Land use monitoring method and system based on remote sensing and big data

Abstract

The application provides a land use monitoring method and system based on remote sensing and big data, and relates to the technical field of land monitoring.The method comprises the following steps: dividing a sub-region, and acquiring multi-time-phase and multi-resolution remote sensing data of the sub-region; performing feature extraction and classification on the remote sensing data based on a deep learning model to generate a land use classification map; identifying a change region based on a double-time-phase difference attention network, and constructing a change driving factor library that fuses meteorological data and human activity data; detecting an abnormal region based on the driving factor library, combining a dynamic early warning threshold to generate an abnormal type label and an attribution analysis report; visually rendering the monitoring result, and outputting an abnormal region early warning map and a disposal suggestion; and achieving high-precision feature extraction and classification, deep analysis of a change region and driving factors, effective detection of an abnormal region and early warning, and improving the accuracy, timeliness and practicality of land use monitoring.

Description

Technical Field

[0001] This application relates to the field of land monitoring technology, and in particular to land use monitoring methods and systems based on remote sensing and big data. Background Technology

[0002] Land use monitoring is crucial for many fields such as resource management, ecological protection, urban planning, and climate change response. With the acceleration of urbanization, population growth, and increased human intervention in the natural environment, land use patterns are changing at an unprecedented rate, making accurate, timely, and comprehensive land use monitoring an urgent need.

[0003] Traditional land use monitoring methods mainly rely on field surveys and a limited amount of aerial photogrammetry. Field surveys not only consume a lot of manpower, resources, and time, but also have a limited monitoring range, making it difficult to cover large areas, especially in remote or inaccessible regions where data acquisition is extremely difficult. While aerial photogrammetry expands the monitoring range to some extent, it is limited by factors such as weather conditions and flight costs, making it difficult to achieve high-frequency, long-term dynamic monitoring, which greatly reduces its timeliness.

[0004] With the rise of remote sensing technology, its advantages, such as large-area synchronous observation, periodic repeated coverage, and lack of limitation by ground conditions, have brought new opportunities for land use monitoring. Early remote sensing monitoring was mostly based on image data from a single time phase and a single sensor, which limited the information obtained and made it difficult to accurately identify complex land use types and their dynamic changes. For example, when distinguishing different land use types under vegetation cover (such as woodland and grassland), single-band or simple multi-band remote sensing images often cannot provide sufficient feature information, resulting in low classification accuracy.

[0005] Furthermore, traditional methods fall short when dealing with cloud cover issues. Cloud obstruction of optical remote sensing imagery severely impacts data integrity and usability, especially in areas with frequent cloud and rain weather, resulting in a significant loss of effective monitoring data and greatly restricting the accuracy and timeliness of land use monitoring. While radar imagery can penetrate clouds to obtain some information, relying solely on radar imagery suffers from relatively low resolution and insensitivity to certain ground features. Moreover, traditional cloud restoration methods are mostly simplistic and crude, failing to achieve high-precision cloud removal and image information recovery.

[0006] In the field of land use change analysis and anomaly detection, traditional methods often lack the ability to effectively integrate and deeply mine multi-source data. Relying solely on a single data source or simple data overlay analysis cannot fully understand the driving mechanisms of land use change, making it difficult to accurately identify areas of abnormal change and their causes. For example, when analyzing the complex relationship between land use change and meteorological factors and human activities, traditional methods cannot effectively capture the spatiotemporal interactions and nonlinear relationships between multiple factors, leading to a one-sided understanding of the driving factors of change, which in turn affects the accurate detection and early warning of anomaly areas.

[0007] In summary, existing land use monitoring methods have significant shortcomings in terms of the comprehensiveness, accuracy, and timeliness of data acquisition, the efficiency of data processing, and the depth and breadth of change analysis and anomaly detection. There is an urgent need for an innovative integrated monitoring method based on remote sensing and big data to overcome these deficiencies and achieve precise, dynamic, and intelligent monitoring of land use. Summary of the Invention

[0008] The purpose of this application is to provide a land use monitoring method and system based on remote sensing and big data. Through multi-source data fusion and standardized storage, high-precision feature extraction and classification through models, in-depth analysis of change areas and driving factors, effective detection of abnormal areas and early warning, and efficient visualization output for decision support, the accuracy, timeliness and practicality of land use monitoring are comprehensively improved.

[0009] The objective of this application is achieved through the following technical solution:

[0010] In a first aspect, the present invention provides a land use monitoring method based on remote sensing and big data, the method comprising:

[0011] S1. Divide the data into sub-regions, acquire multi-temporal and multi-resolution remote sensing data for each sub-region, and standardize the storage of multi-source data using a distributed storage framework.

[0012] S2. Based on a deep learning model, feature extraction and classification of remote sensing data are performed to generate a land use classification map;

[0013] S3. Based on a dual-temporal difference attention network, compare land use classification maps from multiple periods to identify areas of change and construct a library of change-driving factors that integrate meteorological data and human activity data.

[0014] S4. Based on the driving factor library, detect abnormal regions and generate abnormal type labels and attribution analysis reports by combining dynamic early warning thresholds.

[0015] S5. Using geographic information system software and a distributed computing framework, visualize and render the monitoring results, and output early warning maps of abnormal areas and disposal suggestions.

[0016] Preferably, S1 includes:

[0017] Simultaneously acquire satellite optical images, radar images, and UAV hyperspectral data, and access real-time weather station data and human activity data;

[0018] Spatiotemporally adaptive preprocessing of remote sensing data, including cloud restoration based on GAN and radar data fusion;

[0019] The HDFS hierarchical storage architecture is used to achieve spatiotemporal data block management.

[0020] Preferably, the cloud repair based on GAN and radar data fusion includes:

[0021] The cloud-covered area of ​​the optical image is spatiotemporally registered with the radar image, and the backscattering coefficient and texture features of the radar image are extracted.

[0022] A dual-branch generative adversarial network is constructed, including an optical branch and a radar branch. The optical branch is input with registered optical images to repair the semantic information of cloud-covered areas; the radar branch is input with backscattering coefficients and texture features to reconstruct the surface structure under the cloud layer.

[0023] Based on the spectral features of optical images and the backscattering information of radar images, a pixel-level cloud thickness map is generated through a cloud thickness estimation subnetwork.

[0024] Based on the pixel-level cloud thickness map, a dynamic weight is assigned to each pixel, where thin cloud regions are mainly repaired by optical branches and thick cloud regions are mainly reconstructed by radar branches.

[0025] Features from the optical and radar branches are fused through cross-modal attention mechanisms and conditional gated convolutions;

[0026] Design a multi-physical constraint loss function, including radar-spectral correlation loss, terrain consistency loss, and seasonal vegetation index matching loss, to optimize the restoration results.

[0027] Preferably, S2 includes:

[0028] A CNN-Transformer-UNet hybrid network is constructed. The CNN convolutional neural network extracts local features from multispectral images, and the Transformer encoder performs long-term time-series dependency analysis on vegetation index and land surface temperature of multi-temporal images. The UNet decoder fuses spatial context information to generate a pixel-level land use classification map.

[0029] Based on the pre-trained model, a domain adaptation module is introduced to align the feature distribution of the public dataset with that of the target region data through adversarial training, thereby optimizing the classification accuracy in low-sample scenarios.

[0030] Preferably, S3 includes:

[0031] Land use change areas are detected using a dual-temporal differential attention network, and semantic type labels are output.

[0032] The change area is spatially overlaid with meteorological grids and POI kernel density heatmaps to construct a spatiotemporal feature matrix;

[0033] The spatiotemporal propagation effect of factors is analyzed using spatiotemporal graph convolutional networks, and the contribution weights of driving factors are output.

[0034] Preferably, the dual-temporal differential attention network includes:

[0035] Multispectral features were extracted from the two images respectively;

[0036] Seasonal noise regions are suppressed using spatial attention weights, and the weight calculation formula is as follows:

[0037]

[0038] in, The difference lies in the characteristics of the two periods; Use the Sigmoid activation function; Indicates the convolution operation; Spatial attention weights at location;

[0039] Output change type labels based on Softmax.

[0040] Preferably, the spatiotemporal graph convolutional network construction includes:

[0041] Each geographic grid is a node, and the node characteristics include factor mean and variance;

[0042] Edge weights are calculated based on spatial adjacency and temporal lag correlation; the maximum lag step size is determined by Granger causality test.

[0043] Output the contribution weights of the driving factors and the spatiotemporal propagation path diagram.

[0044] Preferably, S4 includes:

[0045] Based on the LSTM model, the baseline value of land use change is predicted, and the residual between the actual value and the predicted value is calculated. If the value exceeds the dynamic threshold, it is marked as an area of ​​intensity anomaly.

[0046] Based on the contradiction between the contribution weight of driving factors and regional attributes, anomaly type labels are generated;

[0047] The SHAP value is used to interpret the main driving factors of the model output, and combined with the spatiotemporal propagation path map to generate high-risk area predictions.

[0048] Preferably, the dynamic threshold setting method includes:

[0049] Based on the sliding window calculation of the mean and standard deviation of land use change area in sub-regions, meteorological data of sub-regions within a preset time window, and POI density change rate of sub-regions within a preset time window, a dynamic threshold is determined.

[0050] The window size is dynamically selected based on the data volatility within the sub-region.

[0051] Secondly, the present invention provides a land use monitoring system based on remote sensing and big data for use in the aforementioned land use monitoring method, the system comprising:

[0052] The data acquisition module is used to divide sub-regions, acquire multi-temporal and multi-resolution remote sensing data of the sub-regions, and standardize the storage of multi-source data through a distributed storage framework.

[0053] The classification module is used to extract features and classify remote sensing data based on a deep learning model to generate a land use classification map.

[0054] The change identification module is used to identify change areas and build a change driver library that integrates meteorological data and human activity data by comparing multiple land use classification maps based on a dual-temporal difference attention network.

[0055] An anomaly detection module is used to detect abnormal regions based on the driving factor library and generate anomaly type labels and attribution analysis reports by combining dynamic early warning thresholds.

[0056] The interactive module is used to visualize and render monitoring results using geographic information system software and a distributed computing framework, and output early warning maps of abnormal areas and handling suggestions.

[0057] The beneficial effects of this invention include: simultaneously acquiring multi-source data such as satellite optical images, radar images, UAV hyperspectral data, real-time meteorological station data, and human activity data, and standardizing storage through a distributed storage framework; adopting an HDFS hierarchical storage architecture to achieve spatiotemporal data block management, which can fully utilize the advantages of different data sources, comprehensively acquire land use-related information, solve the problem of insufficient information from a single data source, and improve the efficiency of data storage and management, facilitating subsequent rapid retrieval and processing; constructing a CNN-Transformer-UNet hybrid network for feature extraction and classification, combining the advantages of CNN in extracting local features, Transformer in analyzing long-term temporal dependencies, and UNet in fusing spatial context information, which can generate pixel-level land use classification maps, improving the accuracy of land use classification; introducing a domain adaptation module, aligning the feature distribution of the public dataset and the target area data through adversarial training, optimizing the classification accuracy in low-sample scenarios, and enabling the model to maintain good performance under different regional and data volume conditions; using a dual-temporal difference attention network to compare land use classification maps from multiple periods, which can accurately identify changing areas and output semantic type labels; and transforming the change This study uses spatial overlay of land use regions with meteorological grids and POI kernel density heatmaps to construct a spatiotemporal feature matrix. Then, a spatiotemporal convolutional network is used to analyze the spatiotemporal propagation effect of factors, outputting the contribution weights of driving factors to uncover the driving mechanisms of land use change. A comprehensive analysis of the spatiotemporal interactions and nonlinear relationships among multiple factors helps to accurately understand the causes of land use change. Based on an LSTM model, baseline values ​​for land use change are predicted. The residuals between actual and predicted values ​​are used to mark areas of intensity anomalies, enabling timely detection of abnormal changes in land use. Anomaly type labels are generated based on the contradiction between the contribution weights of driving factors and regional attributes. SHAP values ​​are used to interpret the main driving factors output by the model. Combined with spatiotemporal propagation path maps, high-risk area predictions are generated, providing strong support for dynamic monitoring and early warning of land use, and facilitating timely measures to address anomalies. Geographic Information System (GIS) software and a distributed computing framework are used to visualize the monitoring results, outputting anomaly area early warning maps and disposal suggestions. Complex monitoring data and analysis results are presented to decision-makers in an intuitive way, providing strong decision support for land resource management and urban planning, and improving the practicality and application value of land use monitoring. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of a land use monitoring method based on remote sensing and big data provided in an embodiment of this application. Detailed Implementation

[0059] The present application will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0060] Example 1: This example provides a land use monitoring method based on remote sensing and big data. The method includes:

[0061] S1. Divide the data into sub-regions, acquire multi-temporal and multi-resolution remote sensing data for each sub-region, and standardize the storage of multi-source data using a distributed storage framework.

[0062] S2. Based on a deep learning model, feature extraction and classification of remote sensing data are performed to generate a land use classification map;

[0063] S3. Based on a dual-temporal difference attention network, compare land use classification maps from multiple periods to identify areas of change and construct a library of change-driving factors that integrate meteorological data and human activity data.

[0064] S4. Based on the driving factor library, detect abnormal regions and generate abnormal type labels and attribution analysis reports by combining dynamic early warning thresholds.

[0065] S5. Using geographic information system software and a distributed computing framework, the monitoring results are visualized and rendered, and an early warning map of abnormal areas and handling suggestions are output; the handling suggestions include automatically generating law enforcement patrol lists and protection plan triggering instructions.

[0066] The working principle and effects of the above technical solution are as follows:

[0067] The monitoring area was divided into multiple sub-regions to facilitate more refined and efficient data processing. Multi-temporal (different time points) and multi-resolution (different levels of clarity and detail) remote sensing data were acquired from these sub-regions using various methods. This multi-source data contains rich information; remote sensing data from different times can reflect dynamic changes in land use, while data at different resolutions can describe land use conditions from both macro and micro perspectives. A distributed storage framework was used to standardize the storage of the multi-source data. This approach fully leverages the storage advantages of distributed systems, improving data storage efficiency and reliability. Furthermore, standardized storage facilitates unified management and processing of the data in the future.

[0068] Deep learning models are used to extract and classify features from stored remote sensing data, such as the spectral and textural features of different land use types (e.g., cultivated land, forest land, and construction land) in remote sensing images. By learning these features, the model can accurately classify remote sensing data, generate land use classification maps, and clearly show the land use types in different regions.

[0069] A dual-temporal differential attention network was used to compare land use classification maps generated over multiple periods. This network can focus on areas where changes have occurred in the land use classification maps, accurately identifying regions of land use change by comparing classification results from different periods. Simultaneously, a database of change-driving factors was constructed by combining meteorological data (such as precipitation and temperature) and human activity data (such as population distribution and economic development indicators). Since land use change is often influenced by both natural factors (meteorological conditions) and human activities, including these factors in the driving factor database provides rich data support for subsequent analysis of the causes of land use change.

[0070] Based on the established driving factor library, anomaly detection is performed in the monitoring area. By analyzing the relationship between driving factors and land use change, it is determined which areas do not conform to normal land use change patterns, thereby identifying anomaly areas. Combined with dynamic early warning thresholds, anomaly type labels are generated to clarify the specific type of anomaly (such as illegal land occupation and construction, over-cultivation, etc.), and attribution analysis is performed to generate an attribution analysis report, explaining the reasons for the anomaly and providing a basis for subsequent decision-making.

[0071] By utilizing Geographic Information System (GIS) software and a distributed computing framework, monitoring results are visualized and rendered. The GIS software can display information such as land use classification maps, change areas, and anomaly areas in an intuitive map format, while the distributed computing framework ensures computational efficiency when processing large-scale data. The final output is an anomaly area early warning map, clearly identifying the location and extent of anomalies, and providing disposal suggestions, thus offering strong support for relevant departments in land use management and decision-making.

[0072] In one possible implementation, S1 includes:

[0073] Based on the diversity and frequency of historical land use types, sub-regions are divided using spatial clustering algorithms;

[0074] Simultaneously acquire satellite optical imagery (Sentinel-2, Gaofen series), radar imagery (Sentinel-1), and UAV hyperspectral data for the sub-region, and access real-time weather station data (precipitation, temperature) and human activity data (nighttime lights, POIs, social media check-in heatmaps).

[0075] Spatiotemporally adaptive preprocessing of remote sensing data, including cloud restoration based on GAN and radar data fusion;

[0076] The HDFS hierarchical storage architecture is used to implement spatiotemporal data block management; data blocks are divided according to spatiotemporal granularity (year / province / data type).

[0077] Based on the current sub-region type, type change frequency, anomaly occurrence frequency, and importance score, the data storage period is set: the type score of the current sub-region type, the normalized change frequency, the normalized anomaly occurrence frequency, and the normalized importance score are weighted to obtain the sub-region comprehensive impact factor; the data storage period of the sub-region is determined by the comprehensive impact factor.

[0078] The working principle and beneficial effects of the above technical solution are as follows:

[0079] Based on the diversity and frequency of historical land use types, spatial clustering algorithms are used to divide the area into sub-regions. High diversity of land use types indicates a complex land use situation in the area, while high frequency of change indicates an unstable land use status. Spatial clustering algorithms can group areas with similar land use characteristics together to form sub-regions, enabling more precise data processing and analysis for sub-regions with different characteristics, thereby improving the efficiency and accuracy of subsequent monitoring.

[0080] Simultaneously acquire multiple types of data, including satellite optical imagery (such as Sentinel-2 and Gaofen series), radar imagery (Sentinel-1), UAV hyperspectral data, as well as real-time meteorological station data (precipitation, temperature) and human activity data (nighttime light, POI, social media check-in heatmap). Different types of remote sensing data have their own advantages. Satellite optical imagery can provide rich surface spectral information, radar imagery is not limited by weather conditions and can acquire data in cloudy or foggy weather, while UAV hyperspectral data has higher spectral resolution and can reflect ground features in more detail. Meteorological data and human activity data can be used as auxiliary information to analyze the driving factors of land use change.

[0081] Spatiotemporally adaptive preprocessing of remote sensing data is performed, including cloud inpainting based on the fusion of GAN (Generative Adversarial Network) and radar data. Since satellite optical imagery is easily affected by cloud cover, some areas may have missing information. GAN has powerful image generation and inpainting capabilities. By fusing it with radar data, the characteristic that radar data is not affected by clouds can be used to repair cloud-covered areas in optical imagery, thereby improving the quality and integrity of remote sensing data and providing a more accurate data foundation for subsequent feature extraction and classification.

[0082] The HDFS (Hadoop Distributed File System) hierarchical storage architecture is adopted to realize spatiotemporal data block management, and data blocks are divided according to spatiotemporal granularity (year / province / data type); this facilitates data organization and management, and makes subsequent data query and analysis easier.

[0083] The data storage period is set according to the current sub-region type, type change frequency, anomaly frequency, and importance score. First, the type score, normalized change frequency, normalized anomaly frequency, and normalized importance score of the current sub-region type are weighted to obtain the sub-region's comprehensive impact factor. Normalization is used to eliminate the influence of different dimensions between indicators, making them comparable. The comprehensive impact factor obtained by weighted summation comprehensively considers multiple characteristics of the sub-region. Then, the data storage period for the sub-region is determined based on the comprehensive impact factor. For sub-regions with high comprehensive impact factors, it means that the land use situation in the region is complex, changes frequently, or is of great significance, requiring a longer data storage period for more in-depth historical data analysis. Conversely, for sub-regions with low comprehensive impact factors, the data storage period can be appropriately shortened to save storage resources, optimizing storage resource utilization while ensuring data availability.

[0084] In one possible implementation, the cloud repair based on GAN and radar data fusion includes:

[0085] The cloud-covered area of ​​the optical image is spatiotemporally registered with the radar image, and the backscattering coefficient and texture features of the radar image are extracted.

[0086] A dual-branch generative adversarial network is constructed, including an optical branch and a radar branch. The optical branch is input with registered optical images to repair the semantic information of cloud-covered areas; the radar branch is input with backscattering coefficients and texture features to reconstruct the surface structure under the cloud layer.

[0087] Based on the spectral features of optical images and the backscattering information of radar images, a pixel-level cloud thickness map is generated through a cloud thickness estimation subnetwork.

[0088] Based on the pixel-level cloud thickness map, a dynamic weight is assigned to each pixel, where thin cloud regions are mainly repaired by optical branches and thick cloud regions are mainly reconstructed by radar branches.

[0089] Features from the optical and radar branches are fused through cross-modal attention mechanisms and conditional gated convolutions;

[0090] Design multi-physics constraint loss functions, including radar-spectral correlation loss, terrain consistency loss, and seasonal vegetation index matching loss, to optimize the restoration results;

[0091] Guided filtering is used to smooth the transition edges between the repaired area and the normal area.

[0092] Spatiotemporal registration includes:

[0093] Seasons are determined based on image acquisition time and geographical location, and an adaptive feature point selection strategy is adopted:

[0094] In summer, feature points at the edge of vegetation are selected and registered using affine transformation.

[0095] In winter, select building corner points and use projection transformation for registration.

[0096] Dynamic weight allocation is achieved in the following ways:

[0097] A seasonal-weather sensing module is introduced, which uses LSTM to encode time-series meteorological data to generate dynamic weighting coefficients;

[0098] The cross-modal attention mechanism includes:

[0099] Calculate the spatial correlation matrix between optical features and radar features, and generate attention weights through Softmax normalization;

[0100] Adversarial physics discriminator:

[0101] Input the restored image and the actual cloudless image, output a physical plausibility score, and the loss function of the discriminator is:

[0102]

[0103] in, is the loss function, used to measure the performance of the discriminator in distinguishing between restored images and real cloudless images; E[] is the mathematical expectation; As an adversarial physics discriminator, The image is a true cloudless image; G is the generator. The original image with clouds;

[0104] The cloud thickness estimation subnetwork is a lightweight U-Net structure. Its inputs are the visible light band of the optical image and the radar backscattering coefficient, and its output is a pixel-level cloud thickness probability map.

[0105] The gating weights of the conditional gated convolution are dynamically adjusted based on the seasonal coding.

[0106] )

[0107] in, The gate weights are calculated using conditional gated convolution. () represents the Sigmoid function; This is the weight matrix; This is a vector concatenation operation. Here, the optical image feature vector and the seasonal state vector are concatenated; c is one-hot encoding; for example, a year is divided into four seasons: spring, summer, autumn and winter. If the current season is spring, the corresponding one-hot encoded vector is [1,0,0,0].

[0108] The working principle of the above technical solution is as follows:

[0109] Optical imagery (such as Sentinel-2 and Landsat-8) is used as the restoration target. For cloud-covered areas, cloudless images of the same area in the historical time-series image database are used as references for filling. Simultaneously, radar imagery (such as Sentinel-1) is used as auxiliary data because it can penetrate the cloud layer to obtain surface information below the cloud cover. The optical and radar data are spatiotemporally registered to the same coordinate system and resolution (e.g., 10m×10m grid). This spatiotemporal registration is achieved through specific geographic coordinate transformation algorithms and resampling techniques. Cloud masks are generated using cloud detection algorithms (such as FMask) to mark cloud-covered areas in the optical imagery. Backscattering coefficients (VV / VH polarization) and texture features (GLCM) are extracted from the radar imagery to characterize the surface type below the cloud cover (e.g., water bodies, bare soil, vegetation).

[0110] Dual-branch GAN network architecture design:

[0111] The generator is divided into two branches: Branch 1 for optical restoration, which takes cloud-containing optical images (RGB + near-infrared bands) and cloud masks as input, and outputs the restored image; Branch 2 for radar assistance, which takes radar feature maps (backscattering coefficients + textures) as input and extracts radar semantic information using convolutional layers. A feature fusion module is used to fuse the feature maps from both branches in an intermediate layer using attention-weighted concatenation and other methods to ensure that the restored area matches the radar surface features.

[0112] Discriminator: Input the generated image and the real cloudless image, use multi-scale convolution to judge the authenticity of the restored area, and introduce spectral normalization to improve training stability.

[0113] Based on the spectral features of optical images and the backscattering information of radar images, a lightweight U-Net structured cloud thickness estimation subnetwork is used to generate a pixel-level cloud thickness probability map, accurately estimating the cloud thickness of each pixel.

[0114] On the one hand, based on the pixel-level cloud thickness map, dynamic weights are assigned to each pixel. The optical branch dominates the restoration in thin cloud areas, while the radar branch dominates the reconstruction in thick cloud areas. For example, in thin cloud areas (grayscale < 100), color and texture are directly restored; in thick cloud areas (grayscale ≥ 100), radar data is used to reconstruct the surface structure (such as building outlines and water body boundaries). On the other hand, a seasonal-weather perception module is introduced, which uses LSTM to encode time-series meteorological data to generate dynamic weight coefficients. Taking into account factors such as cloud thickness, season, and weather, the fusion weights of different regional features are determined.

[0115] By employing a cross-modal attention mechanism, the correlation matrix between optical and radar feature spaces is calculated, and attention weights are generated through Softmax normalization to focus on important features. Combined with conditional gated convolution, the gate weights are dynamically adjusted based on seasonal encoding (the seasonal state vector of one-hot encoding) to flexibly control feature flow and achieve effective fusion of optical and radar branch features.

[0116] We designed a multi-physical constraint loss function (radar-spectral correlation loss, terrain consistency loss, and seasonal vegetation index matching loss) to constrain the rationality of the restoration results from a physical perspective. Using an adversarial physics discriminator, we input the restored image and the real cloudless image, and output a physical rationality score. The loss function measures the discriminator's ability to distinguish between different images, guides the generator to generate restored images that are more in line with physical laws, and optimizes the final restoration effect.

[0117] Guided filtering is used to smooth the transition edges between the repaired and normal areas. Quantitative evaluation is performed by simulating repair in partially cloudless areas, calculating PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) to verify the repair effect in cloudless areas. UAV sampling is conducted in the repaired areas (such as farmland and urban areas) for on-site verification. The UAV sampling plan is rationally planned based on the importance and geographical distribution of the repaired areas.

[0118] The above technical solution achieves the following effects: By fusion of optical and radar images through spatiotemporal registration, radar images, unobstructed by cloud cover, can acquire information on the surface structure beneath the clouds, while optical images provide rich spectral and semantic information. The combination of the two provides comprehensive data support for restoration, improving restoration accuracy. A cloud thickness estimation sub-network generates pixel-level cloud thickness maps, which are then used to assign dynamic weights to pixels. Thin clouds are weighted towards the optical branch, while thick clouds are weighted towards the radar branch, ensuring accurate restoration of areas with varying cloud thicknesses. Seasonal and geographical location classifications are used: in summer, vegetation edge feature points are selected using affine transformation, while in winter, building corner points are selected using projection transformation for spatiotemporal registration, adapting to different seasonal land cover features, improving registration accuracy, and enhancing the model's adaptability to different environments. The seasonal-weather perception module uses LSTM-encoded time-series meteorological data to generate dynamic weight coefficients, combined with dynamic cloud thickness weights, enabling the model to flexibly adjust restoration strategies based on season, weather, and cloud conditions, improving restoration capabilities in complex and variable environments. The correlation matrix between optical and radar feature spaces is calculated and normalized using Softmax to generate attention weights, allowing the model to focus on important features, effectively fusing multimodal features, and improving restoration performance. The gating weights are dynamically adjusted according to seasonal coding, flexibly controlling the flow of optical and radar branch features to make feature fusion more consistent with the actual conditions of different seasons and improve the rationality of restoration. It includes radar-spectral correlation, terrain consistency, and seasonal vegetation index matching losses, constraining the restoration results from multiple physical levels to ensure that the restored image matches reality in terms of physical characteristics, improving the reliability and authenticity of the restoration results. Through the discriminator outputting a physical rationality score and loss function optimization, it prompts the generator to generate physically reasonable restored images, further ensuring that the restoration results conform to physical laws. Guided filtering is used to smooth the transition edges between restored and normal areas, eliminating restoration traces and making the restored image visually more natural and coherent, improving image quality.

[0119] In one possible implementation, S2 includes:

[0120] A CNN-Transformer-UNet hybrid network is constructed. The CNN convolutional neural network extracts local features of multispectral images, the Transformer encoder models temporal dependencies, and the UNet decoder fuses spatial context information to generate a pixel-level land use classification map.

[0121] Based on the pre-trained model, a domain adaptation module is introduced to align the feature distribution of the public dataset with that of the target region data through adversarial training, thereby optimizing the classification accuracy in low-sample scenarios.

[0122] The CNN-Transformer-UNet hybrid network includes:

[0123] Low-level feature extraction module: Extracts local texture and spectral features of multispectral images based on ResNet-50 network;

[0124] Temporal modeling module: Performs long-term temporal dependency analysis on vegetation index and land surface temperature of multi-temporal images using a Transformer encoder;

[0125] Spatial fusion module: The UNet decoder is used to upsample the feature map, and skip connections are used to preserve the details of the ground boundary.

[0126] The implementation of the domain adaptation module includes:

[0127] Add a gradient reversal layer between the generator and the discriminator to force the network to learn domain-invariant features;

[0128] The adversarial loss function calculates the distance between the feature distributions of the source and target domains, and optimizes the model parameters by minimizing the maximum adversarial loss.

[0129] The optimization of classification accuracy in low-sample scenarios includes:

[0130] Only the parameters of the last three layers of the hybrid network are fine-tuned, while the weights of the remaining layers are frozen to reduce the amount of training computation.

[0131] Based on a small amount of labeled data in the target region, the matching degree between the model output and the semi-supervised labels is constrained by consistency regularization.

[0132] The Transformer encoder takes a multi-temporal vegetation index (NDVI) sequence as input and captures seasonal vegetation change patterns through a multi-head attention mechanism.

[0133] In one possible implementation, S2 includes:

[0134] The model is dynamically updated based on changes in sub-region data, including incremental training and full training modes.

[0135] The model update triggering conditions include:

[0136] The sub-region classification accuracy continuously decreased beyond the threshold.

[0137] Unlabeled land cover types were detected, and their proportion exceeded the limit.

[0138] The amount of new data has reached the preset ratio.

[0139] The working principle and effects of the above technical solution are as follows:

[0140] ResNet-50 is used as the underlying feature extraction module to extract local texture and spectral features from multispectral images. Through a series of convolutional layers and residual connections, ResNet-50 can effectively capture low-level features in the images, providing a foundation for subsequent processing.

[0141] The Transformer encoder receives vegetation index (such as NDVI) and land surface temperature data sequences from multi-temporal imagery. Through a multi-head attention mechanism, the Transformer can capture seasonal vegetation change patterns, analyze long-term temporal dependencies between multi-temporal data, and understand the dynamic characteristics of land use changes over time.

[0142] The UNet decoder upsamples the feature maps extracted and analyzed earlier, and combines skip connections to fuse low-level and high-level features, preserving the details of land feature boundaries, and finally generating a pixel-level land use classification map to achieve fine classification of land use types.

[0143] This module is introduced on top of the pre-trained model. By adding a gradient inversion layer between the generator and discriminator, the network is forced to learn domain-invariant features during training. The discriminator calculates the distance between the feature distributions of the source domain (public dataset) and the target domain (target region data), feeding this distance back to the model as an adversarial loss function. The model optimizes its parameters by minimizing the maximum adversarial loss, enabling it to achieve good classification accuracy even on target region data with limited samples, reducing the impact of dataset differences.

[0144] By fine-tuning the parameters of only the last three layers of the hybrid network and freezing the weights of the remaining layers, the computational cost can be reduced when training with a small amount of labeled data in the target region, while avoiding excessive modification of the pre-trained feature extraction part. Furthermore, consistency regularization is used to constrain the matching between the model output and the semi-supervised labels, further improving the model's classification accuracy in cases with few samples.

[0145] The model is dynamically updated based on changes in sub-region data. When the classification accuracy of a sub-region continuously declines beyond a threshold, it indicates that the model's performance in that sub-region has deteriorated, possibly due to changes in data characteristics, requiring a model update. When unlabeled land cover types are detected and their proportion exceeds the limit, it means that new land cover types have appeared that the model had not previously learned, and these types account for a large proportion in the sub-region, affecting the overall performance of the model. In this case, the model needs to be updated to identify the new land cover types. When the amount of new data reaches a preset proportion, it indicates that the sub-region has enough new data, which may contain new features or trends, also requiring a model update.

[0146] Update modes include incremental training and full training. Incremental training is suitable for situations where data changes are relatively minor. It trains only on the existing model using new data, enabling rapid adaptation to data changes and reducing training time and resource consumption. Full training, on the other hand, is used when data changes significantly or model performance deteriorates severely. It retrains on all data (including both existing and new data), allowing the model to fully learn the new data features and improving its adaptability and accuracy.

[0147] In one possible implementation, S3 includes:

[0148] Land use change areas were detected using a dual-temporal differential attention network and semantic type labels were output. Local spectral features were extracted using ResNet-50 and the long-term temporal dependence of multi-temporal vegetation indices on land surface temperature was analyzed using a Transformer encoder.

[0149] Input two images, focus on areas of significant change through an attention mechanism, and suppress seasonal noise (such as crop rotation); output change type labels (such as "forest land → construction land" and "arable land → desert").

[0150] The change area is spatially overlaid with meteorological grids and POI kernel density heatmaps to construct a spatiotemporal feature matrix;

[0151] Spatiotemporal graph convolutional networks are used to analyze the spatiotemporal propagation effect of factors, and output the contribution weights and spatiotemporal path graphs of driving factors.

[0152] The dual-temporal differential attention network includes:

[0153] Multispectral features of the two phases of images are extracted separately; dual-temporal multi-source images (optical + radar) are input, and features are extracted jointly by ResNet-50 and Transformer encoder;

[0154] Seasonal noise regions are suppressed using spatial attention weights, and the weight calculation formula is as follows:

[0155]

[0156] in, The difference lies in the characteristics of the two periods; Use the Sigmoid activation function; Indicates the convolution operation; Spatial attention weights at location;

[0157] Pixel-level change probability maps are generated based on the UNet decoder, and semantic labels are output through Softmax.

[0158] The spatiotemporal feature matrix includes:

[0159] Meteorological data: precipitation and temperature grids generated by Kriging interpolation;

[0160] Human activity data: POI kernel density heatmap, nighttime light intensity growth rate; among which, the spatial clustering degree of POIs is calculated using a Gaussian kernel function, the formula is:

[0161]

[0162] in, for The kernel density value of the POI at the location, where h is the bandwidth and K is the kernel function, which is a non-negative function, such as the Gaussian kernel function. This indicates the coordinates of the target location where the kernel density is to be calculated; Let be the coordinates of the i-th POI (Point of Interest); This represents the total number of POIs within the study area;

[0163] The construction of the spatiotemporal graph convolutional network (ST-GCN network) includes:

[0164] Node definition: Each geographic grid is a node, and node features include factor mean / variance;

[0165] Edge weight calculation: based on spatial adjacency and temporal lag correlation (e.g., changes in POI density in the upstream region affect the reduction of downstream arable land after 1 year).

[0166] Among them, the spatial edge weight is based on the weighting of geographical adjacency and the inverse of distance;

[0167] Time-side weights: The maximum lag step size between factors is determined through Granger causality tests, and the calculation formula is as follows:

[0168]

[0169] in, The time edge weights are used; GC is the Granger causality test statistic. is the lag step size; T is the set maximum lag step size; In time A variable at a given moment (e.g., weather factors, human activity factors, etc.). In time Another variable at any given time (e.g., variables related to land use change); To and The statistical value obtained by performing a Granger causality test; used to measure the lag step size. At time, the causal strength of X on Y; at each lag step The maximum value among all possible Granger causality test statistics;

[0170] Spatial and temporal weight ratio coefficients are generated based on the region attribute encoder. ;

[0171] Output: Contribution weights of driving factors and spatiotemporal propagation path graph;

[0172] Specifically, the contribution weights of driving factors are determined by fusing ST-GCN hidden features and original factor features through a factor attention mechanism.

[0173] The working principle and effects of the above technical solution are as follows:

[0174] Two phases of multi-source imagery (optical + radar) are input, and ResNet-50 is used to extract local spectral features, capturing the spectral details of land features in the images. Simultaneously, a Transformer encoder is used to analyze the long-term temporal dependence of multi-temporal vegetation indices and land surface temperature, uncovering the evolutionary patterns of land use over time. Noise suppression: Feature differences between the two phases of imagery are calculated, and spatial attention weights are generated using a formula. The convolution operation results are mapped to 0 to 1 to highlight areas of significant change and suppress interference from seasonal noise such as crop rotation. The UNet decoder upsamples the processed features to generate pixel-level change probability maps, and then outputs semantic labels (e.g., "forest land → construction land" "cultivated land → desert") using the Softmax function to clearly identify the type of land use change. A spatiotemporal feature matrix is ​​constructed by spatially overlaying the detected land use change areas with meteorological raster data (precipitation and temperature rasteres generated through Kriging interpolation) and human activity data (POI kernel density heatmap, nighttime light intensity growth rate). Each geographic grid is a node, and node features include factor mean and variance (e.g., statistics of meteorological factors and human activity factors). Spatial edge weights are based on geographic adjacency and inverse distance weighting, reflecting spatial proximity influences. Temporal edge weights are determined by the Granger causality test formula to determine the maximum lag step size l, measuring the causal strength of time lags between factors (e.g., the impact of upstream POI density changes on downstream farmland one year later). Weight ratio adjustment: The regional attribute encoder generates spatial and temporal weight ratio coefficients based on geographic grid attributes (e.g., land type, economic development level), dynamically adjusting their importance in the model. Output results: By fusing ST-GCN hidden features and original factor features through a factor attention mechanism, key driving factors are highlighted, the contribution weights of driving factors are determined (e.g., the impact of precipitation changes and POI density increases on land use change), and a spatiotemporal propagation path map is generated, visually demonstrating the propagation effect of factors on land use change in the spatiotemporal dimensions.

[0175] In summary, S3 accurately detects and labels land use change through a dual-temporal differential attention network, constructs a spatiotemporal feature matrix by combining multi-source data, and uses a spatiotemporal graph convolutional network to analyze the spatiotemporal effects of driving factors, providing key evidence for subsequent anomaly detection and attribution analysis, and achieving a deep understanding and analysis of land use change.

[0176] In one possible implementation, S4 includes:

[0177] Based on the LSTM model, the baseline value of land use change is predicted, and the residual between the actual value and the predicted value is calculated. If the value exceeds the dynamic threshold, it is marked as an area of ​​intensity anomaly.

[0178] Based on the contradiction between the contribution weight of driving factors and regional attributes, anomaly type labels are generated;

[0179] The SHAP value is used to interpret the main driving factors of the model output, and combined with the spatiotemporal propagation path map to generate high-risk area predictions.

[0180] The exception type labels include:

[0181] Suspected of violations (dominated by human activity and located in a sensitive area);

[0182] Natural disaster type (dominated by natural factors and abnormal meteorological data);

[0183] Data noise type (driving factors are contradictory and there is no spatial clustering).

[0184] In one possible implementation, the dynamic threshold setting method includes:

[0185] Based on the sliding window calculation of the mean and standard deviation of land use change area in sub-regions, meteorological data of sub-regions within the sliding window, and POI density change rate of sub-regions within the sliding window, a dynamic threshold is determined.

[0186]

[0187] in, For dynamic thresholds, This represents the average area of ​​land use change within the sliding window for each sub-region. The standard deviation of land use change area within the sliding window is the sub-region's land use change area, where k is a coefficient ranging from 2 to 4, determined by the current sub-region type. The Pearson correlation coefficient between meteorological data and residuals; The Pearson correlation coefficient between the rate of change of POI density and the residual; This represents the normalized value of the meteorological data at the current moment. This is the normalized value of the rate of change of POI density at the current moment.

[0188] The sliding window size is dynamically selected based on the data volatility within the sub-region;

[0189] like Then the sliding window length is Tc = Ta / And Tc > Tmin

[0190] like Then the sliding window length is the preset sliding window length Ta;

[0191] in, For coefficients, 2; Tmin is the minimum constraint value; This represents the mean squared error of the data within the sliding window at the previous time step. This represents the historical mean squared deviation.

[0192] Calculate the mean squared error for different types of data; set different coefficients K2 according to different data types; if the mean squared error of any type of data in the previous time window exceeds the corresponding coefficient and the historical mean squared error of the corresponding data, then adjust the length of the sliding window.

[0193] The working principle and beneficial effects of the above technical solution are as follows:

[0194] LSTM (Long Short-Term Memory) models excel at processing time-series data. By learning time-series patterns from historical land use change data, they capture long-term dependencies and trends in land use change, thereby predicting baseline values ​​for land use change. The residuals between actual land use change values ​​and LSTM-predicted baseline values ​​are calculated to measure the degree of deviation between actual and expected changes. The calculation of dynamic thresholds comprehensively considers multiple factors: the mean and standard deviation of land use change area in sub-regions, reflecting the overall level and volatility of change; the Pearson correlation coefficient between meteorological data, POI density change rate, and residuals; the correlation strength between quantitative meteorological factors, human activities, and change residuals; and the normalized values ​​of current meteorological data and POI density change rate, dynamically adjusting the threshold to respond to real-time data.

[0195] If the residual exceeds the limit, the area is marked as an area of ​​intensity anomaly, thus achieving preliminary monitoring of land use change anomalies.

[0196] Suspected violation type: If human activity factors (such as changes in POI density) dominate the change and the area is located in a sensitive area (such as an ecological protection zone), a "suspected violation type" label will be generated, indicating that human activities may exceed the reasonable range.

[0197] Natural disaster type: When natural factors (such as precipitation and temperature anomalies) dominate the changes and meteorological data show anomalies (such as extreme precipitation), it is marked as "natural disaster type" and is attributed to sudden changes in natural factors.

[0198] Data noise type: If the contribution weights of driving factors are contradictory (such as conflicting interpretations of changes by meteorological and human activity factors) and the change area has no spatial clustering (non-continuous distribution), it is judged as "data noise type", which may be caused by data error or random factors.

[0199] The system calculates the root mean square error (RMSE) of different data types (such as land use change area, meteorological data, and POI density) within a sub-region, and adjusts the window length based on the relationship between the MRMSE of the data within the previous sliding window and the historical mean MRMSE. It calculates the mean and standard deviation of land use change area using a sliding window, and combines the correlation between meteorological and POI data and residuals with real-time normalized values. This allows the dynamic threshold to adapt to the data characteristics and change patterns of different sub-regions, avoiding the limitations of fixed thresholds and improving the accuracy and robustness of anomaly detection.

[0200] High-risk area prediction utilizes SHAP (SHapley Additive ex Planations) values ​​to interpret model outputs, quantifying the contribution of each driving factor to land use change anomalies and identifying key driving factors. Combined with spatiotemporal propagation path maps (generated by S3's spatiotemporal graph convolutional network), the propagation trends and impact ranges of driving factors in the spatiotemporal dimensions are analyzed. This predicts the locations and types of areas that may evolve into high-risk areas in the future, providing forward-looking support for land management and decision-making, such as preventing illegal development or natural disasters from damaging land use.

[0201] In one possible implementation, S5 includes:

[0202] Based on a distributed computing framework, monitoring results data are processed in parallel, multi-scale geographic information is aggregated in real time, and layered vector tiles are generated.

[0203] By loading vector tiles using Geographic Information System (GIS) software, multi-level land use maps are rendered, including:

[0204] Heatmap of abnormal areas (colored according to confidence level);

[0205] Contribution distribution map of driving factors (overlaid with POI density and meteorological anomaly layers);

[0206] Dynamic trajectory of spatiotemporal propagation path;

[0207] Build an interactive early warning map interface that supports the following functions:

[0208] Area Penetration Query: Clicking on an abnormal area displays historical image comparisons, attribution analysis reports, and a list of suggested remediation measures;

[0209] Slide the timeline: Dynamically display the temporal evolution of land use change;

[0210] Output an early warning map of the abnormal area and corresponding handling suggestions, including:

[0211] Generate reports in PDF / GeoJSON format according to administrative divisions;

[0212] By connecting with the land law enforcement platform via RESTful API, high-priority abnormal coordinates and patrol plans can be pushed.

[0213] The working principle and effects of the above technical solution are as follows:

[0214] The monitoring results data output by S4 (such as coordinates of abnormal areas, weights of driving factors, spatiotemporal propagation paths, etc.) are processed in parallel using distributed computing frameworks (such as Hadoop and Spark). Through sharded storage and computation, large-scale geographic data (such as raster images and vector polygons) are decomposed into multiple sub-tasks, which are processed simultaneously by cluster nodes, significantly improving data processing efficiency and meeting real-time or near-real-time monitoring requirements.

[0215] The processed data is hierarchically aggregated according to geographic spatial range (such as administrative divisions, grid units) and scale (such as 1:100,000, 1:10,000) to generate multi-scale vector tiles (such as MapboxVectorTiles format).

[0216] The tiles contain the geometric coordinates and attribute information of geographic features (such as anomaly type, confidence level, and contribution of driving factors), supporting fast front-end loading and dynamic rendering, and realizing a "stepless zoom" map browsing experience.

[0217] Based on the confidence level of the abnormal area (e.g., the number of times the residual exceeds the dynamic threshold), the area is divided into different levels (e.g., low, medium, and high risk) using the symbolization function of GIS software (e.g., ArcGIS, QGIS). Gradient colors (e.g., from yellow to red) are used for visualization to intuitively show the spatial clustering degree and risk level of the abnormal area.

[0218] By overlaying POI kernel density heatmaps (reflecting the intensity of human activity), meteorological anomaly raster maps (such as precipitation anomaly percentage), and land use change areas, spatial overlay analysis (such as raster algebra and buffer analysis) is used to generate thematic maps of driving factor contributions. For example, areas with "POI density change rate contribution > 60%" are marked in blue to highlight anomaly areas dominated by human activity.

[0219] Based on the spatiotemporal propagation path map (such as the spatiotemporal causal relationship chain of driving factors) generated by S3, a dynamic timeline is constructed in GIS. The path nodes are connected in chronological order through interpolation algorithms, and the factor propagation process (such as the spatiotemporal diffusion trajectory of a factory construction → the conversion of surrounding farmland into industrial land) is rendered frame by frame using animation techniques (such as WebGL and Canvas), allowing users to intuitively understand the driving mechanism of land use change.

[0220] In the front-end interface (such as a WebGIS platform), clicking on the vector features of the abnormal area triggers an attribute query event. The system then retrieves historical imagery of the area in real time (from multi-temporal remote sensing data stored in S1), attribution analysis reports (such as the main driving factors for SHAP value sorting and anomaly type labels), and a pre-set list of handling suggestions (such as "recommend on-site verification within 3 days" and "activate the ecological restoration plan"), displaying them through pop-ups or sidebars to achieve in-depth information interaction with a "what you see is what you get" approach.

[0221] The system integrates a timeline control, linking the multi-temporal land use classification map generated by S2 with the change detection results of S3. When the user slides the timeline, the system dynamically loads the classification map or change heatmap for the corresponding time node, visualizing the evolution of land use over time (such as the transformation path of "cultivated land → forest land → construction land" in a watershed over 10 years), and supporting trend analysis and anomaly tracing.

[0222] The monitoring results are statistically summarized according to administrative divisions (such as provinces, cities, and counties), generating PDF-format graphic reports (including a schematic diagram of the distribution of abnormal areas, statistical charts of driving factors, and a list of disposal recommendations) and GeoJSON-format vector data (including spatial coordinates and attribute information) to meet the needs of offline archiving and professional analysis.

[0223] Through a RESTful API interface, information such as the coordinates, anomaly type, and driving factor summaries of high-priority anomaly areas (e.g., areas suspected of violations with a confidence level >90%) is pushed to the land enforcement platform in real time. The enforcement platform can then automatically generate patrol plans based on this data (e.g., planning drone patrol routes and assigning tasks to enforcement personnel), achieving closed-loop management of the entire process from monitoring to early warning to response.

[0224] Example 2: This example provides a land use monitoring system based on remote sensing and big data to implement the land use monitoring method described in Example 1. The system includes:

[0225] The data acquisition module is used to divide sub-regions, acquire multi-temporal and multi-resolution remote sensing data of the sub-regions, and standardize the storage of multi-source data through a distributed storage framework.

[0226] The classification module is used to extract features and classify remote sensing data based on a deep learning model to generate a land use classification map.

[0227] The change identification module is used to identify change areas and build a change driver library that integrates meteorological data and human activity data by comparing multiple land use classification maps based on a dual-temporal difference attention network.

[0228] An anomaly detection module is used to detect abnormal regions based on the driving factor library and generate anomaly type labels and attribution analysis reports by combining dynamic early warning thresholds.

[0229] The interactive module is used to visualize and render monitoring results using geographic information system software and a distributed computing framework, and output early warning maps of abnormal areas and handling suggestions.

[0230] The working principle and effect of the above technical solution are the same as those in the method embodiments of this application, and will not be repeated here.

[0231] This application describes the invention from the perspectives of purpose, performance, progress, and novelty, and it meets the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings are merely preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this application, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, shall fall within the scope of protection of this patent application.

Claims

1. A land use monitoring method based on remote sensing and big data, characterized in that, The method includes: S1. Divide the data into sub-regions, acquire multi-temporal and multi-resolution remote sensing data for each sub-region, and standardize the storage of multi-source data using a distributed storage framework. S2. Based on a deep learning model, feature extraction and classification of remote sensing data are performed to generate a land use classification map; S3. Based on a dual-temporal difference attention network, compare land use classification maps from multiple periods to identify areas of change and construct a library of change-driving factors that integrate meteorological data and human activity data. The dual-temporal difference attention network includes: extracting multispectral features from two images respectively; and suppressing seasonal noise regions through spatial attention weights, the weight calculation formula being: in, (x,y) represents the difference in characteristics between the two periods at position (x,y); Use the Sigmoid activation function; Indicates the convolution operation; Spatial attention weights at position (x,y); output change type labels based on Softmax; S4. Detect abnormal regions based on the driving factor library, and generate abnormal type labels and attribution analysis reports by combining dynamic thresholds; The dynamic threshold setting method includes: Based on the sliding window calculation of the mean and standard deviation of land use change area in sub-regions, meteorological data of sub-regions within a preset time window, and POI density change rate of sub-regions within a preset time window, a dynamic threshold is determined. The formula for calculating the dynamic threshold is: in, For dynamic thresholds, This represents the average area of ​​land use change within the sliding window for each sub-region. The standard deviation of land use change area within the sliding window is given by k1, which is a coefficient determined by the current sub-region type. The Pearson correlation coefficient between meteorological data and residuals; The Pearson correlation coefficient is the relationship between the rate of change of POI density and the residual. This represents the normalized value of the meteorological data at the current moment. This is the normalized value of the POI density change rate at the current moment; the window size is dynamically selected based on the data volatility within the sub-region. S5. Using geographic information system software and distributed computing framework, visualize and render the monitoring results, and output early warning maps of abnormal areas and handling suggestions. Wherein, S1 includes: Simultaneously acquire satellite optical images, radar images, and UAV hyperspectral data, and access real-time data from meteorological stations and human activity data; Spatiotemporally adaptive preprocessing of remote sensing data, including cloud restoration based on GAN and radar data fusion; The cloud repair based on GAN and radar data fusion includes: The cloud-covered area of ​​the optical image is spatiotemporally registered with the radar image, and the backscattering coefficient and texture features of the radar image are extracted. A dual-branch generative adversarial network is constructed, including an optical branch and a radar branch. The optical branch is input with registered optical images to repair the semantic information of cloud-covered areas; the radar branch is input with backscattering coefficients and texture features to reconstruct the surface structure under the cloud layer. Based on the spectral features of optical images and the backscattering information of radar images, a pixel-level cloud thickness map is generated through a cloud thickness estimation subnetwork. Based on the pixel-level cloud thickness map, a dynamic weight is assigned to each pixel, where thin cloud regions are mainly repaired by optical branches and thick cloud regions are mainly reconstructed by radar branches. Features from the optical and radar branches are fused through cross-modal attention mechanisms and conditional gated convolutions; Design a multi-physical-constraint loss function to optimize the repair results.

2. The land use monitoring method according to claim 1, characterized in that, The multi-physics constraint loss function includes radar-spectral correlation loss, terrain consistency loss, and seasonal vegetation index matching loss, which optimizes the restoration results.

3. The land use monitoring method according to claim 1, characterized in that, S2 includes: A CNN-Transformer-UNet hybrid network is constructed. The CNN convolutional neural network extracts local features from multispectral images, and the Transformer encoder performs long-term time-series dependency analysis on vegetation index and land surface temperature of multi-temporal images. The UNet decoder fuses spatial context information to generate a pixel-level land use classification map. Based on the pre-trained model, a domain adaptation module is introduced to align the feature distribution of the public dataset with that of the target region data through adversarial training, thereby optimizing the classification accuracy in low-sample scenarios.

4. The land use monitoring method according to claim 1, characterized in that, S3 further includes: The change area is spatially overlaid with meteorological grids and POI kernel density heatmaps to construct a spatiotemporal feature matrix; The spatiotemporal propagation effect of factors is analyzed using spatiotemporal graph convolutional networks, and the contribution weights of driving factors are output.

5. The land use monitoring method according to claim 4, characterized in that, The construction of the spatiotemporal graph convolutional network includes: Each geographic grid is a node, and the node characteristics include factor mean and variance; Edge weights are calculated based on spatial adjacency and temporal lag correlation; the maximum lag step size is determined by Granger causality test. Output the contribution weights of the driving factors and the spatiotemporal propagation path diagram.

6. The land use monitoring method according to claim 1, characterized in that, S4 further includes: Based on the LSTM model, the baseline value of land use change is predicted, and the residual between the actual value and the predicted value is calculated. If the value exceeds the dynamic threshold, it is marked as an area of ​​intensity anomaly. Based on the contradiction between the contribution weight of driving factors and regional attributes, anomaly type labels are generated; The SHAP value is used to interpret the main driving factors of the model output, and combined with the spatiotemporal propagation path map to generate high-risk area predictions.

7. A land use monitoring system based on remote sensing and big data, used to implement the land use monitoring method described in claim 1, characterized in that, The system includes: The data acquisition module is used to divide sub-regions, acquire multi-temporal and multi-resolution remote sensing data of the sub-regions, and standardize the storage of multi-source data through a distributed storage framework. The classification module is used to extract features and classify remote sensing data based on a deep learning model to generate a land use classification map. The change identification module is used to identify change areas and build a change driver library that integrates meteorological data and human activity data by comparing multiple land use classification maps based on a dual-temporal difference attention network. The dual-temporal difference attention network includes: extracting multispectral features from two images respectively; and suppressing seasonal noise regions through spatial attention weights, the weight calculation formula being: in, Let (x,y) represent the difference in characteristics between the two periods at position (x,y). Use the Sigmoid activation function; Indicates the convolution operation; Spatial attention weights at position (x,y); output change type labels based on Softmax; An anomaly detection module is used to detect anomalous regions based on the driving factor library and generate anomaly type labels and attribution analysis reports by combining dynamic thresholds; wherein, the dynamic threshold setting method includes: Based on the sliding window calculation of the mean and standard deviation of land use change area in sub-regions, meteorological data of sub-regions within a preset time window, and POI density change rate of sub-regions within a preset time window, a dynamic threshold is determined. The formula for calculating the dynamic threshold is: in, For dynamic thresholds, This represents the average area of ​​land use change within the sliding window for each sub-region. The standard deviation of land use change area within the sliding window is given by k1, which is a coefficient determined by the current sub-region type. The Pearson correlation coefficient between meteorological data and residuals; The Pearson correlation coefficient is the relationship between the rate of change of POI density and the residual. This represents the normalized value of the meteorological data at the current moment. This is the normalized value of the POI density change rate at the current moment; the window size is dynamically selected based on the data volatility within the sub-region. The interactive module is used to visualize and render monitoring results using geographic information system software and a distributed computing framework, and output early warning maps of abnormal areas and handling suggestions. The data acquisition module is configured as follows: Simultaneously acquire satellite optical images, radar images, and UAV hyperspectral data, and access real-time data from meteorological stations and human activity data; Spatiotemporally adaptive preprocessing of remote sensing data, including cloud restoration based on GAN and radar data fusion; The cloud repair based on GAN and radar data fusion includes: The cloud-covered area of ​​the optical image is spatiotemporally registered with the radar image, and the backscattering coefficient and texture features of the radar image are extracted. A dual-branch generative adversarial network is constructed, including an optical branch and a radar branch. The optical branch is input with registered optical images to repair the semantic information of cloud-covered areas; the radar branch is input with backscattering coefficients and texture features to reconstruct the surface structure under the cloud layer. Based on the spectral features of optical images and the backscattering information of radar images, a pixel-level cloud thickness map is generated through a cloud thickness estimation subnetwork. Based on the pixel-level cloud thickness map, a dynamic weight is assigned to each pixel, where thin cloud regions are mainly repaired by optical branches and thick cloud regions are mainly reconstructed by radar branches. Features from the optical and radar branches are fused through cross-modal attention mechanisms and conditional gated convolutions; Design a multi-physical-constraint loss function to optimize the repair results.

Images