Global land comprehensive regulation monitoring method based on artificial intelligence and remote sensing image

Abstract

This invention relates to a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, belonging to the field of environmental management technology. The invention acquires multi-band high-resolution remote sensing images of the target area before and after consolidation, performs radiometric and atmospheric corrections to ensure data quality, and utilizes a pre-trained deep convolutional neural network land cover classification model to generate land cover classification maps for both time periods. Through pixel-level change detection and confidence analysis, reliable land cover change areas are identified, vectorized, and spatially filtered to extract multi-dimensional change features. The ecological impact and spatial distribution patterns are comprehensively assessed, generating comprehensive monitoring results including spatial distribution maps, statistical reports, and trend predictions. These results are integrated into an interactive geographic information system platform for visualization and early warning. This invention achieves automated, precise, and systematic monitoring and evaluation of the entire land consolidation process.

Description

Technical Field

[0001] This invention relates to a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, belonging to the field of environmental management technology. Background Technology

[0002] In the past, monitoring of comprehensive land consolidation mainly relied on manual field surveys and traditional remote sensing interpretation methods. These methods typically require staff to conduct extensive field verification and map comparison, which is not only tedious and time-consuming, but also heavily dependent on personal experience, resulting in highly subjective and low repeatability of monitoring results, making it difficult to meet the needs of large-scale, high-frequency routine monitoring.

[0003] With technological advancements, some methods have begun to combine remote sensing imagery with computer-aided analysis, such as using simple image interpolation or post-classification comparison methods to identify land surface changes. However, these methods often require extremely high accuracy in image registration and temporal consistency, and the classification algorithms used are mostly traditional models based on pixel spectral characteristics, which have limited ability to handle phenomena such as "same object, different spectra" and "same spectrum, different objects," resulting in insufficient accuracy and reliability in change detection and difficulty in accurately distinguishing complex land cover types from subtle remediation activities.

[0004] Furthermore, existing monitoring systems primarily focus on identifying single targets, such as the expansion of construction land or changes in the amount of arable land, lacking a systematic assessment of the spatial pattern, type transformation path, and comprehensive ecological impact of the remediation areas. Due to the failure to integrate artificial intelligence and multidimensional feature analysis, existing technologies struggle to automatically and efficiently extract deep information such as remediation intensity, diversity, and spatial patterns from massive amounts of remote sensing data, thus failing to provide comprehensive and accurate decision support for land spatial planning and ecological protection. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution: According to a first aspect of the present invention, the present invention claims protection for a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, comprising the following steps: S101: Acquire remote sensing image data of the target area in the first time phase before land consolidation and the second time phase after land consolidation; S102: Input the remote sensing image data of the first time phase and the second time phase into the pre-trained land cover classification model respectively to generate the land cover classification map of the first time phase and the land cover classification map of the second time phase. The land cover classification map is raster data, and each pixel is assigned a land cover category label. S103: Perform pixel-level change detection on the land cover classification maps of the first and second time phases. By comparing the land cover categories of pixels at the same coordinate positions, identify pixels that have changed and construct a land cover change map. The change detection includes calculating the change confidence level and determining the confidence level of the changed pixels using the probability product method based on the category probabilities output by the land cover classification model. S104: Extract land consolidation-related change features from the land cover change map, including the spatial distribution of change areas, change area statistics, and change type conversion matrix. Calculate the category distribution based on the change type conversion matrix to reflect the richness of land consolidation types. S105: Based on the extracted change features, generate comprehensive land consolidation monitoring results for the entire region, including creating a spatial distribution map of consolidation, compiling a report on the area of ​​consolidation, assessing the trend of consolidation type conversion, and integrating the monitoring results into a geographic information system platform for visualization. The monitoring results also include a consolidation effect rating, which is comprehensively evaluated based on the consolidation intensity index and diversity index. The visualization includes dynamic maps and statistical charts to support decision analysis.

[0006] Furthermore, step S102 also includes: The land cover classification model adopts a U-Net network structure. The encoder part includes multiple convolutional layers and pooling layers for downsampling and feature extraction. The decoder part includes multiple deconvolutional layers and skip connections for upsampling to restore spatial resolution and fuse multi-scale features. The skip connections concatenate the feature maps in the encoder with the corresponding feature maps in the decoder. The convolutional layers use rectified linear units as activation functions, the pooling layers use max pooling, and the deconvolutional layers use bilinear interpolation for upsampling. The output layer of the land cover classification model uses the softmax function to generate the class probability distribution for each pixel.

[0007] Furthermore, the pixel-level change detection in step S103 includes the following sub-steps: S1031: Spatial registration is performed on the land cover classification maps of the first and second time phases. The feature point matching algorithm is used to align the classification maps of the two time phases to the same coordinate system. Image resampling is achieved through affine transformation or polynomial transformation to ensure pixel-level alignment accuracy.

[0008] S1032: Compare the land cover category labels of corresponding pixels in the first and second time-phase classification maps pixel by pixel. If the category labels are different, the pixel is initially marked as a candidate pixel for change. The comparison operation is based on a one-to-one correspondence of pixel coordinates and uses a raster calculation tool to record the position coordinates and category change information of each candidate pixel for change. S1033: For each candidate pixel that changes, obtain the class probability distribution of the pixel in the first temporal classification model and the class probability distribution of the pixel in the second temporal classification model; S1034: Calculate the change confidence score based on the category probability distribution. The change confidence score reflects the degree of confidence that the pixel has indeed undergone land cover change. S1035: Set a confidence threshold, which is determined by historical verification data. Compare the change confidence with the threshold. If the change confidence is greater than the threshold, the pixel is confirmed as a changed pixel; otherwise, it is excluded. S1036: Post-process all confirmed changed pixels, including using morphological opening operations to remove small area noise, using morphological closing operations to fill small holes, and generating a smooth land cover change map. S1037: Vectorize the land cover change map into polygon vector data, where each polygon represents a continuous change area, and record the area and perimeter attributes of the polygons. The vectorization process uses a raster-to-vector algorithm. S1038: Spatial filtering is performed on the vectorized change area to remove areas with an area smaller than the minimum remediation unit. The minimum remediation unit is set according to the land remediation policy. After spatial filtering, the final land cover change map is generated for feature extraction.

[0009] Furthermore, step S104, which involves extracting change features related to comprehensive land consolidation, includes the following sub-steps: S1041: Extract all changed pixels from the land cover change map, count the total number of changed pixels, and calculate the total area of ​​the changed region based on the pixel size. The total area of ​​the changed region is obtained by multiplying the number of pixels by the area of ​​a single pixel. The area of ​​a single pixel is calculated based on the spatial resolution of the remote sensing image. S1042: Perform cluster analysis on the change region, aggregate spatially adjacent change pixels into continuous change patches, use the connected component labeling algorithm to identify spatially continuous pixel groups, and assign a unique identifier to each patch; S1043: Record attribute information for each changed patch, including patch area, patch perimeter, patch shape index, and land cover category in the first and second time phases; S1044: Construct a land cover change type transformation matrix. The rows of the matrix represent the land cover category in the first time phase, the columns represent the land cover category in the second time phase, and the matrix elements represent the number or area of ​​patches that change from the row category to the column category. The transformation matrix is ​​used to analyze the type transformation pattern of land consolidation. S1045: Calculate the remediation intensity index and determine the ratio of the total area of ​​the changed area to the total area of ​​the target area; S1046: Calculate the land consolidation diversity index based on the category distribution of the change type transformation matrix, reflecting the richness and evenness of land consolidation types; S1047: Analyze the spatial distribution pattern of changed patches, use spatial statistical methods to assess the degree of clustering or dispersion of changed patches, and identify hotspots for remediation; S1048: Based on the change characteristics, generate a remediation priority assessment. Calculate the priority score for each patch based on its area, shape index, and transformation type. Patches with high scores indicate that they require key monitoring.

[0010] Furthermore, in step S101, the remote sensing image data comes from the same satellite sensor platform to ensure data consistency and comparability. The image acquisition time is selected during the vegetation growing season or the non-growing season to highlight land cover changes. The remote sensing image data also includes radar remote sensing data to penetrate cloud cover, and land cover is analyzed mainly using optical remote sensing images.

[0011] Furthermore, in step S102, the land cover classification model also includes an attention mechanism, which is embedded in the encoder-decoder structure. By calculating the importance weights of the feature map, the feature extraction of key areas is enhanced. The attention mechanism is either self-attention or channel attention. Self-attention calculates the global dependency between pixels, while channel attention adjusts the feature contribution of different channels. The attention layer outputs a weighted feature map to improve classification accuracy.

[0012] Furthermore, in step S103, the change detection also includes time series analysis. When there are remote sensing image data from more than two time phases, continuous time series change detection is performed to identify the dynamic process of land consolidation. The time series analysis uses a time series fitting method, and the continuous time series change detection generates a change trajectory map, which displays the change history of each pixel over time.

[0013] Furthermore, in step S104, the change characteristics also include ecological impact indicators. These indicators are assessed by calculating the change in ecosystem service value caused by land cover change. The ecosystem service value is calculated based on the value coefficient per unit area and the changed area. The value coefficient is referenced to international standards or localized studies. The ecological impact indicators are integrated into the monitoring report for sustainable development assessment.

[0014] Furthermore, in step S105, the monitoring results also include a prediction model, which is trained using time series features to predict future land consolidation trends. The prediction model is based on historical change data and uses regression or classification algorithms to output a consolidation probability map within a preset future time period. The prediction results are used for risk warning and planning support.

[0015] Furthermore, in step S105, the visualization display also includes an interactive network platform, through which users can query monitoring results, customize display layers and statistical indicators. The interactive platform uses Web GIS technology, supports map zooming, panning and attribute querying, and provides data export function. The platform also integrates an alarm module, which automatically sends a notification when abnormal changes are detected.

[0016] This invention relates to a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, belonging to the field of environmental management technology. The invention acquires multi-band high-resolution remote sensing images of the target area before and after consolidation, performs radiometric and atmospheric corrections to ensure data quality, and utilizes a pre-trained deep convolutional neural network land cover classification model to generate land cover classification maps for both time periods. Through pixel-level change detection and confidence analysis, reliable land cover change areas are identified, vectorized, and spatially filtered to extract multi-dimensional change features. The ecological impact and spatial distribution patterns are comprehensively assessed, generating comprehensive monitoring results including spatial distribution maps, statistical reports, and trend predictions. These results are integrated into an interactive geographic information system platform for visualization and early warning. This invention achieves automated, precise, and systematic monitoring and evaluation of the entire land consolidation process. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the workflow of a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, for which this invention is claimed. Figure 2 This is a second workflow diagram of a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, for which protection is claimed in this invention. Figure 3 The third workflow diagram is for a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, which is claimed in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0020] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] According to the first embodiment of the present invention, referring to Figure 1 This invention seeks protection for a comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, comprising the following steps: S101: Acquire remote sensing image data of the target area in the first time phase before land consolidation and the second time phase after land consolidation; S102: Input the remote sensing image data of the first time phase and the second time phase into the pre-trained land cover classification model respectively to generate the land cover classification map of the first time phase and the land cover classification map of the second time phase. The land cover classification map is raster data, and each pixel is assigned a land cover category label. S103: Perform pixel-level change detection on the land cover classification maps of the first and second time phases. By comparing the land cover categories of pixels at the same coordinate positions, identify pixels that have changed and construct a land cover change map. The change detection includes calculating the change confidence level and determining the confidence level of the changed pixels using the probability product method based on the category probabilities output by the land cover classification model. S104: Extract land consolidation-related change features from the land cover change map, including the spatial distribution of change areas, change area statistics, and change type conversion matrix. Calculate the category distribution based on the change type conversion matrix to reflect the richness of land consolidation types. S105: Based on the extracted change features, generate comprehensive land consolidation monitoring results for the entire region, including creating a spatial distribution map of consolidation, compiling a report on the area of ​​consolidation, assessing the trend of consolidation type conversion, and integrating the monitoring results into a geographic information system platform for visualization. The monitoring results also include a consolidation effect rating, which is comprehensively evaluated based on the consolidation intensity index and diversity index. The visualization includes dynamic maps and statistical charts to support decision analysis.

[0022] In this embodiment, the target area is first defined, which can be an administrative division or a natural geographical unit, ranging in size from square kilometers to tens of thousands of square kilometers to accommodate monitoring needs at different scales. The remote sensing image data includes data from the first time phase before land consolidation and the second time phase after consolidation. This data comes from multi-temporal satellite remote sensing imagery, covering the visible light, near-infrared, and short-wave infrared bands to provide rich spectral information. The image spatial resolution is better than 10 meters to ensure the identification of subtle changes in land features.

[0023] The time interval between the first and second time phases is at least one year to cover the entire land consolidation cycle, from planning and implementation to effect evaluation. Data is acquired via satellite platform and processed through radiometric calibration and atmospheric correction. Radiometric calibration converts raw digital values ​​into surface reflectance, eliminating sensor differences; atmospheric correction removes the effects of atmospheric scattering and absorption, ensuring consistency and comparability of data from different time phases. The processed remote sensing imagery is stored in raster format, with each pixel containing spectral values ​​across multiple bands for easy subsequent analysis. Cloud cover effects are also considered during data acquisition; cloudless or low-cloud imagery is selected, and multi-source data are fused when necessary to ensure quality.

[0024] Remote sensing imagery data from the first and second time phases are input into a pre-trained land cover classification model. This model is based on a deep convolutional neural network architecture, employing an encoder-decoder structure. The encoder extracts multi-scale features from the imagery progressively through multiple convolutional and pooling layers, from local details to global context. The decoder upsamples the imagery using deconvolutional layers and skip connections to restore spatial resolution and outputs the land cover category probability for each pixel. Skip connections fuse low-level features from the encoder with high-level features from the decoder to preserve detail. Land cover categories include cultivated land, orchards, forest land, grassland, commercial and service land, industrial and mining land, residential land, public management and public service land, special land, transportation land, water areas and water conservancy facilities land, and other land, covering common land use types. The model output is raster data, with each pixel assigned a land cover category label based on the highest probability category in the probability distribution. While a large-scale land cover annotation dataset is used for pre-training, this method does not involve model training steps; it directly uses the pre-trained model to ensure efficiency and consistency.

[0025] Pixel-level comparisons are performed on land cover classification maps from the first and second time phases to identify pixels showing land cover changes. First, the classification maps from the two time phases are ensured to be spatially aligned, achieved through coordinate system normalization and resampling to achieve pixel-level matching. Then, land cover category labels at the same coordinate locations are compared pixel-by-pixel; if the category labels differ, the pixel is marked as a candidate for change. To improve reliability, change confidence is calculated using a probability product method based on the category probabilities output by the land cover classification model: the maximum probability value is extracted from the probability distributions of the first and second time phases to calculate the change confidence. Only change pixels with confidence scores higher than a preset threshold are retained. This threshold, determined through historical data validation, ranges from 0.5 to 0.9 to accommodate the uncertainty of different land cover categories. Simultaneously, spatial consistency checks are performed, removing isolated change pixels and filling small holes through morphological operations to ensure spatial continuity of the change areas. Finally, a land cover change map is generated, displaying the location and type of changes.

[0026] Features related to integrated land consolidation are extracted from land cover change maps. These include the spatial distribution of changed areas, statistical data on changed areas, and a change type transformation matrix. The rows of the change type transformation matrix represent the land cover category in the first time phase, the columns represent the land cover category in the second time phase, and the matrix elements represent the area proportion that changes from the row category to the column category, used to analyze the type transformation patterns of land consolidation. Simultaneously, a consolidation intensity index is calculated, defined as the ratio of the changed area area to the total area of ​​the target area, reflecting the intensity of consolidation activities. A consolidation diversity index is calculated based on the category distribution of the change type transformation matrix, using the Shannon diversity index to assess the richness and evenness of consolidation types; a higher value indicates greater diversity. These features collectively provide a quantitative assessment of land consolidation, including spatial patterns and type distribution.

[0027] Based on the extracted change characteristics, comprehensive land consolidation monitoring results are generated. This includes creating a spatial distribution map of the consolidation areas, displaying the geographical locations of the changed areas; compiling consolidation area reports, listing the areas of various changes; assessing the conversion trends of consolidation types, and analyzing the main conversion directions. The monitoring results are integrated into a geographic information system platform for visualization, such as dynamic maps and statistical charts, to support decision analysis. Furthermore, consolidation effectiveness is rated, based on a comprehensive assessment using consolidation intensity and diversity indices, resulting in a grade classification such as excellent, good, average, and poor. The visualization includes interactive operations such as zooming, querying, and layer switching, facilitating in-depth analysis of specific areas. The monitoring results can also be exported as report formats for policy evaluation and planning.

[0028] Furthermore, step S102 also includes: The land cover classification model adopts a U-Net network structure. The encoder part includes multiple convolutional layers and pooling layers for downsampling and feature extraction. The decoder part includes multiple deconvolutional layers and skip connections for upsampling to restore spatial resolution and fuse multi-scale features. The skip connections concatenate the feature maps in the encoder with the corresponding feature maps in the decoder. The convolutional layers use rectified linear units as activation functions, the pooling layers use max pooling, and the deconvolutional layers use bilinear interpolation for upsampling. The output layer of the land cover classification model uses the softmax function to generate the class probability distribution for each pixel.

[0029] In this embodiment, the land cover classification model adopts the U-Net network structure. U-Net is an encoder-decoder architecture specifically designed for image segmentation tasks. The encoder part includes multiple convolutional and pooling layers: the convolutional layers use rectified linear units as activation functions to progressively extract local features and global context from the image; the pooling layers use max pooling to reduce the spatial resolution of the feature maps while retaining important features. The decoder part includes multiple deconvolutional layers and skip connections: the deconvolutional layers use bilinear interpolation for upsampling to progressively restore the spatial resolution of the feature maps to match the input image size; the skip connections concatenate the feature maps from the encoder with the corresponding feature maps from the decoder, fusing low-level detail features and high-level semantic features to avoid loss of detail. The output layer uses the softmax function to convert the feature vector of each pixel into a class probability distribution, representing the probability that the pixel belongs to each land cover class. Finally, the class label of each pixel is determined based on the highest probability. The entire network structure learns the feature representation of land cover through pre-training, eliminating the need for retraining in this method and ensuring processing efficiency and consistency.

[0030] Furthermore, referring to Figure 2 The pixel-level change detection in step S103 includes the following sub-steps: Spatial registration is performed on the land cover classification maps of the first and second time phases. The feature point matching algorithm is used to align the classification maps of the two time phases to the same coordinate system. Image resampling is achieved through affine transformation or polynomial transformation to ensure pixel-level alignment accuracy. The land cover category labels of corresponding pixels in the first and second time-phase classification maps are compared pixel by pixel. If the category labels are different, the pixel is initially marked as a candidate pixel for change. The comparison operation is based on a one-to-one correspondence of pixel coordinates and uses a raster calculation tool to record the position coordinates and category change information of each candidate pixel for change. For each candidate pixel that changes, obtain the class probability distribution of the pixel in the first temporal classification model and the class probability distribution of the pixel in the second temporal classification model; Based on the category probability distribution, the change confidence score is calculated, which reflects the degree of confidence that the pixel has indeed undergone land cover change; Set a confidence threshold, which is determined by historical verification data. Compare the change confidence with the threshold. If the change confidence is greater than the threshold, the pixel is confirmed as a changed pixel; otherwise, it is excluded. Post-processing is performed on all confirmed changed pixels, including using morphological opening operations to remove small area noise, using morphological closing operations to fill small holes, and generating a smooth land cover change map. The land cover change map is vectorized into polygon vector data, where each polygon represents a continuous area of ​​change, and the area and perimeter attributes of the polygons are recorded. The vectorization process uses a raster-to-vector algorithm. Spatial filtering is performed on the vectorized change areas to remove areas with an area smaller than the minimum remediation unit, which is set according to the land remediation policy. After spatial filtering, the final land cover change map is generated for feature extraction.

[0031] In this embodiment, the land cover classification maps of the first and second time phases are spatially registered to ensure they are aligned in the same coordinate system. Feature points in the two time-phase images are extracted using feature point matching algorithms such as SIFT or ORB, and then these feature points are matched. Based on the matched points, affine or polynomial transformation parameters are calculated to transform the second-phase classification map to the coordinate system of the first time phase. The resampling process uses interpolation methods such as nearest neighbor or bilinear interpolation to ensure that the registration error is less than one pixel, guaranteeing pixel-level alignment accuracy.

[0032] After registration, the land cover category labels of corresponding pixels in the two temporal classification maps are compared pixel by pixel. If the category labels are different, the pixel is initially marked as a candidate pixel for change. A raster calculation tool is used to efficiently traverse all pixels and record the location coordinates and category change information of each candidate pixel for change.

[0033] For each candidate pixel with a change, the class probability distribution for the first and second time phases is obtained from the output of the land cover classification model. The probability distribution is a vector containing the probability value of the pixel belonging to each land cover class, stored as a multidimensional array.

[0034] The change confidence score is calculated based on the category probability distribution. The maximum probability value is extracted from the first time-phase probability distribution and the second time-phase probability distribution. The change confidence score is then calculated as 1 minus the product of these two maximum probability values. The change confidence score ranges from 0 to 1; a higher value indicates a more reliable change.

[0035] A confidence threshold, ranging from 0.5 to 0.9, is determined using historical validation data. The change confidence of each candidate pixel is compared to the threshold; if it exceeds the threshold, the pixel is confirmed as changed; otherwise, it is excluded. The threshold setting considers the diversity and spatial heterogeneity of land cover categories, and the threshold can be adjusted for different categories to optimize detection accuracy.

[0036] Perform morphological operations on all confirmed changed pixels. Use morphological opening to remove small-area noise: the opening operation first erodes and then dilates, eliminating isolated points. Use morphological closing to fill small holes: the closing operation first dilates and then erodes, connecting broken areas. The morphological operations use circular structuring elements with a radius of 3 to 5 pixels.

[0037] The land cover change map is converted from raster format to vector format. Raster-to-vector algorithms such as boundary tracing or region growing are used to identify continuous change areas and generate polygons. Each polygon represents a change patch, recording its area, perimeter, and change type attributes. The change type is determined based on a combination of categories from the first and second time phases.

[0038] The vectorized change areas are filtered to remove areas smaller than the minimum remediation unit. The minimum remediation unit is set according to land consolidation policies and is typically 100 square meters or more to ensure the monitoring results are meaningful. The filtered data is then used to generate the final land cover change map for feature extraction.

[0039] Furthermore, referring to Figure 3 Step S104, which extracts change features related to comprehensive land consolidation, includes the following sub-steps: Extract all changed pixels from the land cover change map, count the total number of changed pixels, and calculate the total area of ​​the changed region based on the pixel size. The total area of ​​the changed region is obtained by multiplying the number of pixels by the area of ​​a single pixel, and the area of ​​a single pixel is calculated based on the spatial resolution of the remote sensing image. Cluster analysis is performed on the changing regions to aggregate spatially adjacent changing pixels into continuous changing patches. A connected component labeling algorithm is used to identify spatially continuous pixel groups and assign a unique identifier to each patch. Record attribute information for each changed patch, including patch area, patch perimeter, patch shape index, and land cover category for the first and second time phases; Construct a land cover change type transformation matrix. The rows of the matrix represent the land cover category in the first time phase, the columns represent the land cover category in the second time phase, and the matrix elements represent the number or area of ​​patches that change from the row category to the column category. The transformation matrix is ​​used to analyze the type transformation pattern of land consolidation. Calculate the remediation intensity index to determine the ratio of the total area of ​​the changed area to the total area of ​​the target area; The land consolidation diversity index is calculated based on the category distribution of the change type transformation matrix, reflecting the richness and evenness of land consolidation types; Analyze the spatial distribution patterns of changed patches, use spatial statistical methods to assess the degree of clustering or dispersion of changed patches, and identify hotspots for remediation. Based on the change characteristics, a remediation priority assessment is generated. According to the area, shape index and transformation type of the patch, the priority score of each patch is calculated. The patches with high scores indicate that they need to be monitored more closely.

[0040] In this embodiment, all changed pixels are extracted from the land cover change map, and the total number of changed pixels is counted. The area of ​​a single pixel is calculated based on the spatial resolution of the remote sensing image; for example, for a 10-meter resolution image, the area of ​​a single pixel is 100 square meters. Then, the total area of ​​the changed region is calculated by multiplying the number of pixels by the area of ​​a single pixel.

[0041] Cluster analysis is performed on the changing regions to group spatially adjacent changing pixels into continuous changing patches. A connected component labeling algorithm is used to identify spatially continuous pixel groups based on 8-neighbor or 4-neighbor connectivity. Each patch is assigned a unique identifier for easy management.

[0042] Each changed land parcel records attribute information, including parcel area, parcel perimeter, parcel shape index, first-phase land cover category, and second-phase land cover category. The parcel shape index is calculated as the ratio of perimeter to area; a higher value indicates a more complex shape, potentially reflecting the boundary effects of remediation efforts. Attribute information is stored as a vector attribute table for easy querying and analysis.

[0043] Create a land cover change type transformation matrix, where rows represent land cover categories in the first time phase and columns represent land cover categories in the second time phase. Matrix elements represent the number or area of ​​patches that transitioned from a row category to a column category. The transformation matrix is ​​stored in tabular form, and a Sankey diagram is used to visualize the flow relationships between categories, enabling analysis of land consolidation type transformation patterns.

[0044] The remediation intensity index is defined as the ratio of the total area of ​​the changed area to the total area of ​​the target area. The total area of ​​the target area is calculated from the coverage of the remote sensing image. The remediation intensity index is normalized to a range of 0 to 1 for easy cross-regional comparison. A higher value indicates a higher remediation intensity.

[0045] Based on the change type transformation matrix, the area proportion of each change type is extracted. Then, the Shannon diversity index is calculated by summing each proportion and multiplying it by the natural logarithm of the proportion, and then taking the negative value. The higher the value, the higher the diversity, reflecting the richness and evenness of the types of remediation activities.

[0046] Spatial statistical methods were used to assess the clustering degree of change patches. The nearest neighbor index was calculated, which is the ratio of the average nearest neighbor distance to the expected distance of a random distribution; a value less than 1 indicates clustering, and a value greater than 1 indicates dispersion. Simultaneously, Ripley's K-function was used to analyze multi-scale spatial patterns and identify remediation hotspots.

[0047] Based on the area, shape index, and transformation type of the map patch, a priority score is calculated for each patch. The score is obtained by weighted summation of multiple feature indicators, with weights adjusted according to the remediation policy objectives. Map patches with high scores indicate those requiring focused monitoring.

[0048] Furthermore, in step S101, the remote sensing image data comes from the same satellite sensor platform to ensure data consistency and comparability. The image acquisition time is selected during the vegetation growing season or the non-growing season to highlight land cover changes. The remote sensing image data also includes radar remote sensing data to penetrate cloud cover, and land cover is analyzed mainly using optical remote sensing images.

[0049] In this embodiment, in step S101, the remote sensing image data comes from the same satellite sensor platform to ensure data consistency and comparability. For example, Landsat series satellites, Sentinel-2 satellites, or China's Gaofen series satellites are used. These platforms provide multispectral imagery, including visible, near-infrared, and shortwave infrared bands. The image acquisition time is selected during the vegetation growing season or the non-growing season to highlight land cover changes: during the growing season, vegetation cover is high, making it easy to distinguish between vegetated and non-vegetated areas; during the non-growing season, the land cover is more exposed, making it easier to identify changes in built-up land. In addition, radar remote sensing data can be combined to penetrate cloud cover and provide all-weather observation capabilities. However, optical remote sensing imagery is mainly used for land cover analysis because it provides rich spectral information and is more suitable for land cover classification.

[0050] Furthermore, in step S102, the land cover classification model also includes an attention mechanism, which is embedded in the encoder-decoder structure. By calculating the importance weights of the feature map, the feature extraction of key areas is enhanced. The attention mechanism is either self-attention or channel attention. Self-attention calculates the global dependency between pixels, while channel attention adjusts the feature contribution of different channels. The attention layer outputs a weighted feature map to improve classification accuracy.

[0051] In this embodiment, in step S102, the land cover classification model further includes an attention mechanism. This attention mechanism is embedded in the encoder-decoder structure to enhance feature extraction from key regions. The self-attention mechanism calculates global dependencies between pixels, capturing long-range contextual information, for example, by calculating the association weights of each pixel with all other pixels and weighted summing of features to highlight important regions. The channel attention mechanism adjusts the feature contributions of different channels, reweighting the feature map according to the importance of each channel, for example, by generating channel weights through global average pooling and then multiplying them by the original feature map. The attention layer outputs a weighted feature map, which is used in subsequent convolutional layers to improve classification accuracy. The attention mechanism enables the model to focus on key features related to land cover, such as vegetation indices or texture features, improving the model's performance in complex scenarios.

[0052] Furthermore, in step S103, the change detection also includes time series analysis. When there are remote sensing image data from more than two time phases, continuous time series change detection is performed to identify the dynamic process of land consolidation. The time series analysis uses a time series fitting method, and the continuous time series change detection generates a change trajectory map, which displays the change history of each pixel over time.

[0053] In this embodiment, step S103 of the change detection further includes time-series analysis. When remote sensing image data from more than two time phases are available, continuous time-series change detection is performed to identify the dynamic process of land consolidation. For example, multi-year remote sensing images of the target area are acquired, covering multiple time points from before to after consolidation. Time-series analysis uses time series fitting methods, such as linear regression or seasonal decomposition. Linear regression fits the land cover change trend for each pixel, assessing whether it is a continuous change or an abrupt change; seasonal decomposition separates the trend, seasonal, and residual components in the time series, helping to identify long-term changes caused by consolidation. Continuous time-series change detection generates a change trajectory map, showing the change history of each pixel over time, such as the transformation from cultivated land to construction land and then to green space. This helps to understand the dynamic process and phased effects of land consolidation, providing support for long-term monitoring.

[0054] Furthermore, in step S104, the change characteristics also include ecological impact indicators. These indicators are assessed by calculating the change in ecosystem service value caused by land cover change. The ecosystem service value is calculated based on the value coefficient per unit area and the changed area. The value coefficient is referenced to international standards or localized studies. The ecological impact indicators are integrated into the monitoring report for sustainable development assessment.

[0055] In this embodiment, step S104 further includes ecological impact indicators in the change characteristics. These indicators are assessed by calculating the change in ecosystem service value caused by land cover change. First, referring to international standards or localized studies, an ecosystem service value coefficient per unit area is assigned to each land cover category. For example, forest land has a higher ecosystem service value, while built-up land has a lower value. Then, for each changed patch, the change in ecosystem service value is calculated based on the categories of the first and second time phases: the value coefficient of the second time phase category is subtracted from the value coefficient of the first time phase category, and then multiplied by the changed area to obtain the change in ecosystem service value. Positive values ​​indicate ecological improvement, and negative values ​​indicate ecological degradation. These ecological impact indicators are integrated into the monitoring report to assess the sustainability of land consolidation and support ecological protection decisions.

[0056] Furthermore, in step S105, the monitoring results also include a prediction model, which is trained using time series features to predict future land consolidation trends. The prediction model is based on historical change data and uses regression or classification algorithms to output a consolidation probability map within a preset future time period. The prediction results are used for risk warning and planning support.

[0057] In this embodiment, step S105 further includes a prediction model in the monitoring results. A machine learning prediction model is trained using historical time-series change features to predict future land consolidation trends. However, this method does not involve a model training step; instead, it uses an existing pre-trained model. The prediction model is based on historical change data, such as land cover change patches and features from the past few years, using regression or classification algorithms. Regression algorithms predict future consolidation area or intensity, while classification algorithms predict the probability of consolidation type. The output is a consolidation probability map for a future period, showing which areas are likely to be consolidated. The prediction results are used for risk warning and planning support, for example, identifying potentially encroached farmland or ecologically sensitive areas, providing forward-looking guidance for land management.

[0058] Furthermore, in step S105, the visualization display also includes an interactive network platform, through which users can query monitoring results, customize display layers and statistical indicators. The interactive platform uses Web GIS technology, supports map zooming, panning and attribute querying, and provides data export function. The platform also integrates an alarm module, which automatically sends a notification when abnormal changes are detected.

[0059] In this embodiment, step S105 includes an interactive web platform for visualization. Users can access monitoring results through a web interface and customize display layers and statistical indicators. The platform uses Web GIS technology, supporting map zooming, panning, and attribute querying. For example, users can click on a changed patch to view its attribute information, such as area, change type, and priority score. It also provides data export functionality, supporting common formats such as Shapefile or CSV for further analysis. The platform also integrates an alarm module, automatically sending notifications to relevant departments when abnormal changes are detected. Alarm thresholds can be set according to policy requirements to ensure timely response. The interactive platform enhances user experience and promotes data sharing and collaborative analysis.

[0060] The following is a specific example: The target area is a typical administrative unit, covering several thousand square kilometers, with terrain including plains, hills, and water bodies. Remote sensing imagery data of this area were acquired in the first time phase before and the second time phase after comprehensive land consolidation. The consolidation period covered at least one year and involved projects such as farmland protection, construction land consolidation, and ecological restoration. This embodiment demonstrates the entire process from data acquisition to the generation of monitoring results.

[0061] First, remote sensing imagery data of the target area in the first and second time phases were acquired from the satellite sensor platform. The data included visible light, near-infrared, and shortwave infrared bands, with a spatial resolution better than 10 meters to ensure the identification of small ground features. Image acquisition was conducted during the vegetation growing season to highlight changes in vegetation cover. The data underwent radiometric calibration and atmospheric correction: radiometric calibration converted the raw digital values ​​to surface reflectance, eliminating sensor differences; atmospheric correction removed atmospheric scattering and absorption effects, ensuring consistency and comparability of data from different time phases. The processed remote sensing images were stored in raster format, with each pixel containing spectral information across multiple bands. Additionally, radar remote sensing data was used to address cloud cover, but optical remote sensing imagery was the primary focus of the analysis. This step emphasizes that the data must originate from the same satellite platform to ensure consistency.

[0062] Remote sensing image data from the first and second time phases are input into a pre-trained land cover classification model. This model is based on a deep convolutional neural network architecture, employing a U-Net structure. The encoder consists of multiple convolutional and pooling layers: the convolutional layers use rectified linear units as activation functions to extract local features; the pooling layers use max pooling to reduce spatial resolution while preserving important features. The decoder consists of multiple deconvolutional layers and skip connections: the deconvolutional layers use bilinear interpolation for upsampling to restore spatial resolution; the skip connections concatenate the feature maps from the encoder with the corresponding feature maps from the decoder, fusing low-level details and high-level semantic features. The output layer uses the softmax function to generate the land cover category probability distribution for each pixel. Land cover categories include cultivated land, orchards, forest land, grassland, commercial and service land, industrial and mining land, residential land, public management and public service land, special land, transportation land, water area and water conservancy facilities land, and other land. The model output is a raster-formatted land cover classification map, with each pixel assigned a category label. The pre-training process uses a large-scale labeled dataset, but this method does not involve model training. Furthermore, the model incorporates an attention mechanism: self-attention calculates global dependencies between pixels, highlighting key regions; channel attention adjusts the importance of features across different bands, enhancing classification accuracy.

[0063] Performing pixel-level change detection on the land cover classification maps of the first and second time phases includes the following sub-steps: S1031: Spatial Registration: Feature points in the two temporal classification maps are extracted and matched using a feature point matching algorithm such as SIFT. The second temporal classification map is aligned to the coordinate system of the first temporal map through affine transformation. Bilinear interpolation is used for resampling to ensure that the registration error is less than one pixel.

[0064] S1032: Pixel-by-pixel comparison: Compare the class labels at corresponding coordinates in two temporal classification images pixel by pixel. If the class labels are different, they are marked as candidate pixels for change, and the position and class change information are recorded.

[0065] S1033: Obtain the probability distribution: For each candidate pixel with change, obtain the class probability distribution vector form of the first and second phases from the softmax output layer of the land cover classification model.

[0066] S1034: Calculate the confidence level of change: Based on the probability distribution, extract the maximum class probability value from the first and second time phases respectively, and calculate the confidence level of change as 1 minus the product of these two maximum probability values. The confidence level value ranges from 0 to 1, with a higher value indicating a more reliable change.

[0067] S1035: Set the confidence threshold: Determine the confidence threshold range from 0.5 to 0.9 based on historical validation data, and compare the change confidence level with the threshold. Pixels with confidence levels higher than the threshold are confirmed as changed pixels; otherwise, they are excluded. The threshold can be adjusted according to land cover category to optimize detection accuracy.

[0068] S1036: Post-processing: Perform morphological operations on all confirmed changed pixels. Use circular structuring elements with a radius of 3 to 5 pixels to perform opening operations (erosion followed by dilation) to remove isolated noise points, and closing operations (dilation followed by erosion) to fill small holes, generating a smooth land cover change map.

[0069] S1037: Vectorization: The change map is converted into polygon vector data using a raster-to-vector algorithm such as boundary tracing. Each polygon represents a continuously changing region, and the area, perimeter, and change type attributes are recorded based on a combination of categories from the first and second time phases.

[0070] S1038: Spatial Filtering: Remove change areas smaller than the minimum remediation unit, such as 100 square meters, to ensure the monitoring results are meaningful. A land cover change map is then generated for subsequent analysis.

[0071] In addition, this step includes time series analysis: when there are more than two time-phase data, continuous time-series change detection is performed. For example, time series fitting methods such as linear regression are used to analyze the change trend of each pixel, generating a change trajectory map to show the dynamic process of land consolidation.

[0072] Extracting land consolidation-related change features from land cover change maps includes the following sub-steps: S1041: Statistical change area: Count the total number of all changed pixels and calculate the total area of ​​the changed area based on the image spatial resolution. For example, for a 10-meter resolution image, the area of ​​a single pixel is 100 square meters. The total area is obtained by multiplying the number of pixels by the area of ​​a single pixel.

[0073] S1042: Clustering analysis: Using the connected component labeling algorithm, spatially adjacent changing pixels are aggregated into continuously changing patches based on 8-neighborhood connectivity. Each patch is assigned a unique identifier, representing an independent remediation area.

[0074] S1043: Record polygon attributes: Record the area, perimeter, and shape index for each changed polygon, which is the ratio of perimeter to area, reflecting the compactness and land cover category in the first and second time phases. Attributes are stored as a vector attribute table.

[0075] S1044: Constructing a Transition Matrix: Create a matrix where rows represent the first phase category, columns represent the second phase category, and matrix elements represent the area proportion transitioning from the row category to the column category. The matrix is ​​stored in tabular form, and a Sankey diagram is used to visualize the flow relationships between categories.

[0076] S1045: Calculate the remediation intensity index: defined as the ratio of the total area of ​​the changed area to the total area of ​​the target area, normalized to the range of 0 to 1. The larger the value, the higher the remediation intensity.

[0077] S1046: Calculate the remediation diversity index: Based on the change type transformation matrix, extract the area proportion of each change type, calculate the Shannon diversity index by summing each proportion and multiplying it by the natural logarithm of the proportion, and then taking the negative value. The higher the value, the higher the diversity, reflecting the richness of the remediation types.

[0078] S1047: Analyze spatial distribution patterns: Use spatial statistical methods, such as calculating the ratio of the average nearest neighbor distance to the expected distance of the random distribution. A value less than 1 indicates clustering of change patches, while a value greater than 1 indicates dispersion. At the same time, use Ripley's K function to analyze multi-scale spatial patterns and identify hotspots for remediation.

[0079] S1048: Generating a Priority Assessment for Remediation: Based on the area, shape index, and transformation type of the map patch, a priority score is calculated for each map patch through weighted summation. The weights are adjusted according to the remediation policy objectives. Map patches with high scores are marked as key monitoring areas.

[0080] Ecological impact indicators: Based on land cover change, the change in ecosystem service value is calculated. Referring to international standards, a unit area value coefficient is assigned to each land cover category. Then, for each change patch, the value difference between the second and first time phases is calculated multiplied by the area to obtain the change in ecosystem service value. Positive values ​​indicate ecological improvement, and negative values ​​indicate degradation; this is integrated into the feature analysis.

[0081] Step S105: Generate monitoring results Based on the extracted change characteristics, the monitoring results of comprehensive land consolidation across the entire region are generated: Create a spatial distribution map of the remediation area, showing the geographical location and type of the changed patches; compile a report on the remediation area, listing the area proportion of each type of change; assess the trend of remediation type conversion, and analyze the trend of major conversion directions, such as the conversion of farmland to construction land.

[0082] The monitoring results are integrated into a geographic information system platform for visualization. The platform includes an interactive web interface, allowing users to zoom and pan maps, query patch attributes, and customize display layers and statistical indicators. It also provides data export functionality, supporting common formats.

[0083] The effectiveness rating is based on a comprehensive assessment of the remediation intensity index and the diversity index, and is divided into levels such as high, medium and low.

[0084] Predictive Model: This method uses historical time-series change features to train a machine learning model, such as a regression algorithm, to predict future land consolidation trends. It outputs a consolidation probability map, showing areas that may undergo changes in the future, for risk warning and planning support. However, this method does not involve model training steps; it directly uses a pre-trained model.

[0085] The platform integrates an alarm module that automatically sends a notification to the user when abnormal changes are detected, such as illegal occupation of farmland.

[0086] Description of experimental test data In this embodiment, the test data is based on hypothetical simulation results and is presented in a general description: Land cover classification results: The model successfully distinguished various land covers. For example, in densely vegetated areas, the classification accuracy of woodland and high-cover grassland was high; in urban areas, the construction land category was accurately identified. The classification map shows that the first time phase was dominated by cultivated land and woodland, while the second time phase showed a greater number of construction land and water areas.

[0087] Change detection results: Change detection identified multiple changed patches, mainly distributed at the edge of the region and around towns. Change types included conversion of farmland to construction land, conversion of forest land to farmland, and expansion of water areas. False detections were reduced through confidence thresholding and post-processing, such as removing isolated pixels and merging small regions.

[0088] Feature extraction results: The significant proportion of the total area of ​​the changed area to the total area of ​​the target area indicates a high intensity of remediation. The change type transformation matrix shows that the proportion of the area transformed from cultivated land to construction land is relatively large, while the transformation from grassland to forest reflects ecological restoration. The remediation diversity index value is moderate, indicating that the remediation types are relatively concentrated. Spatial distribution analysis shows that the changed patches exhibit a clustering pattern, with hotspots located along transportation routes and near water bodies.

[0089] Monitoring results: The spatial distribution map clearly shows the spatial pattern of the changed areas; the statistical reports summarize the area of ​​various changes; the overall rating of the remediation effect is moderate, with some areas receiving higher ratings due to ecological improvement. The predictive model indicates that future remediation may expand to the northeast, suggesting the need for advance planning.

[0090] Platform Interaction: Users successfully queried the attributes of specific landforms, such as area and change type, through the web interface and exported the data for further analysis. The alarm module issued a notification when rapid changes were detected.

[0091] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0092] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0093] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A comprehensive land consolidation monitoring method based on artificial intelligence and remote sensing imagery, characterized in that, Includes the following steps: S101: Acquire remote sensing image data of the target area in the first time phase before land consolidation and the second time phase after land consolidation; S102: Input the remote sensing image data of the first time phase and the second time phase into the pre-trained land cover classification model respectively to generate the land cover classification map of the first time phase and the land cover classification map of the second time phase. The land cover classification map is raster data, and each pixel is assigned a land cover category label. S103: Perform pixel-level change detection on the land cover classification maps of the first and second time phases. By comparing the land cover categories of pixels at the same coordinate positions, identify pixels that have changed and construct a land cover change map. The change detection includes calculating the change confidence level and determining the confidence level of the changed pixels using the probability product method based on the category probabilities output by the land cover classification model. S104: Extract land consolidation-related change features from the land cover change map, including the spatial distribution of change areas, change area statistics, and change type conversion matrix. Calculate the category distribution based on the change type conversion matrix to reflect the richness of land consolidation types. S105: Based on the extracted change features, generate comprehensive land consolidation monitoring results for the entire region, including creating a spatial distribution map of consolidation, compiling a report on the area of ​​consolidation, assessing the trend of consolidation type conversion, and integrating the monitoring results into a geographic information system platform for visualization. The monitoring results also include a consolidation effect rating, which is comprehensively evaluated based on the consolidation intensity index and diversity index. The visualization includes dynamic maps and statistical charts to support decision analysis. Step S104, which extracts change features related to comprehensive land consolidation, includes the following sub-steps: S1041: Extract all changed pixels from the land cover change map, count the total number of changed pixels, and calculate the total area of ​​the changed region based on the pixel size. The total area of ​​the changed region is obtained by multiplying the number of pixels by the area of ​​a single pixel. The area of ​​a single pixel is calculated based on the spatial resolution of the remote sensing image. S1042: Perform cluster analysis on the change region, aggregate spatially adjacent change pixels into continuous change patches, use the connected component labeling algorithm to identify spatially continuous pixel groups, and assign a unique identifier to each patch; S1043: Record attribute information for each changed patch, including patch area, patch perimeter, patch shape index, and land cover category in the first and second time phases; S1044: Construct a land cover change type transformation matrix. The rows of the matrix represent the land cover category in the first time phase, the columns represent the land cover category in the second time phase, and the matrix elements represent the number or area of ​​patches that change from the row category to the column category. The transformation matrix is ​​used to analyze the type transformation pattern of land consolidation. S1045: Calculate the remediation intensity index and determine the ratio of the total area of ​​the changed area to the total area of ​​the target area; S1046: Calculate the land consolidation diversity index based on the category distribution of the change type transformation matrix, reflecting the richness and evenness of land consolidation types; S1047: Analyze the spatial distribution pattern of changed patches, use spatial statistical methods to assess the degree of clustering or dispersion of changed patches, and identify hotspots for remediation; S1048: Based on the change characteristics, generate a remediation priority assessment. Calculate the priority score for each patch based on its area, shape index, and transformation type. Patches with high scores indicate that they require key monitoring.

2. The method for comprehensive land consolidation monitoring based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, Step S102 further includes: The land cover classification model adopts a U-Net network structure. The encoder part includes multiple convolutional layers and pooling layers for downsampling and feature extraction. The decoder part includes multiple deconvolutional layers and skip connections for upsampling to restore spatial resolution and fuse multi-scale features. The skip connections concatenate the feature maps in the encoder with the corresponding feature maps in the decoder. The convolutional layers use rectified linear units as activation functions, the pooling layers use max pooling, and the deconvolutional layers use bilinear interpolation for upsampling. The output layer of the land cover classification model uses the softmax function to generate the class probability distribution for each pixel.

3. The method for monitoring comprehensive land consolidation based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, The pixel-level change detection in step S103 includes the following sub-steps: S1031: Spatial registration is performed on the land cover classification maps of the first and second time phases. The feature point matching algorithm is used to align the classification maps of the two time phases to the same coordinate system. Image resampling is achieved through affine transformation or polynomial transformation to ensure pixel-level alignment accuracy. S1032: Compare the land cover category labels of corresponding pixels in the first and second time-phase classification maps pixel by pixel. If the category labels are different, the pixel is initially marked as a candidate pixel for change. The comparison operation is based on a one-to-one correspondence of pixel coordinates and uses a raster calculation tool to record the position coordinates and category change information of each candidate pixel for change. S1033: For each candidate pixel that changes, obtain the class probability distribution of the pixel in the first temporal classification model and the class probability distribution of the pixel in the second temporal classification model; S1034: Calculate the change confidence score based on the category probability distribution. The change confidence score reflects the degree of confidence that the pixel has indeed undergone land cover change. S1035: Set a confidence threshold, which is determined by historical verification data. Compare the change confidence with the threshold. If the change confidence is greater than the threshold, the pixel is confirmed as a changed pixel; otherwise, it is excluded. S1036: Post-process all confirmed changed pixels, including using morphological opening operations to remove small area noise, using morphological closing operations to fill small holes, and generating a smooth land cover change map. S1037: Vectorize the land cover change map into polygon vector data, where each polygon represents a continuous change area, and record the area and perimeter attributes of the polygons. The vectorization process uses a raster-to-vector algorithm. S1038: Spatial filtering is performed on the vectorized change area to remove areas with an area smaller than the minimum remediation unit. The minimum remediation unit is set according to the land remediation policy. After spatial filtering, the final land cover change map is generated for feature extraction.

4. The method for monitoring comprehensive land consolidation based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, In step S101, the remote sensing image data comes from the same satellite sensor platform to ensure data consistency and comparability. The image acquisition time is selected during the vegetation growing season or the non-growing season to highlight land cover changes. The remote sensing image data also includes radar remote sensing data to penetrate cloud cover, and land cover is analyzed mainly using optical remote sensing images.

5. The method for monitoring comprehensive land consolidation based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, In step S102, the land cover classification model further includes an attention mechanism, which is embedded in the encoder-decoder structure. By calculating the importance weights of the feature map, the feature extraction of key areas is enhanced. The attention mechanism is either self-attention or channel attention. Self-attention calculates the global dependency between pixels, while channel attention adjusts the feature contribution of different channels. The attention layer outputs a weighted feature map to improve classification accuracy.

6. The method for monitoring comprehensive land consolidation based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, In step S103, the change detection also includes time series analysis. When there are remote sensing image data from more than two time phases, continuous time series change detection is performed to identify the dynamic process of land consolidation. The time series analysis uses a time series fitting method, and the continuous time series change detection generates a change trajectory map, which displays the change history of each pixel over time.

7. The method for monitoring comprehensive land consolidation based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, In step S104, the change characteristics also include ecological impact indicators. These indicators are assessed by calculating the change in ecosystem service value caused by land cover change. The ecosystem service value is calculated based on the value coefficient per unit area and the changed area. The value coefficient is based on international standards or localized studies. The ecological impact indicators are integrated into the monitoring report for sustainable development assessment.

8. The method for monitoring comprehensive land consolidation based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, In step S105, the monitoring results also include a prediction model. The machine learning prediction model is trained using time series features to predict future land consolidation trends. The prediction model is based on historical change data and uses regression or classification algorithms to output a consolidation probability map within a preset future time period for risk warning and planning support.

9. The method for comprehensive land consolidation monitoring based on artificial intelligence and remote sensing imagery according to claim 1, characterized in that, In step S105, the visualization display also includes an interactive network platform, through which users can query monitoring results, customize display layers and statistical indicators. The interactive network platform uses Web GIS technology, supports map zooming, panning and attribute querying, and provides data export function. The interactive network platform also integrates an alarm module, which automatically sends a notification when abnormal changes are detected.

Images