess it. Select remote sensing data from Landsat and Sentinel satellites and perform radiometric, atmospheric, or geometric correction on the data to eliminate interference factors and ensure data consistency and usability.

[0031] Step 2: Use deep learning methods to extract features and annotate remote sensing image data. A U-Net model is used to extract texture, shape, and semantic data from remote sensing images. The LabelMe tool automatically annotates several polygons in the image, generating a JSON file. During annotation, the image objects are framed according to their categories, and the framed shapes are semantically annotated. Semantics are mapped to corresponding objects, and each object can be segmented and labeled. Feature extraction focuses on extracting useful information from the data and requires a combination of domain knowledge and algorithm selection. Data annotation is the foundation of model training and requires a balance between efficiency and accuracy. This method can efficiently extract remote sensing image features while reducing annotation costs, adapting to different task requirements. In practical applications, the solution must be flexibly adjusted based on data characteristics and hardware conditions.

[0032] Step 3: Perform semantic recognition and expansion on the extracted remote sensing image features and annotation data to obtain expanded feature data and expanded annotation data. Perform semantic recognition on the remote sensing image features and annotation data, then search the internet for identical or similar terms. These similar or identical terms are aggregated and categorized based on semantics. Each feature's semantics corresponds to corresponding annotation data, and a mapping table is constructed. One feature corresponds to one or more annotation data, and the semantic meaning of the annotated data is determined based on the scenario in which it is used.

[0033] Step 4: Reverse delineate the expanded annotation data and extended feature data to obtain expanded remote sensing image data. Based on the meaning of each feature, reverse delineate the remote sensing image content, generating a remote sensing image from the semantic meaning, achieving reverse image generation. In reverse delineation, the remote sensing satellite's smallest image unit is directly called, and then the image units are combined to obtain the remote sensing image, achieving remote sensing image amplification. After the image units are combined, they are colored according to the usage scenario.

[0034] Step 5: Perform minimum unit category recognition on the expanded remote sensing image data and the real satellite remote sensing image data to obtain the remote sensing image category dataset. A remote sensing salient recognition classification model is used to memorize all remote sensing images for recognition. During recognition, different objects in each remote sensing image are separated individually. Identical objects are grouped into the same category, but the differences between objects in the same category are distinguished based on their shape and color. Semantic annotation is then used to highlight these differences. Data from all categories is then aggregated to obtain the remote sensing image category dataset.

[0035] Step 6: Randomly combine any two or more minimum unit categories to obtain remote sensing image blocks, and then randomly combine any two or more remote sensing image blocks to obtain a combined remote sensing image. Combine any two or more minimum unit physical images in the remote sensing image category dataset. When combining, the splicing contact positions of the two are random, and different contact edges result in different combined graphics. Then, store the image blocks obtained from each splicing, and then randomly find two image blocks from the stored image blocks to splice. The splicing position and direction of the image blocks are randomly rotated to obtain a single combined remote sensing image. All combined remote sensing images are summarized to obtain a combined remote sensing image dataset.

[0036] Step 7: Use the extended remote sensing image data, combined remote sensing image data, and real satellite remote sensing image data for model training and optimization of model parameters. First, use a convolutional neural network to classify all remote sensing image data. Based on the different classifications, the scenarios used are determined. When performing semantic interpretation, the basic data for the corresponding usage scenarios is searched. When used, training is first performed, and then the model parameters are adjusted according to the training structure. Later, the trained Transformer model is used to identify remote sensing image sequences and semantic data. A mapping table is then constructed with the annotated data to improve the model's prediction accuracy.

[0037] Step 8: Use the trained model to identify remote sensing satellite imagery and construct a knowledge graph based on the identified data. Define an ontology model for the natural resources domain, clarifying concepts, entities, and the relationships between them. Automatically or semi-automatically extract knowledge from remote sensing data analysis results, including feature types, distribution patterns, and changing trends. Integrate data and knowledge from different sources to avoid data redundancy and ensure the integrity and consistency of the knowledge base.

[0038] By performing feature recognition on remote sensing images and then expanding the feature semantics and annotation data of remote sensing images, the use of remote sensing images can be directly increased from the semantic level. When reversely depicting remote sensing images in the later stage, remote sensing image data that has never appeared in reality can be generated, which increases the data source and data field of remote sensing images. Then, the corresponding model recognition is used to obtain a more comprehensive interpretation of remote sensing image data. Then, mapping and upper and lower directory relationships are established in the process of constructing knowledge graphs, providing an accurate knowledge base for subsequent natural resource remote sensing satellite identification.

[0039] Matters not covered by the present invention are known technologies.

[0040] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for constructing a natural resources satellite remote sensing knowledge base for large-scale model applications, characterized by: The method comprises the following steps: Step 1: Obtain satellite remote sensing image data and preprocess the remote sensing image data; Step 2: Use deep learning methods to extract remote sensing image features and label data; Step 3: Perform semantic recognition expansion on the extracted remote sensing image features and annotation data to obtain expanded feature data and expanded annotation data; Step 4: Perform reverse delineation of the remote sensing image on the expanded annotation data and the expanded feature data to obtain the expanded remote sensing image data; Step 5: Perform minimum unit category recognition on the expanded remote sensing image data and the real satellite remote sensing image data to obtain the category dataset of the remote sensing image; Step 6: Randomly combine any two or more minimum unit categories to obtain remote sensing image blocks, and then randomly combine any two or more remote sensing image blocks to obtain a combined remote sensing image; Step 7: Use the extended remote sensing image data, combined remote sensing image data and real satellite remote sensing image data to train the model and optimize the model parameters; Step 8: Use the trained model to identify remote sensing satellite images and build a knowledge graph based on the recognition data.

2. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 1 is: select remote sensing data from Landsat and Sentinel series satellites, perform radiation correction, atmospheric correction or geometric correction on the remote sensing data to eliminate interference factors and ensure data consistency and availability.

3. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 2 is: use the U-Net model to extract the texture, shape features and semantic data of the remote sensing image, and automatically annotate several polygons of the remote sensing image through the LabelMe tool to generate a JSON file. When annotating, frame the image according to the category of the object, and then semantically annotate the framed graphics. At the same time, the semantics correspond to the corresponding targets, each target can be segmented, and point annotation is also performed.

4. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 3 is: semantically identify the remote sensing image features and annotation data, then search the Internet for the same or similar words, summarize the similar or same words, and classify them according to semantics. The semantics of each feature corresponds to the corresponding annotation data, and a mapping table is constructed. One feature corresponds to one or more annotation data, and the semantic meaning of the annotation data is represented by the scenario used.

5. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 4 is as follows: according to the meaning of each feature, reversely describe the remote sensing image content, generate the remote sensing image from the semantic meaning, and realize a reverse image generation. In the reverse description, directly call the remote sensing satellite minimum image unit, and then combine the image units to obtain the remote sensing image, realize the amplification of the remote sensing image, and after the image units are merged, color them according to the usage scenario.

6. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 5 is: use the remote sensing salient recognition classification model to memorize all remote sensing images for recognition. During recognition, different objects on each remote sensing image will be separated separately, and the same objects will be classified into the same category. However, the differences between objects of the same category are distinguished based on the shape and color of the graphics of the real objects. Then, semantic annotation is used to reflect the differences, and the data of all categories are summarized to obtain the category data set of the remote sensing image.

7. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 6 is as follows: any two or more minimum unit physical images in the remote sensing image category dataset are combined. When combining, the splicing contact position of the two is random, and different contact edges result in different combined graphics. Then, the image blocks obtained by each splicing are stored, and then any two image blocks are randomly selected from the stored image blocks for splicing. The splicing position and the direction of the image blocks are randomly rotated to obtain a single combined remote sensing image. All the combined remote sensing images are summarized to obtain a combined remote sensing image dataset.

8. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 7 is: first use the convolutional neural network to classify all remote sensing image data, determine the different scenarios used according to different classifications, and find the basic data of the corresponding usage scenarios when performing semantic interpretation. When using, first train, and then adjust the model parameters according to the training structure. Then, use the trained Transformer model to identify remote sensing image sequences and semantic data, and then build a mapping table with the labeled data to improve the prediction accuracy of the model.

9. The method for constructing a natural resource satellite remote sensing knowledge base for large-scale model applications according to claim 1, characterized in that: The specific process of step 8 is: define the ontology model of the natural resources field, clarify the concepts, entities and their relationships within the field, automatically or semi-automatically extract knowledge from the remote sensing data analysis results, including land feature types, distribution patterns and change trends, integrate data and knowledge from different sources, avoid data redundancy, and ensure the integrity and consistency of the knowledge base.

Images