A remote sensing image change detection method and system based on confidence aggregation of each image spot

Abstract

The present disclosure relates to a remote sensing image change detection method and system based on per-patch confidence aggregation. The method comprises obtaining and preprocessing the pre-stage and post-stage images of remote sensing images, and dividing them into image blocks; calculating the probability of each pixel point changing in the ground object through a base change detection model; calculating the weight of the pixel position in each image block, determining the probability accumulation matrix of the full image size according to the probability of the ground object change and the weight, and calculating the fusion probability map according to the weight accumulation matrix of the full image size; calculating the adaptive threshold according to the probability distribution of the fusion probability map to generate a binary mask; determining all connected domains, calculating the confidence of each connected domain, establishing a confidence grid map, converting each patch to a vector polygon element, adding a confidence field and a geographic area actual area field to each vector polygon element. The present disclosure makes full use of pixel-level probability, aggregates confidence information to patch level, and reduces the burden of subsequent manual review.

Description

Technical Field

[0001] This disclosure relates to the field of remote sensing image processing, and in particular to a method and system for detecting changes in remote sensing images based on patch confidence aggregation. Background Technology

[0002] Remote sensing change detection involves comparing remote sensing images acquired at two different times to identify areas of change in ground features that have occurred between them. Currently, most mainstream methods are based on deep learning-based semantic segmentation models, such as ChangeFormer, BIT, and FC-EF. These models typically output the probability of change for each pixel. Then, by setting a fixed global threshold (like =0.5) Convert the change probability map into a binary change mask:

[0003] (1)

[0004] in, and These represent the x and y coordinates of a pixel in the overall image coordinate system, respectively. Represents pixels The binary change mask value.

[0005] This strategy has the following shortcomings and pain points:

[0006] Poor threshold adaptability: The optimal threshold varies greatly depending on different land types, seasons, and imaging conditions;

[0007] Image patch confidence is missing: The existing processing flow discards pixel probability information, making it impossible to quantify the confidence level of each image patch;

[0008] The lack of confidence-based sorting makes subsequent manual review difficult, wastes manpower, and places a heavy burden on manual reviewers.

[0009] Insufficient information utilization: Some models have high discriminative power at the pixel level, but they are completely binarized in the post-processing stage, resulting in insufficient utilization of the pixel level probability.

[0010] In practical remote sensing change detection operations, fixed thresholds often lead to numerous false positives, resulting in low reliability and placing a significant burden on manual review. Furthermore, large-scale remote sensing images (e.g., tens of thousands of pixels x tens of thousands of pixels) cannot be directly input into deep learning models for processing; a strategy of segmenting and slicing them before model inference is typically employed. However, effectively fusing the probabilistic information from each slice after segmentation and generating a full-image confidence assessment remains a critical technical challenge.

[0011] Therefore, it is necessary to design a remote sensing image change detection method and system based on patch confidence aggregation to solve the above problems. Summary of the Invention

[0012] Therefore, it is necessary to provide a method and system for detecting changes in remote sensing images based on patch confidence aggregation to address the above problems.

[0013] To solve the above problems, the present disclosure adopts the following technical solution:

[0014] In a first aspect, this disclosure provides a method for detecting changes in remote sensing images based on patch-by-patch confidence aggregation, comprising the following steps:

[0015] Acquire remote sensing images from both the earlier and later periods, and perform image preprocessing to ensure that the coordinate systems, resolutions, and spatial overlaps of the earlier and later remote sensing images are consistent.

[0016] All remote sensing images are divided into blocks to obtain image patches, and there are overlapping areas between adjacent image patches;

[0017] The logit value of each pixel belonging to each land cover region category and the probability of land cover change for each pixel are calculated using a deep learning-based change detection model.

[0018] Calculate the weight of pixel positions in each image patch, and determine the probability accumulation matrix of the entire map size based on the probability of ground feature changes and the weights. The weight accumulation matrix for the entire image size is determined based on the weights. Calculate the fusion probability map The for and The ratio;

[0019] according to Probability distribution calculation adaptive threshold ,based on and Generate a binary mask ,Will and Generate candidate probability graphs by multiplying corresponding elements. ;

[0020] Sure All connected components in, according to Calculate the confidence score for each connected component and construct a confidence raster plot. The confidence level is assigned to all pixels within the corresponding connected component.

[0021] Will Each map patch is converted into a vector polygon feature, and a confidence field and an actual area field of the geographic region are added to each vector polygon feature.

[0022] In a preferred embodiment, the weight is calculated using the following formula:

[0023] (9);

[0024] in, Indicates the first Pixel position within an image block The weight, Indicates the first The x-coordinate of the pixel position within each image patch. Indicates the first The ordinate of the pixel position within each image patch. , representing pixels The minimum distance to the boundary of the image patch. Indicates the side length of the image patch. Indicates the number of overlapping pixels.

[0025] In a preferred embodiment, the image preprocessing includes: when EPSG codes are inconsistent, resampling is performed using nearest neighbor interpolation or bicubic interpolation until EPSG codes are consistent; when the spatial resolution of the previously captured remote sensing image and the later captured remote sensing image is inconsistent, resampling is used to make their resolutions consistent.

[0026] In a preferred embodiment, the image block is square; when dividing all remote sensing images into blocks, the formulas for calculating the number of blocks in the height and width directions are as follows:

[0027] (2);

[0028] (3);

[0029] in, Indicates the number of blocks in the height direction. Indicates the number of blocks in the width direction. Indicates rounding down. Indicates the height of the image. Indicates the width of the image. This indicates the size of the square block window. This indicates the step size between blocks.

[0030] In a preferred embodiment, the change of ground features at each pixel is a binary change detection task, and the formula for calculating the probability of a change of ground features at each pixel is as follows:

[0031] (7);

[0032] (8);

[0033] in, Indicates the first Pixel positions of image patches The probability that no changes in ground features have occurred. Indicates the first Pixel positions of image patches The probability of a change in land cover category. Indicates the first Pixels of an image block The logit value of the category belonging to the in-situ feature area category. Indicates the first Pixels of an image block The logit value of a category that belongs to a non-local feature area category.

[0034] In a preferred embodiment, the probability cumulative matrix for determining the overall map size based on the probability of ground feature changes and the weights is... The weight accumulation matrix for the entire image size is determined based on the weights. The specific accumulation process includes:

[0035] (10);

[0036] in, The pixel position in the global coordinate system. This represents the total number of image patches in a remote sensing image. The corresponding position in the local coordinate system of the image patch has the following coordinate transformation relationship:

[0037] (11);

[0038] in, For the first The starting position for writing an image patch in the entire image.

[0039] In a preferred embodiment, the according to Probability distribution calculation adaptive threshold The calculation formula is:

[0040] (13);

[0041] in, The preset base threshold, To calculate the multiple factor, express The probability mean, express The probability standard deviation;

[0042] The basis and Generate a binary mask The calculation formula is:

[0043] (14).

[0044] In a preferred embodiment, the determination All connected components are obtained by applying the 8-connectivity criterion to the binary mask. Perform connected component analysis to determine All connected components in;

[0045] According to The formula for calculating the confidence score of each connected component is as follows:

[0046] (17);

[0047] in, Indicates the confidence level of the connected components. For the first A set of pixels in a connected region. For the first The number of pixels contained in each connected component;

[0048] Confidence Raster Satisfying the formula:

[0049] (18);

[0050] in, The function represents the connected component labeling function. Indicates the connected component number.

[0051] In a preferred embodiment, the will Each map patch in the map is converted into a vector polygon feature. For each vector polygon feature, a confidence field and an actual geographic area field are added, including:

[0052] Filter out Patches with an area smaller than a set threshold;

[0053] Convert each retained patch into a vector polygon feature;

[0054] Add a confidence field and an actual area field for each vector polygon feature.

[0055] Perform topology error correction, buffer deburring, and vector simplification operations on each geometric feature of each vector polygon feature;

[0056] Output a vector map file containing geometric information, a confidence level field, and the actual area field of the geographic region.

[0057] Secondly, this disclosure provides a remote sensing image change detection system based on patch-by-patch confidence aggregation, comprising:

[0058] The acquisition and preprocessing module is used to acquire remote sensing images captured in the early stage and those captured in the later stage, and to perform image preprocessing to ensure that the coordinate systems, resolutions, and spatial overlaps of the remote sensing images captured in the early stage and those captured in the later stage are consistent.

[0059] The segmentation module is used to divide all remote sensing images into blocks to obtain image blocks, with overlapping areas between adjacent image blocks;

[0060] The first calculation module is used to calculate the logit value of each pixel belonging to each land cover region category and the probability of land cover changes occurring at each pixel using a deep learning-based change detection model.

[0061] The second calculation module is used to calculate the weight of pixel positions in each image patch, and determine the probability accumulation matrix of the entire map size based on the probability of ground feature changes and the weights. The weight accumulation matrix for the entire image size is determined based on the weights. Calculate the fusion probability map The for and The ratio;

[0062] The third calculation module is used to calculate based on... Probability distribution calculation adaptive threshold T adaptive ,based on and T adaptive Generate a binary mask ,Will and Generate by multiplying corresponding elements ;

[0063] The confidence level statistics module is used to determine... All connected components in, according to Calculate the confidence score for each connected component and construct a confidence raster plot. The confidence level is assigned to all pixels within the corresponding connected component.

[0064] The vector conversion module is used to convert... Each map patch is converted into a vector polygon feature, and a confidence field and an actual area field of the geographic region are added to each vector polygon feature.

[0065] This disclosure presents a remote sensing image change detection method and system based on patch-by-patch confidence aggregation. It preserves pixel probability information, fully utilizes pixel-level probabilities, ensures consistent pixel-level probability space, quantifies the confidence level of each patch, dynamically assigns confidence values, and assigns an independent confidence field to each changed patch. This aggregates and expresses pixel-level confidence information at the patch level, improving the interpretability of the change detection results. Furthermore, it calculates adaptive thresholds based on probability distributions, enhancing threshold adaptability. The presence of confidence information in patches improves the reliability of the change detection output and reduces the burden of subsequent manual review. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating a method in one embodiment of the present disclosure;

[0067] Figure 2 This is a schematic diagram of the system structure in one embodiment of the present disclosure;

[0068] Figure 3 The diagram shows the extraction results of the traditional change detection model (left) and the extraction results of this invention (right). Detailed Implementation

[0069] The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0070] See Figure 1 A method for detecting changes in remote sensing images based on patch confidence aggregation includes the following steps:

[0071] Acquire early and late images of remote sensing imagery. Preprocess the early and late images to ensure that their coordinate systems, resolutions, and spatial overlap are consistent. The early images are remote sensing images captured earlier, and the late images are remote sensing images captured later. The late images are images captured after the early images, and the early images are images captured before the late images.

[0072] All previous and later images are divided into blocks to obtain image blocks, and there are overlapping areas between adjacent image blocks;

[0073] Each image patch is input into a deep learning-based change detection model. The change detection model calculates the logit value of each pixel belonging to each land cover region category, and (based on the logit value) calculates the probability of land cover change for each pixel.

[0074] Calculate the weight of pixel positions in each image patch, and determine the probability accumulation matrix of the entire map size based on the probability of ground feature changes and the weights. The weight accumulation matrix for the entire image size is determined based on the weights. Calculate the fusion probability map The for and The ratio;

[0075] according to Probability distribution calculation adaptive threshold ,based on and Generate a binary mask ,Will and Generate candidate probability graphs by multiplying corresponding elements. ;

[0076] according to Will Each map patch is converted into a vector polygon feature, and a confidence field and an actual area field of the geographic region are added to each vector polygon feature.

[0077] In this embodiment, dividing all pre-images and post-images into blocks specifically means that pre-images and post-images that are confirmed to have the same coordinate system, the same resolution, and spatial overlap are all divided into blocks. In this embodiment, the weight of the pixel position in each image block is called the fusion weight, and the calculation formula is:

[0078] (9)

[0079] in, Indicates the first Pixel position within an image block The fusion weight, Indicates the first The x-coordinate of the pixel position within each image patch. Indicates the first The ordinate of the pixel position within each image patch. , representing pixels Minimum distance to the slice boundary Indicates the side length of the image patch. Indicates the number of overlapping pixels.

[0080] The method specifically includes the following steps:

[0081] Step 1: Obtain the input pre- and post-images. The input is in GeoTIFF format. Pre-images refer to images captured a certain time before the post-images, covering the same area. For example, an image of area A captured today is a post-image, while an image of area A captured six months ago is a pre-image. Understandably, each post-image usually has a corresponding pre-image. Comparing pre- and post-images is necessary to identify whether changes in terrain features have occurred. Typically, during post-image capture, efforts are made to ensure that the coordinate systems, resolutions, and spatial overlap of the pre- and post-images are consistent. In the following text, achieving consistent coordinate systems, resolutions, and spatial overlap refers to the corresponding pre- and post-images. After inputting two sets of images, the coordinate system consistency is checked. If the EPSG (European Petroleum Survey Group) codes are inconsistent, nearest neighbor or bicubic interpolation resampling is used until the EPSG codes are consistent. Resolution is also checked to ensure consistency between the earlier and later images. If the spatial resolutions of the earlier and later images are inconsistent, resampling is required to make them consistent. The images are then cropped, retaining only the spatially overlapping portions. After ensuring consistent EPSG codes, consistent resolution, and cropped images, the preprocessed earlier and later images are represented uniformly as follows: ,in, and These are the height and width of the image, respectively. For RGB remote sensing images, the number of channels is... =3).

[0082] Step two: First, set the key parameters for the block processing. To segment the entire image into fixed-size image blocks (i.e., slices) suitable for the input of the deep learning model, and to retain a certain overlap between blocks, set the square block window size to [value missing]. (e.g., 512), overlap ratio factor (e.g., 0.2), calculate the step size between blocks based on these parameters. for This design ensures a certain overlap between adjacent slices, preventing the loss of boundary information. The window is moved incrementally in both the horizontal and vertical directions with this step size, recording the coordinates of the top-left corner of all slices. When the edge of the image is less than a full step size, the last slice is automatically aligned to the image boundary, ensuring complete coverage of the entire image.

[0083] Based on the dimensions and block parameters of the original images (i.e., confirmed to have consistent coordinate systems, resolutions, and spatial overlap between pre- and post-images), calculate the number of blocks to be placed in the height and width directions. The formulas for calculating the number of blocks in the height and width directions are as follows:

[0084] (2)

[0085] and

[0086] (3)

[0087] in, Indicates rounding down. Indicates the number of blocks in the height direction. This indicates the number of blocks in the width direction. This calculation method ensures complete coverage of the entire image area.

[0088] To handle image boundary regions, it is checked whether the movable dimension of the image is divisible by the step size. If the remainder after subtracting the tile size from the image height is not divisible by the step size, an additional tile start position is added to the height index list, set to the image height minus the tile size. The same processing logic is applied to the width direction. This boundary handling strategy ensures complete processing of image edge regions.

[0089] Generate two coordinate systems: a read coordinate system and a write coordinate system. The read coordinate system records data from the original image (…). Extract the top-left corner coordinates of each block from the image (in the original image), including overlapping areas. Write the coordinate system record and write the processing result (i.e., the result of step three) back to the large image (the original image). The target position (in the image) has been determined by removing overlapping parts to ensure that there are no duplicate writes during the final stitching.

[0090] Step 3: Input each image patch into a deep learning-based change detection model. The change detection model calculates the logit value of each pixel belonging to each land cover region category. Based on the logit value, the change detection model calculates the probability of land cover change for each pixel. For each image patch, the change detection model outputs the probability of land cover change for each pixel.

[0091] In step two, according to the settings, starting from the top left corner of the image, the original image is sliced ​​(i.e., image blocks) within each sliding window. The sliding window size... When set to 512, 512×512 slices are obtained after cropping the original images from the pre- and post-processing stages.

[0092] Each slice of the preprocessed early-stage remote sensing image and the preprocessed late-stage remote sensing image is input into a deep learning-based change detection model, and forward propagation is performed to obtain the model's inference results:

[0093] (4)

[0094] in, Indicates model parameters, This represents the forward propagation computation function. Indicates the first Image blocks, Indicates the image patch number, It refers to the number of categories. For binary change detection tasks, =2, Indicates the first The logit tensor of each image patch, image patch Medium pixels ( i,j ) belongs to the The class's logit value is used express, The x-coordinate of the pixel position of the image patch. The vertical coordinates of the pixel positions of the image patch. Indicates the category number, i.e., the number Category.

[0095] Understandably, the detection of changes in ground features is usually a binary change detection task, that is, to detect whether changes in ground features have occurred. One category is no change in ground features, and the other category is change in ground features.

[0096] Deep learning-based change detection models can be either BAN (Bi-temporal Adapter Network) or ChangeFormer models; these are just examples and not limitations.

[0097] For each pixel position of the image patch The probability of the change detection model is calculated using the following formula:

[0098] (5)

[0099] in, Represents image blocks pixel position Belongs to the The probability of a class category Indicates the category number.

[0100] The image patch can be obtained from the above formula. Category change probability tensor ,in, Represents pixels Belongs to the The probability of a class category, and satisfying the following axiom:

[0101] (6)

[0102] For a binary change detection task, we have:

[0103] (7)

[0104] (8)

[0105] in, Represents image blocks pixels The probability that the category remains unchanged (no change in ground features). Represents image blocks pixels The probability of a change in category (change in ground features). Represents image blocks pixels The logit value for a category belonging to class 0 (the original category). Represents image blocks pixels The logit value for categories belonging to class 1 (not the original category).

[0106] Step four involves designing a weighting function based on pixel-to-boundary distance to address the fusion problem of multiple prediction results in overlapping regions. Let the side length of the image patch (slice) be... Then pixel The minimum distance to the slice boundary is Set the number of overlapping pixels to... For the first Any pixel position within a slice Its fusion weights are based on linear distance decay. The calculation formula is:

[0107] (9)

[0108] in, , where is the number of overlapping pixels; It represents the minimum distance from a pixel to its four boundaries, with the image patch (slice) being the positive direction.

[0109] The weighting function has the following characteristics: the weight of pixels in the center of the slice is 1, the weight of pixels near the boundary decreases linearly with distance, and the weight at the boundary is close to 0. This design effectively reduces the impact of uncertainty in boundary prediction on the fusion result.

[0110] Construct the probability cumulative matrix for the entire image size. and weight cumulative matrix Both are initially set to 0. For the first... The processing results of each slice are shown in the probability graph. The corresponding weight matrix is Step three The probability map is a matrix formed by representing the probability of change of any point in an image patch. .

[0111] The mathematical description of the accumulation process is as follows:

[0112] (10)

[0113] in, The pixel position in the global coordinate system. This represents the total number of image patches in a remote sensing image. This represents the corresponding position in the local coordinate system of the slice. The coordinate transformation relationship is as follows:

[0114] (11)

[0115] here For the first The starting position for writing each slice in the global coordinate system, and the local coordinates of the pixels inside the slice are marked as follows: .

[0116] After completing the cumulative processing of all slices, the final fusion probability map is calculated. :

[0117] (12)

[0118] The fusion strategy ensures that: (1) the probability value of the overlapping area is obtained by weighted average of the prediction results of multiple slices; (2) the weight allocation is reasonable, and the prediction of the central area receives higher weight; and (3) the fusion result is spatially continuous and numerically stable.

[0119] The fusion probability map is an image of fused probability information, which refers to the weighted accumulation of the predicted probabilities of all image patches according to the fusion weights, and then divided by the weight accumulation matrix. The resulting full-image probability map is the final pixel-level probability map after spatial fusion of all slice prediction results.

[0120] Step 5: Analyze the fusion probability map. Perform statistical analysis to calculate the key statistics of its probability distribution. Key statistics include the probability mean. With probability standard deviation Let's consider the fusion probability graph. The probability mean is The standard deviation is Then the adaptive threshold The calculation formula is:

[0121] (13)

[0122] in, The preset base threshold (e.g., 0.7). To calculate the multiple coefficient (1.0-1.5), formula (13) ensures that the threshold is neither too low nor too low, and can adapt to the probability distribution characteristics of different images. In addition, besides the adaptive threshold strategy, to reduce computation time, the threshold can be customized based on experience or needs (e.g., 0.5).

[0123] Based on adaptive threshold Generate a binary mask :

[0124] (14)

[0125] Fusion probability map With binary mask Perform element-wise multiplication to generate a candidate probability map. :

[0126] (15)

[0127] in, This indicates element-wise multiplication. This operation preserves the original probability information of the high-confidence region while setting the probability value of the low-confidence region to 0.

[0128] The candidate probability map is the fusion probability map after screening, or the fusion probability map that retains high-confidence regions. With binary mask The image obtained by multiplying pixels retains only the probability information of the regions that have passed the threshold screening, while the probability of the remaining regions is set to 0. Therefore, it is called a "candidate probability map".

[0129] Step 6: Apply the 8-connectivity criterion to the binary mask. Perform connected component analysis. 8. Connectivity is defined as: pixel ( x , y ) and its 8 neighboring pixels {( x ±1, y The {±1} have a connectivity relationship.

[0130] The connected component labeling function is defined as follows:

[0131] (16)

[0132] in, The function represents the connected component labeling function. A function representing connected component labeling used in image processing. This means that as long as the foreground pixels are adjacent in any of these 8 directions, they belong to the same connected region.

[0133] This function assigns a displacement identifier {1, 2, …,} to each connected component. The background area is marked as 0. This represents the total number of connected components.

[0134] For the For each connected region, the region confidence score is calculated statistically from the probability values ​​of all pixels within that region. Confidence score calculations include, but are not limited to, the arithmetic mean, variance, and confidence interval. When the statistical method is the mean, the formula for calculating the region confidence score is as follows:

[0135] (17)

[0136] in, For the first A set of pixels in a connected region. Indicates the connected component number. Given the number of pixels contained in the connected component, this formula calculates the arithmetic mean (mean) of the probability values ​​within the region as the overall confidence level of the region. In formula (17), select the coordinate system belonging to the full-image coordinate system. Calculate the connected component confidence of each pixel. .

[0137] Then, the final confidence raster chart is constructed. The calculated region confidence score is then assigned to all pixels within the corresponding region:

[0138] (18)

[0139] The confidence raster image is a raster image carrying confidence information of image patches. It refers to a raster of the same size as the original image (i.e., the preprocessed image), where each pixel value is the confidence score of the connected component to which that pixel belongs. This is used to express the spatial distribution of patch-level confidence at the pixel level.

[0140] This allocation mechanism ensures that all pixels within the same region of variation have the same confidence value, improving the spatial consistency and semantic reasonableness of the results. The calculated confidence map... Write to the specified georeferenced raster file to maintain spatial accuracy.

[0141] Step 7: Convert the confidence level raster image Each polygon in the map is converted into a vector polygon feature, and area filtering is provided (polygons with an area smaller than a set threshold are directly removed), that is, filtering out... For polygons with an area smaller than a set threshold, each retained polygon is converted into a vector polygon feature. For each vector polygon feature, two new fields are added: "confidence" and "area". The "confidence" field equals the confidence level of the polygon, and the "area" field represents the actual area of ​​the geographic region corresponding to the polygon. Then, topology error correction, buffer deburring, and vector simplification are performed on each geometric feature. The output is a vector polygon file (such as Shapefile, GeoJSON, PostGIS, etc.) containing geometric information (information on the processed geometric features), confidence fields, and area fields, facilitating integration with business GIS systems. A confidence field is attached to each polygon extraction level, providing sorting and ranking analysis capabilities. Note that in this example, the output is a continuous, statistical confidence level, not discrete classification labels.

[0142] This disclosure provides a remote sensing image change detection system based on patch-by-patch confidence aggregation, the system comprising:

[0143] The acquisition and preprocessing module is used to acquire the early and late images of remote sensing images, and to preprocess the early and late images to make their coordinate systems consistent, their resolutions consistent, and their spaces overlap.

[0144] The segmentation module is used to divide all images into blocks to obtain image blocks, with overlapping areas between adjacent image blocks;

[0145] The first calculation module is used to calculate the logit value of each pixel belonging to each land cover region category and the probability of land cover changes occurring at each pixel using a deep learning-based change detection model.

[0146] The second calculation module is used to calculate the weight of pixel positions in each image patch, and determine the probability accumulation matrix of the entire map size based on the probability of ground feature changes and the weights. The weight accumulation matrix for the entire image size is determined based on the weights. Calculate the fusion probability map The for and The ratio;

[0147] The third calculation module is used to calculate based on... Probability distribution calculation adaptive threshold T adaptive ,based on and T adaptive Generate a binary mask ,Will and Generate by multiplying corresponding elements ;

[0148] The confidence level statistics module is used to determine... All connected components in, according to Calculate the confidence score for each connected component and construct a confidence raster plot. The confidence level is assigned to all pixels within the corresponding connected component.

[0149] The vector conversion module is used to convert... Each map patch is converted into a vector polygon feature, and a confidence field and an actual area field of the geographic region are added to each vector polygon feature.

[0150] In specific implementation, the remote sensing image change detection system based on patch confidence aggregation can refer to one of the remote sensing image change detection methods based on patch confidence aggregation in any of the above embodiments to achieve rapid iteration of remote sensing image change detection. The specific implementation steps will not be repeated here.

[0151] An electronic device can be implemented according to the method of this disclosure, the electronic device comprising: a memory; one or more processors; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for executing a remote sensing image change detection method based on patch confidence aggregation according to any of the above embodiments.

[0152] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0153] Memory can include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. Furthermore, memory can include any combination of computer-readable storage media; memory can be a semiconductor memory chip, a magnetic disk, or an optical disk.

[0154] The memory stores executable code, which, when processed by the processor, can cause the processor to execute some or all of the methods described above.

[0155] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the functions described above can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0156] This disclosure also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the steps of the remote sensing image change detection method based on patch confidence aggregation as described in any of the above embodiments.

[0157] To verify the applicability and effectiveness of the method and system disclosed herein in practical applications such as urban land use change analysis, Longhua District of Shenzhen City (covering an area of ​​approximately 180 km²) was selected as the verification area, and high-resolution remote sensing images from the second and third quarters were acquired. Through manual interpretation, full-element change patches were constructed from the two images, serving as a real-world control dataset for model accuracy verification. Based on this, the detection performance at the patch level was compared between traditional change detection methods and the proposed "patch-by-patch confidence-based change detection method." The change patch extraction results are as follows: Figure 3 As shown, the left figure is the extraction result of the traditional change detection model, and the right figure is the extraction result of this disclosure. It can be seen that the change patches extracted by the traditional change detection model are single-value results, lacking filterable indicator information, and can only be manually reviewed patch by patch. In contrast, the change patches extracted based on the method of this disclosure, while maintaining the same number and shape of patches, aggregate confidence scores for each patch, increasing the confidence information with differences, and preserving and improving the interpretability of the extraction results.

[0158] The quantitative evaluation results of the changed patches are shown in Table 1.

[0159] Table 1. Validation of the detection accuracy based on the change in confidence level of patch aggregation.

[0160]

[0161] As shown in Table 1, without confidence level filtering, the accuracy of the extracted change patches using traditional methods is 75.28%, with a high proportion of false positives. After applying the patch-by-pattern confidence aggregation and filtering mechanism of this invention, as the confidence threshold increases, low-confidence false positives are gradually eliminated, resulting in a significant increase in overall accuracy. When the threshold is set to 0.7, the accuracy increases to 89.22%; when the threshold reaches 0.8, the accuracy further increases to 95.18%; and when the threshold is 0.9, all patches automatically filtered by confidence level are real change areas, achieving 100% accuracy.

[0162] This experiment fully demonstrates that the mechanism proposed in this disclosure can quickly locate areas suspected of change, significantly reducing the burden of manual review. Especially in large-scale business application scenarios, confidence thresholds can be flexibly set according to actual needs to achieve efficient screening, hierarchical management, priority review, and distribution of map features, thereby greatly improving the usability and business value of remote sensing change detection results.

[0163] The present disclosure discloses a remote sensing image change detection method and system based on patch confidence aggregation. The effect is as follows: The present disclosure designs blocks where there is overlap between previous and later images, calculates the weight of the pixel position in each image block, determines the probability accumulation matrix of the whole map size based on the probability of ground feature change at each pixel, calculates the fusion probability map, generates a binary mask based on the probability distribution of the fusion probability map and an adaptive threshold, determines all connected components in the binary mask, calculates the confidence of each connected component, assigns the confidence to all pixels in the corresponding connected component, establishes a confidence raster map, and converts each patch into a vector polygon feature and adds a confidence field and a geographic area field. This disclosure effectively integrates the probabilistic information of each slice and generates a full-map confidence assessment. Specifically, the processing flow of this disclosure retains pixel probability information, fully utilizes pixel-level probabilities, ensures consistency in the pixel-level probability space, quantifies the confidence level of each patch, assigns an independent confidence field to each changed patch, aggregates and expresses pixel-level confidence information at the patch level, improves the interpretability of ground feature change detection results, and the final result contains confidence information, thus improving the reliability of change detection output results. Adaptive thresholds are calculated based on probability distributions, improving the adaptability of the thresholds and solving the problem of large differences in optimal thresholds under different land types, seasons, and imaging conditions, thereby improving the reliability of change detection. This disclosure fully utilizes the probabilistic output information of deep learning models, transforming the traditional "full-map threshold hard classification" into a "dynamic confidence assignment" soft probability assessment. The results of this disclosure contain a confidence field, allowing subsequent manual review to focus on areas with low confidence levels rather than all areas, facilitating subsequent manual verification and reducing the burden of manual review.

[0164] This disclosure enhances the controllability and business integration of land cover change detection results, and improves the adjustability in post-processing and business workflows. This includes, but is not limited to: controlling the number of land cover patches by adjusting confidence thresholds; managing land cover patches hierarchically based on confidence levels; customizing screening strategies (such as area, shape, and confidence level); and determining the priority of manual review based on confidence level. Therefore, the output volume, review costs, and false detection risks of this disclosure can be flexibly controlled according to business needs.

[0165] This disclosure is decoupled from existing deep learning models and can be deployed as a general post-processing module to meet the accuracy requirements of different application scenarios.

[0166] The final result of this disclosure is a patch, which has the ability to be sorted and its level is controllable, making it suitable for a high-efficiency verification mechanism in large-scale remote sensing change detection scenarios. This disclosure improves the expressive power and reliability of change detection patches. Through the confidence assessment of this disclosure, it can be used for subsequent sorting, level classification and confidence screening, which facilitates the identification and priority processing of high-confidence change areas. For example, high-confidence patches can be processed first, and low-confidence areas can be marked as candidate areas that require manual review. It can better assist in rapid manual review in various business scenarios such as natural resources, agriculture, urban planning, and water conservancy, saving labor and time costs. For example, it can be used to support dynamic threshold judgment of patches to adapt to different business needs. For example, it supports decision-making hierarchy, and patches with different confidence levels can flexibly select different subsequent processing procedures. Due to the simple patch field structure, it can be seamlessly integrated into various remote sensing business platforms or AI annotation platforms. It can be directly integrated with existing mainstream change detection frameworks (such as Open-CD), and has strong versatility.

[0167] The final result of this disclosure not only includes the confidence level, but also filters the confidence level. For example, only areas with a confidence level ≥ 0.7 can be retained or selected. Through the sorting and filtering mechanism based on the confidence level, it is beneficial for manual review and subsequent automated processing based on this disclosure. In subsequent business processes, it is not necessary to manually check all map patches. That is, the priority sorting and automatic filtering processes are automatically completed based on the confidence level of this disclosure, thereby reducing manual intervention.

[0168] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0169] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0170] The embodiments described above are merely illustrative of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these all fall within the protection scope of this disclosure. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for detecting changes in remote sensing images based on patch-by-patch confidence aggregation, characterized in that, Includes the following steps: Acquire remote sensing images from both the earlier and later periods, and perform image preprocessing to ensure that the coordinate systems, resolutions, and spatial overlaps of the earlier and later remote sensing images are consistent. All remote sensing images are divided into blocks to obtain image patches, and there are overlapping areas between adjacent image patches; The logit value of each pixel belonging to each land cover region category and the probability of land cover change for each pixel are calculated using a deep learning-based change detection model. Calculate the weight of pixel positions in each image patch, and determine the probability accumulation matrix of the entire map size based on the probability of ground feature changes and the weights. The weight accumulation matrix for the entire image size is determined based on the weights. Calculate the fusion probability map The for and The ratio; according to Probability distribution calculation adaptive threshold ,based on and Generate a binary mask ,Will and Generate candidate probability graphs by multiplying corresponding elements. ; Sure All connected components in, according to Calculate the confidence score for each connected component and construct a confidence raster plot. The confidence level is assigned to all pixels within the corresponding connected component. Will Each map patch is converted into a vector polygon feature, and a confidence field and an actual area field of the geographic region are added to each vector polygon feature.

2. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 1, characterized in that, The formula for calculating the weight is: (9)； in, Indicates the first Pixel position within an image block The weight, Indicates the first The x-coordinate of the pixel position within each image patch. Indicates the first The ordinate of the pixel position within each image patch. , representing pixels The minimum distance to the boundary of the image patch. Indicates the side length of the image patch. Indicates the number of overlapping pixels.

3. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 1, characterized in that, The image preprocessing includes: when EPSG codes are inconsistent, resampling is performed using nearest neighbor interpolation or bicubic interpolation until EPSG codes are consistent; when the spatial resolution of the previously captured remote sensing images and the later captured remote sensing images is inconsistent, resampling is used to make their resolutions consistent.

4. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 1, characterized in that, The image patch is square; when dividing all remote sensing images into blocks, the formulas for calculating the number of blocks in the height and width directions are as follows: (2)； (3)； in, Indicates the number of blocks in the height direction. Indicates the number of blocks in the width direction. Indicates rounding down. Indicates the height of the image. Indicates the width of the image. This indicates the size of the square block window. This indicates the step size between blocks.

5. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 1, characterized in that, The task of detecting ground feature changes at each pixel is a binary classification task. The formula for calculating the probability of ground feature changes at each pixel is as follows: (7)； (8)； in, Indicates the first Pixel positions of image patches The probability that no changes in ground features have occurred. Indicates the first Pixel positions of image patches The probability of a change in land cover category. Indicates the first Pixels of an image block The logit value of the category belonging to the in-situ feature area category. Indicates the first Pixels of an image block The logit value of a category that belongs to a non-local feature area category.

6. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 1, characterized in that, The probability cumulative matrix for determining the overall map size based on the probability of ground feature changes and the weights. The weight accumulation matrix for the entire image size is determined based on the weights. The specific accumulation process includes: (10)； in, The pixel position in the global coordinate system. This represents the total number of image patches in a remote sensing image. The corresponding position in the local coordinate system of the image patch has the following coordinate transformation relationship: (11)； in, For the first The starting position for writing an image patch in the entire image.

7. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 6, characterized in that, According to Probability distribution calculation adaptive threshold The calculation formula is: (13)； in, The preset base threshold, To calculate the multiple factor, express The probability mean, express The probability standard deviation; The basis and Generate a binary mask The calculation formula is: (14)。 8. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 7, characterized in that, The determination All connected components are obtained by applying the 8-connectivity criterion to the binary mask. Perform connected component analysis to determine All connected components in; According to The formula for calculating the confidence score of each connected component is as follows: (17)； in, Indicates the confidence level of the connected components. For the first A set of pixels in a connected region. For the first The number of pixels contained in each connected component; Confidence Raster Satisfying the formula: (18)； in, The function represents the connected component labeling function. Indicates the connected component number.

9. The remote sensing image change detection method based on patch-by-patch confidence aggregation according to claim 1, characterized in that, The Each map patch in the map is converted into a vector polygon feature. For each vector polygon feature, a confidence field and an actual geographic area field are added, including: Filter out Patches with an area smaller than a set threshold; Convert each retained patch into a vector polygon feature; Add a confidence field and an actual area field for each vector polygon feature. Perform topology error correction, buffer deburring, and vector simplification operations on each geometric feature of each vector polygon feature; Output a vector map file containing geometric information, a confidence level field, and the actual area field of the geographic region.

10. A remote sensing image change detection system based on patch-by-patch confidence aggregation, characterized in that, include: The acquisition and preprocessing module is used to acquire remote sensing images captured in the early stage and those captured in the later stage, and to perform image preprocessing to ensure that the coordinate systems, resolutions, and spatial overlaps of the remote sensing images captured in the early stage and those captured in the later stage are consistent. The segmentation module is used to divide all remote sensing images into blocks to obtain image blocks, with overlapping areas between adjacent image blocks; The first calculation module is used to calculate the logit value of each pixel belonging to each land cover region category and the probability of land cover changes occurring at each pixel using a deep learning-based change detection model. The second calculation module is used to calculate the weight of pixel positions in each image patch, and determine the probability accumulation matrix of the entire map size based on the probability of ground feature changes and the weights. The weight accumulation matrix for the entire image size is determined based on the weights. Calculate the fusion probability map The for and The ratio; The third calculation module is used to calculate based on... Probability distribution calculation adaptive threshold T adaptive ,based on and T adaptive Generate a binary mask ,Will and Generate by multiplying corresponding elements ; The confidence level statistics module is used to determine... All connected components in, according to Calculate the confidence score for each connected component and construct a confidence raster plot. The confidence level is assigned to all pixels within the corresponding connected component. The vector conversion module is used to convert... Each map patch is converted into a vector polygon feature, and a confidence field and an actual area field of the geographic region are added to each vector polygon feature.

Images