Land survey quality monitoring system for remote sensing image intelligent segmentation and ground feature recognition

Abstract

The invention discloses a land investigation quality monitoring system for remote sensing image intelligent segmentation and ground feature recognition, and relates to the technical field of remote sensing, and the system comprises an image recognition module, a label comparison module and a monitoring module, and specifically carries out the remote sensing image acquisition of a to-be-monitored land through a remote sensing sensor, and carries out the structural complexity analysis and image region division. The method comprises the steps of forming a non-uniform region set, analyzing and acquiring category probability vectors of pixel positions of all regions in the formed non-uniform region set to construct a tag map of a whole remote sensing image, comparing pixel-by-pixel categories on a current tag map and a reference tag map to form a change mask map, identifying a structural change region according to the change mask map, and identifying a structural change region through sequential analysis. And determining whether the corresponding candidate tag inconsistent region under different time phase conditions is still a structural change region, so as to obtain a judgment result, and updating the electronic map based on the judgment result.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, in particular to a land survey quality monitoring system for intelligent remote sensing image segmentation and ground object recognition. Background Art

[0002] In the field of land surveying, to accurately acquire and update the distribution of surface features, remote sensing imagery requires intelligent processing to identify land classification labels such as cultivated land, forest land, water bodies, and construction land, and to determine whether features have changed and whether the type of change is authentic and reliable. Therefore, given the complex distribution of features, significant texture differences, and high model classification uncertainty in remote sensing imagery, there is an urgent need for a comprehensive monitoring system that can adapt to regional structural characteristics, assess label credibility, and interpret temporal changes.

[0003] Currently, conventional methods for land surveys and feature identification rely heavily on static remote sensing imagery and classification algorithms based on fixed rules. These methods struggle to fully adapt to the structural complexity of feature distribution, leading to a significant decrease in the accuracy of identifying complex features such as urban-rural transition zones and mixed-use plots. Furthermore, there is a lack of intelligent recognition mechanisms for temporary occupancy, and changes are judged solely based on categorical differences, which can easily lead to misjudgments or omissions. Summary of the Invention

[0004] In view of the above problems existing in the prior art, the present application provides a land survey quality monitoring system for intelligent segmentation of remote sensing images and ground object recognition.

[0005] The disclosed embodiment provides a land survey quality monitoring system for intelligent segmentation of remote sensing images and identification of land objects, including:

[0006] The image recognition module uses remote sensing sensors to acquire remote sensing images of the monitored land. After structural complexity analysis, it divides the image regions into non-uniform regions. The module then analyzes and obtains the category probability vectors of the pixel positions in each region within the non-uniform region set to construct a label map for the entire remote sensing image.

[0007] The label comparison module performs a pixel-by-pixel comparison on the label image under the current phase conditions and the reference label image to form a change mask image. The structural change area is identified based on the change mask image. The structural change area is used to determine the inconsistency of the ground feature category label;

[0008] The monitoring module determines whether the area with inconsistent candidate labels under different time conditions is still a structural change area through temporal analysis to obtain the judgment result. Based on the judgment result, the electronic map is updated. The judgment result includes non-temporary signs and temporary signs.

[0009] Optionally, the image recognition module includes an acquisition unit, a structure analysis unit, a segmentation scale unit, and a label generation unit;

[0010] The acquisition unit acquires remote sensing images of the monitored land through remote sensing sensors and pre-processes the remote sensing images, including standardization processing of radiation correction, atmospheric correction and geometric correction;

[0011] The structural analysis unit divides the remote sensing image into regions to obtain multiple regularized initial grid regions, and analyzes the structural complexity of each initial grid region. The structural change intensity of each pixel in each initial grid region is directionally weighted and averaged to obtain the structural complexity index, which is used to reflect the degree of local structural disturbance in each initial grid region.

[0012] Optionally, the scale unit is divided. If the structural complexity index exceeds a preset complexity threshold, it indicates that the structure of the corresponding initial grid area is complex, and a first-class regional processing mechanism is output. If the structural complexity index does not exceed the preset complexity threshold, it indicates that the structure of the corresponding initial grid area is stable, and a second-class regional processing mechanism is output. If a first-class regional processing mechanism is output, the corresponding initial grid area is subdivided. If a second-class regional processing mechanism is output, the corresponding initial grid area is merged with an initial grid area of ​​the same type and adjacent to it.

[0013] Among them, based on the first-class regional processing mechanism and the second-class regional processing mechanism, the segmentation scale of each initial grid area is determined to form a non-uniform regional set, and the segmentation path of each area in the non-uniform regional set is determined. If the first-class regional processing mechanism is output, the corresponding initial grid area is subdivided into regions, including:

[0014] A minimum granularity limit is preset, and it is determined that the structural complexity index of the corresponding initial grid area exceeds the preset complexity threshold; according to the fact that the structural complexity index is inversely proportional to the block scale, the division granularity of the corresponding initial grid area is set, specifically: Among them, S R is the division granularity of the Rth initial grid area, S min is the minimum granularity limit, k is the scaling factor, It is used to dynamically adjust the division granularity of each initial grid area.

[0015] Optionally, the non-uniform regions are concentrated to determine the segmentation path, including:

[0016] According to the non-homogeneous region set, the regions obtained by the first-type region processing mechanism and the regions obtained by the second-type region processing mechanism are distinguished.

[0017] The regions obtained through the first-class region processing mechanism adopt the deep attention residual path, and the segmentation network structure corresponding to this path is the deep attention residual network. The regions obtained through the second-class region processing mechanism adopt the standard semantic segmentation path, and the segmentation network structure corresponding to this path is the standard fully convolutional semantic segmentation model of U-Net.

[0018] Optionally, the label generation unit obtains the category probability vector of each pixel position according to the segmentation network structure to construct a label map for the entire remote sensing image. The specific steps include:

[0019] The remote sensing images corresponding to each area in the non-uniform area set are input into the corresponding segmentation network structure. The spatial and semantic information are extracted through the convolutional feature extraction layer to output a feature map. The feature map is then upsampled through the deconvolution method. After upsampling, each pixel point in the feature map has a multidimensional vector.

[0020] Through the fully connected mapping layer, the multidimensional vector is mapped into an output vector of length C, where C is the total number of ground object categories;

[0021] Use the Softmax function to normalize the output vector to obtain the category probability vector;

[0022] Take the feature category corresponding to the maximum probability from the category probability vector as the final feature category label;

[0023] Traverse each pixel position and stitch it back into the original image according to the region numbers in the non-uniform region set to construct a label map of the entire remote sensing image.

[0024] Optionally, the label comparison module includes a pixel-level label difference unit and a recognition unit;

[0025] The pixel-level label difference unit performs spatial resampling and geometric registration on the current label map and the reference label map, and performs pixel-by-pixel category comparison on the label map under the current phase conditions and the reference label map. If the labels of the same pixel position are different, it is recorded as a changed pixel to form a change mask map;

[0026] The recognition unit clusters the adjacent changed pixels in the change mask image to extract connected areas as candidate label inconsistent areas. For each candidate label inconsistent area, the ratio of the number of changed pixels in it and the number of pixels in the corresponding clustered area is used to obtain the changed pixel ratio. The changed pixel ratio is used to identify the structural change area.

[0027] Optionally, the monitoring module includes a timing analysis unit and a feedback unit;

[0028] The temporal analysis unit, if the ratio of changed pixels exceeds the preset ratio threshold, will mark the corresponding candidate label inconsistent region under the current phase condition as a structural change region, otherwise no marking will be performed; if the corresponding candidate label inconsistent region under the current phase condition is marked as a structural change region, then through temporal analysis, it is determined whether the candidate label inconsistent region under the current phase condition is still a structural change region under other phase conditions to obtain a judgment result.

[0029] Optionally, the feedback unit, if the candidate label inconsistent area under the current phase conditions is still a structural change area under other phase conditions, the judgment result is a non-temporary mark, otherwise it is a temporary mark. According to the judgment result, a structured attribute record table is generated for the corresponding area in the monitored land, and the structured attribute record table is embedded in the corresponding position in the electronic map to realize the update of the electronic map. The structured attribute record table includes the corresponding area number and its judgment result.

[0030] Beneficial effects of the present invention:

[0031] (1) The system uses a structural analysis unit to rasterize remote sensing images and introduces the structural complexity index (ICI) as a quantitative indicator of the degree of regional disturbance to evaluate each initial grid area, effectively characterizing the intensity of texture changes and edge density within the area. Compared with the traditional fixed-scale segmentation method, the structural complexity index can distinguish the difficulty of structural expression in different areas, thereby supporting the dynamic adjustment of the downstream segmentation scale and the adaptive selection of the recognition path to generate a non-uniform region set. This mechanism enables the system to achieve more refined feature extraction and label discrimination in complex areas without increasing the overall computational complexity.

[0032] (2) By setting up a label generation unit, based on the structural differences corresponding to the non-uniform region set and combining the high-dimensional semantic feature vector output by the differential segmentation network structure, the pixel-level object category recognition and label map construction of the entire remote sensing image are accurately realized, thereby improving the semantic expression ability and classification accuracy of the remote sensing image. The system first inputs each area of ​​the non-uniform region set after the structural complexity driven segmentation processing into the matching segmentation network structure. For each pixel position, there is a corresponding high-dimensional semantic feature vector in the feature map, ensuring that the object judgment at each position is based on the deep representation under full context understanding. Finally, the output vector is converted into a category probability vector in the form of probability. This vector not only reflects the maximum possible label of the object type to which the pixel belongs, but can also be used to obtain the model confidence score. By processing all pixel points one by one and combining the area number information where they are located, the system can spatially reconstruct the classification results according to the non-uniform area structure, and finally generate a whole remote sensing label map with excellent spatial integrity and semantic consistency.

[0033] (3) By setting up a label comparison module and introducing pixel-level difference calculation, mask construction and structural change identification mechanisms, we can achieve high-precision comparison of remote sensing image label maps at different times and automatic extraction of abnormal areas of ground feature changes, thereby improving the traceability, credibility and automatic identification capabilities of ground feature label changes in land surveys. In the pixel-level label difference unit, the label map generated for the current phase and the historical reference label map are spatially resampled and geometrically aligned to ensure that the two maps are highly consistent in terms of spatial coordinates, resolution, rotation angle, etc., thereby effectively avoiding misjudgments and false changes caused by pixel misalignment or uneven scales. Based on coordinate alignment, the system compares the ground feature category label of each pixel position point by point. If the current label is inconsistent with the historical label, it is determined to be a changed pixel, and then a change mask map is generated. The mask map marks the change status of each pixel in a binary manner of 0 / 1, which has a good foundation for visualization and structural processing. Subsequently, in the recognition unit, the system performs connectivity clustering on the changed pixels in the change mask map, aggregating spatially adjacent changed pixels into several candidate label inconsistency regions. This clustering process not only improves the spatial expression capability of change detection, but also provides logical entity support for subsequent regional-level analysis. More importantly, the system further calculates the ratio of changed pixels in each candidate region, using this as a key basis for identifying structural changes. When the change ratio is low, it means that only a few pixels in the region are different, which may be due to occasional misjudgments of the model, edge jitter, and other non-real changes. Such regions will be automatically eliminated by the system to prevent false alarms. When the change ratio is significant, the system will mark it as a structural change region for subsequent ground feature label inconsistency analysis and explanation of the cause of the change.

[0034] (4) By setting up a monitoring module and introducing a mechanism based on temporal change trend analysis and result feedback modeling, the dynamic judgment and interpretation classification of areas with inconsistent remote sensing labels are realized, thereby effectively distinguishing the actual causes of temporary changes and long-term changes, and enhancing the reliability, traceability and interpretation ability of the land survey quality monitoring system. Based on the current structural change area, the system tracks the subsequent time phase data backward and compares the temporal change continuity of the same area horizontally. If it remains a structural change area in multiple subsequent time phases, it means that the land feature change is continuous and stable. The system will judge it as a non-temporary sign, that is, a real and stable land use change; on the contrary, if the area returns to the original label or is no longer marked as a change in the subsequent time phase, it means that the change is temporary or cyclical, such as temporary stacking, agricultural rotation, etc. The system will judge it as a temporary sign to prevent short-term anomalies from being misjudged as long-term land use changes, thereby ensuring the authenticity and timeliness of label judgment. Finally, the feedback unit generates a structured attribute record table for each candidate area based on the above temporal judgment results and binds it to the corresponding spatial position in the electronic map in an embedded manner. The record table contains key attributes such as area number, change status (temporary / non-temporary), evolution phase trajectory, change density, etc. It can be used for click-through query and result tracing on the map side. It also supports conditional filtering, classification statistics and change distribution visualization, providing an efficient data basis for subsequent manual review, anomaly analysis and policy response. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application.

[0036] Figure 1 It is a module diagram of the present invention. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0038] like Figure 1 As shown, the implementation of the present invention proposes a land survey quality monitoring system for remote sensing image intelligent segmentation and object recognition, including an image recognition module, a label comparison module and a monitoring module;

[0039] Specifically, the image recognition module uses remote sensing sensors to acquire remote sensing images of the monitored land. After structural complexity analysis, the image regions are divided to form a non-uniform region set. The class probability vectors of the pixel positions in each region within the non-uniform region set are then analyzed and acquired to construct a label map for the entire remote sensing image.

[0040] The image recognition module includes an acquisition unit, a structure analysis unit, a segmentation scale unit, and a label generation unit;

[0041] The acquisition unit acquires remote sensing images of the monitored land through remote sensing sensors and preprocesses the remote sensing images, including standardization of radiation correction, atmospheric correction and geometric correction, to ensure that the spectral information of different bands has physical consistency and eliminate the effects of atmospheric scattering, solar altitude angle differences, etc.

[0042] The purpose of radiometric correction is to convert the original digital values ​​(DN values) in remote sensing images into radiometric brightness with physical meaning, so as to ensure that images collected at different times or by different sensors are numerically comparable.

[0043] The purpose of atmospheric correction is to eliminate radiation deviations caused by atmospheric aerosols, molecular scattering, water vapor absorption, etc., and to restore the spectral reflectance characteristics of the ground objects in the absence of atmospheric conditions. This is particularly critical in the visible and near-infrared bands. Empirical models (such as the DOS method) can be used to estimate atmospheric path radiation and perform corrections by assuming that dark pixels with near-zero reflectivity (such as deep water) exist in the image.

[0044] The purpose of geometric correction is to correct the position of each pixel in the remote sensing image to the position in the real geographic coordinate system, eliminate the spatial distortion caused by the curvature of the earth, terrain undulations, sensor perspective, attitude error, etc., and achieve consistent spatial alignment with maps or historical images.

[0045] The structural analysis unit divides the remote sensing image into regions to obtain multiple regularized initial grid areas, such as 128×128 pixels per block. Although this division facilitates parallel processing and unified input format, it is not adaptable to the difficulty of identification due to the complex distribution of ground objects.

[0046] Therefore, it is necessary to analyze the structural complexity of each initial grid area. The structural complexity index is obtained by directionally weighting the structural change intensity of each pixel in each initial grid area and averaging it. The structural complexity index is used to reflect the degree of local structural disturbance in each initial grid area.

[0047] The specific method of obtaining the structural complexity index is as follows: Among them, ICI(R) is the structural complexity index of the Rth initial grid area, R is an initial grid area in the remote sensing image, (x, y) is the pixel coordinate, I(x, y) is the brightness of the pixel point, and w(x, y) is the direction weight factor, which reflects whether the point is on the main axis of the structural direction and improves the structural sensitivity of the response. is the Laplace operator value of the remote sensing image, which measures the second-order change rate of the image at that point, that is, the edge strength; It is used to measure the second-order derivative of the grayscale change of each pixel. In the image, the edge position changes most dramatically and its Laplace value is the largest, so this part is used to measure whether the texture structure inside the area is complex and whether there is a clear contour. Its mathematical definition is: Where I is the pixel value at a certain position in the image, θ is the partial derivative symbol, which means to derive a variable in a multivariable function, θ 2 is the second-order partial derivative symbol, which represents the quadratic partial derivative of a variable. It is the second-order derivative of the image in the horizontal direction x, which is used to measure the second-order feature of the image in the horizontal direction to detect the degree of sharp change. It can detect horizontal edges, stripes or horizontal structural mutations. It is the second-order derivative of the image in the vertical direction y, which measures the degree of structural disturbance of the image in the vertical direction and can detect structural features such as vertical edges and textures;

[0048] use The purpose is to enhance the contribution of high-frequency areas and weaken the background or small fluctuations. In an image, there are many small fluctuations, but the area where the structure jumps has a large change amplitude. The square can weightedly amplify the influence of the structure jump area.

[0049] In order to improve the response value of pixels with significant boundary structures in certain specific directions, the system also calculates a directional weight factor for each pixel. The setting of this weight is based on whether the current pixel is in a certain main texture direction or edge direction in the image. If the pixel is in the main structure direction, its weight value is higher; if it is in a blurred edge area or irregular direction, the weight value is lower. By introducing this weight coefficient, the system can more effectively highlight the pixels that contribute to the structural contour, thereby enhancing the response value of complex structural areas.

[0050] The directional weight factor is an evaluation coefficient for the importance or sensitivity of structural disturbances in a certain direction, reflecting the contribution of this direction to the expression of regional texture, edge, arrangement and other structures. The acquisition methods can be divided into the following three categories: empirical setting method, adaptive statistical method (based on local directional distribution), and deep learning driven method (training and learning); among them, the empirical setting method is based on the main direction of common ground structures in remote sensing images (such as roads and water bodies are mostly arranged horizontally or vertically), which is simple and efficient and suitable for large-scale rapid processing; the adaptive statistical method uses the gradient direction histogram (such as HoG) in each initial area to analyze the direction distribution, and the weight is automatically adjusted according to the pixel ratio or energy density distribution in a certain direction. The specific steps are: calculate the gradient direction of each pixel, construct a regional direction histogram, and normalize it into a weight factor according to the frequency. Its acquisition method is suitable for ground object environments with large differences in structural complexity; the deep learning driven rule analyzes the main direction of regional texture based on the trained neural network and automatically outputs the direction response intensity as the weight;

[0051] In the embodiment of the present invention, by setting up a collaborative processing mechanism between the structure analysis unit and the segmentation scale unit in the image recognition module, the structural complexity recognition and dynamic segmentation scale control of complex areas of ground objects in remote sensing images are realized, which effectively solves the problems of classification confusion, boundary misjudgment and label instability in the traditional unified scale processing method when identifying high-complexity areas.

[0052] Traditional methods for remote sensing image segmentation typically use a uniform grid scale, such as a fixed 128×128 pixel block, to perform parallel segmentation across the entire image. However, this ignores the significant differences in texture complexity, edge density, and degree of structural disturbance across different regions in remote sensing images. This leads to frequent boundary misclassification, type confusion, or a jigsaw puzzle effect in areas of high structural complexity, such as urban boundaries, fragmented farmland, or mountain slopes. By introducing the structural complexity index (ICI) as an adaptive basis for region segmentation, the system can identify regions with significant structural disturbances in the image and perform re-segmentation operations, such as further refinement into 64×64 or even 32×32 pixel blocks. For structurally stable regions, such as tracts of cultivated land and water bodies, regions can be merged to form larger-granularity processing units. This allows the system to form a non-uniform set of regions, improving the balance between structural adaptability and computational efficiency in the overall segmentation.

[0053] For example, in a remote sensing image, the upper left corner is a large forest with smooth boundaries and uniform texture, corresponding to a low structural complexity index; the lower right corner is the urban-rural junction, with a large number of buildings and road intersections, corresponding to a high structural complexity index. The system will merge the former into a large area and process it with a simple network; for the latter, the regional granularity will be refined and an enhanced path network will be used to improve boundary recognition accuracy.

[0054] The structural complexity index (ICI) is an indicator used to measure the degree of local texture disturbance in each initial grid area in a remote sensing image. The specific calculation logic is to apply the Laplacian operator to each pixel in the area to extract the structural change response, multiply it by the directional weight factor, and then average it to form a directional comprehensive disturbance value. This indicator can accurately reflect the texture edge strength and structural complexity in the area and is the core basis for segmentation granularity control. The introduction of the structural complexity index provides the system with a discriminability indicator foundation at the regional level. That is, instead of blindly dividing the image, it rationally determines which areas need to be refined and which can be merged. Different semantic segmentation path structures are selected accordingly in the subsequent process, such as the deep attention residual network, to achieve differentiated allocation of model resources, improve the accuracy of complex areas, and reduce computational redundancy in stable areas.

[0055] The aforementioned mechanism in the image recognition module is ultimately used to construct a label map for the entire remote sensing image. Because the system processes structure-sensitive regions of the image and uses different network paths to calculate pixel-level category probabilities for each region, the resulting label map not only exhibits higher boundary consistency and spatial continuity, but also possesses information integrity supported by pixel confidence, providing a robust input foundation for subsequent change detection, label comparison, and monitoring modules. By outputting the category probability vector for each pixel and performing Softmax normalization, the system further provides a confidence score for each predicted label.

[0056] Furthermore, the scale unit is divided. If the structural complexity index exceeds the preset complexity threshold, it means that the structure of the corresponding initial grid area is complex, and a type-one regional processing mechanism is output. If the structural complexity index does not exceed the preset complexity threshold, it means that the structure of the corresponding initial grid area is stable, and a type-two regional processing mechanism is output. If a type-one regional processing mechanism is output, the corresponding initial grid area is subdivided. If a type-two regional processing mechanism is output, the corresponding initial grid area is merged with the initial grid areas of the same type and adjacent to each other.

[0057] Among them, the segmentation scale of each initial grid region is determined based on the first-class region processing mechanism and the second-class region processing mechanism to form a non-uniform region set, and the segmentation path of each region in the non-uniform region set is determined. The segmentation path refers to the neural network processing method used for different regions in the system to extract regional features and realize pixel-level object classification. Different paths usually correspond to different network structures or processing intensities. This division mechanism is used to adaptively select the corresponding network processing path according to the structural feature strength of each image sub-region, thereby effectively controlling the computational complexity and enhancing the regional adaptation capability while ensuring the recognition accuracy.

[0058] If a type of regional processing mechanism is output, the corresponding initial grid area will be subdivided into regions, including:

[0059] A minimum granularity limit is preset, and it is determined that the structural complexity index of the corresponding initial grid area exceeds the preset complexity threshold; according to the fact that the structural complexity index is inversely proportional to the block scale, the division granularity of the corresponding initial grid area is set, specifically: Among them, S R is the division granularity of the Rth initial grid area, S min is the minimum granularity limit, that is, the minimum allowed partition size, to prevent over-subdivision, resulting in over-division and increased system computational burden analysis, k is the scaling factor, used to control the overall scale level, reflecting the reference area granularity (usually determined by the original image size), It is used to dynamically adjust the granularity of each initial grid area. When the ICI(R) is higher, the denominator will be larger, the overall size will be smaller, and the area division will be more detailed. Conversely, the area division will be coarser.

[0060] The selection of the scaling factor is essentially to set the average granularity baseline of the entire image in the block. It can be obtained through the resolution normalization method (image adaptation): it is determined based on the image spatial resolution (for example, how many meters each pixel represents) and the minimum identification unit of the target object: k = the physical size of the desired block in the target application (for example, 20m) divided by the image spatial resolution (for example, 1m / pixel); at the same time, a fixed empirical method (static setting) can also be used. When the system is initialized, a fixed value is set based on the original resolution of the remote sensing image and the target application (such as 10m-level land survey or 0.5m urban monitoring). For example: if the original image is 5120×5120 pixels and the initial block is divided into 128×128, it is recommended that k=128;

[0061] The non-uniform regions are concentrated in each region to determine the segmentation path, including,

[0062] According to the non-homogeneous region set, the regions obtained by the first-type region processing mechanism and the regions obtained by the second-type region processing mechanism are distinguished.

[0063] The regions obtained through the first-class region processing mechanism adopt the deep attention residual path. The segmentation network structure corresponding to this path is the deep attention residual network, which is used to capture microstructures. The regions obtained through the second-class region processing mechanism adopt the standard semantic segmentation path. The segmentation network structure corresponding to this path is the standard U-Net fully convolutional semantic segmentation model, which is used to improve efficiency and suppress overfitting.

[0064] Among them, the deep attention residual path refers to a data processing path adopted by the system when processing areas with high structural complexity. The deep attention residual network is the specific network structure implementation based on the deep attention residual path. This structure integrates the following contents: Residual Block: the input can bypass some convolutional layers and directly fuse with the output; Attention Module: guides the network to focus on key areas or key channels, such as SE modules or CBAM modules; Deep convolution stacking structure: the network depth is usually ≥20 layers, supporting multi-scale semantic information extraction. The standard semantic segmentation path corresponds to the processing strategy for areas with relatively stable structures and relatively homogeneous textures. The standard fully convolutional semantic segmentation model of U-Net is the specific neural network architecture actually used in the standard semantic segmentation path. As a classic fully convolutional network architecture, U-Net is widely used in medical imaging and remote sensing image segmentation tasks due to its symmetrical encoder and decoder structure and important skip connection design. This structure integrates the following contents: the encoder part extracts image features layer by layer and implements spatial downsampling; the decoder part uses upsampling operations to gradually restore spatial resolution, and at the same time fuses high-resolution features with decoded features through skip connections to better restore target boundaries and details.

[0065] The segmentation path refers to the neural network processing method used in the system for different areas to extract regional features and realize pixel-level object classification. Different paths correspond to different network structures or processing intensities. Therefore, the system dynamically decides which segmentation network structure to use based on the structural complexity index of each image area.

[0066] The standardized image undergoes an initial rasterization operation, dividing the entire image into a number of equally sized basic analysis regions. While this division facilitates parallel computation, it fails to capture structural differences within different regions. Therefore, in this system, each initial region is further evaluated for structural complexity, constructing a structure-driven adaptive region system. Structural complexity evaluation is based on metrics such as local edge density, texture directionality, and pixel grayscale variability. Based on the degree of complexity, the system automatically determines whether to refine a region into smaller subregions or merge it with adjacent structurally stable regions to form new, heterogeneous, and adaptive spatial units. The core purpose of this is to retain more detail recognition capabilities in structurally complex regions while improving processing efficiency and recognition robustness in structurally simpler regions. The output of this step is a structural complexity assessment map and a hierarchical region partition map, providing the foundational spatial analysis structure for subsequent segmentation and recognition modules.

[0067] In an embodiment of the present invention, by introducing an adaptive regional processing mechanism guided by structural complexity in the segmentation scale unit, the image segmentation granularity and semantic segmentation path can be dynamically adjusted according to the degree of structural disturbance in different regions in the remote sensing image, effectively solving the problems of difficulty in subdividing structurally complex regions and waste of resources in structurally stable regions in traditional remote sensing segmentation, thereby improving the accuracy of ground object recognition and the overall processing efficiency of the system.

[0068] The system first uses the structural complexity index ICI to determine the degree of structural disturbance of each initial grid area, so that the system has content-sensitive segmentation leading capabilities. If the structural complexity index value of a region exceeds the preset complexity threshold, it indicates that the region has dense texture, interlaced edges or complex semantic boundaries. The system then outputs a type of regional processing mechanism to start the regional re-division process. In this process, according to the formula that the ICI value is inversely proportional to the block scale, a new refinement granularity is dynamically set. For example, if the ICI value of a region is extremely high, the division granularity will be close to S min Limiting the value to reach the subdivision limit, effectively improving the semantic segmentation capability; if the ICI value is in the medium-high range, the system automatically adjusts to medium granularity to maintain a balance between segmentation efficiency and accuracy. Conversely, if the structural complexity index of a region is below the preset threshold, the system determines that its structure is stable, outputs a second-class region processing mechanism, and automatically triggers the adjacent region merging strategy: raster regions with continuous spatial locations and consistent landform types are merged into larger regional units, thereby improving segmentation processing efficiency and reducing model reasoning redundancy.

[0069] After forming a set of non-uniform regions, the system further allocates paths and selects network structures based on the corresponding processing mechanisms for each region. For regions obtained through the first-class mechanism, due to their high structural complexity, the system adopts a deep attention residual path, which has strong perception and boundary preservation capabilities and is suitable for semantic extraction tasks in complex image scenes. For the second-class regions, a standard semantic segmentation path is adopted, namely a fully convolutional neural network based on the U-Net structure, which has advantages such as high computational efficiency and stable convergence, and is suitable for regions with regular structures and stable semantics. For example, in a remote sensing image containing both urban and farmland, the core urban area contains a large number of buildings and roads intertwined, with complex boundaries. Its corresponding ICI value is usually higher than the threshold. Therefore, this area will trigger the first-class mechanism, and the deep attention residual network will be processed after fine-grained segmentation. However, large tracts of farmland in the suburbs have clear boundaries, simple textures, and low ICI values. Therefore, the second-class mechanism will be triggered, and multiple plots will be automatically merged and processed by the U-Net, further improving the model inference speed.

[0070] Through the above strategy, the system reduces the processing overhead of areas with simple structures while ensuring the recognition accuracy of areas with complex structures, forming a collaborative recognition mechanism for precise classification of difficult areas and fast classification of easy areas, enhancing the system's processing capabilities for large-scale remote sensing images, and improving the classification consistency, boundary clarity and system execution efficiency of the final label map.

[0071] Furthermore, the label generation unit obtains the category probability vector of each pixel position according to the segmentation network structure to construct a label map of the entire remote sensing image. The specific steps include:

[0072] The remote sensing images corresponding to each area in the non-uniform area set are input into the corresponding segmentation network structure, and the spatial and semantic information are extracted through the convolutional feature extraction layer to output the feature map. The feature map is then upsampled by the deconvolution method to restore it to the same spatial size as the input image. In this way, each pixel position in the feature map corresponds to a set of high-dimensional semantic feature vectors, ensuring that each pixel has a complete classification basis. After upsampling, each pixel point has a multi-dimensional vector in the feature map.

[0073] Spatial information refers to the position, structure, and boundary morphological characteristics of pixels in an image, such as the spatial arrangement, texture changes, edges, and shapes of objects. Semantic information refers to the high-order category meaning contained in the image. For example, an area belongs to cultivated land rather than forest land, indicating that it is a "building structure" rather than a "natural landscape." Semantic information reflects what it is rather than where it is or what it looks like.

[0074] The feature map is the intermediate image representation of each layer output in the convolutional neural network. It describes the activation response of the image on a specific channel. In layman's terms, the feature map is an abstract description of the original image. Each layer output extracted by the convolution operation will form a set of feature maps.

[0075] A multidimensional vector is a set of high-dimensional numbers output at each pixel after processing by a neural network, representing the semantic description of the pixel. For example, if the system needs to classify objects into seven categories, each pixel will ultimately correspond to a 7-dimensional vector, with each dimension representing the probability value of the pixel belonging to a certain category.

[0076] Through a fully connected mapping layer (usually a 1×1 convolution), the multidimensional vector is mapped to an output vector of length C, where C is the total number of land feature categories. For example, if seven types of land features are set (arable land, forest land, water bodies, buildings, roads, bare land, and grassland), the output is a 7-dimensional vector per pixel, with each dimension representing the score value of the corresponding category.

[0077] Use the Softmax function to normalize the output vector and convert it into a probability form to obtain the category probability vector;

[0078] The feature category corresponding to the maximum probability is taken from the category probability vector as the final feature category label; the feature category label is the surface category identified in the remote sensing image;

[0079] Traverse each pixel position and stitch it back into the original image according to the region numbers in the non-uniform region set to construct a label map of the entire remote sensing image.

[0080] The original image here refers to the result of remote sensing image acquisition of the monitored land by remote sensing sensors;

[0081] Softmax ensures that the sum of the C values ​​of each pixel is 1, and each value is in the interval [0, 1], indicating the relative possibility that the pixel belongs to a certain type of feature.

[0082] Based on the structural segmentation results, multiple discriminative features are extracted for each segmented region and fed into the corresponding segmentation network structure to obtain specific feature labels. Feature extraction encompasses spectral information such as multi-band reflectance values, texture information such as contrast and entropy of the gray-level co-occurrence matrix, geometric information such as shape compactness and boundary curvature, and remote sensing indices such as NDVI, NDBI, and NDWI. Multi-dimensional fusion modeling ensures the system's strong discriminative capabilities. The system dynamically allocates different recognition model paths based on the structural complexity of the region, for example, employing a deep multi-scale fusion network in structurally complex areas and a lightweight model in structurally simpler ones. The model recognition process is a fully supervised semantic segmentation process, which outputs a label map for each region through pixel-level prediction, thereby forming a feature classification result. The system further maps feature labels to standard land use classification systems, such as agricultural land, construction land, and water areas. This mapping can be performed using administrative planning layers or artificial intelligence rules. The output results include pixel label maps, surface vector feature map patterns, and regional feature usage statistics, providing a semantic basis for subsequent quality monitoring and consistency analysis.

[0083] In an embodiment of the present invention, by setting a label generation unit, a process of automatically generating an entire label map from remote sensing images concentrated in a non-uniform area is realized, and the precision and credibility of remote sensing object recognition are improved through pixel-by-pixel category probability calculation and maximum value extraction mechanism. Specifically, the system inputs each remote sensing image segment in a non-uniform area into its corresponding segmentation network structure, such as a deep attention residual network or a U-Net network. First, it passes through a convolutional feature extraction layer to extract a high-dimensional feature map containing spatial structure and semantic context. This step ensures that the model can capture the semantic correlation and local geometric features between pixels, and avoids label misclassification caused by insufficient single-point information. Subsequently, the feature map is restored to the same spatial size as the original image through a deconvolution (upsampling) operation to ensure that each pixel can obtain a complete semantic feature vector in the feature map. This vector represents the model's deep judgment basis for the category to which the pixel belongs. Next, the system uses a 1×1 convolutional layer as a fully connected mapping layer to map the high-dimensional semantic vector of each pixel into an output vector of length C. For example, let C = 7, representing seven categories: cultivated land, forest land, water bodies, buildings, roads, bare land, and grassland. Each dimension of this output vector represents the score value of the corresponding category. The higher the score, the more likely the pixel is to be judged as that type of land feature.

[0084] To ensure that the output has clear attribution probabilities, the system introduces a Softmax normalization function, transforming the score vector into a class probability vector whose sum is 1 and whose values ​​for each dimension are between [0, 1]. Each component of this vector represents the relative probability that the pixel belongs to a certain class of feature. For example, if the output for a pixel is {0.1, 0.05, 0.02, 0.72, 0.03, 0.05, 0.03}, then the feature class corresponding to that pixel is building (Class 4), with a probability of 0.72, indicating a classification confidence of 72%. The system then selects the class with the highest probability value from the class probability vector as the pixel's final label.

[0085] Finally, the system numbers and reassembles all the processing results in the non-uniform area, that is, the classification results are mapped to the original remote sensing image positions according to the regional index, thus forming a complete remote sensing label map with spatial consistency, continuous labels and controllable confidence.

[0086] Through the above steps, the system not only realizes high-precision object recognition at the pixel level, but also establishes an interpretable structure of a set of probabilities for each pixel, so that the label map has high-resolution expression capabilities and quantifiable confidence attributes of the classification results, which is of great significance for land classification supervision, critical area review, label mismatch tracking, etc. For example, in the identification of rural construction land occupation, when the probability of building-type pixels is less than 0.5, it can be judged as low confidence and then enters the queue for manual review, effectively avoiding false alarms and missed judgments. Therefore, the label generation unit of the present invention, combined with the structurally adaptive network path, accurate probability vector output, credible confidence quantification mechanism and regional-level restoration and splicing strategy, improves the comprehensive performance of the remote sensing object recognition system in terms of label accuracy, stability and traceability.

[0087] Specifically, the label comparison module performs a pixel-by-pixel category comparison on the label image under the current phase conditions and the reference label image to form a change mask image. The structural change area is identified based on the change mask image. The structural change area is used to determine the inconsistency of the ground feature category label.

[0088] The reference label map relies on historical remote sensing images or manually annotated land class layers as a reference and ensures that the two label maps have spatial consistency;

[0089] Furthermore, the label comparison module includes a pixel-level label difference unit and a recognition unit;

[0090] The pixel-level label difference unit performs spatial resampling and geometric registration on the current label map and the reference label map to ensure that they have a consistent image coordinate system and avoid misjudgment due to position errors. It also performs a pixel-by-pixel category comparison on the label map under the current phase conditions and the reference label map. If the labels at the same pixel position are different, they are recorded as changed pixels, forming a change mask map.

[0091] Resampling refers to the reinterpolation of image pixels during the geometric transformation of an image to adapt to the target spatial resolution or transformed position. Registration refers to the spatial alignment of two or more images so that the same features have the same position coordinates in different images (i.e., pixel alignment).

[0092] The change mask is a binary image. Specifically, if the current label of a pixel is different from the historical label, the mask value is 1 (changed); if the label is the same, the mask value is 0 (unchanged).

[0093] The recognition unit clusters the adjacent changed pixels in the change mask image to extract connected areas as candidate label inconsistency areas. These areas represent the spatial range where the model recognition results are inconsistent with the historical labels. For each candidate label inconsistency area, the ratio of the number of changed pixels in it to the number of pixels in the corresponding clustered area is used to obtain the ratio of changed pixels, so as to eliminate invalid areas where only individual pixels produce deviations and retain structural change areas. The ratio of changed pixels is used to identify structural change areas.

[0094] The clustering process in the recognition unit mainly includes two steps: connectivity detection and region generation. The specific process is as follows:

[0095] Step 1, scan the change mask map: traverse each pixel in the change mask map. If the mask value of the current pixel is 1, it means it is a changed pixel and is marked as an object to be clustered.

[0096] Step 2: Connectivity analysis of adjacent pixels: For each changed pixel, search for its adjacent pixels (usually using a 4-neighborhood or 8-neighborhood). The 4-neighborhood is: up, down, left, and right; the 8-neighborhood is: up, down, left, and right + up left, up right, down left, and down right. If the adjacent pixel is also a changed pixel (value is 1), it is classified into the same connected area to obtain the candidate label inconsistent area.

[0097] To improve spatial robustness and generalization capabilities, the system will perform neighborhood expansion to derive a region containing core pixels of change and neighboring pixels. This type of technology is commonly found in image analysis systems and is called dilation clustering. The final clustered region includes not only pixels with a value of 1 in the initial change mask, but may also include some neighboring pixels with a value of 0. For example, a group of changed pixels is clustered into a 10×10 region, and there are 50 pixels with a value of 1 in the original change mask. To ensure that the entire suspected region is covered, the system will expand the boundary through a dilation operation. Ultimately, the clustered region of this group of changed pixels contains 80 pixels, and there are 60 pixels with a value of 1 in the change mask. At this point, the ratio of changed pixels is 60 to 80.

[0098] The clustering processing here is not the traditional K-means, but is based on connectivity clustering, such as FloodFill, Connected Component Labeling (CCL) and other methods.

[0099] A connected region is a group of pixels that are spatially connected together, forming a closed or identifiable region.

[0100] Inconsistent candidate label areas may represent real changes in land features, or they may be caused by model errors or label offsets. Therefore, they cannot be directly treated as recognition errors. To improve analysis efficiency, the system further calculates the label change density of each candidate area, that is, the proportion of pixels with label changes in the area. If the change density is too low, the area will be ignored or marked as noise changes, and those areas with significant label change density will be extracted as key detection objects. The system will use these areas as input targets for the next stage of label interpretation analysis to distinguish whether they are temporary occupation or non-temporary occupation, and build a basis for label stability assessment.

[0101] In an embodiment of the present invention, by setting a pixel-level label difference unit and a structural change recognition unit in the label comparison module, and combining spatial alignment, change mask construction and regional clustering analysis mechanisms, the recognition ability of inconsistent areas of land object category labels in remote sensing images and the accuracy of change interpretation are improved, thereby providing a more robust data foundation and semantic support for land use monitoring and land type change determination.

[0102] Specifically, the system first uses the pixel-level label difference unit to spatially resample and geometrically align the label image automatically generated in the current phase with the historical reference label image, ensuring that the two images are aligned at the pixel level in the same coordinate system, eliminating spatial errors caused by sensor posture, track deviation or coordinate system differences. This step is the basis for subsequent difference comparison, which can effectively avoid the problem of labels being actually consistent but being misjudged as changes due to position offset, thereby improving the robustness of the comparison.

[0103] After completing the spatial alignment, the system performs a pixel-by-pixel category comparison operation on each pixel position: if the current phase label is consistent with the historical label, it is judged to be unchanged, and the corresponding pixel in the mask image is assigned a value of 0; if it is inconsistent, it is judged to be changed, and the corresponding pixel in the mask image is assigned a value of 1, thus forming a binary change mask image, which reflects the temporal change distribution of the category labels in the entire image and is the input basis for subsequent regional clustering analysis.

[0104] Furthermore, the recognition unit clusters the spatially adjacent changed pixels in the change mask image to extract regional objects with boundary continuity and spatial connectivity, which are called candidate label inconsistent regions. These candidate regions represent the spatial range where the model classification results diverge from the historical labels, which may be caused by factors such as actual ground object changes, label offset, recognition mismatch or temporary occupation. To avoid misjudgment due to individual pixel prediction deviations, the system calculates the ratio of changed pixels in each candidate region, that is, the number of changed pixels divided by the total number of pixels in the region. If the ratio is lower than the preset threshold, it is determined that the region is composed of only scattered noise pixels and is an invalid change region, and is eliminated; if the ratio is higher, it indicates that the region has the continuity and regionality of the changing spatial structure, and is then identified as a structural change region. This mechanism helps to filter out isolated change points caused by occasional model errors and ensure that the retained change regions have a real and reasonable evolution trend at the spatial and semantic levels. For example, in a multi-temporal remote sensing image, an agricultural plot is changed from cultivated land to construction land, and a large range of pixels in the label map change. Through clustering of the change mask map, obvious connected blocks are formed, and the proportion of changed pixels exceeds 0.75. Based on this, the system identifies it as a structural change area; in the adjacent area, if only a few pixel labels change from grassland to bare land, and the proportion within the area is only 5%, the system will judge it as an accidental error and eliminate it.

[0105] In summary, the label comparison module of the present invention realizes a closed loop of change identification from pixel-level error elimination to regional-level anomaly focusing through key steps such as constructing a change mask map, extracting connected regions, and calculating change density. It improves the system's spatial positioning capability and change interpretation capability for inconsistent land labeling, ensuring that subsequent change cause analysis, temporal stability determination, and structured attribute recording have high accuracy and high explanatory power.

[0106] Specifically, the monitoring module determines whether the area with inconsistent candidate labels under different time conditions is still a structural change area through temporal analysis to obtain a judgment result, and updates the electronic map based on the judgment result. The judgment result includes non-temporary marks and temporary marks.

[0107] Furthermore, the monitoring module includes a timing analysis unit and a feedback unit;

[0108] The temporal analysis unit, if the ratio of changed pixels exceeds the preset ratio threshold, will mark the corresponding candidate label inconsistent region under the current phase condition as a structural change region, otherwise no marking will be performed; if the corresponding candidate label inconsistent region under the current phase condition is marked as a structural change region, then through temporal analysis, it is determined whether the candidate label inconsistent region under the current phase condition is still a structural change region under other phase conditions to obtain a judgment result.

[0109] It should be noted that the other phase conditions here refer to the phases after the current phase, and the phases represent different time points. Through temporal analysis, the fluctuation of time series labels can be effectively judged and the "round trip" phenomenon of labels can be identified.

[0110] The feedback unit determines that if the candidate label inconsistent region under the current time phase condition is still a structural change region under other time phase conditions, the result is a non-temporary mark; otherwise, it is a temporary mark. Based on the judgment result, a structured attribute record table is generated for the corresponding region in the monitored land, and the structured attribute record table is embedded in the corresponding position in the electronic map to realize the update of the electronic map. The structured attribute record table includes the corresponding region number and its judgment result.

[0111] The structured attribute record table is embedded in the corresponding position of the electronic map to explain the content of the corresponding area.

[0112] In an embodiment of the present invention, by setting up a monitoring module and introducing a temporal analysis mechanism and a label change feedback mechanism, the behavior patterns of candidate label inconsistent areas under multiple time phase conditions can be dynamically tracked and reliably judged, thereby realizing intelligent distinction between temporary changes and real changes, and improving the land survey system's ability to discern the causes of label changes and the level of structured expression of change information.

[0113] Specifically, the monitoring module consists of a timing analysis unit and a feedback unit, and its core logic is as follows: First, in the current phase, the system identifies several candidate label-inconsistent regions through the previous label comparison module, and calculates the ratio of changed pixels in each region (that is, the ratio of the number of changed pixels in the region to the total number of pixels). If the ratio exceeds the ratio threshold set by the system (such as 20%), the region is marked as a structural change region in the current phase. Otherwise, the change is considered to be weak and no tracking is required.

[0114] Furthermore, in order to determine whether a certain change is a temporary disturbance or a real transformation, the temporal analysis unit dynamically tracks whether the structural change area continues to show a changing state in multiple subsequent phases. If the area is still a structural change area in two or more subsequent phases, it is considered that the area has a stable change trend and is marked as a non-temporary mark (real change); on the contrary, if the area returns to a stable state in the subsequent phase (that is, the mask value returns to 0), it is determined to be a short-term change or label drift and marked as a temporary mark. For example, in actual applications, the label of a certain area changes from a street to a commercial stall in the current phase, but the changed area is again identified as a street in the subsequent daytime remote sensing image. It can be determined that the change in the area is not continuous, and the system automatically marks it as a temporarily occupied area. This mechanism effectively avoids misjudgment of land use due to one-time changes, thereby ensuring the stability and reliability of label map updates.

[0115] Finally, the system encodes the above judgment results into a structured attribute record table through the feedback unit, and binds it with the spatial information of the changed area to achieve an integrated expression of attributes and space. The structured record table includes key fields such as area number, judgment flag (temporary / non-temporary), time phase range, etc., and is embedded in the corresponding position of the electronic map, so that supervisors can click on any changed area in the map interface to view its change nature, duration and judgment basis, thereby improving the system's interpretability, traceability and user interaction efficiency.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A land survey quality monitoring system based on intelligent segmentation of remote sensing images and object recognition, characterized by: include, The image recognition module uses remote sensing sensors to acquire remote sensing images of the monitored land. After structural complexity analysis, it divides the image regions into non-uniform regions. The module then analyzes and obtains the category probability vectors of the pixel positions in each region within the non-uniform region set to construct a label map for the entire remote sensing image. The label comparison module performs a pixel-by-pixel comparison on the label image under the current phase conditions and the reference label image to form a change mask image. The structural change area is identified based on the change mask image. The structural change area is used to determine the inconsistency of the ground feature category label; The monitoring module determines whether the area with inconsistent candidate labels under different time conditions is still a structural change area through temporal analysis to obtain the judgment result. Based on the judgment result, the electronic map is updated. The judgment result includes non-temporary signs and temporary signs.

2. The land survey quality monitoring system for intelligent remote sensing image segmentation and feature identification according to claim 1 is characterized by: The image recognition module includes an acquisition unit, a structure analysis unit, a segmentation scale unit, and a label generation unit; The acquisition unit acquires remote sensing images of the monitored land through remote sensing sensors and pre-processes the remote sensing images, including standardization processing of radiation correction, atmospheric correction and geometric correction; The structural analysis unit divides the remote sensing image into regions to obtain multiple regularized initial grid regions, and analyzes the structural complexity of each initial grid region. The structural change intensity of each pixel in each initial grid region is directionally weighted and averaged to obtain the structural complexity index, which is used to reflect the degree of local structural disturbance in each initial grid region.

3. The land survey quality monitoring system for intelligent remote sensing image segmentation and object recognition according to claim 2 is characterized by: Split the scale unit. If the structural complexity index exceeds the preset complexity threshold, it means that the structure of the corresponding initial grid area is complex, and the first-class regional processing mechanism is output. If the structural complexity index does not exceed the preset complexity threshold, it means that the structure of the corresponding initial grid area is stable, and the second-class regional processing mechanism is output. If the first-class regional processing mechanism is output, the corresponding initial grid area is subdivided. If the second-class regional processing mechanism is output, the corresponding initial grid area is merged with the initial grid areas of the same type and adjacent to each other. Among them, based on the first-class regional processing mechanism and the second-class regional processing mechanism, the segmentation scale of each initial grid area is determined to form a non-uniform regional set, and the segmentation path of each area in the non-uniform regional set is determined. If the first-class regional processing mechanism is output, the corresponding initial grid area is subdivided into regions, including: Presetting a minimum granularity limit and determining whether the structural complexity index of the corresponding initial grid area exceeds a preset complexity threshold; According to the fact that the structural complexity index is inversely proportional to the block scale, the division granularity of the corresponding initial grid area is set as follows: Among them, S R is the division granularity of the Rth initial grid area, S min is the minimum granularity limit, k is the scaling factor, It is used to dynamically adjust the division granularity of each initial grid area.

4. The land survey quality monitoring system for intelligent remote sensing image segmentation and object recognition according to claim 3 is characterized by: The non-uniform regions are concentrated in each region to determine the segmentation path, including, According to the non-homogeneous region set, the regions obtained by the first-type region processing mechanism and the regions obtained by the second-type region processing mechanism are distinguished. The regions obtained through the first-class region processing mechanism adopt the deep attention residual path, and the segmentation network structure corresponding to this path is the deep attention residual network. The regions obtained through the second-class region processing mechanism adopt the standard semantic segmentation path, and the segmentation network structure corresponding to this path is the standard fully convolutional semantic segmentation model of U-Net.

5. The land survey quality monitoring system for intelligent remote sensing image segmentation and ground feature identification according to claim 4 is characterized by: The label generation unit obtains the category probability vector of each pixel position based on the segmentation network structure to construct the label map of the entire remote sensing image. The specific steps include: The remote sensing images corresponding to each area in the non-uniform area set are input into the corresponding segmentation network structure. The spatial and semantic information are extracted through the convolutional feature extraction layer to output a feature map. The feature map is then upsampled through the deconvolution method. After upsampling, each pixel point in the feature map has a multidimensional vector. Through the fully connected mapping layer, the multidimensional vector is mapped into an output vector of length C, where C is the total number of ground object categories; Use the Softmax function to normalize the output vector to obtain the category probability vector; Take the feature category corresponding to the maximum probability from the category probability vector as the final feature category label; Traverse each pixel position and stitch it back into the original image according to the region numbers in the non-uniform region set to construct a label map of the entire remote sensing image.

6. The land survey quality monitoring system for intelligent remote sensing image segmentation and object recognition according to claim 5 is characterized by: The label comparison module includes a pixel-level label difference unit and a recognition unit; The pixel-level label difference unit performs spatial resampling and geometric registration on the current label map and the reference label map, and performs pixel-by-pixel category comparison on the label map under the current phase conditions and the reference label map. If the labels of the same pixel position are different, it is recorded as a changed pixel to form a change mask map; The recognition unit clusters the adjacent changed pixels in the change mask image to extract connected areas as candidate label inconsistent areas. For each candidate label inconsistent area, the ratio of the number of changed pixels in it and the number of pixels in the corresponding clustered area is used to obtain the changed pixel ratio. The changed pixel ratio is used to identify the structural change area.

7. The land survey quality monitoring system for intelligent remote sensing image segmentation and feature identification according to claim 6 is characterized by: The monitoring module includes a timing analysis unit and a feedback unit; The temporal analysis unit, if the ratio of changed pixels exceeds the preset ratio threshold, will mark the corresponding candidate label inconsistent region under the current phase condition as a structural change region, otherwise no marking will be performed; if the corresponding candidate label inconsistent region under the current phase condition is marked as a structural change region, then through temporal analysis, it is determined whether the candidate label inconsistent region under the current phase condition is still a structural change region under other phase conditions to obtain a judgment result.

8. The land survey quality monitoring system for intelligent remote sensing image segmentation and feature identification according to claim 7 is characterized by: The feedback unit determines that if the candidate label inconsistent region under the current time phase condition is still a structural change region under other time phase conditions, the result is a non-temporary mark; otherwise, it is a temporary mark. Based on the judgment result, a structured attribute record table is generated for the corresponding region in the monitored land, and the structured attribute record table is embedded in the corresponding position in the electronic map to realize the update of the electronic map. The structured attribute record table includes the corresponding region number and its judgment result.

Images