A weak supervision remote sensing image change detection method and system for land resource audit

Abstract

The application discloses a weakly supervised remote sensing image change detection method and system for land resource audit, first extracts double-time features from double-time collected images by using a SegFormer backbone network, then generates double-time specific class masks by taking the double-time collected images and corresponding text prompts as inputs based on a SAM3 large model, and then performs cross-time semantic alignment on the double-time features by using the double-time specific class masks and obtains feature consistency loss; the application realizes the function of introducing feature consistency constraints and activation consistency constraints into a class activation map-based weakly supervised change detection framework to perform weakly supervised remote sensing image change detection, and the feature consistency performs semantic alignment in a double-time feature space by using a pair relation label obtained from a mask to relieve class ambiguity, and the activation consistency combines semantic clues in a class activation map with spatial prior provided by a segmentation model mask to improve change recognition precision in land resource audit.

Description

Technical Field

[0001] This invention relates to the field of image change detection technology, specifically to a method and system for detecting changes in remote sensing images under weak supervision for land resource auditing. Background Technology

[0002] Land resource auditing is an important part of natural resource asset auditing, mainly used to verify farmland protection, construction land expansion, illegal land occupation, land idling, and changes in ecological land use. Remote sensing change detection can identify land cover changes from dual-temporal remote sensing images, providing objective, continuous, and traceable spatial evidence for land resource auditing. It can assist auditors in quickly locating areas suspected of new construction, farmland reduction, and abnormal land use.

[0003] Currently, most existing change detection methods compare differences between two-temporal images using image differencing, image ratio analysis, and change vector analysis. While computationally simple, these methods are easily affected by illumination variations, seasonal differences, image registration errors, and noise, making it difficult to meet the accuracy requirements of land resource auditing for change boundaries and ranges. Deep learning methods, on the other hand, generally rely on a large number of pixel-level annotations, which are not only costly but also unsuitable for large-scale land resource auditing scenarios. Furthermore, the significant differences in appearance between different land cover objects can easily lead to category ambiguity and false detections. Therefore, it is necessary to design a weakly supervised remote sensing image change detection method and system for land resource auditing. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies. To better and more effectively address the problem that most existing change detection methods rely on image difference, image ratio, and change vector analysis to compare differences between two-temporal images, while computationally simple, they are easily affected by illumination changes, seasonal differences, image registration errors, and noise, making it difficult to meet the accuracy requirements of land resource audits for change boundaries and ranges. Deep learning methods generally rely on a large number of pixel-level annotations, which are not only costly but also unsuitable for large-scale land resource audit scenarios. Furthermore, the significant differences in appearance between different land cover objects can easily lead to category ambiguity and false detections. This invention provides a weakly supervised remote sensing image change detection method and system for land resource audits. It achieves the function of introducing feature consistency constraints and activation consistency constraints into a weakly supervised change detection framework based on class activation maps for weakly supervised remote sensing image change detection. Feature consistency utilizes paired relational labels obtained from masks to perform semantic alignment in the two-temporal feature space, thereby alleviating category ambiguity. Activation consistency combines semantic cues in the class activation map with spatial priors provided by the segmentation model mask, improving the accuracy of change identification in land resource audits.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for detecting changes in remote sensing images under weak supervision for land resource auditing includes the following steps: Step A: Extract dual-temporal features from dual-temporal acquired images using the SegFormer backbone network; Step B: Based on the SAM3 large model, dual-temporal acquired images and corresponding text prompts are used as input to generate dual-temporal specific category masks; Step C: Use a dual-temporal-specific category mask to perform cross-temporal semantic alignment on dual-temporal features and obtain feature consistency loss; Step D: Generate pseudo-labels with confidence weights using a dual-temporal category-specific mask; Step E: Obtain activation consistency loss for class activation graph supervision using pseudo-labels with confidence weights; Step F involves incorporating feature consistency loss and activation consistency loss into a weakly supervised change detection framework based on class activation graphs and completing the change detection task for weakly supervised remote sensing images oriented towards land resource auditing.

[0006] The aforementioned method for detecting changes in remote sensing images under weak supervision for land resource auditing, in step A, involves using a SegFormer backbone network to extract dual-temporal features from dual-temporal acquired images. Specifically, this involves setting up dual-temporal images... ,in For spatial dimensions, the SegFormer backbone network consists of four Transformer modules, and the dual-temporal feature is specifically the first... The first output feature of each module Second output features ,in The first output feature Second output features The channel dimensions are respectively , , and The spatial resolutions are respectively , , and .

[0007] In the aforementioned method for detecting changes in remote sensing images under weak supervision for land resource auditing, step B involves generating a dual-temporal-specific category mask based on the SAM3 big model, using dual-temporal acquired images and corresponding text prompts as input. The SAM3 big model is used to identify and segment targets with corresponding semantics in the images based on the input text prompts.

[0008] The aforementioned method for detecting changes in weakly supervised remote sensing images for land resource auditing, step C, involves using a dual-temporal specific category mask to perform cross-temporal semantic alignment of dual-temporal features and obtaining feature consistency loss. The specific steps are as follows: Step C1, the first output feature in the dual-temporal features Second output features The features are obtained by splicing along the spatial dimension, as shown in formula (1). (1) in, The spliced ​​features are the features after being stitched together. ,in The first Channel dimensions and spatial dimensions of each stage feature; Step C2, stitch the features Remodeling And calculate the feature affinity matrix, as shown in formula (2). (2) in, For the characteristic affinity matrix, Features after splicing The A spatial feature vector, Features after splicing The One spatial feature vector; Step C3: Generate a dual-temporal-specific category mask based on the feature affinity matrix and textual prompts, wherein the dual-temporal-specific category mask includes a first specific category mask. Second specific category mask ; Step C4: The dual-temporal specific category masks are stitched together along the spatial dimension to obtain a stitched specific category mask, wherein the stitched specific category mask... ; Step C5, stitching together the specific type of mask Reshape into vector And construct the relation labels as shown in formula (3). (3) in, For paired tags, and if paired tags Then it means the first vectors and the vectors They belong to the same category; Step C6, pairwise relationship tags and eigenaffinity matrix Cross-temporal semantic alignment is performed and feature consistency loss is obtained, as shown in Equation (4). (4) in, This is the feature consistency loss.

[0009] The aforementioned method for detecting changes in remote sensing images under weak supervision for land resource auditing, step D, involves generating pseudo-labels with confidence weights using a dual-temporal, specific-category mask. The specific steps are as follows. Step D1: Fuse the bi-temporal features with the classifier weights to generate a class activation map. Then, thresholding is used to transform the class activation graph. Binarize and adjust to Thus, a binary class activation map is obtained. ; Step D2, let Indicates image-level change labels, if Then the pseudo-tags will be Set as All-zero graph and assign confidence graph ;when At that time, a dual-temporal specific category mask is introduced to guide the generation of pseudo-labels, as shown in formula (5). (5) in, This is a coarse variation diagram, where the coarse variation diagram... ; Step D3: Extract the coarse change graph All connected components in the array are denoted as . , where each connected component For each change instance, the activation region is defined as shown in formula (6). (6) in, The final determined activation region, The index of the connected component. To change the graph The total number of connected components extracted from it; Step D4, if Then directly use the binarized class activation map. As a pseudo-label, specifically as shown in formula (7), (7); Step D5, if Then the pseudo-labels and confidence plots are as shown in formula (8). ; (8) in, , For position Pseudo-labels at the location, For position Confidence level at the location.

[0010] In the aforementioned method for detecting changes in remote sensing images under weak supervision for land resource auditing, step E involves using pseudo-labels with confidence weights to obtain the activation consistency loss for class activation graph supervision, as shown in formula (9). (9) in, To activate consistency loss, For the sigmoid function, Positions in the class activation graph The value of .

[0011] The aforementioned method for detecting changes in remote sensing images under weak supervision for land resource auditing, in step F, incorporates feature consistency loss and activation consistency loss into a weakly supervised change detection framework based on activation maps and completes the change detection operation for remote sensing images under weak supervision for land resource auditing. The weakly supervised change detection framework based on activation maps includes a feature extraction process, a feature fusion process, and a change classification process. The feature extraction process specifically uses feature consistency loss, the feature fusion process specifically uses activation consistency loss, and the change classification process specifically uses change classification loss. The change classification loss is obtained by using the binary cross-entropy between the change prediction probability output by the classifier and the corresponding image-level change label. Specifically, the detection process involves weighting and combining the change classification loss, feature consistency loss, and activation consistency loss to form the overall detection target of the weakly supervised change detection framework. Then, backpropagation is performed based on this overall detection target to optimize the parameters of the weakly supervised change detection framework, as shown in formula (10). (10) in, For the total loss, For change classification loss, and These are the feature loss coefficient and the activation loss coefficient, respectively.

[0012] A remote sensing image change detection system for weak supervision in land resource auditing includes a feature extraction module, a mask generation module, a feature consistency loss acquisition module, a pseudo-label generation module, an activation consistency loss acquisition module, and a detection output module. The feature extraction module extracts bi-temporal features from bi-temporal acquired images using a SegFormer backbone network. The mask generation module generates bi-temporal-specific category masks based on the SAM3 model, using the bi-temporal acquired images and corresponding text prompts as input. The feature consistency loss acquisition module performs cross-temporal semantic alignment of the bi-temporal features using the bi-temporal-specific category masks and obtains the feature consistency loss. The pseudo-label generation module generates pseudo-labels with confidence weights using the bi-temporal-specific category masks. The activation consistency loss acquisition module obtains the activation consistency loss for class-activated graph supervision using the pseudo-labels with confidence weights. The detection output module incorporates the feature consistency loss and activation consistency loss into a class-activated graph-based weak supervision change detection framework to complete the change detection operation for remote sensing images used in land resource auditing.

[0013] The beneficial effects of this invention are as follows: This invention provides a method and system for detecting changes in remote sensing images under weak supervision for land resource auditing. First, a SegFormer backbone network is used to extract dual-temporal features from dual-temporal acquired images. Then, based on the SAM3 large-scale model, dual-temporal acquired images and corresponding text prompts are used as input to generate a dual-temporal specific category mask. Next, the dual-temporal specific category mask is used to perform cross-temporal semantic alignment of the dual-temporal features and obtain feature consistency loss. Then, the dual-temporal specific category mask is used to generate pseudo-labels with confidence weights. Subsequently, the pseudo-labels with confidence weights are used to obtain activation consistency loss for class activation graph supervision. Finally, the feature consistency loss and activation consistency loss are combined... The loss function is incorporated into a weakly supervised change detection framework based on class activation graphs and used to perform change detection on weakly supervised remote sensing images for land resource auditing. This effectively realizes that the image change detection method and system can introduce feature consistency constraints and activation consistency constraints into the weakly supervised change detection framework based on class activation graphs. Feature consistency uses pairwise relational labels obtained from the mask to perform semantic alignment in the dual-temporal feature space, thereby alleviating category ambiguity. Activation consistency can combine semantic cues in the class activation graph with spatial priors provided by the segmentation model mask to generate pseudo-labels for class activation graph regularization, thus better recovering spatial details and improving the change recognition accuracy in land resource auditing. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of a method for detecting changes in remote sensing images under weak supervision for land resource auditing, according to the present invention. Figure 2This is a schematic diagram illustrating the working principle of a remote sensing image change detection system for land resource auditing, based on the present invention. Figure 3 This is a diagram showing the change detection results on the LEVIR-CD dataset in an embodiment of the present invention; Figure 4 This is the change detection result on the WHU-CD dataset in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will now be further described with reference to the accompanying drawings.

[0016] like Figure 1 As shown, the present invention provides a method for detecting changes in remote sensing images under weak supervision for land resource auditing, comprising the following steps: Step A involves using the SegFormer backbone network to extract dual-temporal features from dual-temporal acquired images. Specifically, this involves setting the dual-temporal images... ,in For spatial dimensions, the SegFormer backbone network consists of four Transformer modules, and the dual-temporal feature is specifically the first... The first output feature of each module Second output features ,in The first output feature Second output features The channel dimensions are respectively , , and The spatial resolutions are respectively , , and .

[0017] Step B involves generating a dual-temporal-specific category mask based on the SAM3 big model, using the dual-temporal acquired images and corresponding text prompts as input. The SAM3 big model is used to identify and segment targets with corresponding semantics in the images based on the input text prompts.

[0018] Step C involves using a dual-temporal category-specific mask to perform cross-temporal semantic alignment of the dual-temporal features and obtaining feature consistency loss. The specific steps are as follows: Step C1, the first output feature in the dual-temporal features Second output features The features are obtained by splicing along the spatial dimension, as shown in formula (1). (1) in, The spliced ​​features are the features after being stitched together. ,in The first Channel dimensions and spatial dimensions of each stage feature; Step C2, stitch the features Remodeling And calculate the feature affinity matrix, as shown in formula (2). (2) in, For the characteristic affinity matrix, Features after splicing The A spatial feature vector, Features after splicing The One spatial feature vector; Step C3: Generate a dual-temporal-specific category mask based on the feature affinity matrix and textual prompts, wherein the dual-temporal-specific category mask includes a first specific category mask. Second specific category mask ; Step C4: The dual-temporal specific category masks are stitched together along the spatial dimension to obtain a stitched specific category mask, wherein the stitched specific category mask... ; Step C5, stitching together the specific type of mask Reshape into vector And construct the relation labels as shown in formula (3). (3) in, For paired tags, and if paired tags Then it means the first vectors and the vectors They belong to the same category; Step C6, pairwise relationship tags and eigenaffinity matrix Cross-temporal semantic alignment is performed and feature consistency loss is obtained, as shown in Equation (4). (4) in, This is the feature consistency loss.

[0019] Step D involves generating pseudo-labels with confidence weights using a dual-temporal category-specific mask. The specific steps are as follows: Step D1: Fuse the bi-temporal features with the classifier weights to generate a class activation map. Then, thresholding is used to transform the class activation graph. Binarize and adjust to Thus, a binary class activation map is obtained. ; Step D2, let Indicates image-level change labels, if Then the pseudo-tags will be Set as All-zero graph and assign confidence graph ;when At that time, a dual-temporal specific category mask is introduced to guide the generation of pseudo-labels, as shown in formula (5). (5) in, This is a coarse variation diagram, where the coarse variation diagram... ; Step D3: Extract the coarse change graph All connected components in the array are denoted as . , where each connected component For each change instance, the activation region is defined as shown in formula (6). (6) in, The final determined activation region, The index of the connected component. To change the graph The total number of connected components extracted from it; Step D4, if Then directly use the binarized class activation map. As a pseudo-label, specifically as shown in formula (7), (7); Step D5, if Then the pseudo-labels and confidence plots are as shown in formula (8). ; (8) in, , For position Pseudo-labels at the location, For position Confidence level at the location.

[0020] Step E involves obtaining the activation consistency loss for class activation graph supervision using pseudo-labels with confidence weights, as shown in formula (9). (9) in, To activate consistency loss, For the sigmoid function, Positions in the class activation graph The value of .

[0021] Step F involves incorporating feature consistency loss and activation consistency loss into a weakly supervised change detection framework based on activation maps and completing the change detection operation for weakly supervised remote sensing images for land resource auditing. The weakly supervised change detection framework based on activation maps includes a feature extraction process, a feature fusion process, and a change classification process. Specifically, the feature extraction process uses feature consistency loss, the feature fusion process uses activation consistency loss, and the change classification process uses change classification loss. The change classification loss is obtained by the binary cross-entropy between the change prediction probability output by the classifier and the corresponding image-level change label. The specific detection process involves weighting and combining the change classification loss, feature consistency loss, and activation consistency loss to form the overall detection objective of the weakly supervised change detection framework. Then, backpropagation is performed based on this overall detection objective to optimize the parameters of the weakly supervised change detection framework, as shown in formula (10). (10) in, For the total loss, For change classification loss, and These are the feature loss coefficient and the activation loss coefficient, respectively.

[0022] like Figure 2 As shown, a weakly supervised remote sensing image change detection system for land resource auditing includes a feature extraction module, a mask generation module, a feature consistency loss acquisition module, a pseudo-label generation module, an activation consistency loss acquisition module, and a detection output module. The feature extraction module extracts bi-temporal features from bi-temporal acquired images using a SegFormer backbone network. The mask generation module generates bi-temporal-specific category masks based on the SAM3 model, using the bi-temporal acquired images and corresponding text prompts as input. The feature consistency loss acquisition module performs cross-temporal semantic alignment of the bi-temporal features using the bi-temporal-specific category masks and obtains the feature consistency loss. The pseudo-label generation module generates pseudo-labels with confidence weights using the bi-temporal-specific category masks. The activation consistency loss acquisition module obtains the activation consistency loss for class-activated graph supervision using the pseudo-labels with confidence weights. The detection output module incorporates the feature consistency loss and activation consistency loss into a class-activated graph-based weakly supervised change detection framework and completes the change detection operation for weakly supervised remote sensing images for land resource auditing.

[0023] To better illustrate the effectiveness of the present invention, a specific embodiment of the detection method and system of the present invention is described below.

[0024] (1) Dataset; LEVIR-CD is a large-scale benchmark dataset for building change detection. This dataset contains 637 pairs of high-resolution remote sensing images acquired by Google Earth in Texas, USA, between 2002 and 2018. The images have a spatial resolution of 0.5m and a file size of [size missing]. Pixels. In this embodiment, these images are divided into non-overlapping pixels. The image patch yielded a total of 10,192 samples. The training set, validation set, and test set each contained 7,120, 1,024, and 2,048 samples, respectively.

[0025] WHU-CD is constructed from aerial imagery acquired in Christchurch, New Zealand in 2012 and 2016. The dataset contains images of a size of [size missing]. The large-format image, with a spatial resolution of 0.075m, primarily captures changes in buildings caused by earthquakes. In this embodiment, the original image is divided into non-overlapping... Image patches. The resulting image patches were divided into 5947 training samples, 743 validation samples, and 744 test samples.

[0026] (2) Experimental setup; The method proposed in this invention is denoted as DUEL. DUEL, along with all comparison methods, is implemented in PyTorch and evaluated on an RTX 3090 GPU. DUEL is optimized using AdamW with an initial learning rate of [missing value]. All other hyperparameters remained at their default values. On LEVIR-CD, DUEL was trained for 50,000 iterations with a batch size of 16, and activation consistency loss was introduced after 5,000 iterations. On WHU-CD, DUEL was trained for 30,000 iterations with a batch size of 32, and activation consistency loss was introduced after 2,000 iterations. The CAM threshold used to generate the change map was set to 0.45. , , and The values ​​were set to 0.1, 0.001, 1, and 0.1, respectively. For both datasets, if the change map contained at least one changed pixel, the sample was assigned the image-level label "Changed"; if the change map did not contain any changed pixels, the sample was assigned the label "Unchanged". Since LEVIR-CD and WHU-CD mainly involve building changes, this paper uses three building-related cue words: "house", "building", and "roof" to consider the semantic diversity and top-down perspective features of buildings in remote sensing imagery. The resulting masks were fused into a single building-related mask using a pixel-by-pixel logical OR operation.

[0027] (3) Comparative experiment; To verify the effectiveness of DUEL, this paper compares it with FCD-GAN, TransWCD, AdvCP, ACWCD, MS-Former, and DISep. As shown in Tables 1 and 2, DUEL achieves the best overall accuracy, F1 score, and IoU on both datasets. Compared with the second-best result, DUEL improves the overall accuracy (OA), F1 score, and IoU by 0.71, 4.22, and 5.03 on LEVIR-CD, respectively; and by 0.14, 0.91, and 1.19 on WHU-CD, respectively. Specifically, on the LEVIR-CD dataset, DUEL achieves an accuracy of 69.88%, higher than all compared methods, and its F1 score and IoU reach 72.49 and 56.85, respectively, significantly outperforming other methods, indicating that it can maintain good integrity of the change region while reducing false positives. On the WHU-CD dataset, DUEL's accuracy is close to the best result, only 0.02 lower than DISep. However, its F1 score, overall precision, and intersection-over-union ratio are the highest, reaching 77.21, 98.23, and 62.87 respectively, indicating more stable overall detection performance in complex building change scenarios. This performance improvement may be attributed to DUEL's simultaneous introduction of feature consistency and activation consistency constraints: the former utilizes a mask to guide the semantic alignment of bi-temporal features, enhancing the feature separability between changed and unchanged regions; the latter combines class activation map responses with mask spatial priors to generate more reliable pseudo-labels, thereby refining the spatial boundaries of changed regions and reducing false positives and false negatives. Figure 3 and Figure 4 As shown, the change graph generated by DUEL has clearer boundaries and fewer false positives and false negatives.

[0028] Table 1. Detection results of the LEVIR-CD dataset;

[0029] Table 2. Detection results of the WHU-CD dataset;

[0030] In summary, the present invention provides a method and system for detecting changes in remote sensing images under weak supervision for land resource auditing. First, a SegFormer backbone network is used to extract bi-temporal features from bi-temporal acquired images. Then, based on the SAM3 large-scale model, bi-temporal acquired images and corresponding text prompts are used as input to generate bi-temporal specific category masks. Next, the bi-temporal specific category masks are used to perform cross-temporal semantic alignment of the bi-temporal features and obtain feature consistency loss. Then, pseudo-labels with confidence weights are generated using the bi-temporal specific category masks. Subsequently, the pseudo-labels with confidence weights are used to obtain activation consistency loss for class activation graph supervision. Finally, the feature consistency loss and activation consistency loss are summed. A weakly supervised change detection framework based on class activation graphs was introduced and used to perform change detection on weakly supervised remote sensing images for land resource auditing. This effectively enabled the method and system to incorporate feature consistency constraints and activation consistency constraints into the weakly supervised change detection framework based on class activation graphs. Feature consistency utilizes pairwise relational labels obtained from the mask to perform semantic alignment in the dual-temporal feature space, thereby alleviating category ambiguity. Activation consistency combines semantic cues in the class activation graph with spatial priors provided by the segmentation model mask, thus generating pseudo-labels for class activation graph regularization. This leads to better recovery of spatial details and improves the accuracy of change identification in land resource auditing.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting changes in remote sensing images under weak supervision for land resource auditing, characterized in that: Includes the following steps, Step A: Extract dual-temporal features from dual-temporal acquired images using the SegFormer backbone network; Step B: Based on the SAM3 large model, dual-temporal acquired images and corresponding text prompts are used as input to generate dual-temporal specific category masks; Step C: Use a dual-temporal-specific category mask to perform cross-temporal semantic alignment on dual-temporal features and obtain feature consistency loss; Step D: Generate pseudo-labels with confidence weights using a dual-temporal category-specific mask; Step E: Obtain activation consistency loss for class activation graph supervision using pseudo-labels with confidence weights; Step F involves incorporating feature consistency loss and activation consistency loss into a weakly supervised change detection framework based on class activation graphs and completing the change detection task for weakly supervised remote sensing images oriented towards land resource auditing.

2. The method for detecting changes in remote sensing images under weak supervision for land resource auditing according to claim 1, characterized in that: Step A involves using the SegFormer backbone network to extract dual-temporal features from dual-temporal acquired images. Specifically, this involves setting the dual-temporal images... ,in For spatial dimensions, the SegFormer backbone network consists of four Transformer modules, and the dual-temporal feature is specifically the first... The first output feature of each module Second output features ,in The first output feature Second output features The channel dimensions are respectively , , and The spatial resolutions are respectively , , and .

3. The method for detecting changes in remote sensing images under weak supervision for land resource auditing according to claim 2, characterized in that: Step B involves generating a dual-temporal-specific category mask based on the SAM3 big model, using the dual-temporal acquired images and corresponding text prompts as input. The SAM3 big model is used to identify and segment targets with corresponding semantics in the images based on the input text prompts.

4. The method for detecting changes in weakly supervised remote sensing images for land resource auditing according to claim 3, characterized in that: Step C involves using a dual-temporal category-specific mask to perform cross-temporal semantic alignment of the dual-temporal features and obtaining the feature consistency loss. The specific steps are as follows. Step C1, the first output feature in the dual-temporal features Second output features The features are obtained by splicing along the spatial dimension, as shown in formula (1). （1） in, The spliced ​​features are the features after being stitched together. ,in The first Channel dimensions and spatial dimensions of each stage feature; Step C2, stitch the features Remodeling And calculate the feature affinity matrix, as shown in formula (2). （2） in, For the characteristic affinity matrix, Features after splicing The A spatial feature vector, Features after splicing The One spatial feature vector; Step C3: Generate a dual-temporal-specific category mask based on the feature affinity matrix and textual prompts, wherein the dual-temporal-specific category mask includes a first specific category mask. Second specific category mask ; Step C4: The dual-temporal specific category masks are stitched together along the spatial dimension to obtain a stitched specific category mask, wherein the stitched specific category mask... ; Step C5, stitching together the specific type of mask Reshape into vector And construct the relation labels as shown in formula (3). （3） in, For paired tags, and if paired tags Then it means the first vectors and the vectors They belong to the same category; Step C6, pairwise relationship tags and eigenaffinity matrix Cross-temporal semantic alignment is performed and feature consistency loss is obtained, as shown in formula (4). （4） in, This is the feature consistency loss.

5. A method for detecting changes in weakly supervised remote sensing images for land resource auditing according to claim 4, characterized in that: Step D involves generating pseudo-labels with confidence weights using a dual-temporal category-specific mask. The specific steps are as follows: Step D1: Fuse the bi-temporal features with the classifier weights to generate a class activation map. Then, thresholding is used to transform the class activation graph. Binarize and adjust to Thus, a binary class activation map is obtained. ; Step D2, let Indicates image-level change labels, if Then the pseudo-tags will be Set as All-zero graph and assign confidence graph ;when At that time, a dual-temporal specific category mask is introduced to guide the generation of pseudo-labels, as shown in formula (5). （5） in, This is a coarse variation diagram, where the coarse variation diagram... ; Step D3: Extract the coarse change graph All connected components in the array are denoted as . , where each connected component For each change instance, the activation region is defined as shown in formula (6). （6） in, The final determined activation region, The index of the connected component. To change the graph The total number of connected components extracted from it; Step D4, if Then directly use the binarized class activation map. As a pseudo-label, specifically as shown in formula (7), （7）； Step D5, if Then the pseudo-labels and confidence plot are as shown in formula (8). ； （8） in, , For position Pseudo-labels at the location, For position Confidence level at the location.

6. A method for detecting changes in weakly supervised remote sensing images for land resource auditing according to claim 5, characterized in that: Step E involves obtaining the activation consistency loss for class activation graph supervision using pseudo-labels with confidence weights, as shown in formula (9). （9） in, To activate consistency loss, For the sigmoid function, Positions in the class activation graph The value of .

7. A method for detecting changes in remote sensing images under weak supervision for land resource auditing as described in claim 6, characterized in that: Step F involves incorporating feature consistency loss and activation consistency loss into a weakly supervised change detection framework based on activation maps and completing the change detection operation for weakly supervised remote sensing images for land resource auditing. The weakly supervised change detection framework based on activation maps includes a feature extraction process, a feature fusion process, and a change classification process. Specifically, the feature extraction process uses feature consistency loss, the feature fusion process uses activation consistency loss, and the change classification process uses change classification loss. The change classification loss is obtained by the binary cross-entropy between the change prediction probability output by the classifier and the corresponding image-level change label. The specific detection process involves weighting and combining the change classification loss, feature consistency loss, and activation consistency loss to form the overall detection objective of the weakly supervised change detection framework. Then, backpropagation is performed based on this overall detection objective to optimize the parameters of the weakly supervised change detection framework, as shown in formula (10). （10） in, For the total loss, For change classification loss, and These are the feature loss coefficient and the activation loss coefficient, respectively.

8. A remote sensing image change detection system for weak supervision in land resource auditing, wherein the specific detection process of the remote sensing image change detection system is based on the remote sensing image change detection method for weak supervision as described in any one of claims 1-7, characterized in that: It includes a feature extraction module, a mask generation module, a feature consistency loss acquisition module, a pseudo-label generation module, an activation consistency loss acquisition module, and a detection output module. The feature extraction module is used to extract dual-temporal features from dual-temporal acquired images using the SegFormer backbone network. The mask generation module is used to generate a dual-temporal-specific category mask based on the SAM3 large model, taking dual-temporal acquired images and corresponding text prompts as input. The feature consistency loss acquisition module is used to perform cross-temporal semantic alignment of dual-temporal features using a dual-temporal specific category mask and obtain feature consistency loss. The pseudo-label generation module is used to generate pseudo-labels with confidence weights using a dual-temporal specific category mask; The activation consistency loss acquisition module is used to obtain the activation consistency loss for class activation graph supervision using pseudo-labels with confidence weights. The detection output module is used to incorporate feature consistency loss and activation consistency loss into a weakly supervised change detection framework based on activation graphs and complete the change detection operation of weakly supervised remote sensing images for land resource auditing.

Images