Multi-scale capsule network remote sensing object classification method and system guided by geoscience knowledge
Through the multi-scale capsule network remote sensing geographic classification method guided by geoscience knowledge, the problem of low information extraction accuracy on urban scale of remote sensing technology is solved, and the classification accuracy and automation are achieved, which enhances the interpretability and generalization of the model. ability to transform.
Patent Information
- Application Number
- CN202111663292.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-31
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2041-12-31
AI Technical Summary
Existing remote sensing technology is difficult to effectively extract high-resolution information on the urban scale, and there are "multi-spectrum" and "multi-spectrum" phenomena, and traditional methods cannot fully utilize geographic knowledge, resulting in low accuracy and automation of remote sensing information extraction. Unable to meet urban management needs.
Using the multi-scale capsule network remote sensing geographic classification method guided by geologic knowledge, the geologic knowledge graph and multi-scale convolution capsule network are constructed, combined with deep learning models, and the geologic knowledge constraint training network is used to improve the interpretability and generalization ability of the model. .
It improves the accuracy and automation of remote sensing geographic classification, enhances the interpretability and generalization capabilities of the model, and can more accurately extract the spatiotemporal and spatial spectrum characteristics of urban multi-source heterogeneous data.
Smart Images

Figure CN114219963B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of remote sensing, and in particular relates to a multi-scale capsule network remote sensing object classification scheme guided by geological knowledge. Background Art
[0002] At present, the sustainable development of cities around the world is facing severe challenges. Urban waterlogging, urban heat islands, and urban ecological function degradation, and other "urban diseases" have become global problems. Using remote sensing technology to extract high-resolution information of typical urban elements can provide scientific and technological support for sustainable urban development. However, due to the influence of cultural differences, human activities, and geographical location, the extraction of remote sensing information of typical urban elements on a global scale will encounter serious "same object, different spectrum" and "same spectrum, different object" phenomena. In addition, high-resolution urban images also have information missing caused by building shadows, tree occlusions, and interference from moving targets. The above problems have led to low accuracy and automation in urban remote sensing information extraction, which cannot meet the increasingly sophisticated management needs of cities.
[0003] This patent mainly studies how to obtain high-precision object category information from high-resolution remote sensing images driven by both data and knowledge. With the development of computer technology, computer-aided remote sensing image object classification technology has gradually replaced manual interpretation methods, such as scale-invariant feature transformation, directional gradient histogram, etc. In addition, traditional machine learning methods such as decision trees, random forests, support vector machines, hidden Markov random fields, etc. have also been applied to remote sensing image object classification, improving the object classification effect. However, these traditional methods require manual design and feature extraction, and can only perform several layers of linear or nonlinear processing on the features. They cannot fully learn image features, and have limited ability to express objects in high-resolution remote sensing images in complex urban scenes. When extended to other related tasks, they are prone to underfitting or overfitting, and cannot be truly automated.
[0004] In recent years, deep learning technology has flourished and quickly swept the world. It has also begun to be introduced into the remote sensing field for remote sensing image object classification and has gradually become the mainstream method for remote sensing image interpretation. In remote sensing image object classification, the existing deep neural network is mainly based on the convolutional neural network, and various improvements are made according to the characteristics of remote sensing images, such as constructing a DCNNs semantic segmentation method based on the idea of a full convolutional neural network for remote sensing image classification, and modifying its activation function based on the full convolutional neural network. The remote sensing object classification method based on deep learning has powerful feature learning and expression capabilities, does not require manual feature design and selection, and can achieve automated end-to-end output results. However, deep learning is a data-driven method that is highly dependent on a large number of data samples. It reversely optimizes network parameters by reducing the loss between the output result and the label. It does not make full use of the geological knowledge in the remote sensing field. The output results of the remote sensing image object classification method based on deep learning are poor in reliability and interpretability, and belong to a black box operation. In addition, remote sensing image data contains rich spatial, spectral, temporal and other information. Traditional convolutional neural networks abandon the relative positions and spatial hierarchical relationships between objects during remote sensing image processing, which hinders the further improvement of remote sensing object classification accuracy. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides a multi-scale capsule network remote sensing object classification scheme guided by geological knowledge. By effectively utilizing geological knowledge and establishing a knowledge constraint mechanism, the deep learning model has certain interpretability and good generalization ability.
[0006] To achieve the above object, the technical solution adopted by the present invention is a multi-scale capsule network remote sensing object classification method guided by geological knowledge, comprising the following steps:
[0007] Step a: construct a remote sensing object classification sample library by preprocessing remote sensing images and marking object labels;
[0008] Step b: According to the main land cover types in the city, establish the remote sensing interpretation ontology of five types of surface cover entities: buildings, roads, bare soil, green space, and water bodies. Then, based on the remote sensing interpretation ontology, complete the construction of the geoscience knowledge map by instantiating the ontology and its attributes;
[0009] Step c: explicitly symbolize the geoscientific knowledge contained in the geoscientific knowledge graph, and mine the implicit geoscientific knowledge contained in the geoscientific knowledge graph;
[0010] Step d, constructing a multi-scale convolutional capsule network as a deep learning network, inputting the processed image obtained in step a as sample data into the deep learning network, and constructing a model loss constraint based on the geoscientific knowledge obtained in step c to guide the training of the multi-scale capsule network;
[0011] Step e, extracting and classifying urban multi-source remote sensing information, including training the constructed multi-scale convolutional capsule model based on the loss function established in step d under the constraints of geological knowledge to obtain a training model; and testing the test data based on the trained model, and using the qualified model to classify remote sensing objects after obtaining it.
[0012] Moreover, in step b, the network ontology language OWL is used to describe the remote sensing interpretation ontology and its attribute relationships. The ground objects in the sample set are used as instance objects of the interpretation ontology, and the attributes of the instantiated objects such as the category, shape characteristics, spatial semantics, etc. of the corresponding ground objects are extracted to construct a geological knowledge graph.
[0013] Moreover, in step e, the model test results are evaluated using four indicators: overall accuracy based on the confusion matrix, Kappa coefficient, user accuracy, and producer accuracy.
[0014] The present invention provides a multi-scale capsule network remote sensing object classification system guided by geological knowledge, which is used to implement the multi-scale capsule network remote sensing object classification method guided by geological knowledge as described above.
[0015] Furthermore, the following modules are included,
[0016] The first module is used to construct a remote sensing object classification sample library by preprocessing remote sensing images and labeling object labels;
[0017] The second module is used to establish the remote sensing interpretation ontology of five types of land cover entities, namely buildings, roads, bare soil, green land, and water bodies, according to the main land cover types in the city. Then, based on the remote sensing interpretation ontology, the construction of the geoscience knowledge map is completed by instantiating the ontology and its attributes.
[0018] The third module is used to explicitly symbolize the geoscientific knowledge contained in the geoscientific knowledge graph and to mine the implicit geoscientific knowledge contained in the geoscientific knowledge graph;
[0019] The fourth module is used to construct a multi-scale convolutional capsule network as a deep learning network, input the processed image obtained in step a as sample data into the deep learning network, and build a model loss constraint based on the geoscience knowledge obtained in step c to guide the training of the multi-scale capsule network;
[0020] The fifth module is used for extracting and classifying urban multi-source remote sensing information, including training the constructed multi-scale convolutional capsule model based on the loss function established in step d under the constraints of geological knowledge to obtain a training model; and testing the test data based on the trained model. After obtaining a qualified model, it is used to realize the classification of remote sensing objects.
[0021] Alternatively, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the multi-scale capsule network remote sensing object classification method guided by geological knowledge as described above.
[0022] Alternatively, it includes a readable storage medium having a computer program stored thereon, and when the computer program is executed, the multi-scale capsule network remote sensing object classification method guided by geological knowledge as described above is implemented.
[0023] The beneficial effect of the present invention is as follows: the present invention proposes a multi-scale capsule network remote sensing object classification scheme guided by geological knowledge, which fully takes into account the structure and spatial distribution information between object targets, and effectively utilizes geological domain knowledge to establish a knowledge constraint mechanism to guide the autonomous training of the model, so that the deep learning model has a certain degree of interpretability and good generalization ability, further improving the accuracy and automation of remote sensing object classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 The figure is a flow chart of an embodiment of the present invention.
[0025] Figure 2 This is a network structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION
[0026] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0027] The present invention discloses a multi-scale capsule network remote sensing feature classification scheme guided by geological knowledge, including constructing an urban geological knowledge graph; learning the representation of the domain knowledge of "geographic concept-geographic entity-geographic relationship" contained in the geological knowledge graph into a low-dimensional vector containing semantic information through representation learning of the urban geological knowledge graph; constructing a multi-scale three-dimensional convolutional capsule network remote sensing information extraction model according to the characteristics of remote sensing images and the existing scale effect, and embedding the knowledge vector learned from the geographic knowledge graph into the model, establishing an effective domain knowledge constraint mechanism to guide the learning of the model, so that the model has certain interpretability and good global generalization ability.
[0028] Through the above method, the present invention can uniformly learn the spatiotemporal spectral characteristics of multi-source heterogeneous urban data, effectively utilize geological knowledge, establish an effective knowledge constraint mechanism, make the deep learning model have certain interpretability and good generalization ability, and improve the extraction accuracy of urban remote sensing information.
[0029] See also Figure 1 The embodiment of the present invention provides a multi-scale capsule network remote sensing object classification method guided by geological knowledge, comprising the following steps:
[0030] a. Construction of remote sensing sample library: By preprocessing remote sensing images and labeling objects, a remote sensing object classification sample library is constructed.
[0031] In specific implementation, the acquired remote sensing data can first be preprocessed with geometric correction, radiation correction, and other preprocessing, and then the corresponding existing national geographic census, multi-source survey and monitoring results, and the third national land survey and other historical interpretation data can be collected, and the ground object samples can be selected and annotated in combination with the existing historical interpretation data to build a remote sensing ground object classification sample library. This step can be performed in advance, and the existing remote sensing sample library can be directly input during the process.
[0032] b. Construction of the geological knowledge graph: According to the main land cover types in the city, the remote sensing interpretation ontology of five types of surface cover entities, namely buildings, roads, bare soil, green space and water bodies, is established. Then, based on the remote sensing interpretation ontology, the construction of the geological knowledge graph is completed by instantiating the ontology and its attributes.
[0033] In specific implementation, the urban geoscience knowledge graph can be constructed through knowledge extraction, knowledge fusion and knowledge reasoning from knowledge carriers such as urban geographical location, basic geographic information data, multi-source geographic information and open network text. Preferably, the network ontology language OWL is used to describe the remote sensing interpretation ontology and its attribute relationship, and the ground objects in the sample set are used as instance objects of the interpretation ontology, and the attributes of the instantiated objects such as the category, shape characteristics, spatial semantics, etc. of the corresponding ground object are extracted to construct the geoscience knowledge graph.
[0034] In the embodiment, according to the main types of land objects in general cities, the remote sensing interpretation ontology of five types of surface cover entities, namely buildings, roads, bare soil, green space and water bodies, is constructed, and the ontology and its attribute relationship are described using the ontology network language OWL. In the field of remote sensing, the attribute relationship of the remote sensing interpretation ontology mainly includes upper and lower layer attributes such as subordination, spatial relationship attributes such as adjacent, surrounding, and orientation, and statistical attributes such as majority classes. The images in the remote sensing land object classification sample library have been pre-labeled and contain information such as land object target category, shape characteristics, and spatial semantics. Therefore, when constructing the geological knowledge map, the land object target is used as an instance object of the ontology, and the corresponding land object target category, shape characteristics, spatial semantics and other information instantiation object attributes are extracted to construct the geological knowledge map.
[0035] c. Extraction of geological knowledge: In specific implementation, the Semantic WebRule Language (SWRL) can be used to explicitly symbolize the geological knowledge contained in the geological knowledge graph, and the implicit geological knowledge contained in the geological knowledge graph can be mined through logical reasoning.
[0036] Embodiment Based on the geoscience knowledge graph constructed in step b, the spatial symbiotic knowledge contained in the geoscience knowledge graph is represented by symbiotic conditional probability. The specific steps are as follows:
[0037] Assume that there are two types of features, C i and C j Indicates that the feature category C in the statistical geoscience knowledge graph i The probability of occurrence P(C i ), and then count the number of adjacent target categories of the target C j The probability of the category P(C i ,C j ), and finally calculate C according to formula (1) i The presence of a target in its neighborhood i The probability of class target P(C j |C i ):
[0038]
[0039] d. Construct a multi-scale convolutional capsule network as a deep learning network. Construct a multi-scale convolutional capsule network as a deep learning network. Input the processed image obtained in step a as sample data into the deep learning network. Based on the geological knowledge obtained in step c, construct a model loss constraint to guide the training of the multi-scale capsule network to obtain a training model.
[0040] 1) Considering that the traditional convolutional neural network CNN ignores the position and spatial hierarchy relationship between image pixels, and this information is beneficial to the classification of ground objects. In addition, in remote sensing images, the sizes of ground objects vary, and there is an obvious multi-scale effect. Using a fixed-size convolutional layer will make the features learned by the model too single, which is not conducive to model training and ground object classification. Therefore, this patent uses a multi-scale convolutional capsule network as a deep learning network, see Figure 2 , mainly including multi-scale convolution module, capsule module and fully connected layer.
[0041] In specific implementation, the spatiotemporal spectral features of remote sensing images can be extracted using multi-scale three-dimensional convolution kernels according to the characteristics of remote sensing images and the existing scale effects. Then, based on the shallow features extracted by the convolutional neural network structure, the capsule network can be used to extract deeper features, taking into account the structural and spatial distribution information between urban elements.
[0042] Multi-scale convolution module: In the multi-scale convolution module, the embodiment uses three convolution kernels conv1-3 with sizes of 9×9, 7×7 and 5×5, respectively, denoted as k9, k7, and k5, with the step size set to 1 and the number of channels set to 16, for extracting multi-scale features from remote sensing images. After the three convolution operations, normalized batch processing (BatchNorm) and ReLu activation function are performed to enhance the nonlinear expression ability of the model. Finally, the multi-scale features are connected through the Concatenate operation and input into the capsule module. The operation of the convolution module can be expressed as follows:
[0043]
[0044] In the above formula, X is the input remote sensing image, Conv i×i (·) is the convolution operation with a kernel size of i×i, BN(·) is batch normalization, ReLu(·) is the activation function, Cont(·) is the Concatenate operation, and layer 1 Represents the output of the first layer of the network model, Feature i×i For, i=5,7,9.
[0045] Capsule module: The capsule module consists of an initial capsule layer and a digital capsule layer. In the initial capsule layer, there are n capsules u1, u2, ...un. In the embodiment, n=32, where each capsule is a 1×8 vector, which is equivalent to using 8 groups of convolution kernels with a size of 7×7×32 and a step size of 1 to perform convolution operations on the feature map of the multi-scale convolution block input, and the obtained feature map is used as the input of the digital capsule layer. The output of the digital capsule layer can be obtained according to the dynamic routing algorithm. The number of capsules must be equal to the final number of classifications. The number of capsules is set to 5, and each capsule v1, v2, ...v5 is a 16-dimensional vector. The final output of the digital capsule layer is 5×16. The capsule network layers are connected through a dynamic routing algorithm, as follows:
[0046] Assume that the output of the i-th capsule in the (l-1)th layer is Then, the strength of the connection between the i-th capsule in the (l-1)th layer and the j-th capsule in the lth layer is for:
[0047]
[0048] in, is the weight matrix.
[0049] The input vector of the jth capsule in the lth layer is is the strength of the k capsules in the (l-1)th layer and coupling coefficient The weighted sum of can be expressed as follows:
[0050]
[0051] Assume that there are p capsules in the first layer, then the coupling coefficient The calculation formula is:
[0052]
[0053] in, is the similarity score, and exp(·) is an exponential function with the natural constant e as the base.
[0054] Finally, the output vector of the jth capsule in the lth layer is for:
[0055]
[0056] Here, ||·|| represents the norm.
[0057] During the training iteration of the network, The coupling coefficient is continuously updated by the value of , and the updating process can be expressed as follows:
[0058]
[0059] Fully connected layer: The last layer of the model is the fully connected layer, which plays the role of a classifier and can be expressed as follows:
[0060] y=Wx (8)
[0061] Among them, W is the weight matrix, x is the input, and y is the feature output by the digital capsule layer.
[0062] 2) Establish a loss function constrained by geoscientific knowledge:
[0063] Based on the geoscientific knowledge graph constructed in steps b and c, a loss function constrained by geoscientific knowledge is constructed to guide the training of the multi-scale convolutional capsule network model constructed in 1). Therefore, the loss function designed in this patent includes the edge loss L margin and the geoscientific knowledge constraint loss L knowledge :
[0064]
[0065] Among them, T k is an indicator. When the input vector belongs to category k, T k =1, otherwise T k =0.m - and m + They are the lower and upper limits of the penalty, generally m + =0.9, m - =0.1. ||v k || is the norm of the output vector of the fully connected layer. When category k exists, ||v k ||≥m + , otherwise, ||v k ||≤m - λ is the adjustment coefficient, which is used to adjust the ratio of the influence of the correct result and the wrong result on the final function value. max() is the maximum value function, and n is the number of classification categories.
[0066] In the geoscientific knowledge constraint, according to the category confidence vector Y of the model output i , the maximum value in the confidence vector and its corresponding maximum value number (i.e., category number) are taken as the classification confidence y of the target j Then, taking the target as the center, we take a 7×7 neighborhood, and obtain the category confidence of the target in the current neighborhood space by weighted summation of the classification confidence of the target in the neighborhood and the spatial co-occurrence conditional probability (Formula 1), and use H i Represented, then calculate the category confidence H based on the spatial distribution i And the true category confidence vector Y iThe loss between them can be expressed as follows:
[0067]
[0068] Where N is the total number of samples, i is the sample target (1≤i≤N), m i is the number of targets in the neighborhood, j refers to the target in the neighborhood, c is the number of categories, q refers to the category (1≤q≤C), P(C q |C k ) is the conditional probability of spatial symbiosis.
[0069] The final loss function of the model is L total for:
[0070] L total =L margin +βL knowledge (11)
[0071] Among them, β is a hyperparameter, which is used to adjust L margin and L knowledge The weight of .
[0072] e. Urban multi-source remote sensing information extraction and classification: Based on the sample set constructed in step a, the multi-scale convolutional capsule model constructed is trained under the constraints of geoscientific knowledge based on the loss function established in step d to obtain a training model; and the test data is tested based on the trained model, and the model test results are evaluated using four indicators: overall accuracy based on the confusion matrix, Kappa coefficient, user accuracy, and producer accuracy. After obtaining a qualified model, it can be used to classify remote sensing objects.
[0073] In specific implementation, the method proposed in the technical solution of the present invention can be implemented by technical personnel in this field using computer software technology to realize automatic operation process. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of the present invention and computer equipment running the corresponding computer program, should also be within the scope of protection of the present invention.
[0074] In some possible embodiments, a multi-scale capsule network remote sensing object classification system guided by geoscience knowledge is provided, including the following modules:
[0075] The first module is used to construct a remote sensing object classification sample library by preprocessing remote sensing images and labeling object labels;
[0076] The second module is used to establish the remote sensing interpretation ontology of five types of land cover entities, namely buildings, roads, bare soil, green land, and water bodies, according to the main land cover types in the city. Then, based on the remote sensing interpretation ontology, the construction of the geoscience knowledge map is completed by instantiating the ontology and its attributes.
[0077] The third module is used to explicitly symbolize the geoscientific knowledge contained in the geoscientific knowledge graph and to mine the implicit geoscientific knowledge contained in the geoscientific knowledge graph;
[0078] The fourth module is used to construct a multi-scale convolutional capsule network as a deep learning network, input the processed image obtained in the first module as sample data into the deep learning network, and build a model loss constraint based on the geoscience knowledge obtained in the third module to guide the training of the multi-scale capsule network;
[0079] The fifth module is used for extracting and classifying urban multi-source remote sensing information. It includes training the constructed multi-scale convolutional capsule model based on the loss function established in the fourth module under the constraints of geological knowledge to obtain a training model; and testing the test data based on the trained model. After obtaining a qualified model, it is used to realize the classification of remote sensing objects.
[0080] In some possible embodiments, a multi-scale capsule network remote sensing object classification system guided by geological knowledge is provided, including a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a multi-scale capsule network remote sensing object classification method guided by geological knowledge as described above.
[0081] In some possible embodiments, a multi-scale capsule network remote sensing object classification system guided by geological knowledge is provided, including a readable storage medium, on which a computer program is stored. When the computer program is executed, a multi-scale capsule network remote sensing object classification method guided by geological knowledge as described above is implemented.
[0082] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.
Claims
1. A multi-scale capsule network remote sensing object classification method guided by geological knowledge, characterized by: The method comprises the following steps: step a, constructing a remote sensing object classification sample library by preprocessing remote sensing images and marking object labels; Step b: According to the main land cover types in the city, establish the remote sensing interpretation ontology of five types of surface cover entities: buildings, roads, bare soil, green space, and water bodies. Then, based on the remote sensing interpretation ontology, complete the construction of the geoscience knowledge map by instantiating the ontology and its attributes; Step c: explicitly symbolize the geoscientific knowledge contained in the geoscientific knowledge graph, and mine the implicit geoscientific knowledge contained in the geoscientific knowledge graph; Step d, constructing a multi-scale convolutional capsule network as a deep learning network, inputting the processed image obtained in step a as sample data into the deep learning network, and constructing a model loss constraint based on the geoscientific knowledge obtained in step c to guide the training of the multi-scale capsule network; When constructing the model loss constraint, the loss function set includes the marginal loss L margin and the geoscientific knowledge constraint loss L knowledge , Among them, T k is an indicator. When the input vector belongs to category k, T k =1, otherwise T k =0;m - and m + are the lower and upper limits of the penalty respectively; ||v k || is the norm of the output vector of the fully connected layer. When category k exists, ||v k ||≥m + , otherwise, ||v k ||≤m - ; λ is the adjustment coefficient, which is used to adjust the ratio of the influence of the correct result and the wrong result on the final function value; max() is the maximum value function, and n is the number of classification categories; In the geoscientific knowledge constraint, according to the category confidence vector Y of the model output i , the maximum value in the confidence vector and the corresponding maximum value sequence number are taken as the classification confidence y of the target j and classification category k; then, taking the target as the center, the classification confidence of the target in the neighborhood is obtained by weighted summation of the classification confidence of the target in the neighborhood and the spatial co-occurrence condition probability, and H is used to represent the category confidence of the target in the current neighborhood space. i Represented, then calculate the category confidence H based on the spatial distribution i and the true category confidence vector Y i The loss between is expressed as follows: Where N is the total number of samples, i is the sample target, 1≤i≤N, m i is the number of targets in the neighborhood, j refers to the target in the neighborhood, c is the number of categories, q refers to the category, 1≤q≤C, P(C q |C k ) is the spatial symbiosis conditional probability; The final loss function of the model is L total for: L total =L margin +βL knowledge Among them, β is a hyperparameter used to adjust L margin and L knowledge The weight of Step e, extracting and classifying urban multi-source remote sensing information, including training the constructed multi-scale convolutional capsule model based on the loss function established in step d under the constraints of geological knowledge to obtain a training model; and testing the test data based on the trained model, and using the qualified model to classify remote sensing objects after obtaining it.
2. The multi-scale capsule network remote sensing object classification method guided by geological knowledge according to claim 1 is characterized by: In step b, the web ontology language OWL is used to describe the remote sensing interpretation ontology and its attribute relationships. The ground objects in the sample set are used as instance objects of the interpretation ontology, and the attributes of the instantiated objects such as the category, shape characteristics, spatial semantics, etc. of the corresponding ground objects are extracted to construct a geological knowledge graph.
3. The multi-scale capsule network remote sensing object classification method guided by geological knowledge according to claim 1 or 2 is characterized by: In step e, the model test results are evaluated using four indicators: overall accuracy based on the confusion matrix, Kappa coefficient, user accuracy, and producer accuracy.
4. A multi-scale capsule network remote sensing object classification system guided by geological knowledge, characterized by: Used to implement a multi-scale capsule network remote sensing object classification method guided by geological knowledge as described in any one of claims 1-3.
5. The multi-scale capsule network remote sensing object classification system guided by geoscience knowledge according to claim 4 is characterized by: Includes the following modules, The first module is used to construct a remote sensing object classification sample library by preprocessing remote sensing images and labeling object labels; The second module is used to establish the remote sensing interpretation ontology of five types of land cover entities, namely buildings, roads, bare soil, green land, and water bodies, according to the main land cover types in the city. Then, based on the remote sensing interpretation ontology, the construction of the geoscience knowledge map is completed by instantiating the ontology and its attributes. The third module is used to explicitly symbolize the geoscientific knowledge contained in the geoscientific knowledge graph and to mine the implicit geoscientific knowledge contained in the geoscientific knowledge graph; The fourth module is used to construct a multi-scale convolutional capsule network as a deep learning network, input the processed image obtained in the first module as sample data into the deep learning network, and build a model loss constraint based on the geoscience knowledge obtained in the third module to guide the training of the multi-scale capsule network; The fifth module is used for extracting and classifying urban multi-source remote sensing information. It includes training the constructed multi-scale convolutional capsule model based on the loss function established in the fourth module under the constraints of geological knowledge to obtain a training model; and testing the test data based on the trained model. After obtaining a qualified model, it is used to realize the classification of remote sensing objects.
6. The multi-scale capsule network remote sensing object classification system guided by geoscience knowledge according to claim 4 is characterized by: It includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a multi-scale capsule network remote sensing ground object classification method guided by geological knowledge as described in any one of claims 1 to 3.
7. The multi-scale capsule network remote sensing object classification system guided by geoscience knowledge according to claim 4 is characterized by: It includes a readable storage medium, on which a computer program is stored. When the computer program is executed, it implements a multi-scale capsule network remote sensing object classification method guided by geological knowledge as described in any one of claims 1 to 3.
