Change detection method and device under multiple imaging conditions based on multi-level information baseboard

Abstract

The application provides a change detection method and device under multiple imaging conditions based on a multi-level information baseboard, and belongs to the technical field of image recognition. The method comprises the following steps: encoding remote sensing images to a space-time embedded baseboard, obtaining standardized input data, and then inputting the data into a natural background branch, an artificial background branch and a foreground target branch to obtain a natural background sensitivity function, a geometric descriptor and a target feature fingerprint; filtering the remote sensing images based on the natural background sensitivity function to obtain a deep semantic feature map; comparing the deep semantic feature map with the geometric descriptors at different times in the same geographical position, and if the difference exceeds a structural threshold, it is determined that there is a structural abnormality in the geographical position; and calculating the similarity and spatial position offset of the target feature fingerprints in the two time phases to determine whether there is an increase or decrease in the target. The application realizes accurate differentiation and detection of natural background false changes, artificial facility structural changes and dynamic target complex states.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to a method and apparatus for detecting changes under multiple imaging conditions based on a multi-level information substrate. Background Technology

[0002] With the rapid development of Earth observation technology, remote sensing imagery has become a crucial means of acquiring information about the Earth's surface. Land cover change detection, as one of the core tasks of remote sensing image processing, aims to automatically identify and extract changes in land cover types by analyzing images of the same area acquired at different times. This technology plays a vital role in urban planning and illegal construction monitoring, agricultural yield estimation, ecological environment assessment, disaster emergency response, and national defense security. Accurate and efficient land cover change detection is a key technological prerequisite for achieving dynamic monitoring and intelligent management of the Earth's surface.

[0003] Existing methods for detecting ground feature changes mainly include those based on image algebraic operations, traditional machine learning, and end-to-end feature extraction based on deep learning. However, because the acquisition of remote sensing images is affected by various factors such as sensor type, lighting conditions, seasonal changes, and atmospheric environment, multiple imaging conditions often lead to significant radiometric differences in the same ground feature across different time phases, which can easily cause false detections. In addition, existing methods typically use a single network architecture to process the entire image uniformly, making it difficult to effectively distinguish between pseudo-changes in the natural background caused by environmental interference and real changes in ground features. Furthermore, for man-made facilities, there is a lack of precise description of their geometric structure, making it difficult to accurately determine structural anomalies. For dynamic targets, the reliance on global feature comparison can easily lead to missed detections or false alarms in complex scenarios where targets are added or removed.

[0004] In summary, how to effectively suppress the interference of natural background pseudo-changes under multiple imaging conditions and achieve accurate differentiation and detection of structural anomalies of man-made facilities and complex states of dynamic targets has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a method and apparatus for detecting changes under multiple imaging conditions based on a multi-level information substrate, which solves the shortcomings of the prior art in that it is difficult to distinguish natural background environmental interference and cannot accurately determine the anomalies of man-made facilities and the detection of dynamic targets.

[0006] This application provides a method for detecting changes under multiple imaging conditions based on a multi-level information substrate, including the following steps: The remote sensing images are uniformly encoded onto a spatiotemporal embedding base of a preset dimension to obtain standardized input data; The standardized input data is input into the natural background branch, the artificial background branch, and the foreground target branch, respectively, to obtain the natural background sensitivity function, the geometric descriptor, and the target feature fingerprint. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. The remote sensing image is filtered based on the natural background sensitivity function to obtain a deep semantic feature map; For the deep semantic feature map, compare the geometric descriptors of the same geographical location at different times. If the difference exceeds the structural threshold, it is determined that there is a structural change in the geographical location. Calculate the similarity and spatial offset of the target feature fingerprint in two consecutive time phases. If the similarity is lower than the similarity threshold or there is a mismatch in one time phase, it is determined that there is an increase or decrease in the target.

[0007] According to the multi-level information substrate-based method for change detection under multiple imaging conditions provided in this application, the natural background branch generates a natural background sensitivity function based on the following steps: Spatial heterogeneity quantification is performed on the standardized input data to obtain statistical significance indicators; The standardized input data is subjected to structural regularity quantification to obtain a structural regularity score; The natural background sensitivity function is obtained by weighted summation of the statistical significance index and the structural regularity score.

[0008] According to the multi-level information substrate-based change detection method under multiple imaging conditions provided in this application, the step of quantifying the spatial heterogeneity of the standardized input data to obtain statistical significance indicators includes: Extract features from the local neighborhood, using the center pixel as the core. Calculate the feature mean and feature variance within the local neighborhood; A statistical significance index is determined based on the standardized distance between the pixel features of the center pixel and the mean of the features.

[0009] According to the multi-level information substrate-based change detection method under multiple imaging conditions provided in this application, the step of quantifying the structural regularity of the standardized input data to obtain a structural regularity score includes: Perform a Fast Fourier Transform on the local neighborhood to obtain the spectral energy distribution; The structural regularity score is determined based on the proportion of high-frequency energy.

[0010] According to the multi-level information substrate-based change detection method under multiple imaging conditions provided in this application, the step of filtering the remote sensing image based on the natural background sensitivity function to obtain a deep semantic feature map includes: The standardized input data is nonlinearly mapped to the vector space of the natural background sensitivity function to obtain a deep semantic feature map.

[0011] According to the multi-imaging condition change detection method based on a multi-level information substrate provided in this application, the artificial background branch generates a geometric descriptor based on the following steps: For the deep semantic feature map, the partial derivatives in the horizontal and vertical directions are calculated respectively to obtain the derivative field; Multi-channel fusion is performed on each pixel position in the derivative field to obtain the structure tensor; Eigenvalue decomposition is performed on the structure tensor at each pixel location to obtain the principal energy and secondary energy. Based on the principal energy and the secondary energy, structural invariants are constructed; By fusing structural invariants at different scales, geometric descriptors are obtained.

[0012] According to the multi-level information substrate-based method for change detection under multiple imaging conditions provided in this application, the foreground target branch generates a target feature fingerprint based on the following steps: Calculate the Euclidean distance between each pixel in the standardized input data and the mean of its local neighborhood, obtain the saliency score of each pixel, and then construct a saliency map. The saliency map is binarized based on a dynamic threshold to obtain a set of salient points; Density clustering is performed on the salient point set to obtain multiple target clusters; Feature extraction is performed on each of the target clusters to obtain target feature fingerprints.

[0013] According to the multi-level information substrate-based change detection method under multiple imaging conditions provided in this application, the step of extracting features from each target cluster to obtain target feature fingerprints includes: For each target cluster, the core feature vector is extracted using saliency-weighted attention pooling with the geometric centroid as the central anchor point to obtain the target feature fingerprint.

[0014] This application also provides a change detection device under multiple imaging conditions based on a multi-level information substrate, including the following modules: The data preprocessing module is used to: uniformly encode remote sensing images onto a spatiotemporal embedding base of preset dimensions to obtain standardized input data; The branch parallel module is used to: input the standardized input data into the natural background branch, the artificial background branch and the foreground target branch respectively, and obtain the natural background sensitivity function, the geometric descriptor and the target feature fingerprint respectively. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. The image filtering module is used to: filter the remote sensing image based on the natural background sensitivity function to obtain a deep semantic feature map; The first detection module is used to: compare the geometric descriptors of the same geographical location at different times with the deep semantic feature map; if the difference exceeds the structural threshold, then determine that there is a structural anomaly in the geographical location. The second detection module is used to: calculate the similarity and spatial offset of the target feature fingerprint in two consecutive time phases; if the similarity is lower than the similarity threshold or there is a mismatch in one time phase, then it is determined that there is an increase or decrease in the target.

[0015] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-imaging condition change detection method based on a multi-level information substrate as described above.

[0016] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-imaging condition change detection method based on a multi-level information substrate as described above.

[0017] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-imaging condition change detection method based on a multi-level information substrate as described above.

[0018] The method and apparatus for change detection under multiple imaging conditions based on a multi-level information substrate provided in this application, by uniformly encoding remote sensing images onto a spatiotemporal embedding substrate of a preset dimension, provides a unified and robust feature representation benchmark for subsequent processing, effectively eliminating feature space inconsistencies caused by sensor differences or changes in ambient lighting, thereby significantly improving the anti-interference capability and accuracy of change detection under complex imaging conditions; furthermore, by adaptively labeling and filtering low-weight pseudo-change regions caused by environmental interference through a natural background sensitivity function, false change signals caused by natural factors such as vegetation phenology and shadows are effectively eliminated, reducing the false detection rate and obtaining a high-purity deep semantic feature map; subsequently, for... For man-made structures, this method utilizes geometric descriptors to accurately represent their geometric structure and determines structural anomalies by comparing the differences in geometric descriptors at the same geographical location at different times. This enables the detection of substantial structural changes such as the construction, demolition, or expansion of buildings, avoiding misjudgments caused by changes in surface texture or color, and improving the reliability of man-made feature change detection. For dynamic targets, semantic identification is performed based on target feature fingerprints, and a joint determination is made by combining similarity and spatial position offset to accurately identify the increase or decrease of the target's state and accurately determine whether the target has moved. This overcomes the shortcomings of traditional methods in distinguishing changes in target state and achieves accurate detection of complex motion states of dynamic targets. In summary, this application effectively overcomes interference from multiple imaging conditions through the collaborative work of multi-level information bases, achieving accurate differentiation and detection of pseudo-changes in natural backgrounds, structural anomalies of man-made structures, and complex states of dynamic targets, effectively improving the overall accuracy and robustness of change detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the change detection method under multiple imaging conditions based on a multi-level information substrate provided in this application. Figure 2 This is a schematic diagram of the structure of the multi-imaging condition change detection device based on a multi-level information substrate provided in this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] It should be noted that in the description of the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The terms "upper," "lower," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, or integral connections; they can be mechanical connections or electrical connections; they can be direct connections or indirect connections through an intermediate medium; and they can be internal connections between two elements. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such numbers can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects have an "or" relationship.

[0024] The following is combined with Figures 1-3 This application describes a method and apparatus for detecting changes under multiple imaging conditions based on a multi-level information substrate, as provided in the embodiments of this application.

[0025] Figure 1This is a flowchart illustrating the change detection method under multiple imaging conditions based on a multi-level information substrate provided in this application, as shown below. Figure 1 As shown, the method includes the following: S110 encodes remote sensing images uniformly onto a spatiotemporal embedding base of preset dimensions to obtain standardized input data; S120, the standardized input data is input into the natural background branch, the artificial background branch and the foreground target branch respectively, and the natural background sensitivity function, the geometric descriptor and the target feature fingerprint are obtained accordingly. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. S130, the remote sensing image is filtered based on the natural background sensitivity function to obtain a deep semantic feature map; S140, For the deep semantic feature map, compare the geometric descriptors of the same geographical location at different times. If the difference exceeds the structural threshold, it is determined that there is a structural change in the geographical location. S150, calculate the similarity and spatial offset of the target feature fingerprint in the two time phases. If the similarity is lower than the similarity threshold or there is a mismatch in one time phase, it is determined that there is a target increase or decrease.

[0026] It should be noted that the execution subject of the multi-imaging condition change detection method based on multi-level information substrate provided in this application embodiment can be a server, computer equipment, such as mobile phone, tablet computer, laptop computer, handheld computer, vehicle electronic equipment, wearable device, ultra-mobile personal computer (UMPC), netbook or personal digital assistant (PDA), etc.

[0027] In S110, the spatiotemporal embedding base refers to mapping remote sensing image data from multiple time phases into a unified feature space with a preset dimension, eliminating the differences in feature distribution caused by imaging conditions such as different sensors, different lighting, and different seasons, so that the same ground feature at different times has a high degree of consistency and comparability in the feature space.

[0028] Preferably, the spatiotemporal embedding base has 64 dimensions. For ease of understanding, the following embodiments use a 64-dimensional spatiotemporal embedding base as an example to illustrate the change detection method provided in this application.

[0029] In this embodiment, a pre-trained encoder is used to uniformly encode remote sensing images onto a spatiotemporal embedding base of a preset dimension, transforming the original pixel-level data into high-dimensional feature vectors to obtain standardized input data, effectively aligning the feature space under multiple imaging conditions.

[0030] Specifically, the parameters of the 64-dimensional spatiotemporal embedding base are kept frozen, and an adapter layer with a small number of trainable parameters is used to adjust the gain of the feature response of the long-span structure of the fixed facility and the point salience of the target during the feature extraction stage.

[0031] In S120, the natural background sensitivity function is a mathematical function or feature map used to quantify and evaluate the degree of interference from environmental factors such as changes in lighting, shadows, and seasonal phenology on natural elements in an image, such as vegetation, water bodies, and bare soil. The lower the function value, the greater the environmental interference in the area, and the more likely it is to be a pseudo-change area. Geometric descriptors are feature vectors used to characterize the spatial geometry of man-made structures such as buildings, roads, and bridges, capturing the shape and spatial topological relationships of the target, rather than surface texture or color. Target feature fingerprints are highly unique and discriminative feature representations used for semantic-level identification of dynamic targets such as vehicles, ships, and aircraft. Even with slight changes in viewing angle or lighting, the fingerprint features remain relatively stable, enabling accurate cross-temporal matching.

[0032] In this embodiment, standardized input data is input in parallel into three independent branch networks. The natural background branch extracts features of natural elements such as vegetation and water bodies, and calculates a natural background sensitivity function, which marks low-weight areas caused by lighting, shadows, etc. The artificial background branch extracts the edge, outline, and three-dimensional structure information of artificial facilities such as buildings, and generates a geometric descriptor. The foreground target branch extracts fine-grained features of dynamic targets such as vehicles and ships, and generates a target feature fingerprint for identifying changes in target state (increase, decrease).

[0033] Here, the fixed facility segmentation head employs a multi-scale feature fusion architecture to output pixel-level ground feature masks, accurately defining the geometric contours of buildings, roads, and other facilities. The target detection head uses a lightweight regression network to output the bounding boxes or center positions of target ships, aircraft, and vehicles.

[0034] In S130, the original features are weighted and filtered using the natural background sensitivity function. For regions where the sensitivity function value is lower than the preset threshold, i.e., regions with severe environmental interference, their weight is reduced or set to zero, thereby obtaining a pure deep semantic feature map.

[0035] In S140, structural alterations refer to substantial changes in the physical structure of man-made facilities, such as new construction, demolition, heightening, and expansion.

[0036] In this embodiment of the application, man-made facilities at the same geographical location are located on the deep semantic feature map. The geometric descriptors of the facility in two different time periods are compared, and the difference between the two is calculated, such as cosine distance or Euclidean distance. If the difference exceeds a preset structural threshold, it is determined that a structural change has occurred at the location, such as illegal construction or demolition.

[0037] In S150, spatial position offset refers to the Euclidean distance or geographic distance in the spatial coordinate system between the center point or key anchor point of the same dynamic target in images from two consecutive time phases.

[0038] In this embodiment of the application, for a foreground target, the similarity of the target feature fingerprint in two time phases is calculated, and its spatial position offset is calculated. If the similarity is lower than the similarity threshold, or if there is a complete mismatch at a certain moment, it is determined that the target has been added or disappeared. If the similarity is greater than or equal to the similarity threshold, it is confirmed that it is the same target, and if the spatial position offset is greater than the offset threshold, it is determined that the target has moved.

[0039] In this embodiment, the traditional blind pixel comparison is changed, firstly in and The detection task is performed in parallel within the temporal image. The 64-dimensional features within the target region are spatially pooled and aligned using the mask or bounding box output by the detection head. For fixed facilities, geometric consistency is evaluated by comparing segmentation masks at the same location. If the mask shape undergoes significant deformation or disappearance, it is determined to be a structural anomaly, such as damage or reconstruction. For moving targets, based on the detection results, the number, type, and spatial distribution of targets are statistically analyzed. By calculating the cosine similarity of the corresponding target feature fingerprints, the withdrawal or entry of targets is identified.

[0040] It should be noted that there are no strict timing restrictions between S140 and S150.

[0041] The change detection method based on a multi-level information substrate under multiple imaging conditions provided in this application embodiment encodes remote sensing images uniformly onto a spatiotemporal embedding substrate of a preset dimension, providing a unified and robust feature representation benchmark for subsequent processing. This effectively eliminates feature space inconsistencies caused by sensor differences or changes in ambient lighting, thereby significantly improving the anti-interference capability and accuracy of change detection under complex imaging conditions. Furthermore, by using a natural background sensitivity function to adaptively label and filter low-weight pseudo-change regions caused by environmental interference, false change signals caused by natural factors such as vegetation phenology and shadows are effectively eliminated, reducing the false detection rate. This process yields high-purity deep semantic feature maps. Then, for man-made structures, geometric descriptors are used to accurately represent their geometric structures. By comparing the differences in geometric descriptors of the same geographical location at different times, structural anomalies are determined. This method can capture substantial structural changes such as the construction, demolition, or expansion of buildings, avoiding misjudgments caused by changes in surface texture or color, and improving the reliability of man-made feature change detection. For dynamic targets, a joint determination is made by combining similarity and spatial position offset to accurately identify the increase or decrease in the target's state, thus overcoming the shortcomings of traditional methods in distinguishing changes in target state and achieving accurate detection of dynamic targets. In summary, this application, through the collaborative work of multi-level information baseboards, effectively overcomes interference from multiple imaging conditions, achieving accurate differentiation and detection of pseudo-changes in natural backgrounds, structural anomalies of man-made structures, and complex states of dynamic targets, effectively improving the overall accuracy and robustness of change detection.

[0042] In an optional embodiment, the natural background branch generates a natural background sensitivity function based on the following steps: Spatial heterogeneity quantification is performed on the standardized input data to obtain statistical significance indicators; The standardized input data is subjected to structural regularity quantification to obtain a structural regularity score; The natural background sensitivity function is obtained by weighted summation of the statistical significance index and the structural regularity score.

[0043] In this embodiment, standardized input data is received, and the variance, information entropy, or local binary pattern features of pixels or feature vectors are calculated within the local receptive field to quantify the spatial heterogeneity of the region. Then, combined with the background distribution of the global or neighborhood, a statistical significance index of the heterogeneity value, such as Z-score or p-value, is calculated to eliminate the interference of extreme noise and obtain a reliable statistical significance index.

[0044] In this embodiment of the application, edge detection, line segment extraction or regular texture analysis are performed on the same standardized input data. By calculating the density, directional consistency or symmetry score of regular geometric elements (such as lines and right angles) in the region, the structural regularity quantification result is obtained and normalized into a structural regularity score.

[0045] Understandably, there are no strict time restrictions on the calculation of statistical significance indicators and structural regularity scores.

[0046] Furthermore, the quantitative results from the two dimensions mentioned above are fused. Since both high spatial heterogeneity and low structural regularity point to the natural background, the statistical significance index and the structural regularity score (or its inverse value) are weighted and summed using preset weight coefficients or learnable attention weights to ultimately generate a continuous natural background sensitivity function.

[0047] The change detection method based on a multi-level information substrate provided in this application enhances the robustness of the natural background sensitivity function under complex and extreme imaging conditions by jointly verifying the spatial heterogeneity and structural regularity. This effectively avoids the problem of false changes being missed or misfiltered due to the failure of a single feature.

[0048] In an optional embodiment, the step of quantifying the spatial heterogeneity of the standardized input data to obtain a statistical significance index includes: Extract features from the local neighborhood, using the center pixel as the core. Calculate the feature mean and feature variance within the local neighborhood; A statistical significance index is determined based on the standardized distance between the pixel features of the center pixel and the mean of the features.

[0049] Here, the center pixel is a specific pixel used as the calculation benchmark and evaluation object when performing local feature calculations. The local neighborhood is a specific range extending outwards in space from the center pixel, such as a 3×3 or 5×5 sliding window. The feature mean is the arithmetic mean of the feature values ​​of all pixels within the local neighborhood, reflecting the background baseline level of that local area. The feature variance is the average of the squares of the deviations of the feature values ​​of all pixels within the local neighborhood from their mean. A larger variance indicates a more complex local texture, such as a dense forest; a smaller variance indicates a smoother local texture, such as a water surface or building roof. The standardized distance is the difference between the feature value of the center pixel and the feature mean of the local neighborhood, divided by the square root of the local feature variance, i.e., the standard deviation.

[0050] In this embodiment, a pixel in the standardized input data is taken as the center pixel, and a sliding window of fixed size is used to extract the feature vectors of all pixels in its local neighborhood; for the extracted local neighborhood features, the mean and variance of the feature values ​​within the window are calculated; and the feature value of the center pixel is calculated. F With local mean μ The difference, and using local standard deviation σNormalizing this difference, the standardized distance becomes the statistical significance index. If the absolute value of this index is large, it indicates that the features of the center pixel deviate significantly from its local background, exhibiting extremely high spatial heterogeneity; if the index is close to 0, it indicates that the center pixel is highly consistent with the local background.

[0051] In practical implementation, natural backgrounds in the 64-dimensional feature space typically exhibit textured but irregular, variance-based but statistically stable characteristics. In contrast, man-made structures possess extremely strong directional consistency. (Defining the local neighborhood...) Characteristic variance By calculating the center pixel Its neighborhood mean Standardized distance, construct statistical significance index : ; in, It is a non-zero constant. This index can effectively identify outliers within a local area. Under a uniform natural background, It remains at a low level; however, it will surge significantly once there is structured artificial interference.

[0052] The method for detecting changes under multiple imaging conditions based on a multi-level information substrate provided in this application standardizes the data by using the local standard deviation as the denominator. The scale for measuring heterogeneity is dynamic. In the shadow area, although the absolute difference is small, if the texture of the area is complex, the local standard deviation is also small. The final standardized distance can still accurately reflect its heterogeneity, thereby achieving adaptive normalization of local contrast. This allows the statistical significance index to be completely freed from the constraints of absolute illumination intensity and global contrast, and can stably and accurately quantify the spatial heterogeneity of the natural background under different imaging conditions.

[0053] In an optional embodiment, the step of quantifying the structural regularity of the standardized input data to obtain a structural regularity score includes: Perform a Fast Fourier Transform on the local neighborhood to obtain the spectral energy distribution; The structural regularity score is determined based on the proportion of high-frequency energy.

[0054] Here, the high-frequency energy percentage is the ratio of the total energy of the high-frequency portion in a local neighborhood spectrum to the total energy of the entire spectrum. Man-made structures, due to their sharp edges and regular geometric lines, exhibit significant high-frequency energy concentration in the frequency domain, while the edges of natural backgrounds are usually random, disordered, and blurred, with their energy concentrated in low frequencies or exhibiting a uniform broadband distribution.

[0055] In this embodiment, a local neighborhood feature matrix with the center pixel as the core is extracted, and a two-dimensional fast Fourier transform is applied to it. After the transformation, the pixel matrix in the spatial domain is converted into a complex matrix in the frequency domain, and then its amplitude spectrum or power spectrum is calculated to obtain the spectral energy distribution of the local neighborhood. In the obtained spectral energy distribution, according to the preset frequency division standard, such as taking the center low-frequency point as the center, a circular or rectangular area with a certain radius is set as the low-frequency area, and the rest is the high-frequency area. The total energy of the high-frequency area and the total energy of the entire frequency band are calculated respectively, and the ratio of the two is obtained, that is, the high-frequency energy ratio. The calculated high-frequency energy ratio is normalized or directly used as the structural regularity score. Since man-made facilities have sharp regular edges, their high-frequency energy ratio is significantly higher than that of the natural background. Therefore, this score can accurately characterize the structural regularity of the local area.

[0056] In practice, natural background energy is mainly concentrated in the low-frequency band, and its spectral distribution is isotropic and irregular; while man-made features and key targets exhibit significant high-frequency components and energy concentration in specific directions in the frequency domain. A Fast Fourier Transform (FFT) is performed on local feature blocks to obtain their spectral energy distribution. ,in It is based on pixels local neighborhood centered on Feature blocks. Define structural regularity scoring. : ; in, It is high-frequency energy, representing the energy components in an image that change dramatically and have distinct edges, textures, or structured geometry; This represents the total energy, signifying the sum of the overall energy of the local feature block in the frequency domain. The joint decision-making logic ultimately constructs a natural background sensitivity function through weight fusion. ; in, and These are weighting coefficients used to adjust statistical significance indicators. and structural regularity scoring Its importance in the final decision-making process It is the natural background sensitivity function.

[0057] Low-value areas are identified as natural backgrounds with high randomness and low structure (such as forests and grasslands), and are given lower weights in anomaly detection to suppress seasonal spurious changes. High-value areas are identified as non-natural backgrounds (structured facilities or isolated targets) and are designated as key areas of interest for subsequent branch processing.

[0058] The multi-imaging-condition change detection method based on a multi-level information substrate provided in this application transfers feature extraction to the frequency domain. The Fourier transform has the characteristics of energy conservation and global distribution. Even if the local neighborhood is in shadow, causing the overall brightness to decrease, the relative proportion of high-frequency energy representing regular geometric edges in the total energy will not change fundamentally. By calculating the proportion, it is immune to interference from global illumination changes, shadow occlusion, and differences in surface color of ground objects, so that the structural regularity score remains stable and reliable under extreme imaging conditions.

[0059] In an optional embodiment, filtering the remote sensing image based on the natural background sensitivity function to obtain a deep semantic feature map includes: The standardized input data is nonlinearly mapped to the vector space of the natural background sensitivity function to obtain a deep semantic feature map.

[0060] In this embodiment of the application, the 64-dimensional features are nonlinearly mapped to the natural background basis vector space using the above modeling results. By retaining low-frequency components and filtering out statistical uncertainties, the complex background information is compressed into a set of robust scalar descriptors, providing a clean input benchmark for subsequent calculation of cosine similarity.

[0061] Specifically, a nonlinear mapping function is constructed using the natural background sensitivity function or its high-dimensional feature representation as a condition or guiding signal. This function can be a multilayer perceptron (MLP), a convolutional neural network (CNN), or an attention mechanism module. Standardized input data is then fed into this mapping function. Based on the environmental interference context provided by the natural background sensitivity function, the mapping function nonlinearly distorts, compresses, or redirects the features of the input data. The output of the mapping is the deep semantic feature map. In this output space, the features of the natural background region are nonlinearly pushed to a low-energy or low-weight manifold, while the features of man-made facilities and dynamic targets are preserved on a high-energy or high-discriminative manifold.

[0062] The change detection method under multiple imaging conditions based on a multi-level information base provided in this application projects the input features into a decoupled high-dimensional manifold space, achieving complete separation between pseudo-changes in the natural background and changes in real ground features, and significantly improving the purity of the deep semantic feature map and the recall rate of subsequent change detection.

[0063] In an optional embodiment, the artificial background branch generates a geometric descriptor based on the following steps: For the deep semantic feature map, the partial derivatives in the horizontal and vertical directions are calculated respectively to obtain the derivative field; Multi-channel fusion is performed on each pixel position in the derivative field to obtain the structure tensor; Eigenvalue decomposition is performed on the structure tensor at each pixel location to obtain the principal energy and secondary energy. Based on the principal energy and the secondary energy, structural invariants are constructed; By fusing structural invariants at different scales, geometric descriptors are obtained.

[0064] In this embodiment of the application, for fixed facilities with strong edge and right-angle features such as buildings, roads, and runways, a stable modeling of the spatial topology of ground features is achieved by constructing a multi-scale structural tensor and extracting its geometric invariants. This method can effectively eliminate illumination fluctuations and seasonal spectral changes, and achieve robust detection of structural anomalies.

[0065] Here, the derivative field is a two-dimensional vector field composed of horizontal and vertical partial derivatives, comprehensively depicting the variation trend of each pixel in the feature map within the local space. The principal energy and secondary energy are the larger and smaller eigenvalues ​​obtained after the structure tensor decomposition. The principal energy represents the energy in the dominant direction (such as the edge of a building), and the secondary energy represents the energy perpendicular to the dominant direction.

[0066] In this embodiment, a deep semantic feature map is received, and the partial derivatives of the feature map in the horizontal and vertical directions are calculated to construct a complete derivative field. For each pixel in the derivative field, its horizontal and vertical partial derivatives are combined to calculate the gradient outer product matrix. To enhance robustness and integrate multi-channel information, the outer product matrix is ​​weighted and smoothed in the local neighborhood, such as Gaussian smoothing, to finally obtain the structure tensor at each pixel location. Eigenvalue decomposition is performed on the structure tensor at each pixel location to obtain two eigenvalues. The larger eigenvalue is denoted as the principal energy λ1, and the smaller eigenvalue is denoted as the secondary energy λ2. Geometric features that are invariant to rotation and translation are constructed using λ1 and λ2, for example, by calculating (λ1-λ2). 2 λ1 represents the strength of linear or angular structures, or λ2 represents the overall edge energy, thus obtaining structural invariants. The above steps are repeated at different scales (such as through image pyramids or convolutions with different dilation rates) to extract multi-scale structural invariants. These multi-scale features are then spliced ​​together or weighted and fused through an attention mechanism to finally generate a geometric descriptor with rich geometric details.

[0067] In the specific implementation process, deep semantic feature maps are extracted from the 64-dimensional spatiotemporal embedding base. In the spatial dimension, the partial derivatives in the horizontal and vertical directions are calculated for each channel: ; The gradient operator transforms the original semantic information into a derivative field that represents the boundaries and texture orientation of ground features, providing a basic physical quantity for subsequent geometric structure modeling.

[0068] To accommodate features at different scales (such as the thin edges of a fence and the broad outline of a factory building), a multi-scale Gaussian kernel is introduced. For each pixel location, construct a symmetric positive semi-definite matrix using the channel inner product (summed along the 64-dimensional C-axis), resulting in the structure tensor: ; This tensor describes the energy distribution of the gradient within a local neighborhood. Compared to single-channel gradients, the multi-channel fused structural tensor can more robustly lock the geometric framework of fixed facilities and effectively suppress noise interference in specific bands.

[0069] For each pixel position Perform eigenvalue decomposition to obtain principal energy. With secondary energy To avoid numerical instability caused by directly using eigenvalues, we construct structural invariants that are highly robust to scale and illumination. : ; Corner and edge recognition: If This represents a strongly directional edge, such as the eaves of a house or the edge of a runway; if , representing key geometric corner points. It can effectively filter out unstructured spectral fluctuations. As long as the geometric proportions of the building remain unchanged, the descriptor remains constant even if the image brightness shifts due to cloud shadows or the solar altitude angle.

[0070] To achieve equal-weighted monitoring of facilities of different specifications in complex scenarios, different scales are integrated. The following information is used to obtain the geometric descriptor. : ; In this process, the smaller Used to capture fine details of structures, such as window frames and railings, while larger ones... This is used to lock in the macroscopic architectural configuration, thereby generating a geometric descriptor that takes into account both details and the overall structure.

[0071] Comparison of geometric descriptors generated in two phases and Calculate its time series differences : , ; Set dynamic threshold ,like If the limit is exceeded, it will be judged as Geometric changes indicate anomalies in fixed structures. This logic ensures that monitoring focuses on deformation of ground features rather than color changes. Once a building collapses, an addition is made, or a road is damaged, the intrinsic distribution of the tensor will undergo an irreversible shift, enabling high-precision early warning of structural anomalies.

[0072] The multi-imaging condition change detection method based on a multi-level information base provided in this application not only calculates the gradient of the structural tensor, but also transforms isolated pixel-level gradients into regional-level statistics reflecting the overall trend of the local neighborhood through multi-channel fusion and local smoothing. This filters out high-frequency random noise, enabling the extracted geometric descriptors to stably reflect the true outline and main structure of the building, significantly reducing the misjudgment rate of structural anomalies caused by imaging quality fluctuations. Eigenvalue decomposition of the structural tensor is performed. Since eigenvalues ​​are inherent properties of the matrix and do not change with the rotation of the coordinate system, the geometric descriptors have rotation invariance. Regardless of how the target rotates or slightly misaligns in the preceding and following phases, its geometric descriptors can maintain a high degree of consistency, thus ensuring extremely high accuracy in structural anomaly detection. Finally, structural invariants at different scales are fused, enriching the information dimensions of the structural descriptors.

[0073] In an optional embodiment, the foreground target branch generates a target feature fingerprint based on the following steps: Calculate the Euclidean distance between each pixel in the standardized input data and the mean of its local neighborhood, obtain the saliency score of each pixel, and then construct a saliency map. The saliency map is binarized based on a dynamic threshold to obtain a set of salient points; Density clustering is performed on the salient point set to obtain multiple target clusters; Feature extraction is performed on each of the target clusters to obtain target feature fingerprints.

[0074] In this embodiment of the application, for small but significant dynamic targets such as ships, vehicles, and aircraft, unsupervised saliency modeling and density clustering algorithms are used to aggregate discrete feature responses into target fingerprints with geometric centers as anchor points, effectively solving the problem of decreased anomaly detection accuracy caused by target displacement or background clutter interference.

[0075] Here, each pixel in the standardized input data is traversed, and the Euclidean distance between its feature vector and the mean of the local neighborhood features centered at that point is calculated. This distance reflects the degree to which the point deviates from the background and is used as a saliency score. The saliency scores of all pixels are arranged to generate a saliency map of the entire image. For the generated saliency map, a dynamic threshold is calculated, such as an adaptive threshold based on the current image saliency distribution. Pixels with scores higher than the dynamic threshold are retained, while those lower than the threshold are set to zero, thereby extracting a discrete set of salient points. All points in the salient point set are input into a density clustering algorithm. The algorithm automatically divides clustered points into clusters based on the spatial distance and neighborhood density between points, while automatically removing spatially isolated noise points. For each independent target cluster, the original features of all pixels in the standardized input data within the cluster are extracted, and a fixed-dimensional feature vector is generated through aggregation operations (such as mean calculation, max pooling, or attention mechanisms), which is the target feature fingerprint of the dynamic target.

[0076] In the specific implementation process, 64-dimensional spatiotemporal embedding features are utilized. Calculate each pixel Its local neighborhood mean Spatial heterogeneity, constructing unsupervised significance scores : ; This model utilizes the incoherence of key targets in remote sensing backgrounds (such as uniform water surfaces and flat runways). The background region has high autocorrelation, causing S(p) to tend to extremely low values, while the target region achieves energy focusing due to strong contrast, thus automatically setting the background response to zero without relying on annotation.

[0077] A dynamic threshold T is set to binarize the saliency map and extract the set of high-response pixels. : ; This step compresses the dense 64-dimensional image features into a sparse, salient point cloud that characterizes potential target regions, greatly reducing computational redundancy in subsequent processing.

[0078] For salient point sets Execute the density clustering algorithm (DBSCAN) with noise handling capabilities: ; The neighborhood search radius is This is the minimum number of points required for clustering. The algorithm does not require a preset target number and can automatically determine the number and geometric boundaries of targets based on the point cloud density. It can also effectively identify and remove scattered sea surface spray or land noise interference.

[0079] For each target cluster generated by clustering Feature extraction is performed to obtain the target feature fingerprint.

[0080] Comparison of target sets in two time phases and The target pairs are associated through nearest neighbor search, and a score for symmetric change is calculated. : , ; If the score for symmetrical change is greater than the threshold If the cosine similarity of the matched pairs is low, or if a certain anchor point is completely mismatched in another time phase and has no corresponding target, then it is judged as... As the landscape changes, targets may be added, removed, or replaced. This logic of first identifying targets and then comparing them enables the model to recognize movements such as ships leaving port and vehicles gathering with a much higher accuracy than traditional pixel-level change detection.

[0081] The multi-imaging condition change detection method based on a multi-level information substrate provided in this application calculates the local saliency score of each pixel and then performs binarization using a dynamic threshold. The dynamic threshold can adaptively adjust the cutoff point according to the saliency distribution of the current image or local region, reducing the false negative rate. Density clustering is used to generate target clusters, so that the extraction of target feature fingerprints completely fits the real physical boundary of the target. Even if the target is severely occluded or has an irregular shape, the system can still extract pure target features, avoiding background clutter from contaminating the fingerprint and improving the accuracy of cross-temporal target matching.

[0082] In an optional embodiment, the step of extracting features from each of the target clusters to obtain target feature fingerprints includes: For each target cluster, the core feature vector is extracted using saliency-weighted attention pooling with the geometric centroid as the central anchor point to obtain the target feature fingerprint.

[0083] In this embodiment, for each independent target cluster generated in the preceding steps, the system first calculates the mean of the spatial coordinates of all pixels within the cluster to determine the geometric centroid of the target and sets it as the central anchor point. It then calculates the spatial distance from each pixel within the target cluster to the central anchor point and, combined with the saliency score obtained in the preceding steps, calculates an attention weight. Generally, pixels closer to the centroid and with higher saliency have a greater weight, while pixels located at the target edge or with lower saliency have their weights attenuated. The system then performs element-wise multiplication (weighting) of the feature vectors of each pixel within the target cluster with their corresponding attention weights, and then sums or averages all the weighted feature vectors. Through this weighted aggregation, edge clutter is filtered out, and a fixed-dimensional core feature vector, i.e., the target feature fingerprint of the target, is finally output.

[0084] In the specific implementation process, for each target cluster generated by clustering Calculate its geometric centroid as the central anchor. Using this anchor point as the center, saliency-weighted attention pooling is employed to extract the core feature vector. : ; This central anchoring mechanism condenses dispersed target features into a fixed-length target signature. Even if the target shifts position between two time phases, its signature vector remains consistent. It has translation invariance and can still maintain semantic consistency.

[0085] The method for detecting changes under multiple imaging conditions based on a multi-level information substrate provided in this application ensures the purity of the generated target feature fingerprint through saliency weighting, and reduces the false matching rate caused by background clutter interference when matching across time.

[0086] The following describes the change detection device under multiple imaging conditions based on a multi-level information substrate provided in the embodiments of this application. The change detection device under multiple imaging conditions based on a multi-level information substrate described below can be referred to in correspondence with the change detection method under multiple imaging conditions based on a multi-level information substrate described above.

[0087] Figure 2 This is a schematic diagram of the structure of the multi-imaging condition change detection device based on a multi-level information substrate provided in this application, as shown below. Figure 2 As shown, the multi-imaging condition change detection device based on a multi-level information substrate can include, but is not limited to: Data preprocessing module 210 is used to: uniformly encode remote sensing images onto a spatiotemporal embedding base of preset dimensions to obtain standardized input data; The branch parallel module 220 is used to: input the standardized input data into the natural background branch, the artificial background branch and the foreground target branch respectively, and obtain the natural background sensitivity function, the geometric descriptor and the target feature fingerprint respectively. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. Image filtering module 230 is used to: filter the remote sensing image based on the natural background sensitivity function to obtain a deep semantic feature map; The first detection module 240 is used to: compare the geometric descriptors of the same geographical location at different times with the deep semantic feature map; if the difference exceeds the structural threshold, then determine that there is a structural anomaly in the geographical location. The second detection module 250 is used to: calculate the similarity and spatial position offset of the target feature fingerprint in two consecutive time phases; if the similarity is lower than the similarity threshold or there is a mismatch in one time phase, then it is determined that there is an increase or decrease in the target.

[0088] It should be noted that the multi-imaging condition change detection device based on a multi-level information substrate provided in this application embodiment can execute the multi-imaging condition change detection method based on a multi-level information substrate described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0089] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include: a processor 310, a communications interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communications interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a multi-imaging condition change detection method based on a multi-level information substrate, the method including: The remote sensing images are uniformly encoded onto a spatiotemporal embedding base of a preset dimension to obtain standardized input data; The standardized input data is input into the natural background branch, the artificial background branch, and the foreground target branch, respectively, to obtain the natural background sensitivity function, the geometric descriptor, and the target feature fingerprint. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. The remote sensing image is filtered based on the natural background sensitivity function to obtain a deep semantic feature map; For the deep semantic feature map, compare the geometric descriptors of the same geographical location at different times. If the difference exceeds the structural threshold, it is determined that there is a structural change in the geographical location. Calculate the similarity and spatial offset of the target feature fingerprint in two consecutive time phases. If the similarity is lower than the similarity threshold or there is a mismatch in one time phase, it is determined that there is an increase or decrease in the target.

[0090] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the multi-imaging condition change detection method based on a multi-level information substrate provided by the above methods. The method includes: The remote sensing images are uniformly encoded onto a spatiotemporal embedding base of a preset dimension to obtain standardized input data; The standardized input data is input into the natural background branch, the artificial background branch, and the foreground target branch, respectively, to obtain the natural background sensitivity function, the geometric descriptor, and the target feature fingerprint. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. The remote sensing image is filtered based on the natural background sensitivity function to obtain a deep semantic feature map; For the deep semantic feature map, compare the geometric descriptors of the same geographical location at different times. If the difference exceeds the structural threshold, it is determined that there is a structural change in the geographical location. Calculate the similarity and spatial offset of the target feature fingerprint in two consecutive time phases. If the similarity is lower than the similarity threshold or there is a mismatch in one time phase, it is determined that there is an increase or decrease in the target.

[0092] Furthermore, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-imaging condition change detection method based on a multi-level information substrate provided by the above methods, the method comprising: The remote sensing images are uniformly encoded onto a spatiotemporal embedding base of a preset dimension to obtain standardized input data; The standardized input data is input into the natural background branch, the artificial background branch, and the foreground target branch, respectively, to obtain the natural background sensitivity function, the geometric descriptor, and the target feature fingerprint. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. The remote sensing image is filtered based on the natural background sensitivity function to obtain a deep semantic feature map; For the deep semantic feature map, compare the geometric descriptors of the same geographical location at different times. If the difference exceeds the structural threshold, it is determined that there is a structural change in the geographical location. Calculate the similarity and spatial offset of the target feature fingerprint in two consecutive time phases. If the similarity is lower than the similarity threshold or there is a mismatch in one time phase, it is determined that there is an increase or decrease in the target.

[0093] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting changes under multiple imaging conditions based on a multi-level information substrate, characterized in that, include: The remote sensing images are uniformly encoded onto a spatiotemporal embedding base of a preset dimension to obtain standardized input data; The standardized input data is input into the natural background branch, the artificial background branch, and the foreground target branch, respectively, to obtain the natural background sensitivity function, the geometric descriptor, and the target feature fingerprint. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. The remote sensing image is filtered based on the natural background sensitivity function to obtain a deep semantic feature map; For the deep semantic feature map, compare the geometric descriptors of the same geographical location at different times. If the difference exceeds the structural threshold, it is determined that there is a structural change in the geographical location. Calculate the similarity and spatial offset of the target feature fingerprint in two consecutive time phases. If the similarity is lower than the similarity threshold or there is a mismatch in one time phase, it is determined that there is an increase or decrease in the target.

2. The method for detecting changes under multiple imaging conditions based on a multi-level information substrate according to claim 1, characterized in that, The natural background branch generates the natural background sensitivity function based on the following steps: Spatial heterogeneity quantification is performed on the standardized input data to obtain statistical significance indicators; The standardized input data is subjected to structural regularity quantification to obtain a structural regularity score; The natural background sensitivity function is obtained by weighted summation of the statistical significance index and the structural regularity score.

3. The method for detecting changes under multiple imaging conditions based on a multi-level information substrate according to claim 2, characterized in that, The step of quantifying the spatial heterogeneity of the standardized input data to obtain statistical significance indicators includes: Extract features from the local neighborhood, using the center pixel as the core. Calculate the feature mean and feature variance within the local neighborhood; A statistical significance index is determined based on the standardized distance between the pixel features of the center pixel and the mean of the features.

4. The method for detecting changes under multiple imaging conditions based on a multi-level information substrate according to claim 3, characterized in that, The step of quantifying the structural regularity of the standardized input data to obtain a structural regularity score includes: Perform a Fast Fourier Transform on the local neighborhood to obtain the spectral energy distribution; The structural regularity score is determined based on the proportion of high-frequency energy.

5. The method for detecting changes under multiple imaging conditions based on a multi-level information substrate according to claim 4, characterized in that, The step of filtering the remote sensing image based on the natural background sensitivity function to obtain a deep semantic feature map includes: The standardized input data is nonlinearly mapped to the vector space of the natural background sensitivity function to obtain a deep semantic feature map.

6. The method for detecting changes under multiple imaging conditions based on a multi-level information substrate according to claim 1, characterized in that, The artificial background branch generates a geometric descriptor based on the following steps: For the deep semantic feature map, the partial derivatives in the horizontal and vertical directions are calculated respectively to obtain the derivative field; Multi-channel fusion is performed on each pixel position in the derivative field to obtain the structure tensor; Eigenvalue decomposition is performed on the structure tensor at each pixel location to obtain the principal energy and secondary energy. Based on the principal energy and the secondary energy, structural invariants are constructed; By fusing structural invariants at different scales, geometric descriptors are obtained.

7. The method for detecting changes under multiple imaging conditions based on a multi-level information substrate according to claim 1, characterized in that, The foreground target branch generates a target feature fingerprint based on the following steps: Calculate the Euclidean distance between each pixel in the standardized input data and the mean of its local neighborhood, obtain the saliency score of each pixel, and then construct a saliency map. The saliency map is binarized based on a dynamic threshold to obtain a set of salient points; Density clustering is performed on the salient point set to obtain multiple target clusters; Feature extraction is performed on each of the target clusters to obtain target feature fingerprints.

8. The method for detecting changes under multiple imaging conditions based on a multi-level information substrate according to claim 7, characterized in that, The step of extracting features from each of the target clusters to obtain target feature fingerprints includes: For each target cluster, the core feature vector is extracted using saliency-weighted attention pooling with the geometric centroid as the central anchor point to obtain the target feature fingerprint.

9. A change detection device under multiple imaging conditions based on a multi-level information substrate, characterized in that, include: The data preprocessing module is used to: uniformly encode remote sensing images onto a spatiotemporal embedding base of preset dimensions to obtain standardized input data; The branch parallel module is used to: input the standardized input data into the natural background branch, the artificial background branch and the foreground target branch respectively, and obtain the natural background sensitivity function, the geometric descriptor and the target feature fingerprint respectively. The natural background sensitivity function is used to mark the low-weight pseudo-change region caused by environmental interference, the geometric descriptor is used to characterize the geometric structure of the artificial facility, and the target feature fingerprint is used for semantic identification and matching of dynamic targets. The image filtering module is used to: filter the remote sensing image based on the natural background sensitivity function to obtain a deep semantic feature map; The first detection module is used to: compare the geometric descriptors of the same geographical location at different times with the deep semantic feature map; if the difference exceeds the structural threshold, then determine that there is a structural anomaly in the geographical location. The second detection module is used to: calculate the similarity and spatial offset of the target feature fingerprint in two consecutive time phases; if the similarity is lower than the similarity threshold or there is a mismatch in one time phase, then it is determined that there is an increase or decrease in the target.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for detecting changes under multiple imaging conditions based on a multi-level information substrate as described in any one of claims 1 to 8.

Images