A Geographical Environment Information Management System

Abstract

The present application discloses a geographic environment information management system, which relates to the field of intelligent management technology. It uses computer vision technology based on deep learning to analyze remote sensing images, extracts image feature expressions of target areas of objects in remote sensing images, and performs channel dimension local feature enhancement on target area features of objects based on an attention mechanism, thereby intelligently identifying object types. In this way, the accuracy and reliability of object target recognition can be improved, and the phenomenon of misclassification and omission can be reduced.

Description

Technical Field

[0001] The present application relates to the field of intelligent management technology, and more specifically, to a geographic environment information management system. Background Art

[0002] With the rapid development of information technology, geographic environment information management systems are increasingly used in urban planning, resource monitoring, disaster warning and other fields. At the same time, with the gradual maturity of remote sensing technology, remote sensing images, with their wide coverage, large amount of information, and fast update speed, play an increasingly important role in resource surveys, environmental monitoring, urban planning and other fields, and have become an important means of obtaining geographic environment information.

[0003] In geographic environment information management, automatic recognition and classification of ground objects using remote sensing image data is a key task. However, the traditional pixel-based ground object classification method may face the "salt and pepper phenomenon" when processing high-resolution remote sensing images, that is, a large number of small patches often appear in the classification results, resulting in a decrease in classification accuracy, and there are problems such as large amount of calculation, low classification accuracy, and insufficient use of feature information. Therefore, an optimized geographic environment information management system is expected. Summary of the invention

[0004] In order to solve the above technical problems, this application is proposed.

[0005] Accordingly, according to one aspect of the present application, a geographic environment information management system is provided, which includes: a data acquisition module for collecting geographic environment data, the geographic environment data including remote sensing images, ground monitoring data and meteorological data; a data storage and management module for storing and managing the geographic environment data; an identification module for performing object identification on the ground object in the remote sensing image to obtain an object identification result; a labeling module for labeling the object identification result in the remote sensing image; wherein the identification module includes: a ground object target feature extraction unit for extracting the ground object target area features in the remote sensing image to obtain a ground object target local feature map; a feature dispersion unit for performing feature dispersion processing on the ground object target local feature map along the channel dimension of the ground object target local feature map to obtain a sequence of ground object target pixel-level channel feature vectors; an attention enhancement fusion unit for performing attention enhancement fusion on the sequence of ground object target pixel-level channel feature vectors to obtain a ground object surface global saliency fusion feature vector; and a recognition result generation unit for determining the object recognition result based on the ground object surface global saliency fusion feature vector.

[0006] In the above-mentioned geographic environment information management system, the land object target feature extraction unit includes: a land object target detection subunit, which is used to pass the remote sensing image through a land object target detection network to obtain a land object target area of ​​interest image; a target area feature extraction subunit, which is used to pass the land object target area of ​​interest image through a land object target area feature extractor based on a convolutional neural network model to obtain the land object target local feature map.

[0007] In the above-mentioned geographic environment information management system, the attention enhancement fusion unit is used to: pass the sequence of the pixel-level channel feature vectors of the object target through an adaptive attention weight fusion network to obtain the global significance fusion feature vector of the object surface.

[0008] In the above-mentioned geographic environment information management system, the attention enhancement fusion unit includes: an autocorrelation attention weight calculation subunit, which is used to calculate the autocorrelation attention weight of each ground object target pixel-level channel feature vector in the sequence of the ground object target pixel-level channel feature vector to obtain a sequence of autocorrelation attention weights; a weighted fusion subunit, which is used to perform weighted fusion on the sequence of the ground object target pixel-level channel feature vector based on the sequence of the autocorrelation attention weights to obtain the global significance fusion feature vector of the ground object surface.

[0009] In the above-mentioned geographic environment information management system, the autocorrelation attention weight calculation subunit is used to: respectively calculate the matrix product between each ground object target pixel-level channel feature vector in the sequence of ground object target pixel-level channel feature vectors and each predetermined weight matrix to obtain a sequence of encoded ground object target pixel-level channel feature vectors; add a bias vector to each encoded ground object target pixel-level channel feature vector in the sequence of encoded ground object target pixel-level channel feature vectors and then pass the Selu activation function to obtain a sequence of activated ground object target pixel-level channel feature vectors; respectively calculate the product between each activated ground object target pixel-level channel feature vector in the sequence of activated ground object target pixel-level channel feature vectors and the transposed vector of each preset reference feature vector to obtain the sequence of autocorrelation attention weights.

[0010] In the above-mentioned geographic environment information management system, the weighted fusion subunit is used to: normalize the sequence of autocorrelation attention weights to obtain a sequence of normalized autocorrelation attention weights; use the sequence of normalized autocorrelation attention weights as weights to calculate the vector-by-vector weighted sum of the sequence of pixel-level channel feature vectors of the target object to obtain the global significance fusion feature vector of the object surface.

[0011] In the above-mentioned geographic environment information management system, the recognition result generating unit is used to: pass the global saliency fusion feature vector of the surface of the object through a classifier-based object identifier to obtain the object recognition result, and the object recognition result is a feature type label.

[0012] The above-mentioned geographic environment information management system also includes a training unit for training the land object detection network, the land object area feature extractor based on the convolutional neural network model, the adaptive attention weight fusion network and the classifier-based land object identifier.

[0013] In the above-mentioned geographic environment information management system, the training unit includes: a training data acquisition subunit, which is used to acquire training data, and the training data includes a training remote sensing image and a true value of a ground object recognition result; a training image target detection subunit, which is used to pass the training remote sensing image through the ground object target detection network to obtain a training ground object target area of ​​interest image; a training ground object feature extraction subunit, which is used to pass the training ground object target area of ​​interest image through the ground object target area feature extractor based on the convolutional neural network model to obtain a training ground object target local feature map; a training feature dispersion subunit, which is used to perform feature dispersion processing on the training ground object target local feature map along the channel dimension of the training ground object target local feature map to obtain a training ground object target pixel-level channel feature vector. A sequence of quantities; a training data adaptive fusion subunit, used to pass the sequence of training ground object target pixel-level channel feature vectors through the adaptive attention weight fusion network to obtain a training ground object surface global significance fusion feature vector; a classification loss subunit, used to pass the training ground object surface global significance fusion feature vector through the classifier-based ground object identifier to obtain a classification loss function value; a model training subunit, used to train the ground object target detection network, the ground object target area feature extractor based on the convolutional neural network model, the adaptive attention weight fusion network and the classifier-based ground object identifier with the classification loss function value, wherein in each round of iteration of the training, the training ground object surface global significance fusion feature vector is iteratively corrected.

[0014] Compared with the prior art, the geographic environment information management system provided by this application uses computer vision technology based on deep learning to analyze remote sensing images, extracts image feature expressions of the target area of ​​the object in the remote sensing image, and performs channel dimension local feature enhancement on the target area features of the object based on the attention mechanism, thereby intelligently identifying the type of the object. In this way, the accuracy and reliability of object target recognition can be improved, and the phenomenon of misclassification and omission can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0016] Figure 1 It is a block diagram of an identification module in a geographic environment information management system according to an embodiment of the present application.

[0017] Figure 2 Schematic diagram of the architecture of the identification module in the geographic environment information management system according to an embodiment of the present application.

[0018] Figure 3 It is a block diagram of a ground object target feature extraction unit in a geographic environment information management system according to an embodiment of the present application.

[0019] Figure 4 It is a block diagram of an attention enhancement fusion unit in a geographic environment information management system according to an embodiment of the present application.

[0020] Figure 5 It is a block diagram of a training unit in a geographic environment information management system according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] Below, the embodiments of the present application will be described in more detail in conjunction with the accompanying drawings, and the above-mentioned and other purposes, features and advantages of the present application will become more obvious. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein. At the same time, the accompanying drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, and are used together with the embodiments of the present application to explain the present application, and do not constitute a limitation on the present application. In the drawings, the same reference numerals generally represent the same parts or steps.

[0022] Exemplary Systems

[0023] The present application provides a geographic environment information management system, which includes: a data acquisition module, used to collect geographic environment data, the geographic environment data including remote sensing images, ground monitoring data and meteorological data; a data storage and management module, used to store and manage the geographic environment data; an identification module, used to perform object identification on ground objects in the remote sensing image to obtain object identification results; and a labeling module, used to label the object identification results in the remote sensing image.

[0024] More specifically, at present, when identifying the ground objects in the remote sensing image, the traditional method usually adopts a pixel-based classification method. However, since the pixel-based ground object target classification method mainly relies on the spectral information of a single pixel for classification, it is difficult to overcome the problems of "same spectrum, different objects" and "same object, different spectrum", that is, ground objects with the same spectral characteristics may actually be different objects, and ground objects with different spectral characteristics may be mistakenly classified into the same category. In addition, due to the spatial resolution limitation of the remote sensing image itself, the classification results may have more misclassification and omission phenomena, resulting in low classification accuracy. In addition, this method only uses spectral information for type recognition, ignoring other useful information such as the spatial structure and texture of the ground object, which limits the accuracy and reliability of type recognition.

[0025] In response to the above technical problems, the technical concept of this application is to use computer vision technology based on deep learning to analyze remote sensing images, extract image feature expressions of the target area of ​​the ground object in the remote sensing image, and enhance the local features of the target area of ​​the ground object in the channel dimension based on the attention mechanism, so as to intelligently identify the type of ground object. In this way, the accuracy and reliability of ground object target recognition can be improved, and the phenomenon of misclassification and omission can be reduced.

[0026] Figure 1 It is a block diagram of an identification module in a geographic environment information management system according to an embodiment of the present application. Figure 2 FIG. 1 is a schematic diagram of the architecture of the identification module in the geographic environment information management system according to an embodiment of the present application. Figure 1 and Figure 2 As shown, the recognition module 100 of the geographic environment information management system according to the embodiment of the present application includes: a land object feature extraction unit 110, which is used to extract the land object target area features in the remote sensing image to obtain a land object target local feature map; a feature dispersion unit 120, which is used to perform feature dispersion processing on the land object target local feature map along the channel dimension of the land object target local feature map to obtain a sequence of land object target pixel-level channel feature vectors; an attention enhancement fusion unit 130, which is used to perform attention enhancement fusion on the sequence of land object target pixel-level channel feature vectors to obtain a land object surface global significance fusion feature vector; a recognition result generation unit 140, which is used to determine the object recognition result based on the land object surface global significance fusion feature vector.

[0027] In the identification module 100 of the geographic environment information management system, the ground object feature extraction unit 110 is used to extract the ground object regional features in the remote sensing image to obtain a ground object local feature map. Figure 3 FIG. 1 is a block diagram of a ground object target feature extraction unit 110 in a geographic environment information management system according to an embodiment of the present application. Figure 3As shown, the ground object target feature extraction unit 110 includes: a ground object target detection subunit 111, which is used to pass the remote sensing image through a ground object target detection network to obtain a ground object target area of ​​interest image; a target area feature extraction subunit 112, which is used to pass the ground object target area of ​​interest image through a ground object target area feature extractor based on a convolutional neural network model to obtain the ground object target local feature map.

[0028] Specifically, the ground object detection subunit 111 is used to pass the remote sensing image through the ground object detection network to obtain the ground object target area of ​​interest image. It should be understood that considering that in the remote sensing image, the ground object target usually only occupies a part of the image, in addition to the ground object target, the image also contains a large amount of background information. Directly extracting and analyzing the features of the entire image will not only increase the amount of calculation, but also may introduce interference information. Therefore, the ground object detection network is further used to locate the ground object target area of ​​the remote sensing image to obtain the ground object target area of ​​interest image, so that in the subsequent remote sensing image analysis and processing process, only the ground object target area can be feature analyzed to avoid the interference of background information, improve the accuracy of ground object type recognition, reduce the amount of calculation, and save computing resources.

[0029] Specifically, the target area feature extraction subunit 112 is used to pass the image of the area of ​​interest of the ground object target through the ground object target area feature extractor based on the convolutional neural network model to obtain the local feature map of the ground object target. It should be understood that in remote sensing images, ground objects usually have complex textures, shapes and spectral characteristics, which are crucial to the identification of ground object types. Based on this, in the technical solution of the present application, a convolutional neural network model that performs well in the field of image recognition is used to construct a ground object target area feature extractor to extract features from the image of the area of ​​interest of the ground object target. The ground object target area feature extractor based on the convolutional neural network model can learn the high-level image features of the ground object target area through a combination of network structures such as convolutional layers and pooling layers, extract the spatial structure, texture, shape features, etc. of the ground object target area, and express it in the form of a local feature map of the ground object target, providing powerful feature support for subsequent ground object type identification.

[0030] In the identification module 100 of the above-mentioned geographic environment information management system, the feature dispersion unit 120 is used to perform feature dispersion processing on the local feature map of the ground object target along the channel dimension of the local feature map of the ground object target to obtain a sequence of pixel-level channel feature vectors of the ground object target. It should be understood that in the recognition of ground object targets, the feature information of the ground object target is usually not limited to the spectral information of a single pixel, but is closely related to the spatial relationship between pixels, texture features and other information. Therefore, in order to more comprehensively describe the characteristics of the ground object target, in the technical solution of the present application, the local feature map of the ground object target is further subjected to pixel-level feature dispersion processing. That is, through feature dispersion processing, the corresponding pixel points on the channel dimension in the local feature map of the ground object target are composed of pixel-level channel feature vectors to form a sequence of pixel-level channel feature vectors of the ground object target, so that more ground object details and context information can be captured and expressed, the expression ability of the features is enhanced, and a more detailed and comprehensive description of the feature information of the ground object target is achieved.

[0031] In the identification module 100 of the above-mentioned geographic environment information management system, the attention enhancement fusion unit 130 is used to perform attention enhancement fusion on the sequence of the pixel-level channel feature vectors of the ground object target to obtain the global saliency fusion feature vector of the ground object surface. In a specific example of the present application, the encoding method of performing attention enhancement fusion on the sequence of the pixel-level channel feature vectors of the ground object target to obtain the global saliency fusion feature vector of the ground object surface is to pass the sequence of the pixel-level channel feature vectors of the ground object target through an adaptive attention weight fusion network to obtain the global saliency fusion feature vector of the ground object surface. It should be understood that since the feature distribution of the ground object target in the remote sensing image is often uneven, the ground object target features at different positions may have different saliencies, that is, the ground object target features at certain positions have a significant impact on type recognition, while the ground object target features at other positions may be difficult to identify due to factors such as illumination, shadow, and occlusion. Therefore, considering this uneven feature distribution, it is expected to allocate attention to the ground object target features at different positions to improve the expressiveness of the features. Based on this, in the technical solution of the present application, an adaptive attention weight fusion network is introduced to adaptively weight and fuse the sequence of pixel-level channel feature vectors of the ground object target, so as to highlight important features and suppress minor features, thereby improving the accuracy and reliability of ground object target recognition. Unlike processing each pixel or local feature separately, the adaptive attention weight fusion network can consider the global contextual information of the image. Since ground object targets are often closely related to their surrounding environment and background, the introduction of this global contextual information can more accurately identify ground object targets.

[0032] Figure 4FIG. 1 is a block diagram of an attention enhancement fusion unit in a geographic environment information management system according to an embodiment of the present application. Figure 4 As shown, the attention enhancement fusion unit 130 includes: an autocorrelation attention weight calculation subunit 131, which is used to calculate the autocorrelation attention weight of each ground object target pixel-level channel feature vector in the sequence of ground object target pixel-level channel feature vectors to obtain a sequence of autocorrelation attention weights; a weighted fusion subunit 132, which is used to perform weighted fusion on the sequence of ground object target pixel-level channel feature vectors based on the sequence of autocorrelation attention weights to obtain the global saliency fusion feature vector of the ground object surface.

[0033] More specifically, the autocorrelation attention weight calculation subunit 131 is used to: respectively calculate the matrix product between each ground object target pixel-level channel feature vector in the sequence of ground object target pixel-level channel feature vectors and each predetermined weight matrix to obtain a sequence of encoded ground object target pixel-level channel feature vectors; add a bias vector to each encoded ground object target pixel-level channel feature vector in the sequence of encoded ground object target pixel-level channel feature vectors and then pass the result through a Selu activation function to obtain a sequence of activated ground object target pixel-level channel feature vectors; respectively calculate the product between each activated ground object target pixel-level channel feature vector in the sequence of activated ground object target pixel-level channel feature vectors and the transposed vector of each preset reference feature vector to obtain the sequence of autocorrelation attention weights.

[0034] That is, the autocorrelation attention weight of each ground object target pixel-level channel feature vector in the sequence of ground object target pixel-level channel feature vectors is calculated using the following autocorrelation attention weight calculation formula to obtain the sequence of autocorrelation attention weights, wherein the autocorrelation attention weight calculation formula is: ; ;in, is the first pixel-level channel feature vector in the sequence of the ground object target The pixel-level channel feature vector of ground objects, is the predetermined weight matrix, is the preset reference feature vector, represents the transpose of the eigenvector, is the bias vector, is the activation function, Represents the variable input to the Selu activation function, and is a hyperparameter, For the said The autocorrelation attention weight of the pixel-level channel feature vector of the ground object.

[0035] More specifically, the weighted fusion subunit 132 is used to: normalize the sequence of autocorrelation attention weights to obtain a sequence of normalized autocorrelation attention weights; use the sequence of normalized autocorrelation attention weights as weights, calculate the vector-by-vector weighted sum of the sequence of pixel-level channel feature vectors of the ground object target to obtain the global saliency fusion feature vector of the ground object surface.

[0036] In the recognition module 100 of the above-mentioned geographic environment information management system, the recognition result generation unit 140 is used to determine the object recognition result based on the global significant fusion feature vector of the surface of the object. In a specific example of the present application, the implementation method of determining the object recognition result based on the global significant fusion feature vector of the surface of the object is to pass the global significant fusion feature vector of the surface of the object through a classifier-based object identifier to obtain the object recognition result, and the object recognition result is a type label of the object. That is, the classifier-based object identifier is used to implement the parsing and recognition of the global significant fusion feature vector of the surface of the object, so as to obtain the type label of the object, that is, the category recognition result of the object. In a specific implementation, the classifier-based object identifier receives the global significant fusion feature vector of the surface of the object as input, and maps the input global significant fusion feature vector of the surface of the object to the corresponding object type label according to the classification rules learned in the training process, thereby realizing the type recognition of the object target.

[0037] It should be understood that before using the above-mentioned neural network model, it is necessary to train the ground object target detection network, the ground object target area feature extractor based on the convolutional neural network model, the adaptive attention weight fusion network and the ground object identifier based on the classifier. That is to say, in the geographic environment information management system of the present application, it also includes a training unit for training the ground object target detection network, the ground object target area feature extractor based on the convolutional neural network model, the adaptive attention weight fusion network and the ground object identifier based on the classifier.

[0038] Figure 5 FIG. 1 is a block diagram of a training unit in a geographic environment information management system according to an embodiment of the present application. Figure 5As shown, the training unit 200 includes: a training data acquisition subunit 210, which is used to acquire training data, wherein the training data includes a training remote sensing image and a true value of a ground object recognition result; a training image target detection subunit 220, which is used to pass the training remote sensing image through the ground object target detection network to obtain a training ground object target region of interest image; a training ground object feature extraction subunit 230, which is used to pass the training ground object target region of interest image through the ground object target region feature extractor based on the convolutional neural network model to obtain a training ground object target local feature map; a training feature dispersion subunit 240, which is used to perform feature dispersion processing on the training ground object target local feature map along the channel dimension of the training ground object target local feature map to obtain a sequence of pixel-level channel feature vectors of the training ground object target. The training data adaptive fusion subunit 250 is used to pass the sequence of the training ground object target pixel-level channel feature vectors through the adaptive attention weight fusion network to obtain the training ground object surface global significance fusion feature vector; the classification loss subunit 260 is used to pass the training ground object surface global significance fusion feature vector through the classifier-based ground object identifier to obtain a classification loss function value; the model training subunit 270 is used to train the ground object target detection network, the ground object target area feature extractor based on the convolutional neural network model, the adaptive attention weight fusion network and the classifier-based ground object identifier with the classification loss function value, wherein in each round of iteration of the training, the training ground object surface global significance fusion feature vector is iteratively corrected.

[0039] In the above technical scheme, the sequence of pixel-level channel feature vectors of the training ground object targets represents the image semantic features of the training remote sensing image based on the channel dimension of the convolutional neural network model, and after the sequence of pixel-level channel feature vectors of the training ground object targets passes through the adaptive attention weight fusion network, the features of some pixel-level channel feature vectors of the training ground object targets can be adaptively focused on based on the overall distribution of the image semantic features under the local channel. As a result, the global saliency fusion feature vector of the training ground object surface has a channel dimension local enhanced representation of the image semantic features, thereby resulting in the complexity of the expression of the global image semantic features, that is, the global saliency fusion feature vector of the training ground object surface has rich geometric shape details in the high-dimensional image semantic feature space, so that there is a distortion of the category representation when the classifier is used for classification to map to the class probability domain, that is, the category representation of the global saliency fusion feature vector of the training ground object surface for the category predetermined by the classifier is affected, thereby reducing the accuracy of the classification result.

[0040] Based on this, in the technical solution of the present application, each time the global saliency fusion feature vector of the training surface is classified and iterated by the classifier, the global saliency fusion feature vector of the surface of the training object is corrected by the following correction formula to obtain the corrected global saliency fusion feature vector of the training surface of the training object, wherein the correction formula is: ;in, is the global saliency fusion feature vector of the training object surface, is the global saliency fusion feature vector of the training object surface The class probability value obtained by the classifier, is the global saliency fusion feature vector of the training object surface No. eigenvalues, is the global saliency fusion feature vector of the training object surface The mean of all eigenvalues ​​of Represents the global saliency fusion feature vector of the training object surface The one-norm of is the weight hyperparameter, represents the exponential function processing with e as the base, The first one represents the global significant fusion feature vector of the corrected training surface feature values.

[0041] Specifically, the global saliency fusion feature vector of the training object surface While maintaining the geometric details of the high-dimensional feature manifold under the class probability mapping, the feature value details are fused with the global significance of the training surface feature vector by means of class probability distortion consultation. The variational representation of the feature set is used to perform shape attribute editing, thereby integrating the global saliency fusion feature vector of the training object surface The class probability distortion is mapped to the latent class space feature representation, and then the consultation fusion is performed by supplementing the basic low-rank constrained latent inversion representation to bridge the gap between the edited consultation representation and the original geometric shape details, thereby improving the global saliency fusion feature vector of the training object surface. The category representation of high-dimensional features can be improved, thereby improving the accuracy of the classification results.

[0042] In summary, the geographic environment information management system according to the embodiment of the present application is explained, which uses computer vision technology based on deep learning to analyze remote sensing images, extracts image feature expressions of the target area of ​​the object in the remote sensing image, and performs channel dimension local feature enhancement on the target area features of the object based on the attention mechanism, thereby intelligently identifying the type of the object. In this way, the accuracy and reliability of object target recognition can be improved, and the phenomenon of misclassification and omission can be reduced.

[0043] The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to being implemented by adopting the above specific details.

[0044] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiment described above is only schematic. For example, the module division is only a logical function division, and there may be other division methods in actual implementation. The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0045] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0046] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0047] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference to a figure in a claim should not be considered as limiting the claim to which it relates.

[0048] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0049] Finally, it should be noted that the above description has been given for the purpose of illustration and description. In addition, the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A geographic environment information management system, characterized in that: include: A data acquisition module is used to collect geographic environment data, wherein the geographic environment data includes remote sensing images, ground monitoring data and meteorological data; A data storage and management module, used for storing and managing the geographic environment data; A recognition module, used for performing object recognition on the ground object in the remote sensing image to obtain an object recognition result; A labeling module, used for labeling the object recognition result in the remote sensing image; Wherein, the identification module includes: A ground object target feature extraction unit, used to extract the ground object target regional features in the remote sensing image to obtain a ground object target local feature map; A feature dispersion unit, used for performing feature dispersion processing on the local feature map of the ground object target along the channel dimension of the local feature map of the ground object target to obtain a sequence of pixel-level channel feature vectors of the ground object target; An attention enhancement fusion unit is used to perform attention enhancement fusion on the sequence of pixel-level channel feature vectors of the ground object target to obtain a global saliency fusion feature vector of the ground object surface; A recognition result generating unit, used for determining the object recognition result based on the global saliency fusion feature vector of the surface of the object; Wherein, the ground object target feature extraction unit includes: A ground object detection subunit, used for passing the remote sensing image through a ground object detection network to obtain a ground object target area of ​​interest image; A target region feature extraction subunit, used for passing the image of the region of interest of the ground object target through a ground object region feature extractor based on a convolutional neural network model to obtain a local feature map of the ground object target; Wherein, the attention enhancement fusion unit is used to: The sequence of pixel-level channel feature vectors of the ground object target is passed through an adaptive attention weight fusion network to obtain a global saliency fusion feature vector of the ground object surface; Wherein, the attention enhancement fusion unit includes: An autocorrelation attention weight calculation subunit, used to calculate the autocorrelation attention weight of each ground object target pixel-level channel feature vector in the sequence of ground object target pixel-level channel feature vectors to obtain a sequence of autocorrelation attention weights; A weighted fusion subunit, configured to perform weighted fusion on the sequence of pixel-level channel feature vectors of the ground object target based on the sequence of autocorrelation attention weights to obtain a global saliency fusion feature vector of the ground object surface; Wherein, the autocorrelation attention weight calculation subunit is used to: Respectively calculating the matrix products between each ground object target pixel level channel feature vector and each predetermined weight matrix in the sequence of ground object target pixel level channel feature vectors to obtain a sequence of encoded ground object target pixel level channel feature vectors; Each encoded ground object target pixel level channel feature vector in the sequence of encoded ground object target pixel level channel feature vectors is added with a bias vector and then activated by a Selu activation function to obtain a sequence of activated ground object target pixel level channel feature vectors; The products of each activated ground object target pixel-level channel feature vector and the transposed vector of each preset reference feature vector in the sequence of activated ground object target pixel-level channel feature vector are respectively calculated to obtain the sequence of the autocorrelation attention weights.

2. The geographic environment information management system according to claim 1, characterized in that: The weighted fusion subunit is used for: Normalizing the sequence of autocorrelation attention weights to obtain a sequence of normalized autocorrelation attention weights; The sequence of normalized autocorrelation attention weights is used as weights to calculate the vector-by-vector weighted sum of the sequence of pixel-level channel feature vectors of the ground object target to obtain the global saliency fusion feature vector of the ground object surface.

3. The geographic environment information management system according to claim 2, characterized in that: The recognition result generating unit is used to: The global saliency fusion feature vector of the surface of the object is passed through a classifier-based object identifier to obtain the object recognition result, and the object recognition result is a ground object type label.

4. The geographic environment information management system according to claim 3, characterized in that: It also includes a training unit for training the ground object detection network, the ground object target area feature extractor based on the convolutional neural network model, the adaptive attention weight fusion network and the classifier-based ground object identifier.

5. The geographic environment information management system according to claim 4, characterized in that: The training unit comprises: A training data acquisition subunit is used to acquire training data, wherein the training data includes a training remote sensing image and a true value of a ground object recognition result; A training image target detection subunit, used for passing the training remote sensing image through the ground object detection network to obtain a training ground object target region of interest image; A training ground object feature extraction subunit, used for passing the training ground object target region of interest image through the ground object target region feature extractor based on the convolutional neural network model to obtain a training ground object target local feature map; A training feature dispersion subunit is used to perform feature dispersion processing on the training ground object target local feature map along the channel dimension of the training ground object target local feature map to obtain a sequence of training ground object target pixel-level channel feature vectors; A training data adaptive fusion subunit, used for passing the sequence of the training object target pixel-level channel feature vectors through the adaptive attention weight fusion network to obtain a global saliency fusion feature vector of the training object surface; A classification loss subunit, used for passing the global saliency fusion feature vector of the training surface of the object through the classifier-based object identifier to obtain a classification loss function value; The model training subunit is used to train the ground object detection network, the ground object target area feature extractor based on the convolutional neural network model, the adaptive attention weight fusion network and the classifier-based ground object identifier with the classification loss function value, wherein in each round of iteration of the training, the global significance fusion feature vector of the training ground object surface is iteratively corrected.

Images