Multi-source remote sensing image zero sample change detection method

Abstract

The invention discloses a multi-source remote sensing image zero sample change detection method, and relates to the technical field of remote sensing image processing, and the method comprises the following steps: obtaining remote sensing images collected by two or more remote sensing sensors at different time points in the same geographic area, the image types including optical images and radar images; preprocessing each source image, unifying the spatial resolution and the registration precision, and denoising and standardizing the image; according to the method, the cross-modal shared semantic embedding space is constructed and unsupervised comparative learning is introduced, so that the semantic consistency of the multi-source remote sensing image is effectively improved, and the change recognition capability of the model under the zero sample condition is enhanced; and meanwhile, a difference fusion calculation and structure consistency constraint module is adopted, so that the boundary judgment precision of a change region and the overall structure consistency are improved, and the accuracy and stability of a detection result are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a zero-sample change detection method for multi-source remote sensing images. Background Art

[0002] Zero-sample change detection using multi-source remote sensing imagery refers to a method for automatically identifying and detecting changes in the same area over time using image data from different remote sensing sensors (such as optical, radar, and infrared), without providing labeled samples for the target area (i.e., zero samples). This technology transcends the limitations of traditional change detection, which relies on supervised learning with large amounts of labeled data. It typically employs artificial intelligence methods such as transfer learning, meta-learning, or semantic embedding to fuse cross-modal image information and extract change patterns. It is suitable for scenarios requiring rapid response and where sample acquisition is difficult, such as disaster assessment, environmental monitoring, and urban expansion analysis.

[0003] Existing technologies have the following shortcomings: In the existing zero-sample change detection process of multi-source remote sensing images, the semantic inconsistency problem between cross-source images is a key challenge. Due to the significant differences in the imaging mechanisms of different remote sensing sensors, for example, optical images reflect the spectral information of the ground object, while radar images reflect more of the structural characteristics of the ground object, the same ground object may appear significantly different in different images. This semantic difference makes it difficult for the model to accurately establish cross-source correspondences without sample supervision, resulting in misjudgment of change detection results. This manifests as areas with no change being mistakenly identified as changed, or areas with real changes being omitted, seriously affecting the reliability of detection and the practicality of the application.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present invention is to provide a zero-sample change detection method for multi-source remote sensing images. By constructing a cross-modal shared semantic embedding space and introducing unsupervised comparative learning, the semantic consistency of multi-source remote sensing images is effectively improved, and the change recognition ability of the model under zero-sample conditions is enhanced; at the same time, the difference fusion calculation and structural consistency constraint module are adopted to improve the accuracy of change area boundary judgment and overall structural coherence, significantly improving the accuracy and stability of the detection results, so as to solve the problems in the above-mentioned background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solution: a zero-sample change detection method for multi-source remote sensing images, comprising the following steps: Acquire remote sensing images collected by two or more remote sensing sensors over the same geographical area at different time points. Image types include optical images and radar images. Pre-process each source image to unify the spatial resolution and registration accuracy, and perform denoising and standardization on the image; Construct a cross-modal feature encoding network to extract features from different source images separately, and introduce a shared semantic embedding space to enhance the semantic alignment ability between different modalities; By optimizing feature representation through unsupervised contrastive learning, the embedding vectors of the same feature in different images maintain high similarity, thus improving semantic consistency. Calculate the difference between temporal image features in a shared semantic space and determine the changed area based on a set threshold; The change detection result graph is output, and the structural consistency constraint module is used to further optimize the boundary judgment accuracy and improve the reliability of change area identification.

[0007] Preferably, the acquired multi-source remote sensing images must undergo precise spatiotemporal registration of the same geographical area. The specific steps are as follows: The global positioning system data and attitude information are used for geometric correction, and then a registration algorithm based on maximizing mutual information is used to accurately align the images so that the pixels from different remote sensing sources have a one-to-one correspondence in geographic space. Perform cloud detection and occlusion removal on images to ensure the spatial integrity and comparability of input data; In terms of image time point selection, the acquisition time difference is controlled within no more than three months to maximize the reality of changes in the structure of objects in the target area.

[0008] Preferably, the image preprocessing step includes enhancing spectral and structural features using a dual-domain normalization strategy for multi-source data; When processing optical images, histogram equalization and local contrast enhancement algorithms are introduced to improve the clarity of object boundaries. For radar images, logarithmic transformation and coherent filtering methods are used to eliminate speckle noise and enhance the structural texture of objects. All images are uniformly converted into a pseudo-color format with 3 channels to ensure consistent input to the subsequent deep neural network processing module and prevent degradation of model performance due to channel mismatch.

[0009] Preferably, the cross-modal feature encoding network uses a parallel convolutional architecture to construct two sub-channels, corresponding to optical and radar image inputs respectively, and connects normalization and ReLU activation functions after the output of each convolution layer to avoid gradient vanishing and enhance nonlinear feature extraction capabilities; To enhance the alignment capability of the cross-modal shared semantic space, a residual connection mechanism and a channel attention module are introduced into the network, enabling the model to perceive high-level semantic features while capturing local texture details. A shared transformation module is connected at the end of each sub-channel to map high-dimensional features to the same embedding space, laying the foundation for subsequent unsupervised contrastive learning.

[0010] Preferably, the unsupervised contrastive learning mechanism adopts a loss optimization strategy based on positive and negative sample pairs. The specific steps are as follows: Randomly extract paired samples from the feature vectors of different modalities at the same geographical location as positive sample pairs; Features are collected from non-corresponding positions as negative sample pairs; finally, by optimizing the contrast loss function based on cosine distance, the distance between positive sample pairs is minimized and the distance between negative sample pairs is maximized; To prevent the model from falling into local optimality, a momentum encoder mechanism is introduced during the training process to dynamically update the positive sample representation to ensure the stability and discriminability of the embedding space.

[0011] Preferably, when identifying the changed area in the shared semantic space, a dual threshold strategy is used for pixel-level judgment, including the following methods: Calculate the feature difference between any pair of embedding vectors at different times; The upper and lower limits of the change threshold are set, and pixels with a difference below the lower limit are marked as "unchanged", and pixels with a difference above the upper limit are marked as "changed". Pixels in the middle area are further processed by the edge enhancement model, and their final labels are determined by the gradient boosting algorithm.

[0012] Preferably, the difference between the embedding vectors is calculated using a multi-dimensional fusion algorithm. The specific steps are as follows: Assume that the shared semantic space embedding vectors extracted from the two moments are and , calculate the Euclidean distance between them as the basic change metric, the calculation expression is as follows: , where is the Euclidean difference between the embedded features, It is the two-norm of the vector, also known as the Euclidean distance, which measures the geometric distance between two embedded vectors; Calculate the cosine similarity between vectors and convert it into a difference factor as the basis for judging directional differences. The calculation expression is as follows: , where and Represent the shared semantic space embedding vectors and The Euclidean length of is the cosine "difference"; Calculate the average difference between the two embedding vectors in each dimension as the average trend of local changes. The calculation expression is as follows: , where It is to take the absolute value of the channel difference and then find the average. is the average change amplitude under all channels; Fusion 、 as well as , forming the final difference score, the formula is as follows: , where is the Euclidean difference between the embedded features The weighting coefficient of is the cosine "difference" The weighting coefficient of is the average change amplitude under all channels The weighting coefficient of is the final difference score.

[0013] Preferably, after outputting the change detection result graph, the prediction graph is further post-processed using a structural consistency constraint module; The structural consistency constraint module combines structural similarity indicators with morphological processing methods to reconstruct and merge all adjacent changed areas in the detection image. The specific process includes calculating the structural consistency score of each changed segment and its neighborhood. Those with scores below the set threshold are eliminated or edge-reduced. At the same time, boundary discontinuity areas are repaired through dilation and erosion operations to ensure the continuity and morphological rationality of the changed areas.

[0014] Preferably, the structural consistency constraint module adopts a structural confidence scoring mechanism constructed by combining edge gradient direction and local contrast entropy to perform boundary optimization processing on the change detection result graph. The specific steps are as follows: Implement sliding window analysis on the output change detection map, and set the center pixel of each window to be , calculate the local contrast entropy of the predicted change area in the current window to measure the information complexity of its pixel distribution. The calculation expression is as follows: , where It is based on pixel points in the change detection map The contrast entropy within the local sliding window centered at Is the first The proportion of class pixels, is the number of pixel categories; Extract the edge map from the original multi-source image, and calculate the gradient direction consistency factor between the edge direction vector of the predicted change map and the direction vector of the original edge map. The calculation expression is as follows: , where is the pixel in the predicted change detection map The gradient direction angle of the window, is the true edge direction angle extracted from the corresponding position in the original multi-source remote sensing image, is the gradient direction consistency factor; joint and , construct a structural confidence scoring function to evaluate the credibility of each window change prediction. The calculation expression is as follows: , where is the structural complexity weight, is the boundary direction consistency weight, is the structure confidence score.

[0015] In the above technical solution, the technical effects and advantages provided by the present invention are: The present invention significantly improves the semantic consistency between multi-source remote sensing images by introducing a cross-modal shared semantic embedding space and combining it with an unsupervised comparative learning mechanism, thereby effectively alleviating the semantic mismatch problem caused by differences in the imaging mechanisms of different sensors. By constructing a dual-channel feature encoding network, key semantic features in optical and radar images are extracted respectively, and mapped to a unified semantic space, so that even in the absence of labeled samples, the model can still accurately understand and match the manifestations of the same land object in different modalities. This method provides a stable and high-quality feature basis in the process of determining the changed area, greatly reducing the false positive rate and missed judgment rate, and improving the robustness and reliability of change detection.

[0016] The present invention further improves the boundary accuracy of the changed area and the coherence of the overall image structure through the collaborative optimization of the difference fusion calculation mechanism and the structural consistency constraint module. The multi-dimensional difference calculation formula that integrates Euclidean distance, cosine similarity and mean value difference enhances the model's ability to distinguish different types of changes; while the structural consistency module post-processes the detection results based on contrast entropy and gradient direction consistency, effectively eliminating misjudged areas with blurred boundaries and incomplete structures. This design significantly improves the accuracy and consistency of the final output image at the visual level, and is particularly suitable for remote sensing analysis tasks that require high structural integrity, such as the edges of urban buildings and water bodies. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0018] Figure 1 This is a method flow chart of a zero-sample change detection method for multi-source remote sensing images of the present invention. DETAILED DESCRIPTION

[0019] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0020] The present invention provides Figure 1 The method for zero-sample change detection of multi-source remote sensing images shown in FIG. 1 includes the following steps: Acquire remote sensing images collected by two or more remote sensing sensors over the same geographical area at different time points. Image types include optical images and radar images. The multi-source remote sensing images acquired must go through the precise spatiotemporal registration process of the same geographical area. The specific steps are as follows: The global positioning system data and attitude information are used for geometric correction, and then a registration algorithm based on maximizing mutual information is used to accurately align the images so that the pixels from different remote sensing sources have a one-to-one correspondence in geographic space. Cloud detection and occlusion area removal are performed on the images to ensure the spatial integrity and comparability of the input data; in terms of image time point selection, the acquisition time difference is controlled within no more than three months to maximize the reality of changes in the target area's ground structure.

[0021] This step ensures the consistency of the input basis for subsequent semantic alignment and change detection, and reduces the interference of non-surface change factors on the results.

[0022] The core function of this step is to rigorously align the spatial and temporal data of multi-source remote sensing images, while removing interfering elements such as cloud cover and shadows. This ensures that the input remote sensing images have high spatial consistency and temporal comparability. This process is one of the most fundamental and critical steps in the entire change detection system, and its quality directly determines the accuracy of subsequent feature extraction and change assessment.

[0023] For spatial registration, this step employs a multi-level alignment strategy to ensure image alignment with sub-pixel accuracy. Initial geometric correction is performed using the remote sensing image's inherent metadata (such as GPS coordinates, attitude angle information, and imaging orientation). This stage uses a projective transformation to unify all images into a common geographic coordinate system, forming the foundation for georeferenced images. This step ensures that all modal images are converted from their original acquisition coordinate system to a globally consistent geographic coordinate space, eliminating coordinate discrepancies for subsequent registration.

[0024] Next, based on the initial geometric registration, a local non-rigid registration algorithm based on mutual information (MIT) is implemented to further refine the pixel-level spatial correspondence between the images. This algorithm analyzes the similarity of grayscale statistics between different images within a local window and dynamically adjusts the registration parameters to maximize the structural match between images from different imaging mechanisms, such as optical and radar. This method is particularly effective in areas with dense buildings or complex terrain, effectively resolving significant differences in image texture and perspective, ensuring that all change analysis is performed with precise spatial correspondence.

[0025] For temporal matching, the system requires that the time difference between input image acquisitions be within three months to minimize interference from non-surface feature changes due to seasonal and climatic variations. Data with large time differences will be automatically removed through metadata analysis to ensure the timeliness and realism of the comparison basis.

[0026] This step also includes image cleaning, which primarily identifies and removes interfering areas, such as clouds, shadows, and dense fog, common in remote sensing images under natural conditions. This process utilizes a multi-channel cloud detection algorithm, combining multiple features such as NDVI (vegetation index), BRDF reflectance characteristics, and brightness differences to extract a highly reliable interfering area mask. Identified areas are subsequently ignored or filled with adjacent data to prevent them from affecting the effectiveness of change detection.

[0027] In summary, dependent claim 1 achieves comprehensive spatial, temporal, and structural unification of the input remote sensing imagery through rigorous geometric registration, temporal screening, and interference removal strategies. It lays a solid foundation for the effective integration of multi-source, heterogeneous remote sensing data and is a prerequisite for the successful implementation of the entire change detection method. Without the accuracy assurance of this step, any subsequent feature extraction, embedding comparison, or difference analysis may fail due to "foundation misalignment." Therefore, although this step is at the beginning of the method flow, it determines the upper limit of the overall system performance.

[0028] Pre-process each source image to unify the spatial resolution and registration accuracy, and perform denoising and standardization on the image; The image preprocessing steps include enhancing spectral and structural features using a dual-domain normalization strategy for multi-source data; When processing optical images, histogram equalization and local contrast enhancement algorithms are introduced to improve the clarity of object boundaries. For radar images, logarithmic transformation and coherent filtering methods are used to eliminate speckle noise and enhance the structural texture of objects. All images are uniformly converted into a pseudo-color format with 3 channels to ensure consistent input to the subsequent deep neural network processing module and prevent degradation of model performance due to channel mismatch.

[0029] This normalization strategy effectively retains the key features of various types of images and improves the input quality of cross-modal fusion.

[0030] This step aligns the spectral and structural features of multi-source remote sensing images through specific image preprocessing techniques, ensuring consistent visual representation when fed into the deep learning model. This process is crucial for zero-shot learning, as the model cannot rely on sample labels to correct for discrepancies and must rely on the consistency of the input data to extract reasonable features.

[0031] In practical implementation, histogram equalization is used to enhance the dynamic range of pixel distribution in optical images, thereby improving image contrast, particularly enhancing details in dark areas. Local contrast enhancement methods (such as adaptive histogram equalization (CLAHE)) are then used to further enhance local texture and edge representation, making the structure of objects more distinct. This processing significantly improves the recognition of elements such as vegetation, buildings, and roads in optical images.

[0032] Radar images often suffer from strong speckle noise. Logarithmic transformation is first used to compress the intensity dynamic range, preventing areas of strong reflection from overly dominating the image. Subsequently, a coherent filtering algorithm (such as the Lee filter or the Frost filter) is used to suppress the noise while preserving the geometric and structural information of the radar image as much as possible. This preprocessing approach is crucial for ensuring the stability and effectiveness of radar images.

[0033] Finally, to ensure consistent input to the deep neural network, all preprocessed images are converted to a three-channel pseudo-color format. For grayscale radar images, a three-channel pseudo-color image is generated by selecting specific texture dimensions or duplicating the main channel. This avoids model structural inconsistencies caused by mismatched input channel counts and is essential for improving system compatibility in real-world deployments.

[0034] In summary, this step achieves pre-alignment of optical and radar images at the feature level through differentiated but synergistic processing strategies, significantly enhancing the contrast and model perception capabilities of the subsequent feature encoding stage, and laying a solid foundation for highly robust change detection under zero-shot conditions.

[0035] Construct a cross-modal feature encoding network to extract features from different source images separately, and introduce a shared semantic embedding space to enhance the semantic alignment ability between different modalities; The cross-modal feature encoding network uses a parallel convolutional architecture to construct two sub-channels, corresponding to optical and radar image inputs respectively. Normalization and ReLU activation functions are applied to the output of each convolutional layer to avoid gradient vanishing and enhance nonlinear feature extraction capabilities. To enhance the alignment capability of the cross-modal shared semantic space, a residual connection mechanism and a channel attention module are introduced into the network, enabling the model to perceive high-level semantic features while capturing local texture details. A shared transformation module is connected at the end of each sub-channel to map high-dimensional features to the same embedding space, laying the foundation for subsequent unsupervised contrastive learning.

[0036] The core purpose of this step is to extract meaningful features from multi-source remote sensing imagery and project them into a unified shared semantic space, ensuring that the same features in different modalities have similar high-level semantic representations in this space. This mechanism directly addresses the most critical issue in zero-shot change detection: inter-modal semantic misalignment.

[0037] The specific implementation involves constructing two parallel feature extraction subnetworks to process optical and radar imagery, respectively. These two subnetworks are structurally symmetrical, with each layer consisting of convolutional units, normalization layers, and ReLU activation functions, forming a multi-layered feature extraction path that captures information at various levels, from edge to semantic. During the extraction process, a cross-layer residual connection mechanism is employed to avoid the vanishing gradient problem in deep networks while preserving low-level detail features, helping to preserve ground object details.

[0038] Furthermore, to enhance the network's responsiveness to key semantic features, a channel attention module is introduced. This module dynamically adjusts channel weights by globally analyzing the response to each channel's features, improving its responsiveness to channels that are sensitive to changes. For example, if a vegetation area exhibits significant color changes in optical images but significant texture changes in radar images, this module can assign higher weights to each channel, achieving modality-adaptive perception.

[0039] At the end of both sub-networks, a unified embedding mapping module projects the high-dimensional features output by both sources into a semantic space of the same dimension. This space is constructed based on a shared weight strategy, and through training and optimization, the embedding vectors corresponding to the same feature in different modalities are kept as close as possible. This structure effectively establishes a "translation mechanism" between modalities, laying the semantic foundation for subsequent unsupervised learning mechanisms.

[0040] By optimizing feature representation through unsupervised contrastive learning, the embedding vectors of the same feature in different images maintain high similarity, thus improving semantic consistency. The unsupervised contrastive learning mechanism adopts a loss optimization strategy based on positive and negative sample pairs. The specific steps are as follows: Randomly extract paired samples from the feature vectors of different modalities at the same geographical location as positive sample pairs; Features are collected from non-corresponding positions as negative sample pairs; finally, by optimizing the contrast loss function based on cosine distance, the distance between positive sample pairs is minimized and the distance between negative sample pairs is maximized; To prevent the model from falling into local optimality, a momentum encoder mechanism is introduced during the training process to dynamically update the positive sample representation to ensure the stability and discriminability of the embedding space.

[0041] This mechanism significantly improves the consistent expression of images of different modalities in the semantic embedding space.

[0042] The core function of this step is to construct "positive and negative sample pairs" without labeled samples, forcing the neural network to learn to identify intermodal consistency, thereby achieving semantic alignment and consistent representation across different remote sensing imagery. This is the most groundbreaking aspect of zero-shot change detection methods, aiming to bypass the reliance on large amounts of manually labeled data and improve the model's ability to identify changes in unknown areas.

[0043] In terms of implementation, a "positive sample pair" is first defined as the feature vectors of images from different modalities at the same geographic location, that is, the representation of the same area in optical and radar images. A "negative sample pair" is composed of a combination of image features from different modalities from spatially non-adjacent areas. This comparative construction method is a typical information-discriminative training method in unsupervised learning.

[0044] During training, the model is optimized using a contrastive loss function (such as NT-Xent or Triplet Loss). The goal is to minimize the distance between pairs of positive examples in the embedding space and maximize the distance between pairs of negative examples. This trained embedding space can autonomously identify semantically consistent regions without relying on manual labeling, effectively enabling change analysis.

[0045] To prevent unstable distribution of positive samples during training, the model introduces a "momentum encoder" mechanism. This involves constructing a slowly updating target encoder, using its encoding results as a reference for representing positive samples. This mechanism maintains representation stability during training, preventing learning drift caused by sample fluctuations and resulting in a more coherent representation of the semantic space.

[0046] In general, this step enables the model to complete cross-modal recognition tasks without manual guidance by imitating the logic of human observers in identifying differences in comparison. It is the key to the core capability of the present invention to achieve "zero-sample" detection.

[0047] Calculate the difference between temporal image features in a shared semantic space and determine the changed area based on a set threshold; When identifying the changed regions in the shared semantic space, a dual threshold strategy is used for pixel-level judgment, including the following methods: Calculate the feature difference between any pair of embedding vectors at different times; The upper and lower limits of the change threshold are set, and pixels with a difference below the lower limit are marked as "unchanged", and pixels with a difference above the upper limit are marked as "changed". Pixels in the middle area are further processed by the edge enhancement model, and their final labels are determined by the gradient boosting algorithm.

[0048] This strategy effectively reduces misjudgments in fuzzy edge areas while enhancing the boundary clarity of change detection results. It is particularly suitable for remote sensing scenes with clear ground contours but weak textures.

[0049] This step enhances the model's decision-making ability during the change judgment phase, particularly in situations with blurred edges and uncertain change scales. Essentially, it's a "re-judgment of the intermediate region" strategy that replaces traditional single-threshold classification, improving the refinement of detection results and the reliability of edge judgment.

[0050] The implementation process includes three parts: First, the system extracts the differences in feature vectors at different periods from the embedding space and calculates the degree of change at the pixel level.

[0051] Then two thresholds are set: the upper threshold represents the limit of "significant change" and the lower threshold represents the limit of "significantly unchanged".

[0052] All pixels are classified according to their difference scores: those above the upper limit are directly marked as changed, those below the lower limit are marked as unchanged, and "fuzzy pixels" in between are not directly judged.

[0053] These blurred pixels are then fed into an edge enhancement module, which typically builds a lightweight convolutional network based on image gradients and structural edge information. This model assists with recognition by analyzing contextual edge strength and directional changes. This model not only considers spatial continuity but also incorporates local contrast and image semantic consistency to enhance the confidence of edge judgments.

[0054] This step effectively reduces the "false positives" and "false negatives" caused by factors such as blurred object boundaries and inconsistencies in remote sensing imaging. It is a key means to improve the accuracy and credibility of the model's final change detection map.

[0055] The difference between the embedding vectors is calculated using a multi-dimensional fusion algorithm. The specific steps are as follows: Assume that the shared semantic space embedding vectors extracted from the two moments are and , calculate the Euclidean distance between them as the basic change metric, the calculation expression is as follows: , where It is the Euclidean difference between embedded features and is the basic measure for evaluating whether there is a significant spatial offset between the features at two moments. The larger the value, the more significant the change may be. It is the vector's two-norm (Euclidean norm), also known as the Euclidean distance, which measures the geometric distance between two embedded vectors and reflects the "distance" of the overall features in space;

[0056] The Euclidean difference between embedded features reflects the absolute distance between two embedded vectors in high-dimensional space and is mainly used to identify areas with large changes.

[0057] Calculate the cosine similarity between vectors and convert it into a difference factor as the basis for judging directional differences. The calculation expression is as follows: , where and Represent the shared semantic space embedding vectors and The Euclidean length of , normalized by one, It is the cosine "difference". The closer the value is to 1, the greater the difference in the directions of the two vectors is, that is, the features at the two time points show different semantic tendencies.

[0058] This difference is sensitive to vector direction and can identify local structural changes, such as details of urban expansion boundaries.

[0059] Calculate the average difference between the two embedding vectors in each dimension as the average trend of local changes. The calculation expression is as follows: , where It is to take the absolute value of the channel-by-channel difference and then find the average value to get the average value of the overall difference. It is the average change amplitude under the full channel, which can reflect the small but consistent regional change trends in a balanced manner, such as seasonal changes in cultivated land; Average change in all channels Indicates how balanced the changes are across all channels, helping to distinguish noise disturbances from actual changes.

[0060] Fusion 、 as well as , forming the final difference score, the formula is as follows: , where is the Euclidean difference between the embedded features The weighting coefficient of is the cosine "difference" The weighting coefficient of is the average change amplitude under all channels The weighted coefficient of , the optimal setting is made through cross-validation during model training, It is the final difference score, which comprehensively considers Euclidean distance (spatial difference), cosine direction difference (semantic change) and average channel difference (detail trend), and is the core criterion for determining the changed area.

[0061] This method can jointly evaluate the degree of change between two time points from three dimensions: geometric amplitude, directional consistency, and average trend, greatly improving the detection robustness and classification accuracy under zero-sample conditions.

[0062] This step establishes a highly discriminative, multidimensional change measurement mechanism within the feature embedding space, addressing the inadequacy of traditional single-metric recognition in complex remote sensing imagery. By simultaneously considering geometric distances, directional similarities, and statistical distribution differences between feature vectors, this mechanism significantly improves the model's sensitivity and discernment in areas of fine-grained variation, making it particularly useful in scenarios where semantic mismatches between cross-modal data are a serious concern.

[0063] In its implementation, the mechanism first extracts feature vectors of corresponding pixels between two time-series remote sensing images from the embedding network output. These feature representations contain local spatial, spectral, and textural information of the image. While high in dimensionality, the embedded structure possesses stable semantic meaning. The system then employs multiple submodules to deconstruct these vectors, calculating their vector difference magnitude, angular difference, and mean change intensity.

[0064] These submodules are independently constructed and trained using a multi-layer perceptron network to be more responsive to different types of changes, such as material changes, structural movement, and contour disappearance. The difference scores are then normalized and aggregated by a fusion module using empirical weights to generate a single change assessment score.

[0065] Furthermore, the scoring system is self-learning, adjusting scoring weights based on a small number of known samples, enabling it to generalize and transfer across different remote sensing data sources. The resulting difference score map no longer relies on traditional single-threshold decision-making, but can be used for high-precision decision-making in subsequent modules such as change classification and refinement.

[0066] Through this mechanism, the system can keenly identify subtle changes caused by fuzzy boundaries, small-scale disturbances and texture misalignment, thereby significantly improving the sensitivity, accuracy and interpretability of change detection.

[0067] Output the change detection result graph and use the structural consistency constraint module to further optimize the boundary judgment accuracy and improve the reliability of change area identification; After outputting the change detection result graph, the predicted graph is further post-processed using the structural consistency constraint module; The structural consistency constraint module combines the structural similarity index (SSIM) with morphological processing methods to reconstruct and merge all adjacent changed regions in the detection image. The specific process includes calculating the structural consistency score of each changed segment and its neighborhood. Those with scores below the set threshold are eliminated or edge-reduced. At the same time, boundary discontinuity areas are repaired through dilation and erosion operations to ensure the coherence and morphological rationality of the changed areas.

[0068] This module not only enhances the stability of the detection image, but also effectively suppresses the false positive areas caused by isolated misjudged pixels.

[0069] The purpose of this step is to optimize the boundary clarity and regional integrity of the detection image through structural analysis and image morphological operations based on the preliminary detection result image, reduce false positive areas and broken areas caused by feature ambiguity and noise disturbance, and thus improve the stability and practical application value of the change detection image.

[0070] The specific implementation steps are as follows: First, the system assigns a structural similarity score (e.g., SSIM) to all predicted change regions. It then compares each region with its original image counterpart, identifying "abnormal change segments" with low structural consistency. These segments are likely artifacts or misidentified regions, and the system then suppresses or adjusts their edges accordingly.

[0071] Next, the system performs the "morphological reconstruction" operation, which includes: Dilation processing: used to repair edge breaks and reconnect the changed areas caused by local discontinuity or texture loss; Erosion processing: used to remove isolated pixels or small noise clusters in the detection image to improve the overall cleanliness of the changed area; Closing / opening operations: Combine dilation and erosion to reconstruct the contours of features, smooth edges, and prevent excessive breaks.

[0072] This structural consistency constraint module is equivalent to an "image reviewer", which not only improves the visual quality of the change detection map, but more importantly, enhances the robustness of the detection system in high-noise environments.

[0073] The structural consistency constraint module uses a structural confidence scoring mechanism constructed by combining edge gradient direction and local contrast entropy to perform boundary optimization on the change detection result map. The specific steps are as follows: Implement sliding window analysis on the output change detection map, and set the center pixel of each window to be , calculate the local contrast entropy of the predicted change area in the current window to measure the information complexity of its pixel distribution. The calculation expression is as follows: , where It is based on pixel points in the change detection map The contrast entropy within the local sliding window centered on is used to measure the complexity of the pixel category distribution in the area. Is the first The proportion of pixels of the same type (such as changed, unchanged), the value range is [0, 1], satisfying , is the number of pixel categories; Extract the edge map from the original multi-source image, and calculate the gradient direction consistency factor between the edge direction vector of the predicted change map and the direction vector of the original edge map. The calculation expression is as follows: , where is the pixel in the predicted change detection map The gradient direction angle of the window is usually obtained by extracting the edge through the Sobel or Canny operator. is the true edge direction angle extracted from the corresponding position in the original multi-source remote sensing image (such as optical or radar image), is the gradient direction consistency factor, which measures the similarity of the edge gradient direction between the predicted image and the original multi-source image; Gradient direction consistency factor A value close to 1 indicates a high degree of consistency in direction, which is a valid change boundary; a value close to 0 or a negative value indicates a pseudo boundary, which may be caused by noise.

[0074] joint and , construct a structural confidence scoring function to evaluate the credibility of each window change prediction. The calculation expression is as follows: , where is the structural complexity weight, which indicates the influence of local entropy information on the total score, and its value range is [0, 1]. is the boundary direction consistency weight, which reflects the importance of edge information in the total score and also takes the value range [0, 1], satisfying , which is dynamically set by the structural coincidence optimization model to balance the contribution of local complexity and edge direction reliability, It is a structural confidence score used to filter pixel areas with unstable structures and unreliable edges, and perform edge compression or elimination in the detection image.

[0075] All window areas with score values ​​lower than the preset threshold are subjected to edge compression, morphological culling or area ablation processing to remove isolated errors and unreasonable change blocks, and improve the clarity and spatial continuity of the result boundaries.

[0076] This method comprehensively considers the structural complexity of the region and the consistency of the boundary direction, significantly enhancing the credibility of change detection results under unsupervised conditions. It is particularly suitable for areas where complex landforms intersect, such as the transition zone between cities and farmland.

[0077] The core function of this step is to assess and correct the structural quality of the change detection map initially predicted by the model, ensuring high consistency between its spatial boundaries and the reference image. This prevents "false structure" and "false contours" in the change region caused by modal differences, texture weakening, or model misidentification. By introducing local contrast analysis and edge alignment mechanisms, this step systematically scores the structural rationality of the prediction results, enabling post-processing optimization.

[0078] In practice, the system first uses a sliding window strategy to scan the change map region by region and analyzes the contrast pattern of the pixel intensity distribution within each window. This contrast analysis is not based on overall brightness, but rather on texture comparison of local image blocks to determine whether the region has clear structural boundaries. Insufficient contrast may indicate a false detection.

[0079] Next, the system extracts edge directions from the change map and the original image's edge map, generating a principal direction vector for each pixel. By comparing the edge orientations of each region in the predicted image with those in the original image, the system determines whether the detected region truly matches the reference image's structure. If edge directions are significantly inconsistent, the system identifies the region as structurally unstable, potentially indicating a false positive.

[0080] Ultimately, the two scoring systems—local contrast and edge orientation—are combined into a single structural consistency score map. Regions with scores below a set threshold are removed, compressed, or undergo edge reconstruction, improving the detection map's boundary accuracy, contour coherence, and visual cleanliness.

[0081] This scoring mechanism is particularly well-suited for environments with complex boundaries, weak texture details, and diverse change types in high-resolution remote sensing imagery. It introduces a structural "second judgment" into the change detection system, which is crucial for improving the usability and output quality of change maps.

[0082] The present invention significantly improves the semantic consistency between multi-source remote sensing images by introducing a cross-modal shared semantic embedding space and combining it with an unsupervised comparative learning mechanism, thereby effectively alleviating the semantic mismatch problem caused by differences in the imaging mechanisms of different sensors. By constructing a dual-channel feature encoding network, key semantic features in optical and radar images are extracted respectively, and mapped to a unified semantic space, so that even in the absence of labeled samples, the model can still accurately understand and match the manifestations of the same land object in different modalities. This method provides a stable and high-quality feature basis in the process of determining the changed area, greatly reducing the false positive rate and missed judgment rate, and improving the robustness and reliability of change detection.

[0083] The present invention further improves the boundary accuracy of the changed area and the coherence of the overall image structure through the collaborative optimization of the difference fusion calculation mechanism and the structural consistency constraint module. The multi-dimensional difference calculation formula that integrates Euclidean distance, cosine similarity and mean value difference enhances the model's ability to distinguish different types of changes; while the structural consistency module post-processes the detection results based on contrast entropy and gradient direction consistency, effectively eliminating misjudged areas with blurred boundaries and incomplete structures. This design significantly improves the accuracy and consistency of the final output image at the visual level, and is particularly suitable for remote sensing analysis tasks that require high structural integrity, such as the edges of urban buildings and water bodies.

[0084] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0085] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

[0086] It should be noted that, in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0087] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0088] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0089] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0090] The units described as separate components may or may not be physically separate, and 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 these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0091] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0092] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0093] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. A zero-sample change detection method for multi-source remote sensing images, characterized in that: The following steps are involved: Acquire remote sensing images collected by two or more remote sensing sensors over the same geographical area at different time points. Image types include optical images and radar images. Pre-process each source image to unify the spatial resolution and registration accuracy, and perform denoising and standardization on the image; Construct a cross-modal feature encoding network to extract features from different source images separately, and introduce a shared semantic embedding space to enhance the semantic alignment ability between different modalities; By optimizing feature representation through unsupervised contrastive learning, the embedding vectors of the same feature in different images maintain high similarity, thus improving semantic consistency. Calculate the difference between temporal image features in a shared semantic space and determine the changed area based on a set threshold; The change detection result graph is output, and the structural consistency constraint module is used to further optimize the boundary judgment accuracy and improve the reliability of change area identification.

2. The zero-sample change detection method for multi-source remote sensing images according to claim 1, characterized in that: The multi-source remote sensing images acquired must go through the precise spatiotemporal registration process of the same geographical area. The specific steps are as follows: The global positioning system data and attitude information are used for geometric correction, and then a registration algorithm based on maximizing mutual information is used to accurately align the images so that the pixels from different remote sensing sources have a one-to-one correspondence in geographic space. Perform cloud detection and occlusion removal on images to ensure the spatial integrity and comparability of input data; In terms of image time point selection, the acquisition time difference is controlled within no more than three months to maximize the reality of changes in the structure of objects in the target area.

3. The zero-sample change detection method for multi-source remote sensing images according to claim 1, characterized in that: The image preprocessing steps include enhancing spectral and structural features using a dual-domain normalization strategy for multi-source data; When processing optical images, histogram equalization and local contrast enhancement algorithms are introduced to improve the clarity of object boundaries. For radar images, logarithmic transformation and coherent filtering methods are used to eliminate speckle noise and enhance the structural texture of objects. All images are uniformly converted into a pseudo-color format with 3 channels to ensure consistent input to the subsequent deep neural network processing module and prevent degradation of model performance due to channel mismatch.

4. The zero-sample change detection method for multi-source remote sensing images according to claim 1, characterized in that: The cross-modal feature encoding network uses a parallel convolutional architecture to construct two sub-channels, corresponding to optical and radar image inputs respectively. Normalization and ReLU activation functions are applied to the output of each convolutional layer to avoid gradient vanishing and enhance nonlinear feature extraction capabilities. To enhance the alignment capability of the cross-modal shared semantic space, a residual connection mechanism and a channel attention module are introduced into the network, enabling the model to perceive high-level semantic features while capturing local texture details. A shared transformation module is connected at the end of each sub-channel to map high-dimensional features to the same embedding space, laying the foundation for subsequent unsupervised contrastive learning.

5. The zero-sample change detection method for multi-source remote sensing images according to claim 1, characterized in that: The unsupervised contrastive learning mechanism adopts a loss optimization strategy based on positive and negative sample pairs. The specific steps are as follows: Randomly extract paired samples from the feature vectors of different modalities at the same geographical location as positive sample pairs; Features are collected from non-corresponding positions as negative sample pairs; finally, by optimizing the contrast loss function based on cosine distance, the distance between positive sample pairs is minimized and the distance between negative sample pairs is maximized; To prevent the model from falling into local optimality, a momentum encoder mechanism is introduced during the training process to dynamically update the positive sample representation to ensure the stability and discriminability of the embedding space.

6. The zero-sample change detection method for multi-source remote sensing images according to claim 1, characterized in that: When identifying the changed regions in the shared semantic space, a dual threshold strategy is used for pixel-level judgment, including the following methods: Calculate the feature difference between any pair of embedding vectors at different times; The upper and lower limits of the change threshold are set. Pixels with a difference below the lower limit are marked as "unchanged", and those above the upper limit are marked as "changed". Pixels in the middle area are further processed by the edge enhancement model, and their final labels are determined by the gradient boosting algorithm.

7. The zero-sample change detection method for multi-source remote sensing images according to claim 6, characterized in that: The difference between the embedding vectors is calculated using a multi-dimensional fusion algorithm. The specific steps are as follows: Assume that the shared semantic space embedding vectors extracted from the two moments are and , calculate the Euclidean distance between them as the basic change metric, the calculation expression is as follows: , where is the Euclidean difference between the embedded features, It is the two-norm of the vector, also known as the Euclidean distance, which measures the geometric distance between two embedded vectors; Calculate the cosine similarity between vectors and convert it into a difference factor as the basis for judging directional differences. The calculation expression is as follows: , where and Represent the shared semantic space embedding vectors and The Euclidean length of is the cosine "difference"; Calculate the average difference between the two embedding vectors in each dimension as the average trend of local changes. The calculation expression is as follows: , where It is to take the absolute value of the channel difference and then find the average. is the average change amplitude under all channels; Fusion 、 as well as , forming the final difference score, the formula is as follows: , where is the Euclidean difference between the embedded features The weighting coefficient of is the cosine "difference" The weighting coefficient of is the average change amplitude under all channels The weighting coefficient of is the final difference score.

8. The zero-sample change detection method for multi-source remote sensing images according to claim 1, characterized in that: After outputting the change detection result graph, the predicted graph is further post-processed using the structural consistency constraint module; The structural consistency constraint module combines structural similarity indicators with morphological processing methods to reconstruct and merge all adjacent changed areas in the detection image. The specific process includes calculating the structural consistency score of each changed segment and its neighborhood. Those with scores below the set threshold are eliminated or edge-reduced. At the same time, boundary discontinuity areas are repaired through dilation and erosion operations to ensure the continuity and morphological rationality of the changed areas.

9. The zero-sample change detection method for multi-source remote sensing images according to claim 8, characterized in that: The structural consistency constraint module uses a structural confidence scoring mechanism constructed by combining edge gradient direction and local contrast entropy to perform boundary optimization on the change detection result map. The specific steps are as follows: Implement sliding window analysis on the output change detection map, and set the center pixel of each window to be , calculate the local contrast entropy of the predicted change area in the current window to measure the information complexity of its pixel distribution. The calculation expression is as follows: , where It is based on pixel points in the change detection map The contrast entropy within the local sliding window centered at Is the first The proportion of class pixels, is the number of pixel categories; Extract the edge map from the original multi-source image, and calculate the gradient direction consistency factor between the edge direction vector of the predicted change map and the direction vector of the original edge map. The calculation expression is as follows: , where is the pixel in the predicted change detection map The gradient direction angle of the window, is the true edge direction angle extracted from the corresponding position in the original multi-source remote sensing image, is the gradient direction consistency factor; joint and , construct a structural confidence scoring function to evaluate the credibility of each window change prediction. The calculation expression is as follows: , where is the structural complexity weight, is the boundary direction consistency weight, is the structural confidence score.

Images