Land management system and method based on GIS system and big data analysis

Abstract

The invention relates to the cross technical field of land resource management and information technology, in particular to a land management system and method based on a GIS system and big data analysis, and the system comprises a data layer, a support layer and an application layer. The data layer comprises a multi-source data collection module and a GIS spatial database; the support layer comprises a big data processing engine, a GIS analysis tool set and a machine learning modeling platform, the application layer comprises an intelligent supervision module, a planning auxiliary module and a decision support module, and the system is suitable for intelligent management and decision support of land resource investigation, planning, protection and development of a full life cycle.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of land resource management and information technology, specifically to a land management system and method based on GIS system and big data analysis. Background Technology

[0002] With the acceleration of urbanization and the intensification of the contradiction between land resource supply and demand, traditional land management models face the following challenges: First, land-related data comes from diverse sources and is not uniformly formatted, including remote sensing imagery, GPS mapping data, government business system data, and socio-economic statistics. This data is scattered across different systems, resulting in a serious "data silo" phenomenon, making efficient integration and utilization difficult. Second, traditional analysis methods mainly rely on manual experience or simple statistics, failing to deeply explore the spatiotemporal correlations between multi-dimensional data, such as the coupling relationship between land use change and population migration and economic development, leading to limited decision support capabilities. Third, existing technologies lack the ability to monitor real-time changes in land resources, making it difficult to promptly detect problems such as illegal land use and ecological degradation, often resulting in regulatory lag. Furthermore, the business needs of macro-planning and micro-regulation differ significantly, making it difficult for existing technologies to achieve multi-scale collaborative analysis from territorial spatial planning to parcel management. These problems severely restrict the scientific nature and timeliness of land resource management. To address these issues, existing technologies urgently need improvement. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a land management system and method based on GIS system and big data analysis.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a land management system based on GIS system and big data analysis, comprising a three-layer architecture of data layer, support layer, and application layer:

[0005] The data layer includes a multi-source data acquisition module and a GIS spatial database. The multi-source data acquisition module integrates remote sensing imagery, ground sensor data, land management business system data, socio-economic data, and internet crawler data. The GIS spatial database uses PostGIS or MongoDB geo-extended storage to store spatial and attribute data, and optimizes query efficiency through R-tree spatial indexing. The indexing formula is: O(d) = log d N, where d is the number of spatial dimensions and N is the total amount of data;

[0006] The support layer includes a big data processing engine, a GIS analysis toolset, and a machine learning modeling platform. The big data processing engine is based on the Hadoop / Spark distributed computing framework for data cleaning, fusion, and standardization. The GIS analysis toolset integrates basic algorithms for spatial statistics, spatial interpolation, and change detection. The machine learning modeling platform supports the training and deployment of random forests, long short-term memory networks, and graph neural networks.

[0007] The application layer includes an intelligent monitoring module, a planning assistance module, and a decision support module. The intelligent monitoring module covers farmland protection (monitoring of land occupation and compensation balance), illegal land use identification (screening of suspected land parcels), and ecological red line early warning (detection of changes in ecologically sensitive areas). The planning assistance module is used for multi-scheme simulation of territorial spatial planning and assessment of land resource carrying capacity. The decision support module is used to generate land consolidation priority maps and market trend reports.

[0008] In some embodiments, the remote sensing imagery includes high-resolution satellite data from Gaofen-2 (1 meter / sub-meter level) or Sentinel-2 (10 meter level), and the ground sensor data includes hourly data on soil moisture and ground temperature.

[0009] In some embodiments, the big data processing engine performs batch remote sensing image preprocessing via Hadoop / Spark, including 6S atmospheric correction, RPC model geometric correction, Gram-Schmidt Pan Sharpening image fusion, and removal of images with cloud coverage >10%.

[0010] In some embodiments, the illegal land use identification uses a random forest classifier to screen suspected land parcels, and the ecological red line early warning triggers a threshold alarm based on change detection results.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a land management method based on a GIS system and big data analysis, employing the aforementioned management system, comprising the following steps:

[0012] Step 1: Multi-source data acquisition initialization

[0013] Collect multi-temporal remote sensing images and optimize the acquisition strategy based on the solar altitude angle: select periods with significant differences in solar altitude angle, collect multiple sets of images every quarter for key monitoring areas, and collect multiple sets of images every six months for general areas;

[0014] Step 2: Multi-source data preprocessing and fusion

[0015] A threshold-based segmentation algorithm is used to remove cloud noise from remote sensing images. The formula is used to mark cloud noise as |I(x,y)-μ|>T, where I(x,y) is the pixel value, μ is the neighborhood mean, and T is the preset threshold.

[0016] For attribute data, missing values ​​are filled using rule-based validation (land parcel area > 0) and linear interpolation.

[0017] Formula 2: x t =x t-1 +k·(x t+1 -x t-1 ), where x t x represents the missing value at time t. t-1 x t+1 For adjacent valid data, k is the interpolation weight;

[0018] Spatial data from different sources are unified to the CGCS2000 coordinate system, and a polynomial transformation model is used to correct the offset. The formula is: Where (X) o ,Y o (X) represents the original coordinates. t ,Y t Let C be the target coordinates and C be the polynomial coefficient matrix.

[0019] Step 3: Multi-dimensional analysis and modeling of land resources

[0020] By fusing multi-temporal imagery, a random forest classifier is used to calculate a land use change probability map, employing the following formula:

[0021] Where f i For the output of the i-th decision tree, w i σ is the weight, and σ is the Sigmoid function;

[0022] Construct an AHP-entropy weighted combined weighting model to assess arable land quality, formula:

[0023] Where S k For the normalized values ​​of indicators (organic matter content, slope, etc.), w j As the weights of the criterion layer (soil fertility, topographic conditions, etc.), w k The objective weights of the index layer are calculated using the entropy weight method;

[0024] Based on LSTM neural network to predict land market trends, the formula is:

[0025] P t =LSTM(P t-1 ,V,θ), where P t Let t be the land price at time t, V be the plot ratio, and θ be the policy impact factor;

[0026] Step 4: Intelligent Shadow Restoration of High-Resolution Remote Sensing Image

[0027] a. An improved U-Net+attention mechanism model (ShadowNet) is used to output a shadow probability map, and a binary shadow mask is generated by morphological closing operation and Otsu thresholding method;

[0028] b. Collect multi-temporal non-shadow reflectance values ​​for shadow pixels, and fill all-temporal shadows using Gaussian kernel spatial interpolation.

[0029] c. Use a conditional generative adversarial network (ShadowGAN) to generate shadowless images, with the loss function being: in To combat the losses, In order to perceive loss, Cross-entropy loss;

[0030] Step 5: Dynamic Monitoring and Intelligent Early Warning

[0031] Set up an early warning rule base ("Cultivated land decreases by more than 5 mu for 3 consecutive months", "Construction land expansion rate exceeds the planning limit by 120%).

[0032] The model calculation is triggered by real-time data streams. When the indicator exceeds the threshold, the abnormal area is marked on the GIS map and an early warning information is pushed.

[0033] In some embodiments, the multi-time image acquisition strategy in step 1 includes: selecting time periods with significant differences in solar altitude angle (9:00-10:30 am and 2:00-3:30 pm), acquiring 4-6 sets of images per quarter for key monitoring areas, and acquiring 2-3 sets per six months for general areas.

[0034] In some embodiments, in step 3, the farmland quality assessment determines the criterion layer weights (soil fertility, topographic conditions, irrigation capacity) through AHP, and calculates the objective weights of the index layer (organic matter content, slope, well density) using the entropy weight method.

[0035] In some embodiments, the shadow repair in step 4 needs to be verified by SSIM (≥0.9), PSNR (≥30dB) and classification accuracy (improvement ≥5%), and the unqualified areas need to be supplemented with images or the model needs to be adjusted.

[0036] In some embodiments, the ShadowNet network structure includes a ResNet-18 encoder with channel attention and a transposed convolutional decoder with spatial attention, and the loss function is Dice Loss + Cross-Entropy Loss.

[0037] In some embodiments, the generator of the ShadowGAN uses a U-Net structure, the discriminator uses a PatchGAN structure, and the input includes multi-temporal fused pixel values, shadow context spectrum, and land cover category labels.

[0038] Compared with existing technologies, the beneficial effects of this invention are: by integrating multi-source data through a three-layer architecture, processing massive information through a distributed computing framework, and realizing dynamic monitoring through machine learning models, it solves the problems of data silos, insufficient analytical capabilities, and lagging supervision, and has the advantages of efficiently integrating multi-source heterogeneous data, deeply mining spatiotemporal correlation patterns, and realizing dynamic monitoring and early warning.

[0039] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. The embodiments of this application will provide a detailed description and understanding of the application. Attached Figure Description

[0040] Figure 1 This is a diagram of the overall system architecture.

[0041] Figure 2 A flowchart for multi-source data fusion;

[0042] Figure 3 This is a schematic diagram of a farmland quality assessment model.

[0043] Figure 4 This is a structural diagram of the LSTM land price prediction model. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] In traditional land management systems, data integration faces the challenge of storing mixed structured and unstructured data from multiple heterogeneous sources. Spatial registration accuracy between remote sensing imagery and ground sensor data is insufficient due to coordinate system differences. Furthermore, the lack of unified semantic standards makes it difficult to establish correlation models between land business data and socio-economic data. Dynamic monitoring relies on manual interpretation of changes in remote sensing imagery, resulting in a three-day to one-week response delay in identifying illegal land use. Farmland protection monitoring cannot track the balance of cultivated land occupation and compensation in real time. In cross-scale analysis, differences in storage architecture between macro-level planning data and micro-level land parcel data lead to topological error rates exceeding 8% in spatial overlay analysis.

[0046] For example, a provincial land monitoring platform needs to process 2TB of newly added multi-source data daily, including Gaofen-2 satellite imagery, third national land survey vector data, soil moisture time-series data, and internet land transaction text data. Due to the lack of a unified spatial data engine, remote sensing image preprocessing requires three format conversions between ENVI and ArcGIS platforms, resulting in processing a single image taking over four hours. When assessing land resource carrying capacity, planning departments need to manually link population density raster data with construction land approval forms; errors in attribute field mapping cause the assessment model output to deviate from actual values ​​by 12%-15%. Ecological red line monitoring relies on quarterly remote sensing image comparisons, failing to capture sudden deforestation events; subsequent investigations revealed that illegal activities had been ongoing for two months.

[0047] If the above problems are not addressed, the low efficiency of multi-source data fusion will directly limit the data update frequency of the land supervision system, making it impossible for the dynamic monitoring cycle to meet the real-time requirements of early warning for farmland protection red lines. Accumulated spatial analysis errors will cause land spatial planning schemes to deviate from the actual distribution characteristics of land resources, resulting in an imbalance in the allocation of construction land indicators. The lack of cross-scale data collaboration capabilities will hinder the closed loop of macro-planning and micro-level supervision, leading to spatial decision-making gaps in the investigation and handling of illegal land use and the implementation of land consolidation projects.

[0048] To address the aforementioned challenges, this application first considers establishing a unified data storage architecture to eliminate format differences, aiming to resolve the difficulty of integrating multi-source heterogeneous data. For the pain point of insufficient spatial registration accuracy, it explores improving coordinate consistency through standardized processing procedures. To address the issue of lagging dynamic monitoring, it attempts to introduce automated analysis models to replace manual interpretation. Regarding the difficulties of cross-scale collaboration, it studies multi-level data processing mechanisms to bridge macro and micro needs. Furthermore, it was found that simply improving data storage or analysis modules is insufficient to form a systematic solution; a complete architectural framework is required. By analyzing the coupling relationships among various technical aspects, it was ultimately determined that data acquisition, processing, and analysis functions should be designed in layers to form a scalable technical framework.

[0049] In this regard, such as Figures 1 to 4 As shown, this application proposes a land management system based on GIS and big data analysis, comprising a three-layer architecture: a data layer, a support layer, and an application layer. The data layer includes a multi-source data acquisition module and a GIS spatial database. The multi-source data acquisition module integrates remote sensing imagery, ground sensor data, land management business system data, socio-economic data, and internet crawler data, supporting structured, semi-structured, and unstructured data access. The GIS spatial database uses PostGIS or MongoDB geo-extended storage for spatial and attribute data, and optimizes query efficiency through R-tree spatial indexing. The indexing formula is: O(d) = log dN, where d is the number of spatial dimensions and N is the total amount of data; the support layer includes a big data processing engine, a GIS analysis toolset, and a machine learning modeling platform: the big data processing engine uses the Hadoop / Spark distributed computing framework to clean, integrate, and standardize massive amounts of data; the GIS analysis toolset integrates spatial statistics, spatial interpolation, and change detection algorithms; the machine learning modeling platform supports the training and deployment of random forests, long short-term memory networks, and graph neural networks; the application layer includes an intelligent supervision module, a planning assistance module, and a decision support module: the intelligent supervision module covers farmland protection, illegal land use identification, and ecological red line early warning; the planning assistance module is used for multi-scheme simulation of land spatial planning and land resource carrying capacity assessment; the decision support module generates land consolidation priority maps and market trend reports.

[0050] The GIS spatial database refers to a dedicated database for storing geospatial data. It can be implemented using PostGIS or MongoDB geospatial extensions, supporting the storage of vector data (land parcels, land use features) and raster data (remote sensing imagery, DEM). It optimizes query efficiency through R-tree spatial indexing, addressing the problems of scattered data storage and low query efficiency in traditional land management. The multi-source data acquisition module integrates various heterogeneous data sources. Specifically, it can access remote sensing imagery (optical / radar), ground sensor data (soil moisture, ground temperature), land management business system data (national land survey, real estate registration, planning permits), socio-economic data (population, GDP, industrial distribution), and internet crawler data (land transaction information). It supports structured, semi-structured, and unstructured data access, solving the integration difficulties caused by diverse data sources and inconsistent formats. The big data processing engine refers to a distributed computing framework for processing massive amounts of data. Specifically, it uses Hadoop / Spark to perform data cleaning (duplicate removal, logical error checking), fusion (spatial coordinate registration, semantic alignment), and standardization (unified CGCS2000 coordinate system, UTC timestamps), addressing the problems of low efficiency and poor data quality in traditional methods. The GIS analysis toolset refers to a collection of algorithms providing spatial analysis functions, specifically integrating spatial statistics (hotspot analysis Getis-Ord Gi*), spatial interpolation (Kriging, inverse distance weighted IDW), and change detection (difference analysis, machine learning classification) algorithms, addressing the inability of traditional analysis methods to uncover spatiotemporal correlations. The machine learning modeling platform refers to modules for training and deploying machine learning models, specifically supporting the training and deployment of random forests, long short-term memory networks, and graph neural networks, addressing the limited decision support capabilities of traditional experience-driven methods. Among them, the application layer modules refer to the functional modules oriented towards business needs, including the intelligent supervision module (farmland protection, illegal land use identification, ecological red line early warning), the planning assistance module (territorial spatial planning simulation, land resource carrying capacity assessment), and the decision support module (land consolidation priority map, market trend report), which solve the problems of lagging dynamic monitoring and difficulties in cross-scale collaboration.

[0051] The core innovation of this application lies in the integration of multi-source heterogeneous data through a three-layer architecture design of data layer, support layer and application layer. By combining distributed computing engine, GIS spatial analysis tools and machine learning models, it achieves unified storage, efficient processing and multi-dimensional analysis of land management data, thereby systematically solving the problems of data silos, shallow analysis and weak dynamic response in the traditional model.

[0052] The working process and principle of this application are as follows: the land management system adopts a three-layer architecture design, including a data layer, a support layer, and an application layer. The data layer is responsible for multi-source data collection and storage, the support layer provides data processing and analysis capabilities, and the application layer implements specific business functions.

[0053] The multi-source data acquisition module in the data layer integrates remote sensing imagery, ground sensor data, land management business system data, socioeconomic data, and internet crawler data. This data encompasses structured, semi-structured, and unstructured types and is accessed through a unified interface. The GIS spatial database uses PostGIS or MongoDB geo-extended storage for spatial and attribute data. An R-tree spatial index is used to optimize query efficiency, with a complexity of O(d) = log0. d N, where d is the number of spatial dimensions and N is the total amount of data.

[0054] The supporting big data processing engine is based on the Hadoop / Spark distributed computing framework, performing data cleaning, fusion, and standardization. The GIS analysis toolset integrates spatial statistics, spatial interpolation, and change detection algorithms. The machine learning modeling platform supports the training and deployment of random forest, long short-term memory networks, and graph neural network models.

[0055] The application layer comprises three modules: intelligent supervision, planning assistance, and decision support. The intelligent supervision module enables farmland protection, illegal land use identification, and ecological red line early warning. The planning assistance module is used for multi-scheme simulation of national land spatial planning and land resource carrying capacity assessment. The decision support module generates land consolidation priority maps and market trend reports.

[0056] The various layers work collaboratively: the data layer provides the data foundation for the upper layers, the support layer processes and analyzes the data, and the application layer implements specific business functions based on the processing results. This layered design improves the system's modularity and scalability.

[0057] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0058] Data layer implementation: The multi-source data acquisition module connects to remote sensing satellite data centers, meteorological station networks, land and resources department business systems, statistics bureau databases, and custom web crawlers via API interfaces. The GIS spatial database is implemented using PostGIS extended PostgreSQL, storing vector data (such as land parcel boundaries) and raster data (such as remote sensing imagery). The R-tree index is automatically built and maintained using built-in PostGIS functions.

[0059] Support layer implementation: The big data processing engine is deployed on a Hadoop cluster, using YARN for resource scheduling and Spark for distributed computing. The GIS analysis toolset is developed based on the GDAL / OGR library, integrating algorithms such as Getis-Ord Gi* hotspot analysis and Kriging spatial interpolation. The machine learning modeling platform uses the TensorFlow framework and comes pre-built with model structures such as random forest and LSTM.

[0060] Application layer implementation: The intelligent monitoring module uses the random forest algorithm to identify illegal land use and sets ecological red line early warning thresholds based on change detection results. The planning assistance module uses cellular automata to simulate urban expansion and combines multi-factor overlay analysis to assess land resource carrying capacity. The decision support module uses the AHP method to determine land consolidation priorities and uses the ARIMA model to predict land market trends.

[0061] The system is deployed on a Linux server cluster, employing Docker containerization technology to implement a microservice architecture, and using Kubernetes for container orchestration and management. The front-end uses the Vue.js framework to develop web applications, while the back-end uses the Spring Boot framework to provide RESTful APIs.

[0062] Through the above-described scheme, this application achieves efficient integration and storage of multi-source heterogeneous data, improving the spatiotemporal integrity of land resource data. The analysis methods based on big data and machine learning technologies enhance the predictive ability of land use change trends. The automated dynamic monitoring mechanism shortens the response time for identifying illegal land use and issuing ecological red line warnings. The multi-scale data collaborative analysis framework promotes business collaboration between macro-planning and micro-level supervision. These improvements effectively address the problems faced by traditional land management systems, such as data fragmentation, insufficient analytical depth, lagging dynamic monitoring, and difficulties in cross-scale collaboration, providing technical support for the scientific management and sustainable utilization of land resources.

[0063] In some of the solutions described above in this application, the multi-source data acquisition module integrates various remote sensing images and ground sensor data. However, in practical applications, there are problems such as insufficient image resolution leading to limited accuracy in ground feature identification, and excessively long sensor data acquisition intervals affecting the timeliness of dynamic monitoring.

[0064] This application further proposes remote sensing imagery including data from Gaofen-2 or Sentinel-2 satellites, and ground sensor data including hourly soil moisture and ground temperature data.

[0065] The Gaofen-2 satellite data provides spatial resolution ranging from 1 meter to sub-meter, while the Sentinel-2 satellite data provides 10-meter resolution. Ground sensors collect soil moisture and ground temperature data at a fixed frequency every hour, forming a time-series dataset. Gaofen-2 data is suitable for detailed ground feature identification, while Sentinel-2 data is suitable for large-scale monitoring; the two complement each other in terms of resolution. The hourly sensor data constructs a high-density time sampling point, which can capture the short-period variation patterns of soil parameters.

[0066] Specifically, in land resource monitoring scenarios, Gaofen-2 satellite data, with a 1-meter resolution, can clearly identify microscopic land features such as field boundaries and small structures, while Sentinel-2 data, with a 10-meter resolution, enables regional-level monitoring of land cover changes. The fusion of these two types of satellite data at multiple scales can simultaneously meet the needs of macroscopic trend analysis and microscopic anomaly detection. Soil moisture data collected hourly by ground sensors, through time-series analysis, can establish a dynamic correlation model between soil moisture and crop growth status. For example, in drought early warning scenarios, an irrigation alert can be triggered if the humidity decline exceeds a threshold for six consecutive hours. The hourly acquisition frequency of ground temperature data matches the meteorological data update cycle, providing fundamental data with minute-level time accuracy for permafrost thawing monitoring. This technical solution, through the combined acquisition of multi-resolution satellite data and high-frequency sensor data, enables land management systems to possess both spatial detail capture capabilities and continuous monitoring capabilities in the temporal dimension.

[0067] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0068] The remote sensing imagery includes data from the Gaofen-2 or Sentinel-2 satellites. Gaofen-2 provides optical imagery with a resolution of 1 meter / sub-meter, while Sentinel-2 provides multispectral imagery with a resolution of 10 meters. Ground sensor data includes hourly data on soil moisture and ground temperature. Soil moisture sensors are deployed at depths of 0-20cm, 20-40cm, and 40-60cm in farmland, recording soil moisture content hourly. Ground temperature sensors are deployed at depths of 5cm, 10cm, and 20cm below the surface, recording soil temperature hourly.

[0069] Through the above technical solutions, this application achieves high spatiotemporal resolution land resource monitoring. Gaofen-2 and Sentinel-2 satellite imagery provides large-scale, high-frequency surface observation data, effectively identifying land use changes. The ground sensor network provides refined soil moisture and ground temperature information, aiding in the assessment of farmland quality and ecological conditions. The fusion of multi-source data makes land resource monitoring more comprehensive and accurate, providing reliable data support for land management decisions.

[0070] In some of the solutions mentioned above in this application, although the big data processing engine can achieve the cleaning, fusion and standardization of massive data, it lacks a targeted preprocessing process when processing batches of remote sensing images. This results in problems such as atmospheric interference, geometric distortion, insufficient resolution and cloud noise interference in the image data, which affects the accuracy of subsequent analysis.

[0071] This application further proposes a big data processing engine that uses Hadoop / Spark to perform batch remote sensing image preprocessing, including 6S atmospheric correction, RPC model geometric correction, Gram-Schmidt Pan Sharpening image fusion, and automatic removal of images with cloud coverage >10%.

[0072] Among them, 6S atmospheric correction uses a radiative transfer model to eliminate the influence of aerosol scattering and atmospheric absorption on image reflectivity; RPC model geometric correction uses rational polynomial coefficients to perform terrain projection transformation on satellite images to eliminate geometric distortion caused by orbital inclination; Gram-Schmidt Pan Sharpening image fusion uses orthogonal transformation to improve the texture details of the fused image by taking advantage of the spatial resolution difference between panchromatic and multispectral bands; images with cloud coverage >10% are automatically removed by calculating the proportion of cloud pixels in the image and triggering a data filtering mechanism when it exceeds a threshold.

[0073] Specifically, within the distributed computing framework, remote sensing imagery is segmented into multiple data blocks for parallel processing. The 6S atmospheric correction module calculates the atmospheric correction coefficient for each pixel by inputting satellite transit time, solar zenith angle, and atmospheric parameters. The RPC model geometric correction module loads satellite orbital parameters and a digital elevation model to establish a mapping relationship between ground control points and image coordinates. The Gram-Schmidt fusion module performs covariance matrix decomposition on the panchromatic and multispectral bands to generate a high-resolution fused image. The cloud cover detection module uses a threshold segmentation algorithm to count the number of cloud pixels; when a cloud cover area exceeds 10%, the image is automatically marked as invalid data and excluded from subsequent processing. Through these steps, the radiometric accuracy of the image data is improved to within ±5%, and the geometric error is controlled within one pixel, effectively supporting subsequent spatial analysis and machine learning modeling.

[0074] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0075] The big data processing engine employs the Hadoop distributed computing framework for batch remote sensing image preprocessing. First, raw remote sensing image data is distributed and stored across multiple data nodes using HDFS. Then, parallel processing is performed using the MapReduce programming model: In the Map phase, each data node reads locally stored image blocks, executes the 6S atmospheric correction algorithm to eliminate atmospheric scattering and absorption effects, applies the RPC model for geometric correction to achieve sub-pixel level registration accuracy, and uses the Gram-Schmidt Pan Sharpening algorithm to fuse panchromatic and multispectral bands. In the Reduce phase, the processing results are aggregated, cloud coverage is calculated, and images with cloud coverage exceeding 10% are automatically removed. Finally, the processed image results are written back to HDFS for storage.

[0076] Through the above technical solutions, this application achieves efficient batch preprocessing of massive remote sensing image data. Utilizing a distributed computing framework, a large number of images can be processed in parallel, significantly improving data processing speed. Simultaneously, automated atmospheric correction, geometric correction, and image fusion ensure the accuracy and consistency of the processing results. Furthermore, the cloud cover filtering mechanism effectively removes low-quality images, improving the reliability of subsequent analysis.

[0077] In some of the solutions mentioned above in this application, the restoration of shadow areas in high-resolution remote sensing images has the problem of unsatisfactory restoration effect. Traditional methods have difficulty in accurately distinguishing shadows from changes in real ground features, resulting in a high false alarm rate in dynamic monitoring and affecting the reliability of monitoring results.

[0078] This application further proposes a high-resolution remote sensing image shadow intelligent restoration method, including the following steps: a. Using an improved U-Net+attention mechanism model to output a shadow probability map, and generating a binary shadow mask through morphological closing operation and Otsu thresholding method; b. Collecting multi-temporal non-shadow reflectance values ​​for shadow pixels, and filling all-temporal shadows with Gaussian kernel spatial interpolation; c. Using a conditional generative adversarial network to generate shadow-free images, with the loss function being a weighted combination of adversarial loss, perceptual loss and cross-entropy loss.

[0079] The improved U-Net+attention mechanism model employs a ResNet-18 encoder with channel attention to extract multi-scale features, enhancing shadow edge feature capture capabilities through spatial attention. Morphological closing operations utilize a 5×5 circular kernel to eliminate small holes, and the Otsu thresholding method automatically segments the probability map to generate a binary mask. For multi-temporal non-shadow reflectance values, at least three sets of valid observation data from different temporal phases at the same geographical location are selected. For the full-temporal shadow region, neighborhood-weighted interpolation is performed using a Gaussian kernel with a standard deviation of 1.5. The conditional generative adversarial network (GAN) generator uses a U-Net structure to fuse shadow context spectral features, while the discriminator uses a PatchGAN structure for local texture authenticity judgment. In the loss function, the adversarial loss weight λ_adv is set to 1.0, the perceptual loss weight λ_perc is set to 0.8, and the cross-entropy loss weight λ_CE is set to 0.5.

[0080] Specifically, in the shadow probability map generation stage, pixel-level shadow detection is achieved through an improved U-Net model. The channel attention mechanism prioritizes activating shadow-related spectral bands, while spatial attention focuses on the boundary between buildings and vegetation. Morphological closing operations eliminate discrete noise points caused by cloud interference, and the Otsu thresholding method adaptively determines the segmentation threshold to ensure that the shadow mask boundary matches the true outline of the ground features. In the reflectance value restoration stage, for some shadow areas, the mean value is calculated using at least three sets of valid observations from multi-temporal data to fill the shadows, ensuring that the restoration result conforms to the spectral characteristics of the ground features. For shadow areas across all time phases, Gaussian kernel interpolation is used, with the interpolation range controlled by 1.5 standard deviations to avoid excessive smoothing that could lead to texture distortion. During the training of the generative adversarial network, the generator receives a shadow mask and multispectral images as input. The output restored image must simultaneously meet the local texture realism constraint of the discriminator, the perceptual similarity constraint of the VGG16 feature layer, and the pixel-level classification accuracy constraint. The final generated shadowless image must meet the quality standards of SSIM≥0.9 and PSNR≥30dB, so that the restored area and the surrounding land features remain continuous in spectral and texture features, effectively reducing the misjudgment rate of subsequent change detection.

[0081] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0082] Illegal land use identification employs a random forest classifier to screen suspected land parcels. First, a training dataset is constructed, including known illegal land use samples and normal land use samples. Each sample contains multiple features, such as land type change frequency, building density, and vegetation index. The random forest model consists of multiple decision trees, each trained and predicting independently. In the prediction phase, the land parcel to be detected is input into the model, and the voting results of all decision trees are combined to output the probability of illegal land use. A threshold is set (e.g., 0.7); land parcels with probabilities higher than the threshold are marked as suspected illegal land use.

[0083] The ecological red line early warning system dynamically sets thresholds and triggers alarms based on change detection results. First, multi-temporal remote sensing imagery is used to detect changes and calculate the area changes of various land uses within the ecological red line area. A statistical distribution model of area changes is established based on historical data. The dynamic threshold is set to the mean plus three standard deviations. When the detected change exceeds this threshold, an early warning is triggered. The warning information includes the location, area, and type of change, and is pushed to management personnel via the system interface and SMS.

[0084] Through the above technical solutions, this application achieves intelligent identification of illegal land use and dynamic monitoring of ecological red lines. The random forest model comprehensively considers multi-dimensional features, improving the accuracy of illegal land use identification. The dynamic threshold setting makes ecological red line early warning more sensitive and reasonable, reducing false alarms and missed alarms. These measures effectively improve the level of intelligent land management and provide technical support for the timely detection and handling of land use problems.

[0085] Some of the solutions mentioned above in this application suffer from problems such as low efficiency in integrating multi-source data, delayed dynamic monitoring response, and insufficient cross-scale analysis capabilities, which result in an inability to effectively support the real-time decision-making and early warning needs in land resource management.

[0086] This application further proposes a land management method based on GIS system and big data analysis, including the following steps: Step 1: Multi-source data acquisition and initialization; Step 2: Multi-source data preprocessing and fusion; Step 3: Multi-dimensional analysis and modeling of land resources; Step 4: Intelligent restoration of shadows in high-resolution remote sensing images; Step 5: Dynamic monitoring and intelligent early warning.

[0087] In step 1, multi-temporal remote sensing images are acquired by optimizing the time periods of solar altitude angle differences. 4-6 sets of images are acquired quarterly for key monitoring areas, and 2-3 sets are acquired every six months for general areas. In step 2, a threshold-based segmentation algorithm is used to remove cloud noise. The formula is: when |I(x,y)-μ|>T, it is marked as cloud noise. Missing values ​​in the attribute data are filled in using rule validation and linear interpolation. The formula is: x t =x t-1 +k·(x t+1 -x t-1 Spatial data uniformly adopts a polynomial transformation model to correct coordinate offsets, the formula is: In step 3, the land use change probability map is calculated using a random forest classifier. The formula is as follows: Construct an AHP-entropy weighted model to assess farmland quality, using the following formula: The formula for predicting land market trends based on LSTM neural networks is: P t =LSTM(P t-1In step 4, an improved U-Net+attention mechanism model is used to generate a shadow probability map. A binary shadow mask is generated using morphological closing operations and the Otsu thresholding method. Shadow pixels are filled with multi-temporal non-shadow reflectance values ​​or Gaussian kernel spatial interpolation. A conditional generative adversarial network is used to generate a shadowless image, with the loss function being: In step 5, an early warning rule base is set up. When the real-time data stream triggers an indicator that exceeds the threshold, the abnormal area is marked on the GIS map and an early warning information is pushed.

[0088] Specifically, in the multi-source data acquisition phase, image acquisition is conducted during periods of significant difference in solar altitude angle to obtain data on changes in surface features under different lighting conditions. A acquisition frequency of 4-6 sets per quarter in key monitoring areas can capture seasonal land use changes. In the data preprocessing phase, a threshold segmentation algorithm based on neighborhood mean effectively distinguishes between cloud cover and surface reflectance differences. Linear interpolation formulas fill in missing values ​​through weighted calculations of data from adjacent time points, and a multinomial coordinate transformation model eliminates projection biases from different data sources. During land resource analysis, a random forest classifier outputs land use change probabilities through weighted multi-decision tree calculations, an AHP-entropy weighted model comprehensively assesses arable land quality using both subjective and objective weights, and an LSTM network predicts land price fluctuations through temporal feature learning. In the shadow restoration phase, the attention mechanism U-Net model accurately locates shadow areas, multi-temporal reflectance data and Gaussian interpolation ensure the continuity of spectral restoration in shadow areas, and a generative adversarial network optimizes image restoration quality through adversarial training. During the dynamic monitoring phase, quantitative thresholds are preset for the rate of reduction in arable land area and the rate of expansion of construction land. When the real-time data stream triggers the rules, the GIS map automatically marks abnormal areas and pushes early warning information through multiple channels, thereby achieving closed-loop management of land resource changes.

[0089] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0090] During the initial phase of multi-source data acquisition, multi-temporal remote sensing images were collected. The acquisition strategy was optimized based on the solar elevation angle, selecting periods with significant differences in solar elevation angle (9:00-10:30 AM and 2:00-3:30 PM). 4-6 sets of images were collected per quarter for key monitoring areas, and 2-3 sets per six months for general areas.

[0091] In the multi-source data preprocessing and fusion stage, a threshold-based segmentation algorithm is used to remove cloud noise from remote sensing images. Missing values ​​in attribute data are filled in using rule-based validation and linear interpolation. Spatial data from different sources are unified to the CGCS2000 coordinate system, and a polynomial transformation model is used to correct for offsets.

[0092] In the multi-dimensional analysis and modeling stage of land resources, multi-temporal imagery is integrated, and a random forest classifier is used to calculate the probability map of land use change. An AHP-entropy weighting model is constructed to assess arable land quality. Land market trends are predicted based on an LSTM neural network.

[0093] In the intelligent shadow inpainting stage of high-resolution remote sensing imagery, an improved U-Net+attention mechanism model (ShadowNet) is used to output a shadow probability map. A binary shadow mask is generated through morphological closing operations and the Otsu thresholding method. Multi-temporal non-shadow reflectance values ​​are collected for shadow pixels, and Gaussian kernel spatial interpolation is used to fill shadows across all time phases. A conditional generative adversarial network (ShadowGAN) is then used to generate shadow-free imagery.

[0094] The dynamic monitoring and intelligent early warning stage establishes an early warning rule base (a decrease of more than 5 mu of cultivated land for three consecutive months, and an expansion rate of construction land exceeding the planned upper limit by 120%). When the real-time data stream triggers indicators that exceed the threshold, abnormal areas are marked on the GIS map and early warning information is pushed (SMS / APP notification).

[0095] Through the above technical solutions, this application achieves efficient integration and analysis of multi-source heterogeneous data. Based on big data and machine learning technologies, it improves the analytical depth and predictive capabilities of land resource management. Real-time monitoring and intelligent early warning mechanisms enhance the dynamic response capability to changes in land resources. Collaborative analysis of multi-scale data supports the coordination between macro-planning and micro-level supervision. Therefore, this solution effectively addresses the technical challenges faced by traditional land management models, such as data integration, analytical depth, dynamic monitoring, and cross-scale collaboration, thereby improving the level of intelligence in land resource management.

[0096] In some of the solutions described above in this application, optimizing the data collection strategy may lead to a mismatch between the data collection frequency and the monitoring needs. If the same collection cycle is used for key monitoring areas as for general areas, it may result in insufficient capture of key change information or waste of resources.

[0097] This application further proposes that in step 1, 4-6 sets of images should be collected every quarter in key monitoring areas and 2-3 sets should be collected every six months in general areas.

[0098] The data collection frequency is set based on regional monitoring priorities. Key monitoring areas include ecologically sensitive areas or areas with high rates of illegal land use, requiring multiple data collections within a quarter to cover seasonal variations. General areas refer to regions with relatively stable land use, where a six-month collection cycle is sufficient for basic monitoring needs. The number of image sets ranges from 4 to 6 sets to cover different solar altitude angles within a quarter, ensuring multi-angle data acquisition. 2 to 3 sets are used to establish a semi-annual variation baseline through intermittent collection. Data collection times are selected from 9:00-10:30 AM and 2:00-3:30 PM, utilizing the difference in solar altitude angles to create complementary lighting conditions.

[0099] Specifically, key monitoring areas undergo at least four image acquisitions per quarter, with each acquisition spaced approximately 20-25 days apart, creating three-dimensional observation conditions by combining different solar azimuth angles. For example, in monitoring farmland compensation, five acquisitions per quarter can capture land cover changes during the spring plowing, summer planting, and autumn harvest stages. General areas undergo three acquisitions every six months, in January, April, and October, avoiding cloud interference during the rainy season. The acquisition equipment utilizes the Gaofen-2 satellite to achieve sub-meter resolution, coupled with ground base station data reception. With this frequency setting, key areas achieve 16-24 sets of data acquisitions annually, a 100% increase in data density compared to the traditional 8-12 sets; general areas maintain 4-6 sets annually, a 50% reduction in data redundancy compared to the traditional 12 sets. Data storage employs a distributed architecture, with a separate time-series index created for key area data each quarter, supporting high-frequency change detection algorithms.

[0100] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0101] Key monitoring areas include urban peripheries, ecological protection red line areas, and basic farmland protection areas. For these areas, 4-6 sets of high-resolution satellite imagery will be collected each quarter. Data collection will be conducted in representative months of each of the four seasons: spring (March-May), summer (June-August), autumn (September-November), and winter (December-February). Within each quarter, 2-3 sets of imagery will be collected in two time slots: 9:00-10:30 AM and 2:00-3:30 PM.

[0102] For general areas, including urban built-up areas and rural suburbs, 2-3 sets of medium-resolution satellite imagery will be collected every six months. Collection will be conducted 1-2 times each during the growing season (April-September) and the non-growing season (October-March). The collection times will also be 9:00-10:30 AM and 2:00-3:30 PM.

[0103] The image acquisition strategy is optimized based on the solar altitude angle. Acquisition is conducted during periods of significant difference in solar altitude angle to obtain images with varying shadow effects, which is beneficial for subsequent change detection and feature identification. Differentiated acquisition frequencies effectively control data storage and processing costs while ensuring monitoring effectiveness.

[0104] Through the above technical solution, this application achieves differentiated monitoring of key areas and general areas. High-frequency, high-resolution image acquisition of key areas can promptly capture land use changes and improve monitoring accuracy. A lower-frequency acquisition strategy is adopted for general areas, ensuring basic monitoring needs while reducing the data processing burden. Optimized acquisition timing fully utilizes the differences in image characteristics caused by variations in solar altitude angle, improving the accuracy of subsequent analysis. This differentiated acquisition strategy effectively balances monitoring effectiveness and system resource consumption, improving the efficiency and accuracy of land management.

[0105] In some of the schemes mentioned above in this application, traditional farmland quality assessment methods suffer from excessive subjectivity or insufficient data-driven approaches in determining weights. A single weighting method is difficult to balance expert experience with objective data characteristics, resulting in deviations between the assessment results and the actual land conditions, which affects the scientific nature of subsequent land consolidation and planning decisions.

[0106] This application further proposes that the arable land quality assessment uses the analytic hierarchy process (AHP) to determine the weights of the criteria layer, including soil fertility, topographic conditions, and irrigation capacity, and the entropy weight method to calculate the objective weights of the index layer, including organic matter content, slope, and well density.

[0107] Among them, the Analytic Hierarchy Process (AHP) constructs a criterion-level judgment matrix and uses expert scoring to compare the pairwise importance of soil fertility, topographic conditions, and irrigation capacity. After calculating the eigenvectors, the weights of the criterion-level are normalized, and the consistency test requires the CR value to be less than 0.1. The entropy weight method calculates the information entropy based on the dispersion of each indicator data and determines the indicator-level weights through the information utility value. In specific implementation, the indicators of organic matter content, slope, and well density are standardized to eliminate dimensional differences. The combined weighting model linearly weights the subjective weights obtained by the AHP and the objective weights obtained by the entropy weight method to generate a comprehensive weight for farmland quality assessment.

[0108] Specifically, in the process of assessing arable land quality, a criterion layer system is first established, encompassing soil fertility, topographic conditions, and irrigation capacity. Each criterion layer includes quantifiable indicators such as organic matter content, slope, and well density. The Analytic Hierarchy Process (AHP) is used to rank the importance of the criterion layer elements by experts; for example, soil fertility is assigned a weight of 0.5, topographic conditions 0.3, and irrigation capacity 0.2, meeting the CR (Comparison of Concentrations) requirement of 0.08. The entropy weight method calculates the entropy values ​​of each indicator based on historical arable land data. For example, the well density indicator has a lower information entropy due to large regional differences, thus receiving a higher weight of 0.4, while the organic matter content indicator, with its concentrated data distribution, has a weight of 0.2. The final comprehensive weight is calculated by combining the weights from the AHP and the entropy weight method, each accounting for 50%. For example, the comprehensive weight for the irrigation capacity criterion is 0.2 × 0.5 + 0.3 × 0.5 = 0.25. This combined model retains the experts' empirical judgments on the factors influencing arable land quality while reflecting the actual data distribution characteristics, making the assessment results more closely reflect the true state of regional land resources.

[0109] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0110] The farmland quality assessment uses the analytic hierarchy process (AHP) to determine the weights of the criterion layer, including soil fertility, topographic conditions, and irrigation capacity. Soil fertility is weighted at 0.4, topographic conditions at 0.3, and irrigation capacity at 0.3. Further, the entropy weight method is used to calculate the objective weights of the indicator layer. Specifically, organic matter content is weighted at 0.35, slope at 0.3, and well density at 0.35. Therefore, the comprehensive score calculation formula is:

[0111] S 综合 =0.4(0.35S) 有机质 +0.3S 坡度 +0.35S 机井 )·0.3(0.35S 有机质 +0.3S 坡度 +0.35S 机井 )·0.3(0.35S 有机质 +0.3S 坡度 +0.35S 机井 )

[0112] Among them, S 有机质 S 坡度 S 机井 These are the normalized scores for organic matter content, slope, and well density, respectively.

[0113] Through the above technical solution, this application achieves objectivity and comprehensiveness in arable land quality assessment. The criterion layer weights determined by the analytic hierarchy process (AHP) reflect expert experience, while the indicator layer weights calculated by the entropy weight method reflect the information content of the data itself. The combination of the two preserves the guiding role of expert knowledge while avoiding biases that may arise from subjective judgment. Furthermore, the comprehensive score of multi-dimensional indicators more comprehensively reflects the arable land quality status, providing a reliable basis for land resource management and planning decisions.

[0114] In some of the schemes described above in this application, the shadow restoration process may fail to meet the restoration standards due to image quality or model parameter deviations, affecting the accuracy of subsequent land use classification and change detection.

[0115] This application further proposes that shadow restoration should be verified by SSIM (≥0.9), PSNR (≥30dB) and classification accuracy (improvement ≥5%), and that unqualified areas should be supplemented with images or the model should be adjusted.

[0116] SSIM measures the structural similarity between the restored image and the original image, PSNR assesses the peak signal-to-noise ratio, and the improvement in classification accuracy is calculated by comparing the land cover classification results before and after restoration. For areas where SSIM or PSNR does not meet the standards, it is necessary to collect additional multi-temporal images or adjust the hyperparameters of the generative adversarial network. If the improvement in classification accuracy is insufficient, it is necessary to optimize the model training samples or adjust the weight allocation of the attention mechanism module. Image supplementation should prioritize data from non-shaded periods, and model adjustments include retraining the ShadowNet encoder or optimizing the weight ratio of the loss function in ShadowGAN.

[0117] Specifically, the verification process first calculates the SSIM and PSNR values ​​of the restored image. If SSIM ≥ 0.9 and PSNR ≥ 30dB, the image quality is deemed acceptable; otherwise, an image supplementation process is triggered. Acceptable images are input into a pre-trained land use classification model, and the difference in classification accuracy before and after shadow area restoration is statistically analyzed. When the accuracy improvement is less than 5%, the channel attention module of ShadowNet needs to be adjusted or the number of residual connection layers in the generator needs to be increased. For image areas that fail verification three times consecutively, a multi-temporal non-shadow reflectance database is automatically invoked for interpolation restoration, while the training dataset of the generative adversarial network is updated. This verification mechanism ensures that shadow restoration meets the needs of land supervision operations in three dimensions: visual quality, signal fidelity, and practical application effectiveness.

[0118] As a preferred embodiment, the specific implementation of this application's solution is as follows: After completing the shadow restoration processing of the high-resolution remote sensing image, the restoration result is input into the quality assessment module for verification. First, the structural similarity index (SSIM) is used to calculate the similarity between the restored area and the real ground object area. When the SSIM value is lower than 0.9, the image supplementation process is triggered, and multi-angle image data of the area are re-acquired by a drone. For restored areas where the peak signal-to-noise ratio (PSNR) does not reach 30dB, the training dataset of the generative adversarial network is expanded to include sample data of the same type of ground object under different lighting conditions. For areas where the classification accuracy improvement is less than 5%, a transfer learning mechanism is introduced during the model optimization stage. The pre-trained weights are jointly trained with the newly added samples, and the receptive field parameters of the discriminator are adjusted to enhance the local feature extraction capability. Restoration areas that fail verification are marked as pending units, and the system automatically calls non-shadow data from adjacent time nodes in the multi-temporal image library to perform secondary restoration.

[0119] Through the above technical solutions, this application effectively solves the problem of accumulated monitoring errors caused by unclear image restoration quality assessment standards. Multi-dimensional quantitative verification based on structural similarity, signal-to-noise ratio, and classification accuracy ensures the reliability of restored images in terms of texture detail, radiometric accuracy, and land cover identification, avoiding the risk of misjudgment due to the limitations of a single indicator. A differentiated processing strategy is adopted for areas that do not meet the standards, ensuring restoration efficiency while improving the spatial consistency of overall image data, providing a high-precision data foundation for subsequent dynamic land use monitoring.

[0120] In some of the solutions mentioned above in this application, there are problems with insufficient spectral consistency and ground feature structure between the generated image and the real scene during the shadow restoration process, which may lead to texture blurring or misclassification of ground feature categories in the restored image.

[0121] This application further proposes that the generator adopts a U-Net structure and the discriminator adopts a PatchGAN structure, with inputs including multi-temporal fused pixel values, shadow context spectrum and land cover category labels.

[0122] The generator employs a U-Net structure, fusing feature maps from the encoder and decoder via skip connections to preserve spatial details in shadow areas. The discriminator uses a PatchGAN structure to distinguish between real and fake images in local regions, enhancing the quality of local texture generation through a sliding window approach. The input multi-temporal fusion pixel values ​​integrate reflectance data from different time points, providing a temporal reference for shadow area restoration. The shadow context spectrum extracts spectral features from shadow boundaries and adjacent areas to help distinguish shadows from real land features. Land feature category labels constrain the category attributes of the generated image through semantic segmentation results, ensuring that the restored pixel values ​​are consistent with the actual land feature types.

[0123] Specifically, the generator's U-Net structure extracts multi-scale features through downsampling and restores spatial resolution by incorporating skip connections during upsampling, effectively avoiding detail loss during shadow restoration. The discriminator's PatchGAN structure divides the input image into multiple sub-regions and calculates the adversarial loss for each sub-region, forcing the generator to produce highly realistic textures at local scales. Multi-temporal fusion pixel values ​​provide non-shadow temporal reflectance references for shadow pixels by overlaying remote sensing images of the same area at different times, addressing the restoration difficulties caused by a lack of reference data in full-temporal shadow areas. Shadow context spectrum establishes association rules between shadow masks and real land features by calculating the spectral differences between shadow areas and adjacent areas, such as the spectral response differences between water shadows and building shadows. Land feature category labels are pre-acquired through a semantic segmentation network and concatenated with multi-temporal data at the generator input, guiding the restoration process to conform to land feature category characteristics; for example, maintaining a reasonable vegetation index is necessary after restoring farmland shadows. Therefore, by optimizing the structure and fusing multi-source data, the visual quality of shadow restoration and the accuracy of land cover classification are improved, meeting the technical indicators of SSIM≥0.9 and PSNR≥30dB.

[0124] As a preferred embodiment, the scheme of this application is implemented as follows: The generator of the generative adversarial network adopts a U-Net structure, and its encoder consists of five downsampling modules, each containing a convolutional layer and a ReLU activation function. The decoder consists of five upsampling modules, each containing a transposed convolutional layer and a skip connection operation. The discriminator adopts a PatchGAN structure, consisting of a four-layer convolutional network. Each convolutional layer is followed by a LeakyReLU activation function and a batch normalization operation. The output feature map is divided into 16×16 local receptive fields for real and fake discrimination. The input data includes multi-temporal fused pixel values, which are generated by fusing reflectance data from non-shaded phases within the past three months using a weighted average algorithm; the shadow context spectrum extracts the spectral feature vector of the 50×50 pixel region surrounding the shadow pixel through a sliding window; the land cover category label adopts a One-Hot encoding format, including eight land cover classification results such as cultivated land, buildings, and water bodies.

[0125] Through the above technical solutions, this application achieves end-to-end generation of high-precision shadow restoration. The U-Net structure effectively preserves spatial details of the image, and the PatchGAN discriminator enhances the realism of local textures. The introduction of multi-temporal fusion pixel values ​​solves the problem of missing data in shadow areas across all time periods. The fusion of shadow context spectra ensures spatial consistency between the restoration result and the spectral characteristics of surrounding land features, while the embedding of land feature category labels ensures that the reflectance differences of different land feature types are accurately modeled during the generation process.

[0126] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0127] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A land management system based on GIS system and big data analysis, characterized in that: The architecture includes a three-layer structure: data layer, support layer, and application layer. The data layer includes a multi-source data acquisition module and a GIS spatial database. The multi-source data acquisition module integrates remote sensing imagery, ground sensor data, land management business system data, socio-economic data, and internet crawler data. The GIS spatial database uses PostGIS or MongoDB geo-extended storage to store spatial and attribute data, and optimizes query efficiency through R-tree spatial indexing. The indexing formula is: O(d) = log d N, where d is the number of spatial dimensions and N is the total amount of data; The support layer includes a big data processing engine, a GIS analysis toolset, and a machine learning modeling platform. The big data processing engine is based on the Hadoop / Spark distributed computing framework for data cleaning, fusion, and standardization. The GIS analysis toolset integrates basic algorithms for spatial statistics, spatial interpolation, and change detection. The machine learning modeling platform supports the training and deployment of random forests, long short-term memory networks, and graph neural networks. The application layer includes an intelligent monitoring module, a planning assistance module, and a decision support module. The intelligent monitoring module covers farmland protection (monitoring of land occupation and compensation balance), illegal land use identification (screening of suspected land parcels), and ecological red line early warning (detection of changes in ecologically sensitive areas). The planning assistance module is used for multi-scheme simulation of territorial spatial planning and assessment of land resource carrying capacity. The decision support module is used to generate land consolidation priority maps and market trend reports.

2. A land management system based on GIS system and big data analysis according to claim 1, characterized in that: The remote sensing images include high-resolution satellite data from Gaofen-2 (1 meter / sub-meter level) or Sentinel-2 (10 meter level), and the ground sensor data includes soil moisture and ground temperature data collected hourly.

3. A land management system based on GIS system and big data analysis according to claim 1, characterized in that: The big data processing engine uses Hadoop / Spark to perform batch remote sensing image preprocessing, including 6S atmospheric correction, RPC model geometric correction, Gram-Schmidt Pan Sharpening image fusion, and removal of images with cloud coverage >10%.

4. A land management system based on GIS system and big data analysis according to claim 1, characterized in that: The illegal land use identification uses a random forest classifier to screen suspected land parcels, and the ecological red line early warning triggers a threshold alarm based on the change detection results.

5. A land management method based on GIS system and big data analysis, characterized in that: The management system according to any one of claims 1 to 4 includes the following steps: Step 1: Multi-source data acquisition initialization Collect multi-temporal remote sensing images and optimize the acquisition strategy based on the solar altitude angle: select periods with significant differences in solar altitude angle, collect multiple sets of images every quarter for key monitoring areas, and collect multiple sets of images every six months for general areas; Step 2: Multi-source data preprocessing and fusion A threshold-based segmentation algorithm is used to remove cloud noise from remote sensing images. The formula is used to mark cloud noise as |I(x,y)-μ|>T, where I(x,y) is the pixel value, μ is the neighborhood mean, and T is the preset threshold. For attribute data, missing values ​​are filled using rule-based validation (land parcel area > 0) and linear interpolation, employing Formula 2: x t =x t-1 +k·(x t+1 -x t-1 ), where x t x represents the missing value at time t. t-1 x t+1 For adjacent valid data, k is the interpolation weight; Spatial data from different sources are unified to the CGCS2000 coordinate system, and offsets are corrected using a polynomial transformation model, formula: Where (X) o ,Y o (X) represents the original coordinates. t ,Y t Let C be the target coordinates and C be the polynomial coefficient matrix. Step 3: Multi-dimensional analysis and modeling of land resources By fusing multi-temporal imagery, a random forest classifier is used to calculate a land use change probability map, employing the following formula: Where f i For the output of the i-th decision tree, w i σ is the weight, and σ is the Sigmoid function; Construct an AHP-entropy weighted combined weighting model to assess arable land quality, formula: Where S k For the normalized values ​​of indicators (organic matter content, slope, etc.), w j As the weights of the criterion layer (soil fertility, topographic conditions, etc.), w k The objective weights of the index layer are calculated using the entropy weight method; Based on LSTM neural network to predict land market trends, the formula is: P t =LSTM(P t-1 ,V,θ), where P t Let t be the land price, V be the plot ratio, and θ be the policy impact factor; Step 4: Intelligent Shadow Restoration of High-Resolution Remote Sensing Image a. An improved U-Net+attention mechanism model (ShadowNet) is used to output a shadow probability map, and a binary shadow mask is generated by morphological closing operation and Otsu thresholding method; b. Collect multi-temporal non-shadow reflectance values ​​for shadow pixels, and fill all-temporal shadows using Gaussian kernel spatial interpolation. c. Use a conditional generative adversarial network (ShadowGAN) to generate shadowless images, with the loss function being: in To combat the losses, In order to perceive loss, Cross-entropy loss; Step 5: Dynamic Monitoring and Intelligent Early Warning Set up an early warning rule base ("Farmland decreases by more than 5 mu for 3 consecutive months", "Construction land expansion rate exceeds the planning limit by 120%); The model calculation is triggered by real-time data streams. When the indicator exceeds the threshold, the abnormal area is marked on the GIS map and an early warning information is pushed.

6. A land management method based on GIS system and big data analysis according to claim 5, characterized in that: The multi-time image acquisition strategy in step 1 includes: selecting time periods with significant differences in solar altitude angle (9:00-10:30 am and 2:00-3:30 pm), acquiring 4-6 sets of images per quarter for key monitoring areas, and acquiring 2-3 sets per six months for general areas.

7. A land management method based on GIS system and big data analysis according to claim 5, characterized in that: In step 3, the farmland quality assessment uses AHP to determine the weights of the criterion layer (soil fertility, topographic conditions, irrigation capacity) and the entropy weight method to calculate the objective weights of the index layer (organic matter content, slope, well density).

8. A land management method based on GIS system and big data analysis according to claim 5, characterized in that: The shadow repair in step 4 needs to be verified by SSIM (≥0.9), PSNR (≥30dB) and classification accuracy (improvement ≥5%). For unqualified areas, supplementary images or model adjustments should be made.

9. A land management method based on GIS system and big data analysis according to claim 8, characterized in that: The ShadowNet network structure includes a ResNet-18 encoder with channel attention and a transposed convolutional decoder with spatial attention, and the loss function is Dice Loss + Cross-Entropy Loss.

10. A land management method based on GIS system and big data analysis according to claim 8, characterized in that: The generator of ShadowGAN adopts the U-Net structure, and the discriminator adopts the PatchGAN structure. The input includes multi-temporal fused pixel values, shadow context spectrum, and land cover category labels.

Images