A grassland type space-time succession intelligent identification and monitoring system for complex terrain

Abstract

The application provides a grassland type space-time succession intelligent identification and monitoring system for complex terrain, relates to the fields of satellite remote sensing image identification and computer vision technology, and comprises an image processing module, which is used for acquiring a first time phase satellite remote sensing image and a second time phase satellite remote sensing image of a same complex terrain target region, performing spatial registration on the satellite remote sensing images of the two time phases, extracting corresponding image semantic features, performing pixel-level grassland type semantic segmentation, and obtaining a first time phase grassland type label and a second time phase grassland type label. The grassland type space-time succession intelligent identification and monitoring system for complex terrain separates the complex terrain disturbance and the real grassland type transfer by screening stable topological anchor points, constructing a bidirectional local terrain appearance migration relationship and a non-succession reference feature, identifies the grassland type transfer according to the bidirectional succession residual error, and reduces the misidentification and missed identification caused by the apparent change of the cross-time phase satellite image.

Description

Technical Field

[0001] This invention relates to the fields of satellite remote sensing image recognition and computer vision technology, specifically to an intelligent identification and monitoring system for the spatiotemporal evolution of grassland types in complex terrain. Background Technology

[0002] Currently, remote sensing monitoring of grassland types mainly utilizes multi-temporal satellite remote sensing imagery. Technicians first perform geometric correction, radiometric normalization, and spatial registration on the images, then extract image features such as spectral density, texture, shape, and neighborhood structure. Some methods also incorporate elevation, slope, aspect, and solar illumination parameters to perform terrain correction on mountain images. Grassland type identification typically employs random forest, support vector machine, or neural network models. The models perform image classification and semantic segmentation, generating pixel-level grassland type label maps for each temporal phase. The monitoring system then compares the label results or semantic features across different temporal phases to determine the spatial changes and transitional relationships of grassland types.

[0003] Existing methods typically apply uniform correction or overall feature alignment to image differences caused by complex terrain, without distinguishing between the perturbation component caused by changes in terrain illumination and the semantic differences caused by the actual transfer of grassland types. When both types of differences occur simultaneously in the same pixel or local region, existing processing methods struggle to preserve the true type transfer features and suppress terrain perturbation features, leading to misidentification or underidentification of grassland type spatiotemporal succession results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain. The technical problem this invention aims to solve is how to address the difficulty in separating complex terrain disturbances from the actual transfer of grassland types, which can lead to misidentification or missed identification, through stable topological anchor points, two-way terrain apparent migration, and succession residual verification.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain, comprising: The image processing module is used to acquire first-phase and second-phase satellite remote sensing images of the same complex terrain target area, perform spatial registration on the two-phase satellite remote sensing images, extract corresponding image semantic features, and perform pixel-level grassland type semantic segmentation to obtain first-phase grassland type labels and second-phase grassland type labels. The stable topological anchor point determination module is used to determine the corresponding pixels in the local neighborhood of the pixel to be identified, where the topological structure of the grass patch remains stable between two temporal phases, the grass type label is the same, and the image appearance features change, as stable topological anchor points. The terrain appearance migration construction module is used to construct a first local terrain appearance migration operator from the first time phase to the second time phase and a second local terrain appearance migration operator from the second time phase to the first time phase based on the image semantic features of the stable topological anchor point in two time phases. The first local terrain appearance migration operator and the second local terrain appearance migration operator satisfy grassland type preservation constraint and bidirectional migration consistency constraint. The succession residual generation module is used to map the first temporal image semantic features of the pixel to be identified using the first local terrain appearance migration operator to generate a second temporal non-succession reference feature, and to map the second temporal image semantic features of the pixel to be identified using the second local terrain appearance migration operator to generate a first temporal non-succession reference feature. Based on the difference between the image semantic features of the two temporal phases and the corresponding non-succession reference features, the module generates a first succession residual and a second succession residual. The succession identification and monitoring module is used to determine the corresponding semantic transfer direction based on the two grassland type labels when the grassland type labels are different in two time phases; when the first succession residual and the second succession residual both satisfy the succession change conditions and are consistent with the corresponding semantic transfer direction, a directed grassland type transfer label containing the grassland type before transfer and the grassland type after transfer is generated, and a pixel-level grassland type spatiotemporal succession monitoring map is generated based on the directed grassland type transfer label.

[0006] Preferably, the stable topology anchor point determination module uses connected grassland patches within the local neighborhood as topology nodes and the adjacency relationship between adjacent connected grassland patches as topology edges. It constructs a first-phase grassland patch topology map and a second-phase grassland patch topology map, respectively. Topology nodes in the two phases whose intersection-union ratio reaches a preset patch overlap threshold and are the largest intersection-union ratios are determined as a set of corresponding nodes. Two topology edges whose endpoints in the first-phase topology edge and the second-phase topology edge form two sets of corresponding nodes are determined as a set of corresponding edges. The node correspondence rate is determined based on the ratio of twice the number of corresponding node groups to the sum of the number of topology nodes in the two phases, and the topology edge correspondence rate is determined based on the ratio of twice the number of corresponding edge groups to the sum of the number of topology edges in the two phases.

[0007] Preferably, the stable topological anchor point determination module determines the brightness gradient difference, local contrast difference, and minimum angle difference of the main texture direction between two corresponding pixel neighborhoods, respectively. Within the local neighborhood, the range normalization of the three types of differences is performed, and the arithmetic mean of the normalized three types of differences is determined as the change in image appearance features.

[0008] Preferably, the stable topology of the grass patch means that the node correspondence rate reaches a preset node correspondence threshold and the topological edge correspondence rate reaches a preset topological edge correspondence threshold; the change in image appearance features means that the change in image appearance features reaches a preset appearance change threshold.

[0009] Preferably, the terrain appearance migration construction module forms anchor point feature pairs by combining the first temporal image semantic features and the second temporal image semantic features of each stable topological anchor point. The anchor point weights are determined using a Gaussian distance function that decreases as the spatial distance between the stable topological anchor point and the pixel to be identified increases. With the goal of minimizing the weighted mapping error of each anchor point feature pair, under the constraints of grassland type preservation and bidirectional migration consistency, the module solves for the first local affine mapping from the first temporal image semantic features to the second temporal image semantic features, and the second local affine mapping from the second temporal image semantic features to the first temporal image semantic features. The first local affine mapping and the second local affine mapping are then used as the corresponding local terrain appearance migration operators.

[0010] Preferably, the grassland type preservation constraint is used to input the image semantic features of each stable topological anchor point and the image semantic features mapped by the corresponding local terrain appearance migration operator into the same pixel-level grassland type semantic segmentation model, and to restrict the grassland type label output by the pixel-level grassland type semantic segmentation model to remain the same.

[0011] Preferably, the bidirectional migration consistency constraint is used to limit the L2 distance between the semantic features of the first temporal image of the stable topological anchor point and the original semantic features of the first temporal image after passing through the first local terrain appearance migration operator and the second local terrain appearance migration operator in sequence, which does not exceed a preset cyclic error threshold, and to limit the L2 distance between the semantic features of the second temporal image of the stable topological anchor point and the original semantic features of the second temporal image after passing through the second local terrain appearance migration operator and the first local terrain appearance migration operator in sequence, which does not exceed the preset cyclic error threshold.

[0012] Preferably, the first succession residual is a first difference vector obtained by subtracting the second temporal non-succession reference feature from the semantic features of the second temporal image, and the second succession residual is a second difference vector obtained by subtracting the first temporal non-succession reference feature from the semantic features of the first temporal image; when the L2 norm of the first difference vector is not lower than a preset first residual threshold, and the L2 norm of the second difference vector is not lower than a preset second residual threshold, it is determined that both the first succession residual and the second succession residual satisfy the succession change condition.

[0013] Preferably, the succession identification and monitoring module takes the average of the image semantic features of stable topological anchors in the first time phase and the image semantic features of stable topological anchors in the second time phase according to the grassland type label, respectively, to obtain the first grassland type semantic prototype corresponding to each grassland type in the first time phase and the second grassland type semantic prototype corresponding to each grassland type in the second time phase; the vector obtained by subtracting the second grassland type semantic prototype corresponding to the grassland type label in the first time phase from the second grassland type semantic prototype corresponding to the grassland type label in the second time phase is determined as the semantic transfer direction corresponding to the first succession residual; the vector obtained by subtracting the first grassland type semantic prototype corresponding to the grassland type label in the second time phase from the first grassland type semantic prototype corresponding to the grassland type label in the first time phase is determined as the semantic transfer direction corresponding to the second succession residual.

[0014] Preferably, when the cosine similarity between the first succession residual and the corresponding semantic transfer direction reaches a preset first direction consistency threshold, and the cosine similarity between the second succession residual and the corresponding semantic transfer direction reaches a preset second direction consistency threshold, it is determined that the first succession residual and the second succession residual are consistent with the corresponding semantic transfer direction, respectively.

[0015] This invention provides an intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain. It offers the following advantages:

[0016] This intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrains separates complex terrain disturbances from actual grassland type transfers by screening stable topological anchor points, constructing two-way local terrain appearance migration relationships and succession-free reference features, and identifying grassland type transfers based on two-way succession residuals, thereby reducing misidentification and omission caused by cross-temporal satellite image appearance changes.

[0017] By employing grassland type preservation constraints, bidirectional migration consistency constraints, and semantic transfer direction verification, the semantics of grassland type are restricted from being altered during the topographic migration process. This approach preserves the true succession features and suppresses topographic illumination disturbances, thereby improving the accuracy and stability of grassland type spatiotemporal succession identification results. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart illustrating the process of determining stable topology anchor points in this invention. Figure 3 This is a flowchart illustrating the terrain appearance migration construction process of the present invention. Figure 4 This is a flowchart of the succession residual generation and determination process of the present invention; Figure 5 This is a flowchart of the succession identification and monitoring output of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention have been clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 like Figure 1-5 As shown, this embodiment of the invention provides an intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain, comprising: The image processing module is used to acquire first-phase and second-phase satellite remote sensing images of the same complex terrain target area, perform spatial registration on the two-phase satellite remote sensing images, extract corresponding image semantic features, and perform pixel-level grassland type semantic segmentation to obtain first-phase grassland type labels and second-phase grassland type labels.

[0021] The stable topological anchor point determination module is used to determine stable topological anchor points within the local neighborhood of the pixel to be identified, where the topological structure of the grass patch remains stable between two temporal phases, the grass type label is the same, and the image appearance features change. The module uses connected grass patches within the local neighborhood as topological nodes and the adjacency relationship between adjacent connected grass patches as topological edges. It constructs topological graphs for the first and second temporal phases respectively. Topological nodes in the two temporal phases whose intersection-union ratio (IU) reaches a preset overlap threshold and are the largest IU for each other are determined as a set of corresponding nodes. Two topological edges whose endpoints form two sets of corresponding nodes are determined as a set of corresponding edges. The node correspondence rate is determined based on the ratio of twice the number of corresponding node groups to the sum of the number of topological nodes in the two temporal phases, and the topological edge correspondence rate is determined based on the ratio of twice the number of corresponding edge groups to the sum of the number of topological edges in the two temporal phases. The stable topological anchor point determination module determines the brightness gradient difference, local contrast difference, and minimum angle difference of the main texture direction between two corresponding pixel neighborhoods. Within the local neighborhood, it performs range normalization on the three types of differences and determines the arithmetic mean of the normalized three types of differences as the change in image appearance features. The stable topological structure of the grass patch means that both the node correspondence rate and the topological edge correspondence rate reach a preset threshold. Changes in image appearance features mean that the change in image appearance features reaches a preset threshold.

[0022] The terrain appearance migration construction module is used to construct a first local terrain appearance migration operator from the first time phase to the second time phase and a second local terrain appearance migration operator from the second time phase to the first time phase, based on the image semantic features of stable topological anchor points in two time phases. The first and second local terrain appearance migration operators satisfy grassland type preservation constraints and bidirectional migration consistency constraints. The terrain appearance migration construction module forms anchor point feature pairs by combining the first and second time phase image semantic features of each stable topological anchor point. Anchor point weights are determined using a Gaussian distance function that decreases as the spatial distance between the stable topological anchor point and the pixel to be identified increases. With the goal of minimizing the weighted mapping error of each anchor point feature pair, under the grassland type preservation and bidirectional migration consistency constraints, the module solves the first local affine mapping from the semantic features of the first time phase image to the semantic features of the second time phase image, and the second local affine mapping from the semantic features of the second time phase image to the semantic features of the first time phase image, respectively. The first and second local affine mappings are then used as the corresponding local terrain appearance migration operators. Grassland type preservation constraints are used to input the image semantic features of each stable topological anchor point and the image semantic features mapped by the corresponding local terrain appearance transfer operator into the same pixel-level grassland type semantic segmentation model, and to restrict the grassland type labels output by the pixel-level grassland type semantic segmentation model to remain the same. Bidirectional transfer consistency constraints are used to restrict the L2 distance between the first-phase image semantic features of the stable topological anchor point after sequentially passing through the first and second local terrain appearance transfer operators and the original first-phase image semantic features to not exceed a preset cycle error threshold, and to restrict the L2 distance between the second-phase image semantic features of the stable topological anchor point after sequentially passing through the second and first local terrain appearance transfer operators and the original second-phase image semantic features to not exceed a preset cycle error threshold.

[0023] The succession residual generation module is used to map the first temporal image semantic features of the pixel to be identified using a first local terrain appearance migration operator to generate a second temporal image without succession reference features, and to map the second temporal image semantic features of the pixel to be identified using a second local terrain appearance migration operator to generate a first temporal image without succession reference features. Based on the differences between the image semantic features of the two temporal phases and their corresponding without succession reference features, a first succession residual and a second succession residual are generated. The first succession residual is a first difference vector obtained by subtracting the second temporal image without succession reference features from the second temporal image semantic features, and the second succession residual is a second difference vector obtained by subtracting the first temporal image without succession reference features from the first temporal image semantic features. When the L2 norm of the first difference vector is not lower than a preset first residual threshold, and the L2 norm of the second difference vector is not lower than a preset second residual threshold, it is determined that both the first and second succession residuals satisfy the succession change condition.

[0024] The succession identification and monitoring module is used to determine the corresponding semantic transfer direction based on the two grassland type labels when the grassland type labels are different in two time phases. When both the first succession residual and the second succession residual satisfy the succession change condition and are consistent with the corresponding semantic transfer direction, a directed grassland type transfer label containing the grassland type before and after the transfer is generated, and a pixel-level grassland type spatiotemporal succession monitoring map is generated based on the directed grassland type transfer label. The succession identification and monitoring module takes the average of the image semantic features of the stable topological anchor points in the first time phase and the image semantic features of the stable topological anchor points in the second time phase according to the grassland type label, respectively, to obtain the first grassland type semantic prototype corresponding to each grassland type in the first time phase and the second grassland type semantic prototype corresponding to each grassland type in the second time phase. The vector obtained by subtracting the second grassland type semantic prototype corresponding to the grassland type label in the first time phase from the second grassland type semantic prototype corresponding to the grassland type label in the second time phase is determined as the semantic transfer direction corresponding to the first succession residual. The vector obtained by subtracting the semantic prototype of the first grassland type corresponding to the first grassland type label from the semantic prototype of the second grassland type label is determined as the semantic transfer direction corresponding to the second succession residual. When the cosine similarity between the first succession residual and the corresponding semantic transfer direction reaches a preset first direction consistency threshold, and the cosine similarity between the second succession residual and the corresponding semantic transfer direction reaches a preset second direction consistency threshold, it is determined that the first succession residual and the second succession residual are consistent with the corresponding semantic transfer directions, respectively.

[0025] This invention establishes a two-way terrain appearance migration relationship based on locally stable topological anchor points, generating non-successive reference features for the pixels to be identified that are consistent with the target temporal imaging conditions, so that changes in terrain lighting, shadows and local textures can be distinguished from the real transfer of grassland types, reducing misjudgment and missed judgment in cross-temporal recognition.

[0026] This invention employs grassland type preservation constraints, bidirectional migration consistency constraints, and succession residual direction verification to limit changes in grassland type semantics during the migration process and to perform bidirectional verification of the true succession direction, thereby improving the accuracy and stability of grassland type spatiotemporal succession identification results.

[0027] Example 2 This embodiment is used to verify whether false grassland type transfers caused by changes in slope shadows, local illumination, and texture response can be identified and excluded.

[0028] 1. Images and verification data A hilly grassland area of ​​12.6 square kilometers was selected. The first temporal image was acquired on July 18, 2023, and the second temporal image was acquired on July 20, 2024. Both images are from the same multispectral satellite sensor and include blue, green, red, near-infrared, and two shortwave infrared bands, with a spatial resolution of 10 meters.

[0029] The two temporal images are processed by the same feature encoder with shared parameters, outputting 64-dimensional image semantic features with consistent dimensions and channel order. Before solving the transfer operator and calculating the succession residual, the image semantic features are normalized to L2 norm.

[0030] After radiometric calibration, orthorectification, and spatial registration, the root mean square error of the registration for the two images was 0.43 pixels. No slope illumination correction based on a uniform coefficient was performed during the preprocessing stage.

[0031] The verification data included contemporaneous 0.8-meter spatial resolution imagery and 48 field quadrats. The field quadrats were surveyed within 7 days before and after the acquisition dates of the two time-phase images. The verification data was not used for model training, threshold determination, or semantic prototype calculation.

[0032] Pixel-level semantic segmentation yielded 1462 candidate pixels with different labels, of which 913 pixels changed from the initial label of typical grassland to meadow steppe. The candidate pixels were mainly distributed along the shadow edges of the northwest slope and in local low-light areas.

[0033] 2. Stable topology anchor point selection A 21×21 pixel local neighborhood is set centered on a candidate pixel at the shadow edge. The first-phase topology graph contains 12 nodes and 17 topological edges, and the second-phase topology graph contains 13 nodes and 18 topological edges. Node matching yields 11 pairs of corresponding nodes, and edge matching yields 15 pairs of corresponding edges.

[0034] Node correspondence rate: The topological edge correspondence rate is: .

[0035] The node correspondence rate and topological edge correspondence rate are higher than 0.80 and 0.75, respectively, and the local grass patch topology is determined to be stable.

[0036] After range normalization, the differences in brightness gradient, local contrast, and minimum angle of the main texture direction were 0.72, 0.61, and 0.47, respectively. The arithmetic mean of the three differences was 0.60, which is higher than the apparent change threshold of 0.45.

[0037] Within a local neighborhood, 38 corresponding pixels simultaneously satisfy the conditions of topological stability, identical grassland type labels, type probability not less than 0.85, and apparent feature variation, and are thus identified as stable topological anchor points.

[0038] 3. Solving the apparent migration operator for terrain The Gaussian scale parameter is set to 6 pixels. The two local terrain appearance migration operators adopt channel-independent diagonal affine mapping. The objective of solving includes anchor point weighted mapping error, grassland type preservation loss, bidirectional cyclic consistency loss and L2 regularization term.

[0039] The objective function is jointly applied to 38 sets of dual-temporal anchor point features, and the first local terrain apparent migration operator and the second local terrain apparent migration operator are obtained respectively.

[0040] After forward and reverse mapping, the grassland type label remains unchanged for stable topological anchor points. The average L2 error of the first-phase features after bidirectional cyclic mapping is 0.074, and the average L2 error of the second-phase features after bidirectional cyclic mapping is 0.069, both below the cyclic error threshold of 0.12.

[0041] 4. No succession reference and residual determination The semantic features of the first temporal image of the pixels to be identified are mapped by the first local terrain appearance transfer operator to obtain the second temporal unsuccessful reference features. The L2 norm of the first successional residual obtained by subtracting the second temporal unsuccessful reference features from the second temporal actual features is 0.083.

[0042] The semantic features of the second temporal image of the pixels to be identified are mapped by the second local terrain appearance transfer operator to obtain the first temporal unsuccessful reference features. The L2 norm of the second successional residual obtained by subtracting the first temporal unsuccessful reference features from the first temporal actual features is 0.079.

[0043] The first and second residual thresholds determined by the independent validation set are both 0.18. Neither residual reaches its corresponding threshold, therefore no directed transfer label for "typical grassland-meadow steppe" is generated.

[0044] 5. Result Verification Independent verification data categorized 1462 candidate change pixels into 1120 pixels representing apparent topographic changes and 342 pixels representing actual type changes. The system excluded 1093 pixels representing apparent topographic changes and retained 329 pixels representing actual type changes. The exclusion rate for pixels representing apparent topographic changes was 97.6%, and the retention rate for pixels representing actual type changes was 96.2%.

[0045] The final output included 356 pixels showing directional grassland type transfer, of which 329 were consistent with the independent verification results. False transfer zones with continuous distribution at the edges of slope shadows were not included in the spatiotemporal succession monitoring map.

[0046] Local migration relationships based on stable topological anchors can provide succession-free references for target temporal phases, reducing the impact of slope shading and local illumination variations on type identification. The processed results preserve actual type change areas confirmed by independent data, making them suitable for monitoring scenarios where grassland types are generally stable but topographical disturbances are significant.

[0047] Example 3 This embodiment is used to verify whether a directional grassland type transfer path can be confirmed when apparent topographic disturbances and actual grassland type transfers occur simultaneously.

[0048] 1. Candidate areas and verification materials The selected area is a valley-slope junction region covering 8.4 square kilometers. The first temporal image was acquired on July 15, 2022, and the second temporal image was acquired on July 17, 2024. Both images are from the same multispectral satellite sensor, with a spatial resolution of 5 meters and using the same band combination. The root mean square error of spatial registration is 0.38 pixels.

[0049] Independently verified data includes 0.5-meter spatial resolution imagery and 54 field quadrats. Field survey records indicate a transition from meadow-steppe to shrub-grassland on the southeastern gentle slope, and significant shadows and localized illuminance variations in the northern valley.

[0050] The initial semantic segmentation yielded 684 candidate change pixels with different labels, of which 427 pixels were labeled as meadow steppe in the first phase and as shrub grassland in the second phase.

[0051] 2. Establishment of migration relationships A local neighborhood of 31×31 pixels was set centered on the candidate pixel. 52 stable topological anchor points for meadow-steppe, 41 stable topological anchor points for shrub-grassland, and 19 stable topological anchor points for typical grassland were obtained through screening.

[0052] The Gaussian scaling parameter is set to 8 pixels, and the cyclic error threshold is set to 0.15. After solving the joint objective function, the average cyclic error of the first local terrain appearance migration operator is 0.091, and the average cyclic error of the second local terrain appearance migration operator is 0.096. All stable topological anchors involved in the solution maintain consistent grassland type labels before and after mapping.

[0053] 3. Judgment of conditions for succession and change A candidate pixel whose initial label changes from meadow steppe to shrubland is selected. The first succession residual is obtained by subtracting the second-phase non-succession reference features from the semantic features of the actual image in the second phase, with a L2 norm of 0.41.

[0054] Subtracting the first-phase unrelated reference features from the semantic features of the actual image in the first phase yields the second succession residual, with a second norm of 0.37.

[0055] The first residual threshold is 0.22, and the second residual threshold is 0.20. When both succession residuals reach their respective thresholds, the candidate pixel enters the semantic transfer direction verification.

[0056] 4. Semantic transfer direction verification The semantic prototypes of grassland types in the first and second time phases are obtained by averaging the features of stable topological anchor points with the same grassland type label and a classification probability of not less than 0.90 in the corresponding time phases.

[0057] Within the second temporal feature space, the semantic transfer direction corresponding to the first succession residual is obtained by subtracting the semantic prototype of meadow steppe from the semantic prototype of shrub steppe. Within the first temporal feature space, the semantic transfer direction corresponding to the second succession residual is obtained by subtracting the semantic prototype of shrub steppe from the semantic prototype of meadow steppe.

[0058] The cosine similarity between the first succession residual and the corresponding semantic transfer direction is 0.86, and the cosine similarity between the second succession residual and the corresponding semantic transfer direction is 0.81, both higher than the direction consistency threshold of 0.70. The system generates a directed grassland type transfer label of "meadow-steppe-shrubland" for the corresponding pixel.

[0059] Another candidate pixel located in the valley shadow region has L2 norms of 0.35 and 0.32 for its first and second succession residuals, respectively, both reaching the residual threshold. The cosine similarities between the two residuals and their corresponding semantic transfer directions are 0.28 and 0.19, respectively, neither reaching the direction consistency threshold. Therefore, no directed transfer label is generated for this pixel.

[0060] 5. Result Verification After processing, 312 of the 684 candidate change pixels were used to generate the label "meadow steppe - shrub grassland", and 47 pixels were used to generate the label "typical grassland - meadow steppe", for a total of 359 directional transfer pixels.

[0061] Independent verification data confirmed 374 pixels of genuine grassland type transfer, and 310 pixels of topographical appearance changes or other non-type transfer differences. Of the 359 pixels output by the system, 341 were consistent with the independent verification results, and of the 325 pixels not output, 33 were verified to be genuine type transfers.

[0062] The directed transfer results achieved a precision of 94.99% and a recall of 91.18%. The 684 candidate pixels obtained directly based on the difference between the two time-phase labels achieved a precision of 54.68%.

[0063] Two-way succession residuals are used to confirm whether actual features deviate from the absence of a succession reference, while semantic transfer direction is used to verify whether the deviation conforms to the change path between specific grassland types. Joint judgment excludes perturbed pixels with large differences but inconsistent semantic directions, and preserves the true transfer areas from meadow steppe to shrubland, etc.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain, characterized in that, include: The image processing module is used to acquire first-phase and second-phase satellite remote sensing images of the same complex terrain target area, perform spatial registration on the two-phase satellite remote sensing images, extract corresponding image semantic features, and perform pixel-level grassland type semantic segmentation to obtain first-phase grassland type labels and second-phase grassland type labels. The stable topological anchor point determination module is used to determine the corresponding pixels in the local neighborhood of the pixel to be identified, where the topological structure of the grass patch remains stable between two temporal phases, the grass type label is the same, and the image appearance features change, as stable topological anchor points. The terrain appearance migration construction module is used to construct a first local terrain appearance migration operator from the first time phase to the second time phase and a second local terrain appearance migration operator from the second time phase to the first time phase based on the image semantic features of the stable topological anchor point in two time phases. The first local terrain appearance migration operator and the second local terrain appearance migration operator satisfy grassland type preservation constraint and bidirectional migration consistency constraint. The succession residual generation module is used to map the first temporal image semantic features of the pixel to be identified using the first local terrain appearance migration operator to generate a second temporal non-succession reference feature, and to map the second temporal image semantic features of the pixel to be identified using the second local terrain appearance migration operator to generate a first temporal non-succession reference feature. Based on the difference between the image semantic features of the two temporal phases and the corresponding non-succession reference features, the module generates a first succession residual and a second succession residual. The succession identification and monitoring module is used to determine the corresponding semantic transfer direction based on the two grassland type labels when the grassland type labels are different in two time phases; when the first succession residual and the second succession residual both satisfy the succession change conditions and are consistent with the corresponding semantic transfer direction, a directed grassland type transfer label containing the grassland type before transfer and the grassland type after transfer is generated, and a pixel-level grassland type spatiotemporal succession monitoring map is generated based on the directed grassland type transfer label.

2. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 1, characterized in that: The stable topology anchor point determination module uses connected grassland patches within the local neighborhood as topology nodes and the adjacency relationship between adjacent connected grassland patches as topology edges. It constructs a first-phase grassland patch topology map and a second-phase grassland patch topology map, respectively. Topology nodes in the two phases whose intersection-union ratio reaches a preset overlap threshold and are the largest intersection-union ratios are determined as a set of corresponding nodes. Two topology edges whose endpoints in the first-phase topology edge and the second-phase topology edge form two sets of corresponding nodes are determined as a set of corresponding edges. The node correspondence rate is determined based on the ratio of twice the number of corresponding node groups to the sum of the number of topology nodes in the two phases. The topology edge correspondence rate is determined based on the ratio of twice the number of corresponding edge groups to the sum of the number of topology edges in the two phases.

3. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 2, characterized in that: The stable topological anchor point determination module determines the brightness gradient difference, local contrast difference, and minimum angle difference of the main texture direction between two corresponding pixel neighborhoods. Within the local neighborhood, the range of the three types of differences is normalized, and the arithmetic mean of the normalized three types of differences is determined as the change in image appearance features.

4. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 3, characterized in that: The stability of the topology of the grass map features means that the node correspondence rate reaches a preset node correspondence threshold and the topological edge correspondence rate reaches a preset topological edge correspondence threshold; the change in image appearance features means that the amount of change in image appearance features reaches a preset appearance change threshold.

5. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 1, characterized in that: The terrain appearance migration construction module forms anchor point feature pairs by combining the first and second temporal image semantic features of each stable topological anchor point. It uses a Gaussian distance function that decreases as the spatial distance between the stable topological anchor point and the pixel to be identified increases to determine the anchor point weights. With the goal of minimizing the weighted mapping error of each anchor point feature pair, under the constraints of grassland type preservation and bidirectional migration consistency, it solves the first local affine mapping from the first temporal image semantic features to the second temporal image semantic features, and the second local affine mapping from the second temporal image semantic features to the first temporal image semantic features. The first local affine mapping and the second local affine mapping are then used as the corresponding local terrain appearance migration operators.

6. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 1, characterized in that: The grassland type preservation constraint is used to input the image semantic features of each stable topological anchor point and the image semantic features mapped by the corresponding local terrain appearance migration operator into the same pixel-level grassland type semantic segmentation model, and to restrict the grassland type label output by the pixel-level grassland type semantic segmentation model to remain the same.

7. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 1, characterized in that: The bidirectional migration consistency constraint is used to limit the L2 distance between the semantic features of the first temporal image of the stable topological anchor point and the original semantic features of the first temporal image after passing through the first local terrain appearance migration operator and the second local terrain appearance migration operator in sequence, which does not exceed a preset cyclic error threshold. It also limits the L2 distance between the semantic features of the second temporal image of the stable topological anchor point and the original semantic features of the second temporal image after passing through the second local terrain appearance migration operator and the first local terrain appearance migration operator in sequence, which does not exceed the preset cyclic error threshold.

8. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 1, characterized in that: The first succession residual is a first difference vector obtained by subtracting the second temporal non-succession reference feature from the semantic features of the second temporal image, and the second succession residual is a second difference vector obtained by subtracting the first temporal non-succession reference feature from the semantic features of the first temporal image; when the L2 norm of the first difference vector is not lower than a preset first residual threshold, and the L2 norm of the second difference vector is not lower than a preset second residual threshold, it is determined that both the first succession residual and the second succession residual satisfy the succession change condition.

9. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 8, characterized in that: The succession identification and monitoring module takes the average of the image semantic features of stable topological anchors in the first and second time phases according to the grassland type labels, respectively, to obtain the first grassland type semantic prototype corresponding to each grassland type in the first phase and the second grassland type semantic prototype corresponding to each grassland type in the second phase. The vector obtained by subtracting the second grassland type semantic prototype corresponding to the grassland type label in the first phase from the second grassland type semantic prototype corresponding to the grassland type label in the second phase is determined as the semantic transfer direction corresponding to the first succession residual. The vector obtained by subtracting the first grassland type semantic prototype corresponding to the grassland type label in the second phase from the first grassland type semantic prototype corresponding to the grassland type label in the first phase is determined as the semantic transfer direction corresponding to the second succession residual.

10. The intelligent identification and monitoring system for the spatiotemporal succession of grassland types in complex terrain according to claim 9, characterized in that: When the cosine similarity between the first succession residual and the corresponding semantic transfer direction reaches a preset first direction consistency threshold, and the cosine similarity between the second succession residual and the corresponding semantic transfer direction reaches a preset second direction consistency threshold, it is determined that the first succession residual and the second succession residual are consistent with the corresponding semantic transfer direction, respectively.

Images