A remote sensing sample generation and incremental learning method based on vector variation driving

Abstract

The application discloses a remote sensing sample generation and incremental learning method based on vector change driving, and belongs to the technical field of artificial intelligence and remote sensing image processing. The method first acquires a remote sensing image to be processed and performs vectorization processing to obtain a vector extraction result of a current time phase, compares the vector extraction result with a historical vector baseline in space, and generates a ground surface change area graph patch; importance of the change area is evaluated to form a sample expansion priority list, and an incremental training sample set is collected or constructed according to the sample expansion priority list; subsequently, the incremental training sample set is input into a pre-trained remote sensing model to perform incremental learning updating; finally, performance of the updated model is evaluated, and the historical vector baseline is updated or sample generation and model updating are triggered again according to the evaluation result. The application can automatically drive sample generation and model adaptive updating by using ground surface vector change, and improves interpretation accuracy and continuous adaptation capability of the remote sensing model in a time-varying scene.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and remote sensing image processing technology, specifically involving a remote sensing sample generation and incremental learning method based on vector change-driven method, which is applicable to the identification of surface change areas, adaptive construction of training samples, and continuous updating of remote sensing interpretation models in time-varying remote sensing scenarios. Background Technology

[0002] Intelligent interpretation of remote sensing imagery is a crucial technical means in land resource surveys, urban planning, disaster monitoring, and ecological environment supervision. With the development of deep learning technology, methods for target detection, semantic segmentation, land cover classification, and vector extraction based on remote sensing imagery have been widely applied. Existing remote sensing interpretation models typically rely on offline training with a large number of labeled samples. After training, they are deployed to real-world scenarios to automatically identify and vectorize new remote sensing images.

[0003] However, remote sensing applications exhibit significant time-varying characteristics. Influenced by factors such as seasonal changes, urban development, road expansion, crop growth, disaster damage, and differences in sensor imaging conditions, the distribution of ground features, target morphology, and image characteristics of the same area can vary significantly across different time periods. When the actual scene differs from the distribution of training samples, deployed models are prone to problems such as decreased recognition accuracy, inaccurate boundary extraction, and increased false positives and false negatives.

[0004] To address the aforementioned issues, existing technologies primarily employ manual sample addition, data relabeling, and periodic model retraining for updates. This approach relies on human experience, making it difficult to promptly identify areas of change, resulting in low sample expansion efficiency and long model update cycles, which fails to meet the application requirements of large-scale, continuously changing remote sensing scenarios. Furthermore, existing sample generation methods often utilize random cropping, conventional data augmentation, or manual selection of difficult samples, failing to fully leverage the relationship between historical vector baselines and current temporal vector results. This makes it difficult to effectively convert newly added areas, disappearing areas, boundary offset areas, and areas with attribute changes into incremental training samples.

[0005] Furthermore, existing incremental learning methods are prone to knowledge forgetting in remote sensing interpretation tasks. When learning new samples in changed regions, the model may weaken its ability to identify stable regions, lacking a joint constraint and evaluation mechanism for adaptability to changed regions and stability preservation. Therefore, there is an urgent need for a method that can automatically drive sample generation using vector changes and combine incremental learning to achieve continuous updates of remote sensing models, in order to improve the interpretation accuracy and adaptability of models in time-varying scenarios. Summary of the Invention

[0006] This invention provides a remote sensing sample generation and incremental learning method based on vector change-driven approach. The method spatially compares the vector extraction results of the current temporal remote sensing image with the historical vector baseline to identify surface change areas, generates a sample expansion priority list based on the importance of the change areas, further constructs an incremental training sample set, and performs incremental learning and updating on the pre-trained remote sensing model, thereby achieving adaptive optimization of the remote sensing interpretation model in time-varying scenarios.

[0007] To achieve the above objectives, this invention provides a method for remote sensing sample generation and incremental learning based on vector transformation, comprising the following steps: Step S1: Obtain the remote sensing image data to be processed, and perform vectorization processing on the remote sensing image data to obtain the vector extraction result of the current time phase; Step S2: Spatially compare the vector extraction results of the current time phase with the historical vector baseline, calculate the vector geometric differences and attribute differences, and generate surface change area patches; Step S3: Based on the surface change area patches, assess the importance of the change areas and generate a sample expansion priority list; Step S4: According to the sample expansion priority list, collect or construct training samples from the change area to form an incremental training sample set; Step S5: Input the incremental training sample set into the pre-trained remote sensing model, perform incremental learning and update the model, and obtain the updated remote sensing interpretation model. Step S6: Evaluate the performance of the updated remote sensing interpretation model, and update the historical vector baseline or trigger sample expansion and model update again based on the evaluation results.

[0008] Further, step S1 specifically includes: Step S1.1: Acquire remote sensing image data of the current time phase and preprocess the remote sensing image data. The preprocessing includes radiometric correction, geometric correction, image registration, spatial reference unification, cropping and normalization. Step S1.2: Input the preprocessed remote sensing image into the pre-trained remote sensing interpretation model to obtain the recognition results of ground objects. The recognition results include target detection results, semantic segmentation results, or instance segmentation results. Step S1.3: Extract boundaries and vectorize the recognition results to obtain a set of vector objects for the current time phase; Step S1.4: Assign category attributes, spatial boundaries, area information, spatial location information, and model prediction confidence to each vector object in the vector object set to form the vector extraction result for the current time phase.

[0009] Further, step S2 specifically includes: Step S2.1: Unify the vector extraction results of the current time phase and the historical vector baseline to the same spatial coordinate system, spatial reference and vector expression format to ensure that the two can be spatially superimposed and difference calculated; Step S2.2: Perform spatial overlay analysis on the current time-phase vector object and the historical vector object to determine the overlapping area, disappearing area, newly added area and boundary offset area between them; Step S2.3: Calculate the geometric differences, attribute differences, and spatial topological differences of the vector objects to obtain the region-level vector change characteristics; wherein, the geometric differences include area changes, boundary offsets, and shape differences, the attribute differences include category changes and attribute value changes, and the spatial topological differences include changes in adjacency relationships, containment relationships, or connectivity relationships. Step S2.4: Calculate the change confidence level based on the regional vector change characteristics. When the change confidence level is greater than the preset change threshold, the corresponding region is determined as a surface change region patch. Step S2.5: Use the surface change area patches as candidate regions for subsequent sample expansion and incremental learning.

[0010] Furthermore, step S3 specifically includes: Step S3.1: For each land surface change area patch, extract its change type, patch area, boundary offset, attribute change intensity, model prediction confidence, and historical error records to construct a change area feature vector; Step S3.2: Based on the feature vector of the changed area, calculate the geometric change score, semantic change score, model uncertainty score and historical error score respectively. The geometric change score is used to characterize the degree of spatial morphological change, and the semantic change score is used to characterize the degree of change of land cover category or attribute. Step S3.3: Perform a weighted fusion of the geometric change score, semantic change score, model uncertainty score, and historical error score to obtain the importance score of the changed region; Step S3.4: Based on the importance score of the changed areas, sort all the surface changed area patches and generate a sample expansion priority list; Step S3.5: According to the sample expansion priority list, select the change regions with importance scores higher than the preset threshold as sample generation candidate regions, and assign the number of samples to be collected and the corresponding incremental learning priority to the candidate regions.

[0011] Furthermore, step S4 specifically includes: Step S4.1: Read the candidate change regions sequentially according to the sample expansion priority list, and construct a sample clipping window containing the change region and its surrounding neighborhood based on the spatial position or geometric center of the candidate change region patch. Step S4.2: Based on the area, shape complexity, and boundary offset of the candidate change region patch, adaptively adjust the scale of the sample cropping window so that the sample cropping window simultaneously includes the change target region, the boundary transition region, and the stable background region. Step S4.3: Map the sample cropping window to the corresponding remote sensing image data for cropping to obtain the corresponding image sample, and generate the corresponding sample label based on the surface change area patch and historical vector baseline; Step S4.4: Divide the image samples into newly added target samples, disappeared target samples, boundary change samples, attribute change samples, and stable background samples according to the change type; Step S4.5: Perform quality screening on the image samples and remove samples that are severely obscured by clouds and fog, have registration errors exceeding a preset threshold, or have inconsistent labels. Step S4.6: Based on the category distribution, the importance score of the change region, and the confidence of the model prediction, resample and hard example mining are performed on the screened samples to form an incremental training sample set for incremental learning.

[0012] Furthermore, step S5 specifically includes: Step S5.1: Select historical samples from the historical sample library that are related to the incremental training sample set in terms of category distribution or spatial distribution, and construct a historically preserved sample set; Step S5.2: Mix the incremental training sample set with the historical retention sample set according to a preset ratio to form an incremental learning training set; Step S5.3: Using the pre-trained remote sensing interpretation model as the initial model, iteratively train the incremental learning training set and update the model parameters; Step S5.4: During model training, different training weights are assigned to newly added target samples, disappearing target samples, boundary change samples, and attribute change samples to enhance the model's ability to learn about changing regions. Step S5.5: Construct history preservation constraints to suppress the forgetting phenomenon of the model in the incremental learning process by ensuring the consistency of the model's output in the stable region before and after the update. Step S5.6: Construct adaptation constraints for changing regions. By optimizing the model's prediction error in changing regions, the model's ability to adapt to new changing patterns is improved. Step S5.7: Construct a joint incremental learning loss function based on the historical preservation constraint and the change area adaptation constraint, and optimize and update the model based on the joint loss function to obtain the updated remote sensing interpretation model.

[0013] Furthermore, step S6 specifically includes: Step S6.1: Apply the updated remote sensing interpretation model to the validation sample set of the changed region and the validation sample set of the stable region respectively, and obtain the corresponding model output results; Step S6.2: Based on the model output results of the verification sample set of the changed region, calculate the adaptation performance index of the changed region, including the vector extraction accuracy of the changed region, the boundary matching accuracy, and the category recognition accuracy. Step S6.3: Based on the model output results of the stable region validation sample set, calculate the stable region preservation performance index, including stable region vector extraction accuracy, false detection rate, and false negative rate; Step S6.4: Compare the performance metrics of the model before and after the update in the changing and stable regions to obtain the performance improvement in the changing region and the performance decline in the stable region. Step S6.5: When the performance improvement of the changed region is greater than the first preset threshold and the performance decline of the stable region is less than the second preset threshold, the updated model is determined as a deployable model, and its output results are used to update the historical vector baseline. Step S6.6: When the performance improvement of the changed region does not reach the first preset threshold or the performance decline of the stable region is greater than the second preset threshold, retain the model before the update, and regenerate the sample expansion priority list based on the failed region to trigger a new round of sample generation and incremental learning process. Step S6.7: Store the changed region patches, incremental training sample set, and model performance evaluation results in the historical record library during this model update process for subsequent sample generation and model optimization strategy matching in similar change scenarios.

[0014] Beneficial Effects: This invention introduces a vector change-driven remote sensing sample generation and incremental learning method. It automatically identifies surface change areas by utilizing the change relationship between the current temporal vector extraction results and historical vector baselines, thereby reducing reliance on manual annotation and sample selection, and improving the automation and accuracy of change area acquisition. Simultaneously, by constructing a change area importance assessment mechanism that integrates geometric changes, semantic changes, model uncertainty, and historical errors, it prioritizes change areas, making sample generation more targeted and improving sample utilization efficiency. Furthermore, through diversified sample construction methods based on change types, it identifies newly added targets, disappearing targets, and... Information such as boundary changes and attribute changes is transformed into structured training samples, enhancing the model's expressive and adaptive capabilities to complex time-varying remote sensing scenarios. Furthermore, by introducing a joint incremental learning mechanism that combines history preservation constraints and change region adaptation constraints, the model's ability to identify change regions is improved while effectively suppressing the forgetting of old knowledge, thereby enhancing the model's continuous learning and generalization performance. By establishing a dual-domain performance evaluation mechanism for change regions and stable regions, as well as a model deployment decision and historical vector baseline update strategy, a closed-loop optimization process is formed from change perception, sample generation to model update, thus significantly improving the remote sensing interpretation model's continuous adaptive capability and overall interpretation accuracy in time-varying scenarios. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a flowchart of data acquisition and vectorization processing in an embodiment of the present invention; Figure 3 This is a flowchart of vector change detection and surface change area generation in an embodiment of the present invention; Figure 4 This is a flowchart of the process for assessing the importance of changing regions and generating sample priorities in an embodiment of the present invention; Figure 5 This is a flowchart of the change-driven sample generation and incremental training sample construction in an embodiment of the present invention; Figure 6 This is a flowchart of the incremental learning model update and anti-forgetting training process in an embodiment of the present invention; Figure 7 This is a flowchart of model evaluation and closed-loop update in an embodiment of the present invention. Detailed Implementation

[0016] The invention will now be further described with reference to the accompanying drawings. Example

[0017] like Figure 1 As shown, a remote sensing sample generation and incremental learning method based on vector transformation driving includes the following steps: Step S1: Obtain the remote sensing image data to be processed, and perform vectorization processing on the remote sensing image data to obtain the vector extraction result of the current time phase; Step S2: Spatially compare the vector extraction results of the current time phase with the historical vector baseline, calculate the vector geometric differences and attribute differences, and generate surface change area patches; Step S3: Based on the surface change area patches, assess the importance of the change areas and generate a sample expansion priority list; Step S4: According to the sample expansion priority list, collect or construct training samples from the change area to form an incremental training sample set; Step S5: Input the incremental training sample set into the pre-trained remote sensing model, perform incremental learning and update the model, and obtain the updated remote sensing interpretation model. Step S6: Evaluate the performance of the updated remote sensing interpretation model, and update the historical vector baseline or trigger sample expansion and model update again based on the evaluation results.

[0018] Furthermore, such as Figure 2 As shown, step S1 specifically includes: Step S1.1: Acquire remote sensing image data of the current time phase and preprocess the remote sensing image data. The preprocessing includes radiometric correction, geometric correction, image registration, spatial reference unification, and normalization to eliminate imaging errors and unify data scales. The preprocessing process can be expressed as follows: ; in, Represents the original remote sensing image. This represents a preprocessing function that includes radiometric correction, geometric correction, image registration, and normalization. This indicates the preprocessed image; Step S1.2: Input the preprocessed remote sensing image into the pre-trained remote sensing interpretation model to obtain the recognition results of ground targets. The recognition results include target detection results, semantic segmentation results, or instance segmentation results; their expression form is as follows: ; in, This represents a pre-trained remote sensing interpretation model. This represents the rasterized prediction result output by the model; Step S1.3: Perform boundary extraction processing on the recognition results and convert the pixel-level results into vector polygon structures to obtain an initial vector object set. The pixel-level recognition results are one or more of the following: target detection results, semantic segmentation results, or instance segmentation results; the vectorization process is represented as follows: ; in, This represents a vectorized transformation function based on contour extraction and polygon fitting. A collection of vector objects representing the current time phase; Step S1.4: Assign attribute values ​​and perform structured expression on the vector object set to obtain the vector extraction result of the current time phase; the vector object set can be represented as: ; in, The spatial boundary of the i-th vector object is represented by a polygon or polyline. Represents category attributes, This represents the area information of the corresponding vector object. This indicates the confidence level of the model prediction. Indicates the number of vector objects; Furthermore, such as Figure 3 As shown, step S2 specifically includes: Step S2.1: Obtain the current phase vector extraction result and historical vector baseline data, and unify them to the same spatial coordinate system, spatial reference, and vector expression format to complete the vector data standardization process, so as to ensure the consistency of subsequent spatial comparisons. The standardization process can be represented as follows: ; in, Represents the current time phase vector set. Represents the historical vector baseline. Represents a spatial coordinate transformation function; Step S2.2: Perform spatial matching and overlay analysis on the standardized current time-phase vector set and the historical vector set. First, match the vector objects one by one based on spatial distance or intersection-over-union ratio (IoU) to obtain the corresponding set of vector object pairs, which is represented as: ; in, This represents the i-th vector object in the current time phase. This represents the i-th vector object that matches the historical data; spatial relationship analysis is performed on the matched vector object pairs to obtain the following change regions: ; ; ; in, Indicates a newly added area. Indicates the area where the message disappears. The overlapping areas are represented by the intersection and union operation, which calculates the IoU value of all matching vector object pairs one by one to obtain the global set of surface change areas, which serves as the input data for the subsequent generation of change patches. Step S2.3: Calculate the geometric differences, attribute differences, and topological differences of the vector object to construct a regional-level change feature representation. Geometric differences describe the degree of spatial morphological change, attribute differences describe category changes, and topological differences describe changes in spatial structural relationships. The comprehensive difference representation is as follows: ; in, This represents the geometric differences calculated based on the rate of change of area, boundary offset, and shape similarity. This represents the attribute difference calculated based on the probability of category change. This represents the topological differences calculated based on changes in adjacency and containment relationships. These are the weighting coefficients; Step S2.4: Calculate the change confidence score based on the comprehensive differences. The change confidence score is used to characterize whether real surface changes have occurred in the current area. The calculation formula is as follows: ; in, This represents the Sigmoid normalization function, used to map difference values ​​to the interval [0,1], when the confidence level changes. Greater than the preset threshold When the area is identified as a land surface change area, the preset change threshold is determined in advance based on the strictness of the determination of land surface change areas. The higher the threshold, the stricter the requirement for the area to be identified as a land surface change area. Step S2.5: Spatial clustering and region merging are performed on the regions that meet the change conditions. The clustering process is based on spatial distance or adjacency relationship to perform connected component analysis in order to eliminate fragmented change regions, generate the final set of surface change region patches, and use them as candidate inputs for subsequent sample generation and incremental learning.

[0019] Furthermore, such as Figure 4 As shown, step S3 specifically includes: Step S3.1: For each land surface change area patch, extract its multidimensional change features. The multidimensional change features include the degree of geometric change, the degree of semantic change, model prediction uncertainty, and historical error information, and construct a change area feature vector, represented as follows: ; in, Indicates geometric change characteristics, Indicates semantic change features, This indicates the uncertainty in the model's predictions. Indicates historical error information; Step S3.2: Calculate the geometric change features to characterize the degree of change in the regional spatial morphology. The calculation method includes area change rate, boundary offset distance, and shape similarity, expressed as: ; in, Indicates the rate of change of area. Indicates the boundary offset distance. This represents the intersection-union ratio (IU / U) between the current and historical vectors. These represent the weighting coefficients corresponding to the area change rate, boundary offset distance, and shape difference, respectively, and are used to adjust the contribution of each geometric change index to the geometric change characteristics. Step S3.3: Calculate the semantic change features to characterize changes in land cover categories or attributes, expressed as follows: ; in, This indicates the probability that the current time phase matches the category of a historical time phase. Step S3.4: Calculate the model prediction uncertainty to characterize the model's confidence level in predicting the current changing region. The calculation method is the prediction probability entropy: ; in, Indicates the first The predicted probability of a class, where K represents the total number of classes.

[0020] Step S3.5: Statistically analyze historical error information to characterize the model's performance in historically similar regions. The calculation method is as follows: ; in, This represents the true labeled value corresponding to the k-th historical sample, where the true labeled value is the supervision label information stored corresponding to that historical sample; This represents the model prediction value obtained by the remote sensing interpretation model for the k-th historical sample. The model prediction value is obtained by inferring from the corresponding historical sample into the current remote sensing interpretation model; N represents the number of historical samples participating in the historical error statistics. Step S3.6: Based on the geometric change features, semantic change features, model uncertainty, and historical error information, a weighted fusion is performed to obtain the importance score of the changed region, expressed as: ; in, These are weighting coefficients, which can be dynamically adjusted according to the training process and satisfy normalization constraints. ; Step S3.7: Sort all changed region patches according to the importance score, generate a sample expansion priority list, and select changed regions with importance scores higher than a preset threshold as candidate regions for sample generation based on the sample expansion priority list. Assign corresponding sample collection quantity and incremental learning priority to the candidate regions. The preset threshold is used to select changed regions that meet the sample expansion requirements and is predetermined based on the distribution of the importance scores of the changed regions and the sample expansion quantity requirements.

[0021] Furthermore, such as Figure 5 As shown, step S4 specifically includes: Step S4.1: According to the sample expansion priority list, sort and read the surface change area patches, and use the spatial geometric center of the patch or the center of the smallest bounding rectangle as the sample generation center point, which is expressed as: ; in, This indicates the center position of the i-th changed patch. represents the spatial coordinate points that constitute the boundary of the patch, and n represents the number of spatial coordinate points that constitute the boundary of the i-th changed patch. Step S4.2: Construct an adaptive clipping window based on the center point. The scale of the clipping window is dynamically adjusted according to the area and shape complexity of the changing patch. The calculation method is as follows: ; in, Indicates the size of the clipping window. Indicates the area of ​​the patch. Indicates the perimeter of the boundary or the complexity of the shape. These are the weighting coefficients; Step S4.3: Map the cropping window onto the original remote sensing image for sample cropping to obtain corresponding image samples, and generate supervision labels based on the surface change area patches and historical vector baselines. The label generation function is expressed as: ; in, Indicates sample label, Indicates the current vector. Represents the historical vector. This represents the tag generation mapping function; Step S4.4: Classify the image samples according to the type of change, including newly added target samples, disappeared target samples, boundary change samples, attribute change samples, and stable background samples, to form a structured sample set: ; Step S4.5: Perform quality screening on the image samples, removing those with severe cloud and fog obscuration, geometric registration errors exceeding a preset threshold, or inconsistent labels. The screening criteria are as follows: ; in, Indicates sample error. To preset the quality threshold, This is an indicator function, which takes the value of 1 when the condition in parentheses is true, and takes the value of 0 otherwise; the quality threshold is used to limit the maximum allowable geometric registration error of the image sample, and it is predetermined based on the spatial resolution of the remote sensing image, the image registration accuracy, and the spatial alignment requirements of the sample label. Step S4.6: Based on the sample importance score, resample and weight the screened samples, and the sampling probability is defined as: ; in, S3 represents the sample importance score calculated in S3, where N represents the total number of samples; Step S4.7: Summarize the processed sample set to form an incremental training sample set, and input it into the incremental learning module for subsequent model update training.

[0022] Furthermore, such as Figure 6 As shown, step S5 specifically includes: Step S5.1: Select historical samples from the historical sample library that are related to the incremental training sample set in terms of category distribution or spatial distribution, and construct a historical preservation sample set to maintain the model's recognition ability in stable regions.

[0023] Step S5.2: Merge the incremental training sample set with the historical stored sample set according to a preset ratio to form an incremental learning training set, which is represented as follows: ; in, This represents the incremental training sample set. This indicates that the historical sample set is preserved. >0 indicates a preset ratio coefficient when fusing the incremental training sample set and the historically maintained sample set. Based on the number of incremental training samples, the ratio of the number of incremental training samples to the number of historically maintained samples is... ; This refers to the historical samples selected from the historical sample set and participating in the fusion according to the preset ratio; Step S5.3: Using the pre-trained remote sensing interpretation model as the initial model, iteratively train the fused training set and update the model parameters. The optimization process is expressed as follows: ; in, Indicates model parameters, Represents the overall loss function; Step S5.4: Construct a change region learning loss function to enhance the model's ability to learn new change regions, which is expressed as: ; in, This represents the true label corresponding to the i-th incremental training sample. This represents the model prediction result obtained by the remote sensing interpretation model for the i-th incremental training sample, and N represents the number of incremental training samples involved in the calculation of the change area learning loss. Step S5.5: Construct a history-preserving constraint loss function to suppress the forgetting phenomenon in the incremental learning process of the model, which is expressed as: ; in, This represents the i-th historically preserved sample in the historically preserved sample set. This represents the output of the remote sensing interpretation model before the update, obtained for the i-th historical preserved sample. This represents the updated remote sensing interpretation model output for the i-th historically preserved sample, and M represents the number of historically preserved samples involved in the historical preservation loss calculation. Step S5.6: Construct a change region adaptation loss function to improve the model's ability to discriminate change regions, which is expressed as: ; in, This represents the model's predicted probability for the i-th type of change region, where K represents the number of categories; Step S5.7: Construct a joint loss function based on the changed region learning loss, history preservation loss, and change adaptation loss, and optimize and update the model, which is expressed as: ; in, The weights are used to balance stability and adaptability. The updated remote sensing interpretation model is obtained through iterative optimization and output to the model evaluation module for subsequent performance verification and closed-loop update.

[0024] Furthermore, such as Figure 7 As shown, step S6 specifically includes: Step S6.1: Apply the updated remote sensing interpretation model to the change area validation sample set and the stable area validation sample set respectively to obtain the corresponding prediction results. The change area validation sample set is used to evaluate the model's ability to adapt to new changes in targets, and the stable area validation sample set is used to evaluate the model's ability to maintain the original structure.

[0025] Step S6.2: Calculate the adaptation performance index of the changed region based on the validation sample set of the changed region, which is expressed as: ; in, represents the intersection-union ratio (IU / U) of the i-th sample in the changed region, used to measure the model's accuracy in recognizing the changed region, and N represents the number of samples in the validation sample set of the changed region.

[0026] Step S6.3: Calculate the stable region preservation performance index based on the stable region validation sample set, which is expressed as: ; in, This represents the true annotation result corresponding to the j-th sample in the stable region validation sample set. This represents the model prediction result obtained by the updated remote sensing interpretation model for the j-th stable region validation sample, M represents the number of samples in the stable region validation sample set, and the stable region preservation performance index is used to measure the model's ability to preserve stable regions; Step S6.4: Compare and analyze the performance indicators of the model before and after the update to obtain the performance improvement in the changed region and the performance decline in the stable region, which are expressed as follows: ; ; in, , These represent the performance metrics of the updated model on the validation set in the changed region and the validation set in the stable region, respectively. , These represent the performance metrics of the corresponding regions in the model before the update; Indicates the extent of performance improvement in the changed region. Indicates the magnitude of performance change in the stable region; Step S6.5: The first preset threshold represents the minimum performance improvement required in the changing region, which is predetermined based on the performance improvement requirements on the validation sample set of the changing region; the second preset threshold represents the maximum performance degradation allowed in the stable region, which is predetermined based on the performance maintenance requirements on the validation sample set of the stable region; when the performance improvement in the changing region is greater than the first preset threshold and the performance degradation in the stable region is less than the second preset threshold, the updated model is determined as a deployable model, and the historical vector baseline is updated using the output of the model. Step S6.6: When the performance improvement of the changed region does not reach the first preset threshold or the performance decline of the stable region is greater than the second preset threshold, retain the model before the update, and regenerate the sample expansion priority list based on the underperforming region to trigger a new round of sample generation and incremental learning process. Step S6.7: Store the changed region patches, incremental training sample set and model evaluation results during this model update process for future reference in optimizing sample generation strategies and updating models in similar changing scenarios. Furthermore, to illustrate the technical advantages of the present invention over existing methods, in a specific embodiment, high-resolution remote sensing images of the same region at different time phases and their corresponding historical vector baselines are selected as test objects to construct a change area verification sample set and a stable area verification sample set. The change area verification sample set includes changed features such as newly added buildings, road expansion, farmland changes, and boundary shifts; the stable area verification sample set includes unchanged features such as buildings, roads, water bodies, and forest land. Using the same pre-trained remote sensing interpretation model as the initial model, and maintaining consistency in input image resolution, training rounds, learning rate, and validation set partitioning, the pre-update remote sensing interpretation model, the method of manually supplementing samples and overall fine-tuning, the method of randomly cropping samples and conventional incremental training, the method of mining difficult samples and incremental fine-tuning, and the method of the present invention are compared and verified. The comparison results are shown in Table 1: Table 1. Quantitative comparison results between the method of the present invention and existing methods:

[0027] As shown in Table 1, under the same validation conditions, compared with the original remote sensing interpretation model, the method of this invention improves the accuracy of vector extraction of changed regions by 15.3 percentage points, the accuracy of boundary matching by 14.7 percentage points, and the accuracy of category recognition by 14.2 percentage points; compared with the random sample cropping and conventional incremental training methods, the accuracy of vector extraction of changed regions is improved by 6.7 percentage points, the accuracy of boundary matching by 7.1 percentage points, the accuracy of category recognition by 6.7 percentage points, and the accuracy of maintaining stable regions by 3.6 percentage points; compared with the manual sample supplementation and overall fine-tuning methods, the sample construction time is reduced from 42.0 h to 5.6 h, a reduction of approximately 86.7%.

[0028] Furthermore, on a stable region validation sample set, the false positive rate of the method of this invention is 4.8%, and the false negative rate is 5.6%, which is lower than the 8.4% and 8.8% of the random sample pruning method and the conventional incremental training method, respectively, and also lower than the 6.9% and 7.3% of the difficult sample mining and incremental fine-tuning method, respectively. These results demonstrate that the present invention can improve the adaptation performance of changing regions while maintaining the interpretation performance of stable regions, reduce the cost of manual sample construction, and verify the comprehensive technical effectiveness of vector change-driven sample generation, changing region importance ranking, history preservation constraints, and closed-loop update mechanism.

[0029] The above results are only used to illustrate the technical effects of the present invention and do not limit specific application scenarios or performance value ranges.

Claims

1. A remote sensing sample generation and incremental learning method based on vector transformation driven, characterized in that, Includes the following steps: Step S1: Obtain the remote sensing image data to be processed, and perform vectorization processing on the remote sensing image data to obtain the vector extraction result of the current time phase; Step S2: Spatially compare the vector extraction results of the current time phase with the historical vector baseline, calculate the vector geometric differences and attribute differences, and generate surface change area patches; Step S3: Based on the surface change area patches, assess the importance of the change areas and generate a sample expansion priority list, specifically including: Step S3.1: For each land surface change area patch, extract its change type, patch area, boundary offset, attribute change intensity, model prediction confidence, and historical error records to construct a change area feature vector; Step S3.2: Calculate the geometric change score, semantic change score, model uncertainty score, and historical error score based on the feature vector of the changed region. Step S3.3: Perform a weighted fusion of the geometric change score, semantic change score, model uncertainty score, and historical error score to obtain the importance score of the changed region; Step S3.4: Based on the importance score of the changed areas, sort all the surface changed area patches in descending order to obtain a sample expansion priority list; Step S3.5: According to the sample expansion priority list, select the change regions with importance scores higher than the preset threshold as candidate regions for training samples, and assign the number of samples collected, sample category labels and incremental learning priority to each candidate region. Step S4: According to the sample expansion priority list, collect or construct training samples from the change area to form an incremental training sample set; Step S5: Input the incremental training sample set into the pre-trained remote sensing model, perform incremental learning and update the model, and obtain the updated remote sensing interpretation model. Step S6: Evaluate the performance of the updated remote sensing interpretation model, and update the historical vector baseline or trigger sample augmentation and model update again based on the evaluation results. Specifically, this includes: Step S6.1: Apply the updated remote sensing interpretation model to the validation sample set of the changed region and the validation sample set of the stable region respectively to obtain the remote sensing interpretation results of the updated model in the changed region and the stable region. Step S6.2: Based on the remote sensing interpretation results of the verification sample set of the changed region, calculate the adaptation performance index of the changed region. The adaptation performance index of the changed region includes the vector extraction accuracy of the changed region, the boundary matching accuracy, and the category recognition accuracy. Step S6.3: Based on the remote sensing interpretation results of the stable region verification sample set, calculate the stable region preservation performance index, which includes the stable region vector extraction accuracy, false detection rate, and false negative rate. Step S6.4: Compare the performance indicators of the updated remote sensing interpretation model with those of the unupdated remote sensing interpretation model in the changed and stable regions to obtain the performance improvement in the changed regions and the performance decline in the stable regions. Step S6.5: When the performance improvement of the changed area is greater than the first preset threshold and the performance decline of the stable area is less than the second preset threshold, the updated remote sensing interpretation model is determined as a deployable model, and the historical vector baseline is updated using the interpretation results of the updated remote sensing interpretation model. Step S6.6: When the performance improvement of the changed area does not reach the first preset threshold, or the performance decline of the stable area is greater than the second preset threshold, retain the remote sensing interpretation model before the update, and regenerate the sample expansion priority list according to the failed evaluation area to trigger sample generation and incremental learning update again. Step S6.7: Store the surface change area patches, incremental training sample set, and model performance evaluation results in the historical record library during this model update process for subsequent sample generation and incremental learning strategy matching in similar change scenarios.

2. The remote sensing sample generation and incremental learning method based on vector transformation driven according to claim 1, characterized in that, Step S2 specifically includes: Step S2.1: Unify the vector extraction results of the current time phase and the historical vector baseline to the same spatial coordinate system, spatial reference and vector expression format; Step S2.2: Perform spatial overlay analysis on the current time-phase vector object and the historical vector object to determine the overlapping area, disappearing area, newly added area and boundary offset area between them; Step S2.3: Calculate the geometric differences, attribute differences, and spatial topological differences of the vector objects respectively to obtain the regional vector change characteristics; Step S2.4: Calculate the change confidence level based on the regional vector change characteristics. When the change confidence level is greater than the preset change threshold, the corresponding region is determined as a surface change region patch. Step S2.5: Use the surface change area patches as candidate regions for subsequent sample expansion and incremental learning.

3. The remote sensing sample generation and incremental learning method based on vector transformation driven according to claim 1, characterized in that, Step S4 specifically includes: Step S4.1: Read the candidate change regions sequentially according to the sample expansion priority list, and construct a sample clipping window containing the change region and its neighborhood background based on the bounding rectangle or center point of the candidate change region patch. Step S4.2: Based on the area, shape complexity, and boundary offset of the candidate change region patch, adaptively adjust the scale of the sample clipping window so that the sample clipping window simultaneously includes the change target region, the boundary transition region, and the surrounding unchanged region. Step S4.3: Map the sample cropping window to the corresponding remote sensing image area, crop the image sample, and generate the corresponding sample label based on the surface change area patch and historical vector baseline; Step S4.4: Divide the image samples into newly added target samples, disappeared target samples, boundary change samples, attribute change samples, and stable background samples according to the change type; Step S4.5: Perform quality screening on the image samples, and remove samples that are severely obscured by clouds and fog, have geometric registration errors exceeding a preset threshold, or have inconsistent labels, to obtain candidate incremental samples; Step S4.6: Based on the category distribution, the importance score of the change region, and the confidence of the model prediction, the candidate incremental samples are resampled and hard example mining is performed to form an incremental training sample set for incremental model learning.

4. The remote sensing sample generation and incremental learning method based on vector transformation driven according to claim 1, characterized in that, Step S5 specifically includes: Step S5.1: Select historical samples from the historical sample library that are related to the incremental training sample set in terms of category distribution or spatial distribution, and construct a historically preserved sample set; Step S5.2: Mix the incremental training sample set with the historical retention sample set according to a preset ratio to form an incremental learning training set; Step S5.3: Using the pre-trained remote sensing model as the initial model, input the incremental learning training set into the pre-trained remote sensing model to update the parameters; Step S5.4: During the parameter update process, different training weights are assigned to newly added target samples, disappearing target samples, boundary change samples, and attribute change samples, so that the model prioritizes learning the newly added features and boundary features caused by the vector change region. Step S5.5: Calculate the historical region preservation loss based on the historical preservation sample set, and ensure that the output of the updated model in the unchanged region remains consistent with that of the model before the update. Step S5.6: Calculate the change region adaptation loss based on the incremental training sample set, and constrain the updated model to improve the remote sensing interpretation accuracy of the change region; Step S5.7: Construct a joint incremental learning loss function based on the historical region preservation loss and the change region adaptation loss, and iteratively update the model based on the joint incremental learning loss function to obtain the updated remote sensing interpretation model.

Images