System and method for tracking and supervising real-time change of image spots of satellite image
Through the real-time change tracking and supervision system of satellite image spectra, the feature extraction and fusion of multi-scale convolutional neural network and deep convolutional layer is solved, and the problems of low monitoring efficiency and untimely data updates in natural resource supervision are achieved, real-time monitoring is achieved and analysis improves the accuracy and real-time nature of natural resource management.
Patent Information
- Application Number
- CN202410860161.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-28
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing technology relies on periodic ground investigations and manual analysis in the supervision of natural resources, resulting in low monitoring efficiency and untimely updates of data, which cannot meet the real-time regulatory needs, especially the inability to detect and stop illegal natural resources, and the inability to effectively protect farmland. and mineral resources.
A real-time change tracking and supervision system for satellite image spectra is designed, including data collection module, multi-scale feature extraction module, multi-feature fusion module and training module, and uses multi-scale convolutional neural network and deep convolutional layer to extract and fusion features. Through the adaptive weight optimization loss function, real-time change monitoring and analysis of satellite image spectra are realized.
It improves the accuracy and efficiency of satellite image spot monitoring, can monitor image spot changes in real time or near real time, provide timely and accurate information for related decisions, reduces the influence of human factors, and enhances the real-time and accuracy of natural resource management.
Smart Images

Figure CN119091310A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image analysis, and in particular to a system and method for tracking and supervising real-time changes of satellite image spots. Background Art
[0002] Effective supervision of natural resources is crucial to ensure the rational use of land and the protection of cultivated land. Existing technologies mostly rely on periodic ground surveys and manual analysis, which have problems such as low monitoring efficiency and untimely data updates, and cannot meet the needs of real-time supervision. Early detection, early prevention, and strict investigation and punishment of various natural resource violations should be carried out to maintain the order of natural resource management, strictly protect cultivated land and mineral resources, especially focusing on the protection of permanent basic farmland, cultivated land that can be used stably in the long term, and strategic minerals such as rare earths. Summary of the invention
[0003] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a system and method for tracking and supervising real-time changes of satellite image spots.
[0004] In order to achieve the above-mentioned object of the present invention, the present invention provides a satellite image spot real-time change tracking and supervision system, comprising:
[0005] The data collection module uses I1, I2, ..., I N To represent the satellite image data of different time periods, where N is the number of image spots;
[0006] Multi-scale feature extraction module, for satellite image monitoring images, satellite image status database, and real-time construction of satellite image data images;
[0007] The multi-feature fusion module is used to perform multi-feature fusion and use adaptive weights to perform weighted fusion of features of different modalities and scales;
[0008] The training module is used to carry out the training process of satellite image spots changing, and optimize the loss function by adjusting the weight of the regression loss term.
[0009] Preferably, the multi-scale feature extraction module comprises:
[0010] Use multi-scale convolutional neural network to extract features:
[0011] F s =Φ s (I N ),s=1,2,...,S,
[0012] where F s is the feature map extracted by multi-scale s, S is the total number of multi-scales, Φ s (I N) represents the satellite image data I with different time periods at multiple scales s N Convolution operation;
[0013] The multi-scale feature extraction formula is to apply convolution operations Φ at different scales s. s (I N ) from the image data I N Extract the feature map F s , which aims to capture information of different scales in the image. Because the required features are local details in the satellite image, multi-scale feature extraction can meet the needs well.
[0014] Preferably, the multi-scale feature extraction module comprises:
[0015] Through feature mapping, at each scale s, a fully connected layer is used to map the features to a common feature space:
[0016] Z s =M s ·F s +c s , where Z s As the irrefutable evidence after mapping, M s is the feature weight matrix, c s is the bias parameter;
[0017] Through adaptive weight allocation, in order to adaptively allocate feature weights of different modalities and scales, weight adaptive allocation is performed by learning the weight vector β, and the softmax function is used after the fully connected layer to calculate the output probability distribution:
[0018] β=softmax(δ(Z s )),
[0019] Among them, δ(·) is the mapping function, which is a fully connected layer. The original output of the fully connected layer is converted into a probability distribution, and the softmax(·) function ensures that the sum of the elements of the weight vector β is 1.
[0020] Preferably, the multi-feature fusion module comprises:
[0021]
[0022] Where X is the fused feature representation, and β is the weight corresponding to multi-feature N and multi-scale S.
[0023] To further integrate the features from the satellite images, deep convolutional layers are used to integrate and refine the fused features:
[0024] H = h(X);
[0025] Where H is the feature after deep integration, and h represents the deep integration operation.
[0026] Preferably, the multi-feature fusion module comprises:
[0027] Execute the specific layer of the satellite image patch change tracking task. In order to compare and identify the changed satellite image patches, the output layer is used to identify the final detection results:
[0028] Y = σ(M·H+c), where Y is the satellite image spot detection result, σ(·) is the activation function, M is the weight of the satellite image spot change tracking output layer, and c is the bias of the satellite image spot change tracking output layer.
[0029] Preferably, the training module comprises:
[0030] Calculate the loss function Γ to train the network, let Γ be the total loss, Γ cls is the classification loss, Γ reg For the regression loss:
[0031] Γ=Γ cls (Y,Y true )+ε·Γ reg ·(Y,Y true )
[0032] Among them, Y true is the true label of the satellite image patch, and ε is the weight used to balance the regression loss term.
[0033] Preferably, the training module comprises:
[0034] Optimize the loss function by adjusting the weight of the regression loss term:
[0035] The formula for calculating the weights using the balanced regression loss is where ε base is the baseline weight for balancing regression loss, α is the image learning hyperparameter, and ΔΓ reg is the difference between the regression loss in the current iteration and the regression loss in the previous iteration, which is used to reflect the change of regression loss; Γ reg_initial is the initial value of the regression loss, which is used to normalize the loss change.
[0036] The present invention also discloses a method for tracking and supervising real-time changes of satellite image spots, comprising the following steps:
[0037] S1, using I1, I2, ..., I N To represent the satellite image data of different time periods, where N is the number of image spots;
[0038] S2, for satellite image monitoring images, satellite image status database, and real-time satellite image data images;
[0039] S3, performs multi-feature fusion, uses adaptive weights to weightedly fuse features of different modalities and scales;
[0040] S4 is used for the training process of satellite image patches changing, and optimizes the loss function by adjusting the weight of the regression loss term.
[0041] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0042] The data collection module represents the satellite image data of different time periods, ensuring the comprehensive collection and management of data. It extracts features from satellite image monitoring images, current status databases, and real-time construction data images, and can capture multiple aspects of information about the image at different times and in different states, thereby improving the accuracy of the analysis. The adaptive weights perform weighted fusion of features of different modes and scales, which fully utilizes the advantages of different features while avoiding the limitations that may be brought about by a single feature. Finally, the training module uses the weight of the regression loss term to optimize the loss function.
[0043] The system includes the ability to process satellite image data in real time, which means that the system can monitor the changes of image spots in real time or near real time, and provide timely and accurate information support for relevant decision-making. In addition, the system can also be applied to many fields such as urban planning, environmental monitoring, disaster warning, etc., with high practical value.
[0044] Through the application of technologies such as deep learning, the system can automatically complete tasks such as feature extraction, feature fusion and model training without human intervention. This not only improves processing efficiency, but also reduces the impact of human factors on the results, making the results more objective and reliable.
[0045] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:
[0047] Figure 1 It is an overall schematic diagram of the present invention;
[0048] Figure 2 It is a schematic diagram of the working of the present invention;
[0049] Figure 3This is a schematic diagram of obtaining the satellite image spot status database of the present invention;
[0050] Figure 4 This is a schematic diagram of the delineation and superposition of historical data of satellite image spots of the present invention;
[0051] Figure 5 It is another schematic diagram of historical data superposition of satellite image spots of the present invention;
[0052] Figure 6 This is a schematic diagram of satellite image spot planning data superposition in the present invention;
[0053] Figure 7 This is a schematic diagram of superimposing the real-time status of satellite image spots of the present invention;
[0054] Figure 8 This is a schematic diagram of the expected use of satellite image spots in the present invention;
[0055] Fig. 9 This is a schematic diagram of limiting the superposition of satellite image spots supply in the present invention;
[0056] Fig.10 and 11 It is a schematic diagram of comparison before and after the satellite image patch construction of the present invention. DETAILED DESCRIPTION
[0057] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.
[0058] like Figure 1 As shown, the present invention discloses a satellite image spot real-time change tracking and monitoring system, comprising:
[0059] Collect remote sensing images of the land, satellite images obtained in the process of natural resource management, and collect images and management data.
[0060] The data collection module uses I1, I2, ..., I N To represent satellite image data of different time periods, where N is the number of image features, such as optical satellite image historical land data images, optical satellite image early construction data, optical satellite image implementation construction data images, etc.
[0061] The multi-scale feature extraction module uses a multi-scale convolutional neural network to extract features from satellite image monitoring images, satellite image status database, and real-time construction satellite image data images:
[0062] F s =Φs (I N ),s=1,2,...,S,
[0063] where F s is the feature map extracted by multi-scale s, S is the total number of multi-scales, Φ s (I N ) represents the satellite image data I with different time periods at multiple scales s N Convolution operation;
[0064] The multi-scale feature extraction formula is to apply convolution operations Φ at different scales s. s (I N ) from the image data I N Extract the feature map F s , which aims to capture information of different scales in the image. Because the required features are local details in the satellite image, multi-scale feature extraction can meet the needs well.
[0065] Through feature mapping, at each scale s, a fully connected layer is used to map the features to a common feature space:
[0066] Z s =M s ·F s +c s , where Z s As the irrefutable evidence after mapping, M s is the initial feature weight matrix of the satellite image patch, c s It is the initial bias parameter of the satellite image patch, which is used to adjust the offset of the output feature;
[0067] Since this calculation is a linear transformation, a nonlinear activation function needs to be used after the fully connected layer in order to learn nonlinear relationships in the neural network.
[0068] Through adaptive weight allocation, in order to adaptively allocate feature weights of different modalities and scales, weight adaptive allocation is performed by learning the weight vector β, and the softmax function is used after the fully connected layer to calculate the output probability distribution:
[0069] β=softmax(δ(Z s )),
[0070] Where δ(·) is a mapping function, which is a fully connected layer. The original output of the fully connected layer is converted into a probability distribution. The softmax(·) function ensures that the sum of the elements of the weight vector β is 1; β is the weight vector, δ(Z s ) The probability value after the softmax function conversion reflects the feature space Z after mapping s middle.
[0071] Code example:
[0072] import numpy as np
[0073] #Assume Z s is the output of the fully connected layer, i.e. logitsZ s =np.array([3.0,1.0,0.2])
[0074] #Use softmax function to calculate probability distribution
[0075] exp_Z s =np.exp(zs)
[0076] softmax_output = exp_Z s / np.sum(exp_Z s )
[0077] print(softmax_output)
[0078] Perform multi-feature fusion and use adaptive weights to weightedly fuse satellite images of different features and scales:
[0079]
[0080] Where X is the fused feature representation, β NS is the weight corresponding to the satellite image feature N and the multi-scale S.
[0081] To further integrate the features from the satellite images, deep convolutional layers are used to integrate and refine the fused features:
[0082] H = h(X);
[0083] Where H is the feature after deep integration, and h represents the deep integration operation.
[0084] In satellite image analysis, when using deep convolutional layers to integrate and refine fusion features X, the formula H = h(X) is a key step.
[0085] The deep integration operation h can capture more complex and abstract patterns in the fused feature X. As the number of convolutional layers increases, the network can learn more and more advanced feature representations. Since the fused feature X has considered information at different scales, the deep integration operation h can further refine this scale information and learn scale-invariant feature representations.
[0086] Execute the specific layer of the satellite image spot change tracking task. In order to compare and identify the changed satellite image spots, the output layer is used to identify the final detection results:
[0087] Y = σ(M·H+c), where Y is the satellite image spot detection result, σ(·) is the activation function, M is the weight of the satellite image spot change tracking output layer, and c is the bias of the satellite image spot change tracking output layer;
[0088] Calculate the loss function Γ to train the network, let Γ be the total loss function, Γ cls is the classification loss, Γ reg For the regression loss:
[0089] Γ=Γ cls (Y,Y true )+ε·Γ reg (Y,Y true );
[0090] Among them, Y true is the true label of the satellite image patch, and ε is the weight used to balance the regression loss;
[0091] The classification loss function measures the performance of the model on the classification task. The classification part in Y is compared with the classification part in Ytrue, and the difference or loss is calculated to determine the accuracy of whether the satellite image spots have changed. The other regression loss function measures the performance of the model on the regression task and predicts the accuracy of the position or shape changes of the satellite image spots. The regression part in Y (such as position coordinates, shape parameters, etc.) is compared with the corresponding part in Ytrue to calculate the difference or loss.
[0092] The formula for calculating the weights using the balanced regression loss is where ε base is the baseline weight for balancing regression loss, α is the image learning hyperparameter, and ΔΓ reg It is the difference between the regression loss in the current iteration and the regression loss in the previous iteration, which is used to reflect the change of regression loss. reg_initial is the initial value of the regression loss, which is used to normalize the loss change.
[0093] Use ε to measure the relative importance of regression loss in the total loss. By adjusting the value of ε, the emphasis of the model on the regression task can be controlled. If ε is large, the model will pay more attention to the accuracy of the regression task; if ε is small, the model will pay more attention to the accuracy of the classification task.
[0094] By minimizing Γ, the performance of the model in classification and regression tasks is optimized, making it more accurate in detecting and tracking changes in satellite image patches.
[0095] This formula combines the loss functions of classification and regression tasks to measure the performance of the model in detecting and tracking satellite image spot changes. By adjusting the weight factor ε, the emphasis of the model on classification and regression tasks is balanced, thereby achieving more accurate spot change detection and tracking.
[0096] First, the data collection module can efficiently collect and organize satellite image data from different time periods to ensure the richness and accuracy of the data, providing a solid foundation for subsequent analysis. Secondly, the multi-scale feature extraction module can extract multi-scale and multi-modal feature information from satellite image monitoring images, current status databases, and real-time construction data. These feature information can comprehensively describe the characteristics and change trends of the image. The multi-feature fusion module weightedly fuses these features through adaptive weights, effectively integrating features from different sources and scales, and improving the system's comprehensive analysis capabilities. Finally, the training module uses backpropagation and gradient descent algorithms to train and optimize the model to ensure that the model can accurately identify changes in satellite image spots and has good generalization and real-time performance. Overall, the four modules of the system work together to greatly improve the accuracy and efficiency of satellite image monitoring and analysis, and provide strong support for decision-making in related fields.
[0097] These steps and formulas together constitute a complete multimodal fusion network, which is used to extract features from satellite image data of different periods, perform effective fusion and analysis, and realize real-time change tracking and supervision of satellite image data.
[0098] By combining multi-source data and multi-scale convolutional neural network technology, an efficient and accurate system and method for tracking and supervising real-time changes in satellite image patches is provided, which significantly improves the real-time and accuracy of natural resource supervision.
[0099] The specific implementation contents include: Figure 3 As shown in the figure, by overlaying with the 2018 land use status database, the land occupation of satellite image spots, especially the occupation of cultivated land, is analyzed, which is of great significance for the calculation of illegal cultivated land occupation and illegal cultivated land occupation ratio in cities and the investigation and punishment of illegal land use. Take the satellite image spots of a certain district on a certain day in 2022 as an example.
[0100] like Figure 4 As shown, by superimposing with the 2018 permanent basic farmland demarcation database, the area of permanent basic farmland occupied by satellite images was analyzed, which is of great significance for the determination standards of facility agricultural land and the investigation and punishment of illegal land use.
[0101] like Figure 5As shown, by superimposing the characteristics and scales with the 2022 "Three Zones and Three Lines" permanent basic farmland demarcation database (permanent basic farmland has not been demarcated in this map and its surrounding areas), the area of permanent basic farmland occupied by the satellite image spots was analyzed, which is mainly of great significance for the determination standards of facility agricultural land and the investigation and punishment of illegal land use.
[0102] like Figure 6 As shown, by superimposing features and scales with the land planning database and analyzing the planning compliance of satellite image patches, we can check whether the punishment methods for illegal land use in counties and districts are accurate, whether illegal land that does not comply with the plan is dealt with by demolition and reclamation, and whether illegal land that complies with the plan is dealt with by confiscation. This is of great significance for verifying whether the methods of investigating and handling illegal land use in counties and districts are accurate.
[0103] like Figure 7 As shown, by superimposing features and scales with the "Three National Land Surveys" database, the actual status and use of satellite image spots can be analyzed, which is of great significance for assisting the city-level verification of legality and verifying whether the legality judgment of districts and counties is accurate.
[0104] like Figure 8 As shown, by superimposing features and scales with the expropriation and conversion database, the legal land use data and area of the issued satellite images can be analyzed. This can be used to verify whether the district and county's determination of legal land use is correct, whether the reported area of legal approval documents is accurate, and whether there is any situation of applying legal approval documents. This is of great significance for calculating the proportion of illegal occupation of farmland.
[0105] like Fig. 9 As shown, by overlaying analysis with the land supply red line characteristics and scale, the legal land use data and area of the issued satellite image spots are analyzed, which is of great significance for providing legal procedures for existing construction land that cannot find agricultural conversion procedures.
[0106] like Fig.10 and 11 As shown, by overlaying the time-series remote sensing monitoring images, the before and after phases are extracted respectively, and combined with the field photos of the spots provided by the districts and counties, the overlaid management data information, etc., it is of great significance for conducting independent city-level legality determination.
[0107] Online evidence submission and results submission service
[0108] Based on the construction of the land survey cloud system, we will further optimize and improve the "Internet +" online evidence service, improve the overall architecture and detailed functional design, construction and testing, and deploy map reporting and field evidence platforms.
[0109] Deploy summary and analysis functions such as real-time statistics to provide technical support for quickly grasping the city's satellite image law enforcement work situation.
[0110] Based on the data preparation and analysis, combined with the city's planning management data, basic survey data and land use approval data, as well as district and county satellite image spots and online evidence and other information, a city-level review is conducted on the legality of the satellite image law enforcement spots reported by districts and counties on a spot-by-spot basis.
[0111] Extract the areas of doubtful city-level legality determinations and send them to districts and counties for online evidence collection.
[0112] Municipal audit and data aggregation
[0113] Data review is carried out based on the reporting of each district and county in the satellite image law enforcement system. Mainly based on the reporting results of the district and county system and the online evidence materials, the satellite image system reported by the district and county is reviewed map by map.
[0114] The reporting status of the satellite image law enforcement spots in each district and county is extracted in real time, and key verification spots are extracted for mid-term supervision and municipal-level acceptance. After the municipal-level acceptance, the satellite image system is reviewed based on the municipal-level acceptance and certification results.
[0115] According to the specific embodiment, it can be known that:
[0116] a. Overwater viaduct: The existing land type is the area where an elevated road or bridge is built on the surface of a river, lake or reservoir.
[0117] b. Reservoir surface. According to the requirements for construction land, the reservoir surface formed by the reservoir inundation area of the newly built water conservancy and hydropower projects does not need to go through the procedures for agricultural land conversion.
[0118] c. Riverbank regulation. Riverbank regulation refers to the implementation of shoreline reinforcement and regulation projects within the reasonable range of river, lake and reservoir embankments. Riverside (river) parks built in the name of riverbank regulation and drainage ditches (channels) of non-agricultural water conservancy facilities shall not be included in the scope of riverbank regulation.
[0119] d. Land for electric towers. Land for electric towers refers to the land used for the foundations of poles and towers of overhead power lines.
[0120] e. Agricultural production ancillary facilities built by rural villagers. The sun-drying dams and simple agricultural production ancillary facilities reasonably built by rural villagers for production and living needs shall be determined as unchanged on the ground. The main part of the house, private villas, supporting landscape facilities, parking lots and agricultural land with facilities that should be filed shall not be reported as unchanged on the ground. After the relevant regulations on land management in Chongqing clarify the scope and standards of homesteads, they shall be implemented in accordance with the new regulations.
[0121] f. For illegal land use that has been reported in previous years’ satellite image enforcement, if there are no new illegal behavior spots, it can be determined that there is no change on the ground. The year, quarter (or month) and spot number of previous years’ satellite images must be filled in the judgment basis.
[0122] g. On the existing construction land, the only non-construction changes are caused by the stacking of materials, which can be judged as no actual changes.
[0123] h. Special land. Land used by special units can be determined to be unchanged after approval by the municipal government and managed in strict accordance with the requirements of satellite image tracking and supervision.
[0124] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.
Claims
1. A satellite image spot real-time change tracking and monitoring system, characterized in that: include: Data collection module, using I1, I2, ..., I N To represent the satellite image data of different time periods, where N is the number of image spots; Multi-scale feature extraction module, for satellite image monitoring images, satellite image status database, and real-time construction of satellite image data images; The multi-feature fusion module is used to perform multi-feature fusion and use adaptive weights to perform weighted fusion of features of different modalities and scales; The training module is used to carry out the training process of satellite image spots changing, and optimize the loss function by adjusting the weight of the regression loss term.
2. The satellite image spot real-time change tracking and monitoring system according to claim 1 is characterized in that: The multi-scale feature extraction module comprises: Use multi-scale convolutional neural network to extract features: F s =Φ s (I N ),s=1,2,...,S, where F s is the feature map extracted by multi-scale s, S is the total number of multi-scales, Φ s (I N ) represents the satellite image data I with different time periods at multiple scales s N Convolution operation; The multi-scale feature extraction formula is to apply convolution operations Φ at different scales s. s (I N ) from the image data I N Extract the feature map F s , which aims to capture information of different scales in the image. Because the required features are local details in the satellite image, multi-scale feature extraction can meet the needs well.
3. The satellite image spot real-time change tracking and monitoring system according to claim 1 is characterized in that: The multi-scale feature extraction module comprises: Through feature mapping, at each scale s, a fully connected layer is used to map the features to a common feature space: Z s =M s ·F s +c s , where Z s As the ironclad evidence after mapping, M s is the feature weight matrix, c s is the bias parameter; Through adaptive weight allocation, in order to adaptively allocate feature weights of different modalities and scales, weight adaptive allocation is performed by learning the weight vector β, and the softmax function is used after the fully connected layer to calculate the output probability distribution: β=softmax(δ(Z s )), Among them, δ(·) is the mapping function, which is a fully connected layer. The original output of the fully connected layer is converted into a probability distribution, and the softmax(·) function ensures that the sum of the elements of the weight vector β is 1.
4. The satellite image spot real-time change tracking and monitoring system according to claim 1 is characterized in that: The multi-feature fusion module includes: Where X is the fused feature representation, and β is the weight corresponding to multi-feature N and multi-scale S. To further integrate the features from the satellite images, deep convolutional layers are used to integrate and refine the fused features: H = h(X); Where H is the feature after deep integration, and h represents the deep integration operation.
5. The satellite image spot real-time change tracking and monitoring system according to claim 1 is characterized in that: The multi-feature fusion module includes: Execute the specific layer of the satellite image patch change tracking task. In order to compare and identify the changed satellite image patches, the output layer is used to identify the final detection results: Y = σ(M·H+c), where Y is the satellite image spot detection result, σ(·) is the activation function, M is the weight of the satellite image spot change tracking output layer, and c is the bias of the satellite image spot change tracking output layer.
6. The satellite image spot real-time change tracking and monitoring system according to claim 1 is characterized in that: The training module includes: Calculate the loss function Γ to train the network, let Γ be the total loss, Γ cls is the classification loss, Γ reg For the regression loss: C=C cls (Y,Y true )+e·C reg ·(Y,Y true ) Among them, Y true is the true label of the satellite image patch, and ε is the weight used to balance the regression loss term.
7. The satellite image spot real-time change tracking and monitoring system according to claim 6 is characterized in that: The training module includes: Optimize the loss function by adjusting the weight of the regression loss term: The formula for calculating the weights using the balanced regression loss is where ε base is the baseline weight for balancing regression loss, α is the image learning hyperparameter, and ΔΓ reg is the difference between the regression loss in the current iteration and the regression loss in the previous iteration, which is used to reflect the change of regression loss; Γ reg_initial is the initial value of the regression loss, which is used to normalize the loss change.
8. A method for tracking and supervising real-time changes of satellite image spots, characterized in that: The steps include: S1, using I1, I2, ..., I N To represent the satellite image data of different time periods, where N is the number of image spots; S2, for satellite image monitoring images, satellite image status database, and real-time satellite image data images; S3, performs multi-feature fusion, uses adaptive weights to weightedly fuse features of different modalities and scales; S4 is used for the training process of satellite image patches changing, and optimizes the loss function by adjusting the weight of the regression loss term.
Citation Information
Cited By
Multi-source high-resolution remote sensing image change detection method and system
CN119360230A
