A high-resolution DEM semantic object recognition method using bag-of-words model

Abstract

The present invention discloses a high-resolution DEM semantic object recognition method using a bag-of-words model, which belongs to the field of geospatial data processing technology and specifically includes the following steps: step 1, landform feature vector expression, step 2, landform bag-of-words model construction, step 3, target landform semantic mapping, and step 4, landform object classification and recognition. In the present invention, landform variables are extracted from a high-resolution DEM data set to generate a landform feature vector, and then an external open data source is used to enrich landform semantic information to form a landform bag-of-words and generate a landform weighted feature vector. Secondly, a mapping is formed between target landform feature elements and high-level concepts in the bag-of-words. Finally, the weighted features of the target data and the weighted features of the training data are compared to realize landform recognition. By integrating landform variables, based on regional features and high-level landform descriptions, automatic landform recognition is realized, significantly improving the recognition accuracy and efficiency of derived landform objects.

Description

Technical Field

[0001] The invention belongs to the technical field of geospatial data processing, and in particular relates to a high-resolution DEM semantic object recognition method using a bag-of-words model. Background Art

[0002] Digital elevation models (DEMs) provide the possibility of characterizing detailed features of the Earth's surface in three-dimensional space, supporting a variety of applications, including vulnerability estimation for natural disasters, urban landscapes, urban planning, and ecological sustainability. Considering that landforms are always too complex to complete land parcel segmentation due to fuzzy boundaries and semantic definitions, previous studies on landform characteristics have shown that landform variables are very meaningful in depicting the main structure of the landform surface. However, landform variables are only represented as points or linear features and cannot accurately describe the structure of regional landforms (such as craters). The limitations of landform variables in landform characteristics have attracted attention and have led to some research on landform characteristics, integrating not only landform variables but also regional-based parameters such as shape, texture, and background.

[0003] So far, many scholars have studied various regional-based geomorphic parameter extraction methods. There is still room for improvement in these geomorphic feature methods. First, there is no large-scale ground truth or benchmark dataset on geomorphic variables and geomorphic objects in relevant terrain analysis. The lack of well-prepared training datasets is a key obstacle to the widespread implementation of machine learning algorithms. Another challenge in using machine learning algorithms comes from the heterogeneity between the features processed by machine learning algorithms and the high-level semantics defined by humans, which is also a major problem in object recognition in the field of pattern recognition and machine intelligence. And few studies have used geomorphic semantics and open data resources to achieve high-resolution geomorphic object recognition.

[0004] Therefore, given that only limited training DEM data is available and that the features of a landform object derived from a DEM may differ from its description, this patent proposes a landform recognition method that requires minimal training data. The present invention proposes a comprehensive model, a bag-of-words model, that automatically recognizes landforms by integrating landform variables, region-based features such as shape and texture, and high-level landform descriptions. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem that there is no large-scale ground truth or benchmark dataset of relevant geomorphic variables and geomorphic objects. The lack of well-prepared training datasets is a key obstacle to the widespread implementation of machine learning algorithms. A high-resolution DEM semantic object recognition method using a bag-of-words model is proposed.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a high-resolution DEM semantic object recognition method using a bag-of-words model, specifically comprising the following steps:

[0007] Step 1: Geomorphic feature vector expression: geomorphic variables are extracted using a high-resolution DEM dataset, and domain-level geomorphic features are detected based on the extracted geomorphic variables to generate vectorized expressions of geomorphic feature elements.

[0008] Step 2: Constructing a bag-of-words model: Selecting landform keywords from the landform domain ontology, enriching the landform semantics using external open resources to form a landform bag-of-words, performing text post-processing and filtering based on the keywords, and performing frequency statistics on the processed keywords to generate a landform weighted feature vector.

[0009] Step 3, target landform semantic mapping: Map the target landform variables and domain-level features to the semantic data in the word bag to form a mapping relationship between the target landform semantics and the high-level semantic concepts in the word bag;

[0010] Step 4: Classification and recognition of landform objects: Compare the weighted features of the training data with the weighted features of the target data to realize landform recognition, classify each object into a predetermined class, and output the landform recognition result.

[0011] As a further description of the above technical solution:

[0012] In step 1, the landform feature vector expression includes the following sub-steps:

[0013] Step 1.1, extracting geomorphic variables using high-resolution DEM datasets, includes the following sub-steps:

[0014] Step 1.1a, predefine the geomorphic variables, determine the center pixel, adjacent pixels, and adjacent pixels beyond a certain distance of the geomorphic variables to be extracted, and let d and i represent the distance and direction indices;

[0015] Step 1.1b, using the spatial context method, measure the aspect difference and elevation difference between the central pixel and its neighboring pixels on each direction axis within multiple distances;

[0016] Step 1.1c, fuse the results of aspect difference and elevation difference to determine whether the central pixel belongs to the predefined geomorphic variable;

[0017] Step 1.2, detecting domain-level geomorphic features based on the extracted geomorphic variables, includes the following sub-steps:

[0018] Step 1.2a, measure height variation in Chinese moments, including the first-order origin moment mean, second-order central moment variance or standard deviation, third-order central moment tilt, and fourth-order central moment kurtosis;

[0019] Step 1.2b: Slope represents the steepness of the ground in both vertical and horizontal dimensions, and curvature represents the slope of the slope.

[0020] In step 1.2c, the local binary pattern is used to calculate the orientation of each pixel based on the gradient histogram. The LBP is calculated based on the elevation gradient. Assume that the orientation of a pixel is represented by the vector [d1, d2, d3, d4, d5, d6, d7, d8], where d1-d8 refer to the differences in eight directions between the central pixel and its neighboring pixels, respectively.

[0021] Step 1.2d, use Hough transform to determine circular landforms from the landform variable extraction results;

[0022] In step 1.2e, a rectangular shape is determined from the geomorphic variable extraction results using contour approximation. Rectangular landforms such as canyons or karsts can be seen on the surface formed by erosion and deposition.

[0023] As a further description of the above technical solution:

[0024] The bag-of-words model construction includes the following sub-steps:

[0025] Step 2.1, selecting landform keywords from landform domain ontology and open resources, includes the following sub-steps:

[0026] Step 2.1a, derive keywords from the annotations and definitions of each category in the domain ontology and classification system;

[0027] Step 2.1b, based on the derived keywords, collect all sentences containing these keywords and their synonyms in open resources;

[0028] Step 2.2: Post-process and filter the keywords. The filtered keywords include the following:

[0029] Step 2.2a, stop words;

[0030] Step 2.2b, behavior;

[0031] Step 2.2c, process;

[0032] Step 2.2d, Example;

[0033] Step 2.2e, specific time and location;

[0034] Step 2.2f, causality;

[0035] Step 2.2g, irrelevant words;

[0036] Step 2.2h, hidden word;

[0037] Step 2.2i, repeat the words;

[0038] Step 2.3, generate weighted feature vector:

[0039] The frequency of each keyword is counted based on the results of text post-processing and screening, and the frequency of the keyword is used to weight the vector.

[0040] As a further description of the above technical solution:

[0041] In step 3, a target landform semantic mapping is established: the variable features and domain-level features of the target landform to be detected are mapped to the high-level semantic data in the word bag and displayed in a list to form a mapping relationship between the target landform semantics and the high-level semantic concepts in the word bag.

[0042] As a further description of the above technical solution:

[0043] In step 4, landform object classification and identification: the purpose of classification is to learn the weighted features of landform types and predict the landform category to which the detected area belongs, including the following sub-steps:

[0044] Step 4.1, learning part:

[0045] Step 4.1a, labeling the minimum bounding boxes (MBBs) of multiple objects belonging to predefined terrain classes based on the high-resolution DEM;

[0046] Step 4.1b, calculate the frequency of keywords for this predefined terrain class and create a reference weighted vector.

[0047] In step 4.1c, a reference feature vector is created based on the MBB of this predefined terrain class.

[0048] Step 4.1d, create a reference weighted feature vector by combining the reference weight vector and the reference feature vector. The reference weighted feature vector is calculated as shown in step 2.3;

[0049] Step 4.2, experimental part:

[0050] Step 4.2a, detecting geomorphic variables;

[0051] Step 4.2b, generating multiple MBBs based on the geomorphic variable detection results;

[0052] Step 4.2c, create a feature vector for each MBB generated in step 4.2b;

[0053] Step 4.2d, creating a reference weighted feature vector for each MBB generated in step 4.2b by merging the reference weighted vector obtained in step 4.1b and the feature vector generated in step 4.2c;

[0054] Step 4.3, prediction part:

[0055] Classification is performed using the SVM classifier: the training data is the weighted feature vector obtained in step 4.1d, and the test data is the weighted feature vector obtained in step 4.2d.

[0056] As a further description of the above technical solution:

[0057] In step 1.1a, the direction axes include an east-west axis, a north-south axis, a northeast-southwest axis, and a northwest-southeast axis.

[0058] As a further description of the above technical solution:

[0059] In step 1.2c, if the central pixel is lower, similar, or higher than its neighboring pixels, the value of d will be assigned to -1, 0, and 1 respectively.

[0060] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0061] In the present invention, geomorphic variables are extracted from a high-resolution DEM dataset to generate a geomorphic feature vector. Then, an external open data source is used to enrich the geomorphic semantic information to form a geomorphic word bag and generate a geomorphic weighted feature vector. Secondly, a mapping is formed between the target geomorphic feature elements and the high-level concepts in the word bag. Finally, the weighted features of the target data and the weighted features of the training data are compared to realize geomorphic recognition. By integrating geomorphic variables, regional characteristics and high-level geomorphic descriptions, automatic geomorphic recognition is realized, which significantly improves the recognition accuracy and efficiency of derived geomorphic objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Schematic diagram of the principle of the method of the present invention;

[0063] Figure 2 Flow chart of the steps of the method of the present invention;

[0064] Figure 3 Schematic diagram of the processing process of an embodiment of the method of the present invention. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0066] See also Figure 1-3The present invention provides a technical solution: a high-resolution DEM semantic object recognition method using a bag-of-words model. In the object recognition method, first, landform variables are extracted from a high-resolution DEM dataset to generate a landform feature vector; then, external open data sources are used to enrich landform semantic information to form a landform bag-of-words and generate a landform weighted feature vector; secondly, a mapping is formed between target landform feature elements and high-level concepts in the bag-of-words; finally, the weighted features of the target data are compared with the weighted features of the training data to realize landform recognition.

[0067] The technical solution of the present invention comprises the following steps:

[0068] Step 1: Geomorphic feature vector expression: geomorphic variables are extracted using a high-resolution DEM dataset, and domain-level geomorphic features are detected based on the extracted geomorphic variables to generate vectorized expressions of geomorphic feature elements.

[0069] Step 2: Constructing a bag-of-words model: Selecting landform keywords from the landform domain ontology, enriching the landform semantics using external open resources to form a landform bag-of-words, performing text post-processing and filtering based on the keywords, and performing frequency statistics on the processed keywords to generate a landform weighted feature vector.

[0070] Step 3, target landform semantic mapping: Map the target landform variables and domain-level features to the semantic data in the word bag to form a mapping relationship between the target landform semantics and the high-level semantic concepts in the word bag;

[0071] Step 4: Classification and recognition of landform objects: Compare the weighted features of the training data with the weighted features of the target data to realize landform recognition, classify each object into a predetermined class, and output the landform recognition result.

[0072] Furthermore, in step 1, the landform feature vector expression includes the following sub-steps:

[0073] Step 1.1, extracting geomorphic variables using high-resolution DEM datasets, includes the following sub-steps:

[0074] Step 1.1a: Predefine the geomorphic variables and determine the center pixel (CP), adjacent pixels (AP), and adjacent pixels (NP) at multiple distances. Let d and i represent the distance and direction indices.

[0075] Step 1.1b: Use the spatial context method to measure the aspect difference and elevation difference between the central pixel (CP) and its neighboring pixels (NP) in each direction axis within multiple distances. The direction axes include the east-west axis, the north-south axis, the northeast-southwest axis, and the northwest-southeast axis.

[0076] In step 1.1c, the results of aspect difference and elevation difference are fused to determine whether the central pixel (CP) belongs to the predefined geomorphic variable.

[0077] Step 1.2, detecting domain-level geomorphic features based on the extracted geomorphic variables, includes the following sub-steps:

[0078] Step 1.2a, the Chinese moments measure height variation, including the first-order origin moment (mean), the second-order central moment (variance or standard deviation), the third-order central moment (skewness), and the fourth-order central moment (kurtosis). The expressions of these four moments are shown below:

[0079]

[0080] Where V represents the elevation of a pixel in the DEM, and N represents the number of pixels in the entire DEM or a local area of ​​the DEM.

[0081] Step 1.2b, Slope describes the steepness of the ground in both vertical and horizontal dimensions. Curvature describes the "slope" of the slope. In detail, the profile curvature describes the convexity and concavity of the slope in the vertical dimension, and the plateau curvature describes the convexity and concavity of the slope in the horizontal dimension.

[0082] In step 1.2c, the Local Binary Pattern (LBP) calculates the orientation of each pixel based on the Histogram of Gradients (HOG). The present invention calculates LBP based on elevation gradients. Assume that the orientation of a pixel is represented by the vector [d1, d2, d3, d4, d5, d6, d7, d8], where d1-d8 refer to the eight directional differences between the center pixel and its neighboring pixels, respectively. If the center pixel is lower, similar, or higher than its neighboring pixels, the value of d is assigned to -1, 0, and 1, respectively.

[0083] In step 1.2d, circular landforms, such as craters, are determined from the landform variable extraction results using Hough transform.

[0084] In step 1.2e, the rectangular shape is determined from the geomorphic variable extraction results using contour approximation. Rectangular landforms such as canyons and karsts can be seen on the surface formed by erosion and deposition.

[0085] According to the above steps, the structure of the landform feature vector is as follows: F y =[fD mom fD lbp fD shp fD slp fD carv ]y∈γ

[0086] Where y is the index of the landform category, fD mom Represents the moment characteristics of the four moments in step 1.2a: dim(fD mom )=4,fD lbpRefers to the LBP pattern diagram, which includes the LBP in eight directions for each pixel: dim(fD lbp )=1,fD shp Refers to the fusion result of Hough transform and contour approximation. If a circle or rectangle can be detected, then fD shs is 1, otherwise it is 0. slp , fD carv are the results for slope and mean curvature, respectively.

[0087] Moreover, in step 2, the bag-of-words model construction includes the following sub-steps:

[0088] Step 2.1, selecting landform keywords from landform domain ontology and open resources, includes the following sub-steps:

[0089] Step 2.1a, derive keywords from the annotations and definitions of each category in the domain ontology and classification system;

[0090] In step 2.1b, based on the derived keywords, all sentences containing these keywords and their synonyms are collected in open resources (such as Wikipedia).

[0091] Step 2.2: Post-process and filter the keywords. The filtered keywords include the following:

[0092] Step 2.2a, stop words: auxiliary verbs, modal words, suggestions, etc.;

[0093] Step 2.2b, behavior: verbs, agents, causes, processes, ongoing verbs, etc.;

[0094] Step 2.2c, process: consists of, in the process of, formed by, comes from, continues, gradually, etc.;

[0095] Step 2.2d, Examples: For example, to show, including, in these instances, specific spaces, specific quantities, etc.;

[0096] Step 2.2e, specific time and place: specific period, recent, new, no longer, when, where, specific geographical name, etc.;

[0097] Step 2.2f, causal relationship: caused by, etc.;

[0098] Step 2.2g, irrelevant words: in contrast, old, young, etc.;

[0099] Step 2.2h, hidden words: underground, seabed, etc.;

[0100] Step 2.2i, repeat the word.

[0101] Step 2.3, generate weighted feature vector:

[0102] The frequency of each keyword is counted after the text is processed and filtered. The frequency of the keyword is used to weight the vector, as shown in the following formula:

[0103] W y =[wD mom ,wD lbp ,wD shp wD slp wD carv ]y∈γ

[0104] Among them wD mom ,wD lhp ,wD shp ,wD slp ,wD carv D mom , D lbp , D shp , D slp and D carv Weighted. Then, the weighted feature vector is expressed as follows:

[0105] WF y =[wD mom × D mom wD lbp ×D lbp wD shp ×D shp wD slp xD slp wD carv ×D carv ],y∈Y

[0106] Moreover, in step 3, a target landform semantic mapping is established: the target landform variables (features) and domain-level features to be detected are mapped to the high-level semantic data in the word bag and displayed in a list to form a mapping relationship between the target landform semantics and the high-level semantic concepts in the word bag.

[0107] Furthermore, in step 4, landform object classification and recognition: the purpose of classification is to learn the weighted features of landform types and predict the landform category to which the detected area belongs, including the following sub-steps:

[0108] Step 4.1, learning part:

[0109] Step 4.1a, labeling the minimum bounding box (MBB) of multiple objects belonging to predefined terrain classes based on the high-resolution DEM;

[0110] Step 4.1b, calculate the frequency of keywords for this predefined terrain class and create a reference weighted vector.

[0111] In step 4.1c, a reference feature vector is created based on the MBB of this predefined terrain class.

[0112] In step 4.1d, a reference weighted feature vector is created by combining the reference weight vector and the reference feature vector. The reference weighted feature vector is calculated as in step 2.3.

[0113] Step 4.2, experimental part:

[0114] Step 4.2a, detecting geomorphic variables;

[0115] Step 4.2b, generating multiple MBBs based on the geomorphic variable detection results;

[0116] Step 4.2c, create a feature vector for each MBB generated in step 4.2b;

[0117] Step 4.2d, create a reference weighted feature vector for each MBB generated in step 4.2b by merging the reference weight vector obtained in step 4.1b and the feature vector generated in step 4.2c.

[0118] Step 4.3, prediction part:

[0119] Classification is performed using the SVM classifier: the training data is the weighted feature vector obtained in step 4.1d, and the test data is the weighted feature vector obtained in step 4.2d.

[0120] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A high-resolution DEM semantic object recognition method using a bag-of-words model, characterized in that: The specific steps include: Step 1: Geomorphic feature vector expression: geomorphic variables are extracted using a high-resolution DEM dataset, and domain-level geomorphic features are detected based on the extracted geomorphic variables to generate vectorized expressions of geomorphic feature elements. Step 2: Constructing a bag-of-words model: Selecting landform keywords from the landform domain ontology, enriching the landform semantics using external open resources to form a landform bag-of-words, performing text post-processing and filtering based on the keywords, and performing frequency statistics on the processed keywords to generate a landform weighted feature vector. Step 3, target landform semantic mapping: Map the target landform variables and domain-level features to the semantic data in the word bag to form a mapping relationship between the target landform semantics and the high-level semantic concepts in the word bag; Step 4: Landform object classification and recognition: Compare the weighted features of the training data with the weighted features of the target data to achieve landform recognition, classify each object into a predetermined class, and output the landform recognition result. In step 1, the landform feature vector expression includes the following sub-steps: Step 1.1, extracting geomorphic variables using high-resolution DEM datasets, includes the following sub-steps: Step 1.1a, predefine the geomorphic variables, determine the center pixel, adjacent pixels, and adjacent pixels beyond a certain distance of the geomorphic variables to be extracted, and let d and i represent the distance and direction indices; Step 1.1b, using the spatial context method, measure the aspect difference and elevation difference between the central pixel and its neighboring pixels on each direction axis within multiple distances; Step 1.1c, fuse the results of aspect difference and elevation difference to determine whether the central pixel belongs to the predefined geomorphic variable; Step 1.2, detecting domain-level geomorphic features based on the extracted geomorphic variables, includes the following sub-steps: Step 1.2a, measure height variation in Chinese moments, including the first-order origin moment mean, second-order central moment variance or standard deviation, third-order central moment tilt, and fourth-order central moment kurtosis; Step 1.2b: Slope represents the steepness of the ground in both vertical and horizontal dimensions, and curvature represents the slope of the slope. In step 1.2c, the local binary pattern is used to calculate the orientation of each pixel based on the gradient histogram. The LBP is calculated based on the elevation gradient. Assume that the orientation of a pixel is represented by the vector [d1, d2, d3, d4, d5, d6, d7, d8], where d1-d8 refer to the differences in eight directions between the central pixel and its neighboring pixels, respectively. Step 1.2d, use Hough transform to determine circular landforms from the landform variable extraction results; In step 1.2e, the rectangular shape is determined from the geomorphic variable extraction results using contour approximation. Rectangular landforms can be seen on the surface formed by erosion and deposition.

2. The high-resolution DEM semantic object recognition method using the bag-of-words model according to claim 1 is characterized in that: The bag-of-words model construction includes the following sub-steps: Step 2.1, selecting landform keywords from landform domain ontology and open resources, includes the following sub-steps: Step 2.1a, derive keywords from the annotations and definitions of each category in the domain ontology and classification system; Step 2.1b, based on the derived keywords, collect all sentences containing these keywords and their synonyms in open resources; Step 2.2: Post-process and filter the keywords. The filtered keywords include the following: Step 2.2a, stop words; Step 2.2b, behavior; Step 2.2c, process; Step 2.2d, Example; Step 2.2e, specific time and location; Step 2.2f, causality; Step 2.2g, irrelevant words; Step 2.2h, hidden word; Step 2.2i, repeat the words; Step 2.3, generate weighted feature vector: The frequency of each keyword is counted based on the results of text post-processing and screening, and the frequency of the keyword is used to weight the vector.

3. The high-resolution DEM semantic object recognition method using the bag-of-words model according to claim 1 is characterized in that: In step 3, a target landform semantic mapping is established: the variable features and domain-level features of the target landform to be detected are mapped to the high-level semantic data in the word bag and displayed in a list to form a mapping relationship between the target landform semantics and the high-level semantic concepts in the word bag.

4. The high-resolution DEM semantic object recognition method using the bag-of-words model according to claim 1 is characterized in that: In step 4, landform object classification and identification: the purpose of classification is to learn the weighted features of landform types and predict the landform category to which the detected area belongs, including the following sub-steps: Step 4.1, learning part: Step 4.1a, labeling the minimum bounding boxes (MBBs) of multiple objects belonging to predefined terrain classes based on the high-resolution DEM; Step 4.1b, calculate the frequency of keywords of this predefined terrain class and create a reference weighted vector; Step 4.1c, creating a reference feature vector based on the MBB of this predefined terrain class; Step 4.1d, creating a reference weighted feature vector by combining the reference weight vector and the reference feature vector; Step 4.2, experimental part: Step 4.2a, detecting geomorphic variables; Step 4.2b, generating multiple MBBs based on the geomorphic variable detection results; Step 4.2c, create a feature vector for each MBB generated in step 4.2b; Step 4.2d, creating a reference weighted feature vector for each MBB generated in step 4.2b by merging the reference weighted vector obtained in step 4.1b and the feature vector generated in step 4.2c; Step 4.3, prediction part: Classification is performed using the SVM classifier: the training data is the weighted feature vector obtained in step 4.1d, and the test data is the weighted feature vector obtained in step 4.2d.

5. The high-resolution DEM semantic object recognition method using the bag-of-words model according to claim 1 is characterized in that: In step 1.1a, the direction axes include an east-west axis, a north-south axis, a northeast-southwest axis, and a northwest-southeast axis.

6. The high-resolution DEM semantic object recognition method using the bag-of-words model according to claim 1 is characterized in that: In step 1.2c, if the central pixel is lower than its neighboring pixels, the value of d will be assigned to -1 accordingly; If the center pixel is similar to its neighboring pixels, the value of d will be assigned to 0 accordingly; If the center pixel is higher than its neighboring pixels, the value of d will be assigned to 1 accordingly.

Images