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

Abstract

The invention proposes a land utilization monitoring method and system based on remote sensing and big data, and relates to the technical field of land monitoring, and the method comprises the steps: dividing sub-regions, and obtaining the multi-temporal and multi-resolution remote sensing data of the sub-regions; performing feature extraction and classification on the remote sensing data based on a deep learning model to generate a land utilization classification map; based on the dual-temporal difference attention network, identifying a change area and constructing a change driving factor library fusing meteorological data and human activity data; detecting an abnormal area based on the driving factor library, and generating an abnormal type label and an attribution analysis report in combination with a dynamic early warning threshold; performing visual rendering on the monitoring result, and outputting an abnormal region early warning map and a disposal suggestion; high-precision feature extraction and classification are realized, change areas and driving factors are deeply analyzed, abnormal areas are effectively detected and early warning is performed, and the accuracy, timeliness and practicability of land utilization monitoring are improved.

Description

Technical Field

[0001] The present application relates to the field of land monitoring technology, and in particular to a land use monitoring method and system based on remote sensing and big data. Background Art

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

[0003] Traditional land use monitoring methods rely primarily on field surveys and limited aerial photogrammetry. Field surveys are not only labor-intensive, material-intensive, and time-consuming, but also have limited coverage, making it difficult to cover large areas. Data acquisition is particularly challenging in remote or inaccessible areas. While aerial photogrammetry has expanded the monitoring scope 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, significantly reducing its timeliness.

[0004] The rise of remote sensing technology, with its advantages of large-scale, simultaneous observations, periodic, and repeated coverage, and its independence from ground conditions, has brought new opportunities for land use monitoring. Early remote sensing monitoring relied on image data from a single temporal phase and sensor. This limited the information available and made it difficult to accurately identify complex land use types and their dynamic changes. For example, when distinguishing between different land use types under vegetation cover (such as woodland and grassland), single-band or simple multi-band remote sensing images often fail to provide sufficient feature information, resulting in low classification accuracy.

[0005] Furthermore, traditional methods are particularly vulnerable to cloud obstruction. Cloud obstruction of optical remote sensing imagery severely impacts data integrity and availability. This is particularly true in areas prone to frequent cloud and rain events, where significant amounts of effective monitoring data are lost, significantly limiting 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 land features. Furthermore, traditional cloud restoration methods are often crude and simplistic, making it difficult 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 fails to fully understand the driving mechanisms of land-use change and makes it difficult to accurately identify areas of anomalous change and their causes. For example, when analyzing the complex relationship between land-use change, meteorological factors, and human activities, traditional methods are unable to effectively capture the spatiotemporal interactions and nonlinear relationships between these multiple factors, resulting in a one-sided understanding of the drivers of change, which in turn affects the accurate detection and early warning of anomalous areas.

[0007] In summary, existing land use monitoring methods have obvious defects in the comprehensiveness, accuracy, timeliness of data acquisition, the efficiency of data processing, and the depth and breadth of change analysis and anomaly detection. An innovative integrated monitoring method based on remote sensing and big data is urgently needed to make up for these shortcomings and realize accurate, 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, which comprehensively improves the accuracy, timeliness and practicality of land use monitoring through multi-source data fusion and standardized storage, model implementation of high-precision feature extraction and classification, in-depth analysis of change areas and driving factors, effective detection of abnormal areas and early warning, and efficient visual output decision support.

[0009] The purpose of this application is achieved by the following technical solutions:

[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 sub-regions, obtain 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;

[0012] S2. Extract and classify features of remote sensing data based on deep learning models to generate land use classification maps;

[0013] S3. Based on a dual-temporal difference attention network, we compare multiple land use classification maps, identify change areas, and construct a change driving factor library that integrates meteorological data and human activity data.

[0014] S4. Detect abnormal areas based on the driving factor library and generate abnormal type labels and attribution analysis reports in combination with dynamic warning thresholds;

[0015] S5. Use geographic information system software and distributed computing framework to visualize the monitoring results and output warning maps and disposal suggestions for abnormal areas.

[0016] Preferably, the S1 includes:

[0017] Synchronously acquire satellite optical images, radar images, and drone hyperspectral data, and access real-time data from weather stations and human activity data;

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

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

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

[0021] Performing spatiotemporal registration of the cloud coverage area of ​​the optical image with the radar image, and extracting the backscatter coefficient and texture features of the radar image;

[0022] A dual-branch generative adversarial network is constructed, including an optical branch and a radar branch. The optical branch inputs the registered optical image to restore the semantic information of the cloud-covered area; the radar branch inputs the backscatter coefficient and texture features to reconstruct the surface structure under the cloud layer.

[0023] Based on the spectral characteristics of optical images and the backscatter information of radar images, a cloud thickness estimation sub-network is used to generate pixel-level cloud thickness maps.

[0024] According to the pixel-level cloud thickness map, a dynamic weight is assigned to each pixel, where thin cloud areas are mainly repaired by the optical branch, and thick cloud areas are mainly reconstructed by the radar branch;

[0025] The features of the optical branch and the radar branch are fused through a cross-modal attention mechanism and conditional gated convolution;

[0026] A multi-physics constraint loss function is designed, including radar-spectral correlation loss, terrain consistency loss, and seasonal vegetation index matching loss, to optimize the restoration results.

[0027] Preferably, the S2 includes:

[0028] A CNN-Transformer-UNet hybrid network was constructed. The CNN convolutional neural network was used to extract local features of multispectral images. The Transformer encoder was used to analyze the long-term dependencies of vegetation indices and surface temperatures in multi-temporal images. The UNet decoder was used to fuse spatial context information and generate pixel-level land use classification maps.

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

[0030] Preferably, the S3 includes:

[0031] Detect land use change areas through a bi-temporal difference attention network and output semantic type labels;

[0032] The change area is spatially superimposed with the meteorological grid and POI kernel density heat map to construct a spatiotemporal feature matrix;

[0033] The spatiotemporal graph convolutional network is used to analyze the spatiotemporal propagation effect of factors and output the contribution weight of driving factors.

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

[0035] Extract the multispectral features of the two phases of images respectively;

[0036] Seasonal noise areas are suppressed by spatial attention weights, and the weight calculation formula is:

[0037]

[0038] in, is the difference in characteristics between the two periods; is the Sigmoid activation function; Represents the convolution operation; is the spatial attention weight at position;

[0039] Based on the Softmax output change type label.

[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] Calculating edge weights based on spatial adjacency and time-lagged correlation; determining the maximum lag step size through Granger causality test for the time-lagged correlation;

[0043] Output driving factor contribution weights and spatiotemporal propagation path diagram.

[0044] Preferably, the S4 includes:

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

[0046] Generate anomaly type labels based on the contradiction between the driving factor contribution weight and regional attributes;

[0047] The SHAP value interpretation model outputs the main driving factors and combines them 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, the mean and standard deviation of the land use change area in the sub-region, the meteorological data of the sub-region within the preset time window, and the POI density change rate of the sub-region within the preset time window are calculated to determine the dynamic threshold;

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

[0051] In a second aspect, the present invention provides a land use monitoring system based on remote sensing and big data, which is used in the aforementioned land use monitoring method, and the system comprises:

[0052] The data acquisition module is used to divide the sub-regions, obtain 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 and classify features of remote sensing data based on deep learning models to generate land use classification maps;

[0054] A change identification module is used to compare multiple land use classification maps based on a bi-temporal difference attention network, identify change areas, and construct a change driving factor library that integrates meteorological data and human activity data;

[0055] An anomaly detection module, used to detect abnormal areas based on the driving factor library and generate anomaly type labels and attribution analysis reports in combination with dynamic warning thresholds;

[0056] The interactive module is used to use geographic information system software and distributed computing framework to visualize the monitoring results and output warning maps and disposal suggestions for abnormal areas.

[0057] The beneficial effects of the present invention include: synchronously acquiring multi-source data such as satellite optical images, radar images, drone hyperspectral data, real-time data from weather stations, and human activity data, and performing standardized storage through a distributed storage framework, and using HDFS hierarchical storage architecture to achieve spatiotemporal data block management, which can fully utilize the advantages of different data sources to comprehensively acquire land use related information, solve the problem of insufficient information from a single data source, and at the same time 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 dependencies, and UNet in fusing spatial context information, which can generate pixel-level land use classification maps and improve the accuracy of land use classification; introducing a domain adaptation module, aligning the feature distribution of public datasets and target area data through adversarial training, optimizing the classification accuracy in small sample scenarios, and enabling the model to maintain good performance in different regions and data volume conditions; using a dual-phase difference attention network to compare multi-period land use classification maps, which can accurately identify changed areas and output semantic type labels; The system spatially overlays the land use area with meteorological grids, POI kernel density heat maps, and other data to construct a spatiotemporal feature matrix. A spatiotemporal graph convolutional network is then used to analyze the spatiotemporal propagation effects of factors, output the contribution weights of driving factors, and explore the driving mechanisms of land use change. This comprehensive analysis of the spatiotemporal interactions and nonlinear relationships between multiple factors helps accurately understand the causes of land use change. An LSTM model is used to predict baseline land use change values, and areas of abnormal intensity are marked by calculating the residual between the actual and predicted values, enabling timely detection of abnormal land use changes. Anomaly type labels are generated based on the contradiction between the driving factor contribution weights and regional attributes. A SHAP value interpretation model is used to output the main driving factors. Combined with the spatiotemporal propagation path map, high-risk area predictions are generated, providing strong support for dynamic land use monitoring and early warning, and facilitating timely response to land use anomalies. Using geographic information system software and a distributed computing framework, the system visualizes the monitoring results, outputting warning maps of abnormal areas and recommended actions. This intuitively presents complex monitoring data and analysis results to decision makers, providing strong decision support for land resource management and urban planning, and enhancing the practicality and application value of land use monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[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 the present application. DETAILED DESCRIPTION

[0059] Below, the present application is further described in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of no 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 comprising:

[0061] S1. Divide the sub-regions, obtain 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;

[0062] S2. Extract and classify features of remote sensing data based on deep learning models to generate land use classification maps;

[0063] S3. Based on a dual-temporal difference attention network, we compare multiple land use classification maps, identify change areas, and construct a change driving factor library that integrates meteorological data and human activity data.

[0064] S4. Detect abnormal areas based on the driving factor library and generate abnormal type labels and attribution analysis reports in combination with dynamic warning thresholds;

[0065] S5. Use geographic information system software and a distributed computing framework to visualize the monitoring results and output warning maps of abnormal areas and disposal suggestions; the disposal suggestions include automatically generating law enforcement inspection lists and protection plan trigger instructions.

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

[0067] The monitoring area is divided into multiple sub-areas to facilitate more refined and efficient data processing. Remote sensing data from multiple sub-areas is acquired through various methods, covering multiple temporal phases (at different time points) and resolutions (at varying clarity and levels of detail). This multi-source data contains a wealth of information. Remote sensing data collected over time can reflect dynamic changes in land use, while data at different resolutions can describe land use from both macro and micro perspectives. A distributed storage framework is used to standardize the storage of multi-source data. This not only fully leverages the storage advantages of distributed systems, improving data storage efficiency and reliability, but also facilitates unified management and processing of the data.

[0068] Deep learning models are used to extract and classify features from stored remote sensing data. For example, they identify the spectral and texture characteristics of different land use types (such as cultivated land, forest land, and construction land) in remote sensing images. By learning these features, the model can accurately classify remote sensing data and generate land use classification maps that clearly display land use types in different regions.

[0069] A bi-temporal differential attention network was used to compare land-use classification maps generated over multiple periods. The bi-temporal differential attention network was able to focus on areas of change within the land-use classification maps. By comparing classification results across different periods, it accurately identified areas of land-use change. Furthermore, a library of change drivers was constructed by combining meteorological data (such as precipitation and temperature) with human activity data (such as population distribution and economic development indicators). Because land-use change is often influenced by both natural factors (meteorological conditions) and human activities, incorporating these factors into the driver library provides rich data support for subsequent analysis of the causes of land-use change.

[0070] Based on the constructed driving factor library, abnormal areas in the monitoring area are detected; by analyzing the relationship between driving factors and land use changes, it is determined which areas have land use changes that do not conform to normal change patterns, thereby identifying abnormal areas; combined with dynamic warning thresholds, abnormal type labels are generated to clarify the specific types of abnormalities (such as illegal land occupation and construction, excessive reclamation, etc.), and attribution analysis is performed to generate an attribution analysis report to explain the causes of the abnormalities and provide a basis for subsequent decision-making.

[0071] Utilizing geographic information system software and a distributed computing framework, monitoring results are visualized. The GIS software displays land use classification maps, areas of change, and abnormal areas in an intuitive map format, while the distributed computing framework ensures computational efficiency when processing large amounts of data. Ultimately, the output is an abnormal area warning map, clearly identifying the location and extent of abnormal areas and providing recommendations for mitigation, providing strong support for land use management and decision-making by relevant departments.

[0072] In a possible implementation, S1 includes:

[0073] Based on the diversity and change frequency of historical land use types, the sub-regions were divided using a spatial clustering algorithm;

[0074] Simultaneously acquire sub-regional satellite optical imagery (Sentinel-2, Gaofen series), radar imagery (Sentinel-1), and drone hyperspectral data, and access real-time data from weather stations (precipitation, temperature), and human activity data (nighttime lights, POIs, and social media check-in trends);

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

[0076] 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] The data storage period is set according to the current sub-region type, type change frequency, anomaly occurrence frequency, and importance score: 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 based on the comprehensive impact factor.

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

[0079] Based on the historical diversity of land use types and the frequency of change, a spatial clustering algorithm is used to divide sub-regions. A high diversity of land use types means that the land use situation in the region is complex, while a high frequency of change indicates that the land use status is unstable. The spatial clustering algorithm can cluster areas with similar land use characteristics together to form sub-regions, and can perform more accurate data processing and analysis on sub-regions with different characteristics, thereby improving the efficiency and accuracy of subsequent monitoring.

[0080] Multiple types of data are acquired simultaneously, including satellite optical imagery (such as Sentinel-2 and Gaofen series), radar imagery (Sentinel-1), drone hyperspectral data, as well as real-time data from weather stations (precipitation, temperature) and human activity data (nighttime lights, POIs, and social media check-in trends). Different types of remote sensing data have their own advantages. Satellite optical imagery can provide rich surface spectral information, radar imagery is not restricted by weather conditions and can acquire data in cloudy and foggy weather, and drone hyperspectral data has higher spectral resolution and can reflect the characteristics of land objects 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] Remote sensing data is subjected to spatiotemporal adaptive preprocessing, including cloud restoration based on the fusion of GAN (Generative Adversarial Network) and radar data. Satellite optical imagery is susceptible to cloud obstruction, resulting in missing information in some areas. GAN has powerful image generation and restoration capabilities. By fusing it with radar data, the characteristic of radar data being unaffected by clouds can be exploited to restore cloud-obstructed 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 used to implement spatiotemporal data block management, and data blocks are divided according to spatiotemporal granularity (year / province / data type); this facilitates data organization and management, and facilitates subsequent data query and analysis.

[0083] The data storage period is set based on the current sub-region type, type change frequency, anomaly frequency, and importance score. First, the sub-region type score, normalized change frequency, normalized anomaly frequency, and normalized importance score are weighted to obtain the sub-region's comprehensive impact factor. Normalization is performed to eliminate the impact of different dimensions between indicators and make them comparable. The weighted summation of the comprehensive impact factor takes into account multiple sub-region characteristics. The comprehensive impact factor is then used to determine the data storage period for the sub-region. Sub-regions with high comprehensive impact factors indicate complex land use, frequent changes, or significant land use, requiring longer data storage periods for more in-depth historical data analysis. Sub-regions with low comprehensive impact factors can, on the other hand, be appropriately shortened to conserve storage resources, ensuring data availability while optimizing storage resource utilization.

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

[0085] Performing spatiotemporal registration of the cloud coverage area of ​​the optical image with the radar image, and extracting the backscatter coefficient and texture features of the radar image;

[0086] A dual-branch generative adversarial network is constructed, including an optical branch and a radar branch. The optical branch inputs the registered optical image to restore the semantic information of the cloud-covered area; the radar branch inputs the backscatter coefficient and texture features to reconstruct the surface structure under the cloud layer.

[0087] Based on the spectral characteristics of optical images and the backscatter information of radar images, a cloud thickness estimation sub-network is used to generate pixel-level cloud thickness maps.

[0088] According to the pixel-level cloud thickness map, a dynamic weight is assigned to each pixel, where thin cloud areas are mainly repaired by the optical branch, and thick cloud areas are mainly reconstructed by the radar branch;

[0089] The features of the optical branch and the radar branch are fused through a cross-modal attention mechanism and conditional gated convolution;

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

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

[0092] Spatiotemporal registration includes:

[0093] The seasons are divided according to the image acquisition time and geographical location, and an adaptive feature point selection strategy is adopted:

[0094] In summer, feature points on the vegetation edge are selected and registered using affine transformation;

[0095] In winter, building corners are selected and registration is performed using projection transformation.

[0096] Dynamic weight distribution is achieved through the following methods:

[0097] Introducing a seasonal-weather perception module, using LSTM to encode time-series meteorological data and generate dynamic weight 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 Physical Discriminator:

[0101] Input the restored image and the real cloud-free image, and output the physical plausibility score. The loss function of the discriminator is:

[0102]

[0103] in, is the loss function used to measure the performance of the discriminator in distinguishing the restored image from the real cloud-free image; E[] is the mathematical expectation; is the adversarial physical discriminator, is a real cloud-free image; G is the generator; is the original image with clouds;

[0104] The cloud thickness estimation subnetwork is a lightweight U-Net structure, which takes the visible light band of the optical image and the radar backscatter coefficient as input and outputs a pixel-level cloud thickness probability map.

[0105] The gating weights of the conditional gated convolution are dynamically adjusted according to the seasonal encoding.

[0106] )

[0107] in, The gating weights calculated for the conditional gated convolution; () is the Sigmoid function; is the weight matrix; is a vector concatenation operation. Here, the optical image feature vector and the seasonal state vector are concatenated. c is a 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 encoding vector is [1, 0, 0, 0].

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

[0109] Optical imagery (such as Sentinel-2 and Landsat-8) is used as the inpainting target. For cloud-covered areas, cloud-free images of the same area from a historical time-series image library are used as reference. Radar imagery (such as Sentinel-1) is used as auxiliary data because it can penetrate clouds and obtain information about the underlying surface. The optical and radar data are spatiotemporally aligned to the same coordinate system and resolution (e.g., 10m×10m grid). This spatiotemporal alignment is achieved using a specific geographic coordinate conversion algorithm and resampling techniques. A cloud mask is generated using a cloud detection algorithm (such as FMask) to annotate cloud-covered areas in the optical imagery. Backscatter coefficients (VV / VH polarization) and texture features (GLCM) are extracted from the radar imagery to characterize the underlying surface type (e.g., water, bare soil, vegetation).

[0110] Dual-branch GAN network architecture design:

[0111] Generator: Branch 1 is set up for optical restoration, taking as input an optical image containing clouds (RGB + near-infrared bands) and a cloud mask, and outputting the restored image. Branch 2 is radar-assisted, taking as input a radar feature map (backscatter coefficient + texture) and extracting radar semantic information using convolutional layers. The feature fusion module fuses the feature maps from both branches at an intermediate layer using methods such as attention-weighted splicing to ensure that the restored area is consistent with the radar surface features.

[0112] Discriminator: It takes as input the generated image and the real cloud-free image, uses multi-scale convolution to determine the authenticity of the repaired area, and introduces spectral normalization to improve training stability.

[0113] Based on the spectral characteristics of optical images and the backscatter information of radar images, a cloud thickness estimation subnetwork with a lightweight U-Net structure is used to generate a pixel-level cloud thickness probability map to accurately estimate the cloud thickness of each pixel.

[0114] On the one hand, dynamic weights are assigned to each pixel based on the pixel-level cloud thickness map. The optical branch takes the lead in restoration in thin cloud areas, while the radar branch takes the lead in 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), surface structures (such as building outlines and water body boundaries) are reconstructed based on radar data first. On the other hand, a seasonal-weather perception module is introduced, and LSTM is used to encode time-series meteorological data to generate dynamic weight coefficients. The feature fusion weights of different regions are determined by comprehensively considering factors such as cloud thickness, season and weather.

[0115] Through the cross-modal attention mechanism, the spatial correlation matrix of optical features and radar features is calculated, and attention weights are generated through Softmax normalization to focus on important features. Combined with conditional gated convolution, its gating weights are dynamically adjusted according to the seasonal code (one-hot encoded seasonal state vector), flexible control of feature flow, and effective fusion of optical and radar branch features.

[0116] A multi-physics constraint loss function (radar-spectral correlation loss, terrain consistency loss, and seasonal vegetation index matching loss) is designed to constrain the rationality of the restoration results from a physical perspective. An adversarial physics discriminator is used to input the restoration image and the real cloud-free image, and output a physical rationality score. Its loss function measures the discriminator's ability to distinguish, guiding the generator to produce restoration images that are more in line with physical laws, thereby optimizing the final restoration effect.

[0117] A guided filter was used to smooth the transition between the repaired area and the normal area. A quantitative assessment was conducted by masking out portions of cloud-free areas to simulate the repair. Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) were calculated to verify the repair effectiveness in the cloud-free areas. Drone sampling was conducted in repaired areas (such as farmland and urban areas) for field verification. The drone sampling plan was rationally planned based on the importance and geographic distribution of the repaired areas.

[0118] The above technical solution achieves the following: Optical and radar images are fused through spatiotemporal registration. Radar images, unobstructed by clouds, can capture information about the underlying surface structure, while optical images provide rich spectral and semantic information. The combination of these two provides comprehensive data support for restoration, improving restoration accuracy. The cloud thickness estimation subnetwork generates a pixel-by-pixel cloud thickness map and assigns dynamic weights to pixels based on this map. Thin clouds are prioritized by the optical branch, while thick clouds are prioritized by the radar branch, enabling accurate restoration across regions of varying cloud thickness. Spatiotemporal registration is performed based on seasonal and geographic location. In summer, vegetation edge feature points are selected using an affine transformation, while in winter, building corners are selected using a projective transformation. This adapts to seasonal features, improves registration accuracy, and enhances the model's adaptability to diverse environments. The seasonal-weather perception module utilizes LSTM-encoded time-series meteorological data to generate dynamic weight coefficients. Combined with the dynamic cloud thickness weights, the model can flexibly adjust its restoration strategy based on season, weather, and cloud conditions, enhancing restoration capabilities in complex and changing environments. The spatial correlation matrix of optical and radar features is calculated and then normalized using Softmax to generate attention weights. This allows the model to focus on important features, effectively integrating multimodal features and improving restoration effectiveness. Gating weights are dynamically adjusted based on seasonal coding, flexibly controlling the flow of optical and radar branch features. This ensures that feature fusion is more consistent with seasonal conditions and improves the rationality of the restoration. Radar-spectral correlation, terrain consistency, and seasonal vegetation index matching losses are included, constraining the restoration results from multiple physical perspectives to ensure that the physical properties of the restored image are consistent with reality, thereby improving the reliability and authenticity of the restoration results. The discriminator outputs a physical rationality score and optimizes the loss function, prompting the generator to produce physically plausible restoration images, further ensuring that the restoration results conform to physical laws. Guided filtering is used to smooth the transition between the restored and normal areas, eliminating restoration artifacts and making the restored image more visually natural and coherent, thereby improving image quality.

[0119] In a possible implementation, S2 includes:

[0120] A CNN-Transformer-UNet hybrid network was constructed. The CNN convolutional neural network was used to extract local features of multispectral images, the Transformer encoder was used to model temporal dependencies, and the UNet decoder was used to fuse spatial context information to generate pixel-level land use classification maps.

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

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

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

[0124] Time series modeling module: Uses the Transformer encoder to analyze the long-term dependency of vegetation indices and surface temperatures in multi-temporal images;

[0125] Spatial fusion module: The UNet decoder is used to upsample the feature map and the skip connection is combined to retain the boundary details of the objects.

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

[0127] Add a gradient reversal layer (GradientReversalLayer) 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 domain and the target domain, and optimizes the model parameters by minimizing the maximum adversarial loss.

[0129] Optimizing the classification accuracy of the few-sample scenario includes:

[0130] Only fine-tune the parameters of the last three layers of the hybrid network and freeze the weights of the remaining layers to reduce the amount of training calculations;

[0131] Based on a small amount of labeled data in the target area, consistency regularization is used to constrain the matching of the model output with the semi-supervised label.

[0132] The input of the Transformer encoder is a multi-temporal vegetation index (NDVI) sequence, and the seasonal vegetation change pattern is captured through a multi-head attention mechanism.

[0133] In a possible implementation, S2 includes:

[0134] Dynamically trigger model updates based on sub-region data changes, including incremental training and full training modes.

[0135] The model update triggering conditions include:

[0136] The sub-region classification accuracy continuously decreases beyond the threshold;

[0137] An unlabeled feature type is detected and its proportion exceeds the limit;

[0138] The amount of new data reaches 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 imagery. Through a series of convolutional layers and residual connections, ResNet-50 effectively captures low-level features in the image, providing a foundation for subsequent processing.

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

[0142] The UNet decoder performs an upsampling operation on the feature map extracted and analyzed previously, and combines skip connections to fuse the underlying features with the high-level features, retaining the boundary details of the land objects, and finally generates a pixel-level land use classification map to achieve fine classification of land use types.

[0143] This module is introduced based on the pre-trained model. By adding a gradient reversal layer between the generator and the 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), and feeds this back to the model in the form of an adversarial loss function. The model optimizes parameters by minimizing the maximum adversarial loss, achieving good classification accuracy even on a small number of target region data samples, reducing the impact of dataset differences.

[0144] Fine-tuning only the parameters of the last three layers of the hybrid network and freezing the weights of the remaining layers reduces computational effort when training with a small amount of labeled data from the target region, while also avoiding over-modification of the pre-trained feature extraction. Furthermore, consistency regularization is used to constrain the model output to match the semi-supervised labels, further improving the model's classification accuracy in the case of few-shot training.

[0145] Dynamically decide whether to update the model based on changes in sub-region data. When the classification accuracy of a sub-region continuously decreases by more than a threshold, it indicates that the model's performance in that sub-region has deteriorated, possibly due to changes in data characteristics, and the model needs to be updated. Detecting unlabeled feature types and exceeding the limit indicates the emergence of new feature types that the model has not previously learned, and their proportion in the sub-region is large, affecting the model's overall performance. In this case, the model needs to be updated to identify the new features. When the amount of new data reaches a preset ratio, it indicates that the sub-region has sufficient new data, which may contain new features or changing trends, and the model also needs to be updated.

[0146] Update modes include incremental training and full training. Incremental training is suitable for situations where data changes are relatively small. Training is performed only on new data based on the existing model. This allows for rapid adaptation to data changes, reducing training time and resource consumption. Full training is used when data changes are significant or model performance degrades significantly. It retrains on all data (both existing and new) to enable the model to fully learn the new data features, improving its adaptability and accuracy.

[0147] In a possible implementation, S3 includes:

[0148] A dual-temporal difference attention network is used to detect land use change areas and output semantic type labels; a ResNet-50 is used to extract local spectral features, and a Transformer encoder is used to analyze the long-term dependencies between multi-temporal vegetation indices and surface temperature;

[0149] It inputs two images and uses an attention mechanism to focus on areas with significant changes, suppressing seasonal noise (such as crop rotation). It also outputs labels for the types of changes (such as "forest land → construction land" or "arable land → desert").

[0150] The change area is spatially superimposed with the meteorological grid and POI kernel density heat map to construct a spatiotemporal feature matrix;

[0151] The spatiotemporal graph convolutional network is used to analyze the spatiotemporal propagation effect of factors, and the contribution weights of driving factors and the spatiotemporal path diagram are output.

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

[0153] Extract multispectral features of the two phases of imagery separately; input dual-phase multi-source imagery (optical + radar), and extract features jointly using ResNet-50 and Transformer encoders;

[0154] Seasonal noise areas are suppressed by spatial attention weights, and the weight calculation formula is:

[0155]

[0156] in, is the difference in characteristics between the two periods; is the Sigmoid activation function; Represents the convolution operation; is the spatial attention weight at position;

[0157] Generate pixel-level change probability maps based on the UNet decoder and output semantic labels through Softmax.

[0158] The spatiotemporal feature matrix includes:

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

[0160] Human activity data: POI kernel density heat map, nighttime light intensity growth rate; POI spatial aggregation is calculated using the Gaussian kernel function, using the formula:

[0161]

[0162] in, for The POI kernel density value at the location, h is the bandwidth, K is the kernel function, which is a non-negative function, such as the Gaussian kernel function; Indicates the target position coordinates for calculating the kernel density; is the coordinate of the i-th POI (point of interest); is the total number of POIs in 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 the node characteristics include factor mean / variance;

[0165] Edge weight calculation: based on spatial adjacency and time lag correlation (e.g., changes in POI density in the upstream region affect the reduction of cultivated land downstream one year later);

[0166] Among them, spatial edge weight: weighted based on geographical adjacency and inverse distance;

[0167] Time edge weight: The maximum lag step between factors is determined by Granger causality test. The calculation formula is:

[0168]

[0169] in, is the time edge weight; GC is the Granger causality test statistic, is the hysteresis step length; T is the set maximum hysteresis step length; For in time A variable at a certain moment (such as meteorological factors, human activity factors, etc.); For in time another variable at the time (e.g., a variable related to land-use change); For and The statistic value obtained by performing the Granger causality test is used to measure the lag step length. The causal strength of X on Y at each lag step The maximum value among all possible Granger causality test statistics under ;

[0170] Generate spatial and temporal weight ratio coefficients based on region attribute encoder ;

[0171] Output: driving factor contribution weight and spatiotemporal propagation path diagram;

[0172] Among them, the ST-GCN hidden features and the original factor features are fused through the factor attention mechanism to determine the contribution weight of the driving factor.

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

[0174] The system takes two multi-source imagery (optical and radar) as input and uses ResNet-50 to extract local spectral features, capturing spectral details of the imaged features. Simultaneously, a Transformer encoder analyzes the long-term dependencies between multi-temporal vegetation indices and surface temperature to uncover patterns in land use evolution over time. To suppress noise, the system calculates feature differences between the two imagery phases and generates spatial attention weights using a formula. The convolution operation results are mapped to a range of 0 to 1, highlighting areas of significant change and suppressing the interference of seasonal noise, such as crop rotation, on change detection. The UNet decoder upsamples the processed features to generate a pixel-level change probability map. A softmax function is then used to output semantic labels (e.g., "forestland → construction land," "arable land → desert") to identify the type of land use change. A spatiotemporal feature matrix is ​​constructed to spatially overlay the detected land use change areas with meteorological raster data (precipitation and temperature rasters generated through kriging interpolation) and human activity data (point of interest kernel density heatmaps and nighttime light intensity growth rates). Each geographic grid cell is a node, and node features include factor means and variances (e.g., statistics of meteorological and human activity factors). Spatial edge weights are weighted based on geographic adjacency and the inverse of distance, reflecting the influence of spatial proximity. Temporal edge weights, using the Granger causality test formula to determine the maximum lag step size, l, measure the causal strength of time lags between factors (e.g., the impact of changes in POI density in an upstream region on cultivated land downstream one year later). Weight scaling: The regional attribute encoder generates spatial and temporal weight scaling coefficients based on geographic grid attributes (e.g., land type, economic development level), dynamically adjusting their importance in the model. Output: The ST-GCN hidden features are integrated with the original factor features through a factor attention mechanism to highlight key driving factors, determine the contribution weights of these driving factors (e.g., the impact of precipitation changes and POI density growth on land use change), and generate a spatiotemporal propagation path map, visually demonstrating the propagation effect of factors on land use change in both spatial and temporal dimensions.

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

[0176] In a possible implementation, the S4 includes:

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

[0178] Generate anomaly type labels based on the contradiction between the driving factor contribution weight and regional attributes;

[0179] The SHAP value interpretation model outputs the main driving factors and combines them with the spatiotemporal propagation path map to generate high-risk area predictions.

[0180] The exception type labels include:

[0181] Suspected violation type (dominated by human activities and located in sensitive areas);

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

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

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

[0185] The dynamic threshold is determined by calculating the mean and standard deviation of the land use change area in the sub-region, the meteorological data of the sub-region within the sliding window, and the POI density change rate of the sub-region within the sliding window.

[0186]

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

[0188] Dynamically select the sliding window size 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, is the coefficient, 2; Tmin is the minimum constraint value; is the mean square error of the data in the sliding window at the previous moment; is the mean of the historical mean square error.

[0192] The mean square error is calculated for different types of data respectively. Different coefficients K2 are set according to different data types. If the mean square error of any type of data in the sliding window at the previous moment exceeds the corresponding coefficient and the mean of the historical mean square error of the corresponding data, the sliding window length is adjusted.

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

[0194] The LSTM (Long Short-Term Memory) model excels at processing time series data. By learning the time series patterns of historical land-use change data, it captures the long-term dependencies and trends of land-use change, thereby predicting a baseline value for land-use change. The residual between the actual land-use change value and the LSTM-predicted baseline value is calculated to measure the degree of deviation between the actual change and the expected change. The dynamic threshold calculation takes into account multiple factors: the mean and standard deviation of the land-use change area in the subregion, reflecting the overall level and volatility of change; the Pearson correlation coefficient between meteorological data, the rate of change in POI density, and the residuals; the strength of the correlation between quantitative meteorological and human activity factors and the residuals; and the normalized values ​​of the current meteorological data and the rate of change in POI density to dynamically adjust the threshold in response to real-time data.

[0195] If the residual exceeds, the area is marked as an intensity anomaly area, achieving preliminary monitoring of abnormal land use changes.

[0196] Suspected Violation: If human activity factors (such as changes in POI density) dominate the changes and the area is located in a sensitive area (such as an ecological protection area), a "Suspected Violation" label is generated, indicating that human activities may exceed reasonable limits.

[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 attributed to sudden changes in natural factors.

[0198] Data noise type: If the contribution weights of driving factors contradict each other (for example, the interpretations of changes by meteorological and human activity factors conflict) and the changing area has no spatial clustering (distributed in non-continuous patches), it is judged as "data noise type", which may be caused by data errors or accidental factors.

[0199] The mean square error (MSE) of different data types within a subregion (such as land use change area, weather, and POI density) is calculated, and the window length is adjusted based on the relationship between the mean square error of the data within the sliding window at the previous moment and the mean of the historical mean square errors. The mean and standard deviation of the land use change area are calculated using a sliding window. Combined with the correlation between weather and POI data and residuals, along with real-time normalization, a dynamic threshold can adapt to the data characteristics and change patterns of different subregions, avoiding the limitations of fixed thresholds and improving the accuracy and robustness of anomaly detection.

[0200] High-risk area prediction uses SHAP (SHapley Additive exPlanations) values ​​to interpret model output, quantifying the contribution of each driver to anomalies in land-use change and identifying the primary driver. Combined with a spatiotemporal propagation path map (generated by S3's spatiotemporal graph convolutional network), this analysis analyzes the propagation trends and impact of drivers across time and space, predicting the locations and types of areas likely to become high-risk in the future. This provides forward-looking support for land management and decision-making, such as preventing land use disruptions from illegal development or natural disasters.

[0201] In a possible implementation, S5 includes:

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

[0203] Load vector tiles through Geographic Information System (GIS) software to render multi-level land use maps, including:

[0204] Heat map of abnormal areas (colored by confidence level);

[0205] Driving factor contribution distribution map (overlaid with POI density and meteorological anomaly layers);

[0206] Dynamic trajectory of space-time propagation paths;

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

[0208] Regional penetration query: Click on the abnormal area to display historical image comparison, attribution analysis report and disposal suggestion list;

[0209] Timeline sliding: Dynamically display the temporal evolution of land use changes;

[0210] Output abnormal area warning map and disposal suggestions, including:

[0211] Generate PDF / GeoJSON format reports by administrative division;

[0212] Connect with the land law enforcement platform through RESTful API to push high-priority abnormal coordinates and inspection plans.

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

[0214] Distributed computing frameworks (such as Hadoop and Spark) are used to perform parallel processing on the monitoring result data output by S4 (such as abnormal area coordinates, driving factor weights, spatiotemporal propagation paths, etc.); through sharded storage and computing, large-scale geographic data (such as raster images and vector polygons) are decomposed into multiple subtasks, which are processed simultaneously by cluster nodes, significantly improving data processing efficiency and meeting real-time or near real-time monitoring needs.

[0215] The processed data is layered and aggregated according to geographic spatial scope (such as administrative divisions, grid cells) and scale (such as 1:100,000, 1:10,000) to generate multi-scale vector tiles (such as Mapbox Vector Tiles format).

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

[0217] Based on the confidence level of the abnormal area (such as the multiple of the residual exceeding the dynamic threshold), the area is divided into different levels (such as low, medium, and high risk) through the symbolization function of GIS software (such as ArcGIS and QGIS), and visualized using gradient colors (such as from yellow to red) to intuitively show the spatial aggregation degree and risk level of the abnormal area.

[0218] Overlaying a POI kernel density heat map (reflecting the intensity of human activity), a meteorological anomaly raster (such as precipitation anomaly percentage), and land use change areas, and performing spatial overlay analysis (such as raster algebra and buffer analysis), we generate a thematic map of driving factor contributions. For example, areas with a POI density change contribution greater than 60% are colored blue to highlight anomalies dominated by human activity.

[0219] Based on the spatiotemporal propagation path diagrams generated by S3 (e.g., the spatiotemporal causal chains of driving factors), a dynamic timeline is constructed in GIS. Interpolation algorithms are used to connect path nodes in chronological order. Animation technologies (e.g., WebGL and Canvas) are used to render the factor propagation process frame by frame (e.g., the spatiotemporal diffusion trajectory of a factory construction and the conversion of surrounding farmland to industrial land), enabling users to intuitively understand the driving mechanisms of land use change.

[0220] In a front-end interface (such as a WebGIS platform), clicking on a vector feature in an anomaly area triggers an attribute query event. The system then retrieves historical imagery of the area (derived from multi-temporal remote sensing data stored in S1), an attribution analysis report (such as the main drivers of SHAP value ranking and anomaly type labels), and a list of pre-set action recommendations (such as "recommend on-site verification within 3 days" and "initiate ecological restoration plan") in real time. These recommendations are displayed via pop-up windows or sidebars, enabling in-depth "what you see is what you get" information interaction.

[0221] An integrated timeline control links the multi-temporal land use classification map generated by S2 with the change detection results from S3. As the user slides the timeline, the system dynamically loads the classification map or change heat map for the corresponding time node, visualizing the evolution of land use over time (e.g., the conversion path from cultivated land to forest land to construction land in a watershed over 10 years), supporting trend analysis and anomaly tracing.

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

[0223] Through a RESTful API, the coordinates, anomaly type, and driver factor summary of high-priority anomaly areas (such as those with a confidence level greater than 90%) are pushed to the national land law enforcement platform in real time. Based on this data, the platform automatically generates inspection plans (such as planning drone inspection routes and assigning enforcement personnel tasks), achieving a closed-loop management process of "monitoring, early warning, and disposal."

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

[0225] The data acquisition module is used to divide the sub-regions, obtain 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 and classify features of remote sensing data based on deep learning models to generate land use classification maps;

[0227] A change identification module is used to compare multiple land use classification maps based on a bi-temporal difference attention network, identify change areas, and construct a change driving factor library that integrates meteorological data and human activity data;

[0228] An anomaly detection module, used to detect abnormal areas based on the driving factor library and generate anomaly type labels and attribution analysis reports in combination with dynamic warning thresholds;

[0229] The interactive module is used to use geographic information system software and distributed computing framework to visualize the monitoring results and output warning maps and disposal suggestions for abnormal areas.

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

[0231] This application is explained from the perspectives of purpose of use, effectiveness, progress and novelty, and has complied with the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings of this application are only preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc. that are similar or identical to those of this application, that is, all equivalent replacements or modifications made in accordance with the scope of the patent application of this application, should fall within the scope of protection of the patent application of this application.

Claims

1. A land use monitoring method based on remote sensing and big data, characterized in that: The method comprises: S1. Divide the sub-regions, obtain 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; S2. Extract and classify features of remote sensing data based on deep learning models to generate land use classification maps; S3. Based on a dual-temporal difference attention network, we compare multiple land use classification maps, identify change areas, and construct a change driving factor library that integrates meteorological data and human activity data. S4. Detect abnormal areas based on the driving factor library and generate abnormal type labels and attribution analysis reports in combination with dynamic warning thresholds; S5. Use geographic information system software and distributed computing framework to visualize the monitoring results and output warning maps and disposal suggestions for abnormal areas.

2. The land use monitoring method according to claim 1, characterized in that: Said S1 comprises: Synchronously acquire satellite optical images, radar images, and drone hyperspectral data, and access real-time data from weather stations and human activity data; Spatiotemporal adaptive preprocessing of remote sensing data, including cloud restoration based on GAN and radar data fusion; HDFS hierarchical storage architecture is used to implement spatiotemporal data block management.

3. The land use monitoring method according to claim 2, characterized in that: The cloud restoration based on GAN and radar data fusion includes: Performing spatiotemporal registration of the cloud coverage area of ​​the optical image with the radar image, and extracting the backscatter coefficient and texture features of the radar image; A dual-branch generative adversarial network is constructed, including an optical branch and a radar branch. The optical branch inputs the registered optical image to restore the semantic information of the cloud-covered area; the radar branch inputs the backscatter coefficient and texture features to reconstruct the surface structure under the cloud layer. Based on the spectral characteristics of optical images and the backscatter information of radar images, a cloud thickness estimation sub-network is used to generate pixel-level cloud thickness maps. According to the pixel-level cloud thickness map, a dynamic weight is assigned to each pixel, where thin cloud areas are mainly repaired by the optical branch, and thick cloud areas are mainly reconstructed by the radar branch; The features of the optical branch and the radar branch are fused through a cross-modal attention mechanism and conditional gated convolution; A multi-physics constraint loss function is designed, including radar-spectral correlation loss, terrain consistency loss, and seasonal vegetation index matching loss, to optimize the restoration results.

4. The land use monitoring method according to claim 1, wherein: The S2 includes: A CNN-Transformer-UNet hybrid network was constructed. The CNN convolutional neural network was used to extract local features of multispectral images. The Transformer encoder was used to analyze the long-term dependencies of vegetation indices and surface temperatures in multi-temporal images. The UNet decoder was used to fuse spatial context information and generate pixel-level land use classification maps. Based on the pre-trained model, a domain adaptation module is introduced to align the feature distribution of the public dataset and the target area data through adversarial training, thereby optimizing the classification accuracy in the few-sample scenario.

5. The land use monitoring method according to claim 1, characterized in that: The S3 includes: Detect land use change areas through a bi-temporal difference attention network and output semantic type labels; The change area is spatially superimposed with the meteorological grid and POI kernel density heat map to construct a spatiotemporal feature matrix; The spatiotemporal graph convolutional network is used to analyze the spatiotemporal propagation effect of factors and output the contribution weight of driving factors.

6. The land use monitoring method according to claim 5, characterized in that: The dual-temporal difference attention network includes: Extract the multispectral features of the two phases of images respectively; Seasonal noise areas are suppressed by spatial attention weights, and the weight calculation formula is: in, is the difference in characteristics between the two periods; is the Sigmoid activation function; Represents the convolution operation; is the spatial attention weight at position; Based on the Softmax output change type label.

7. The land use monitoring method according to claim 5, 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; Calculating edge weights based on spatial adjacency and time-lagged correlation; determining the maximum lag step size through Granger causality test for the time-lagged correlation; Output driving factor contribution weights and spatiotemporal propagation path diagram.

8. The land use monitoring method according to claim 1, characterized in that: The S4 includes: The baseline value of land use change is predicted based on the LSTM model, and the residual between the actual value and the predicted value is calculated. If the residual exceeds the dynamic threshold, it is marked as an intensity abnormal area. Generate anomaly type labels based on the contradiction between the driving factor contribution weight and regional attributes; The SHAP value interpretation model outputs the main driving factors and combines them with the spatiotemporal propagation path map to generate high-risk area predictions.

9. The land use monitoring method according to claim 8, characterized in that: The dynamic threshold setting method includes: Based on the sliding window, the mean and standard deviation of the land use change area in the sub-region, the meteorological data of the sub-region within the preset time window, and the POI density change rate of the sub-region within the preset time window are calculated to determine the dynamic threshold; The window size is dynamically selected based on the data volatility within the sub-region.

10. A land use monitoring system based on remote sensing and big data, used to implement the land use monitoring method according to claim 1, characterized in that: The system comprises: The data acquisition module is used to divide the sub-regions, obtain 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 and classify remote sensing data based on deep learning models to generate land use classification maps; A change identification module is used to compare multiple land use classification maps based on a bi-temporal difference attention network, identify change areas, and construct a change driving factor library that integrates meteorological data and human activity data; An anomaly detection module, used to detect abnormal areas based on the driving factor library and generate anomaly type labels and attribution analysis reports in combination with dynamic warning thresholds; The interactive module is used to use geographic information system software and distributed computing framework to visualize the monitoring results and output warning maps and disposal suggestions for abnormal areas.

Images