Land natural resource dynamic supervision system and method

Abstract

The invention relates to the technical field of land resource management, in particular to a land natural resource dynamic supervision system and method. The method comprises the following steps: carrying out multi-source data acquisition and preprocessing on a land natural resource area to obtain an initial data set; performing intelligent land classification processing according to the initial data set to obtain a land classification map; performing change detection graph extraction according to the land classification graph to obtain a change detection graph; constructing a three-dimensional land resource model according to the change detection graph to obtain the three-dimensional land resource model; and performing early warning trigger processing according to the three-dimensional land resource model and the change detection graph, and pushing resource supervision information so as to realize a dynamic supervision task of the land natural resources. According to the land natural resource dynamic supervision method based on multi-source data, deep learning, time sequence analysis and three-dimensional modeling, automation, intellectualization and dynamics of land resource supervision are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of land resource management, and in particular to a system and method for dynamic monitoring of land natural resources. Background Art

[0002] Land natural resources are an important foundation for human survival and economic development. Their rational use and protection are related to the sustainable development of the ecological environment, food security, urbanization process and the steady operation of the social economy. Traditional land resource supervision methods mainly rely on manual field surveys and manual interpretation of remote sensing images. These methods have the following disadvantages: low efficiency and high cost: manual field surveys require a lot of manpower and material resources, are costly, and take a long time, making it difficult to conduct large-scale and high-frequency monitoring. Manual interpretation of remote sensing images also requires professionals, which is inefficient and easily affected by human factors, resulting in inconsistent results. It is difficult to dynamically supervise large areas of land resources: Due to the limitations of manpower and time, traditional methods are difficult to conduct real-time and dynamic monitoring of large areas of land resources. For some remote areas or areas with inconvenient transportation, it is even more difficult to conduct field surveys. Outdated data acquisition and processing methods: Traditional methods mainly rely on manual recording and measurement, and the data acquisition methods are backward, and the data accuracy and reliability are difficult to guarantee. Data processing also mainly relies on manual calculation and statistics, which is inefficient and prone to errors. Summary of the invention

[0003] Based on this, it is necessary to provide a land natural resource dynamic supervision system and method to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a land natural resource dynamic supervision method includes the following steps:

[0005] Step S1: collect multi-source data of land natural resource areas to obtain original remote sensing image data and auxiliary data sets; perform multi-source data preprocessing based on the original remote sensing image data and auxiliary data sets to obtain an initial data set;

[0006] Step S2: constructing a land classification model based on a convolutional neural network according to the initial data set to obtain a land classification model; performing intelligent land classification processing using the land classification model to obtain a land classification map;

[0007] Step S3: Perform LSTM-based change detection according to the land classification map to obtain LSTM change detection results; perform object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection map; perform change analysis on the preliminary change detection map to obtain a change analysis result; generate a change detection map according to the change analysis result to obtain a change detection map;

[0008] Step S4: constructing a digital elevation model according to the initial data set to obtain a digital elevation model; constructing a basic three-dimensional model according to the digital elevation model and the land classification map to obtain a basic three-dimensional model; refining and assigning attributes to the basic three-dimensional model according to the change detection map to obtain a three-dimensional land resource model;

[0009] Step S5: Dynamically monitor the indicators of the three-dimensional land resource model and the change detection map to obtain dynamic monitoring results; perform early warning trigger processing on the dynamic monitoring results according to the pre-built early warning rule library to obtain early warning information; push resource supervision information according to the early warning information to realize the dynamic supervision task of land natural resources.

[0010] The present invention obtains a high-quality initial data set containing rich surface information through multi-platform and multi-sensor data collection and preprocessing, providing a reliable data foundation for subsequent land classification, change detection and three-dimensional modeling, and improving the accuracy and reliability of the entire process. The deep learning technology is used to realize the automatic and high-precision classification of land resources, which is more efficient and accurate than traditional methods, especially for the classification effect of complex terrain areas. The application of multi-source data feature fusion and deep learning model effectively improves the classification accuracy and generalization ability. The change detection method based on LSTM and attention mechanism can effectively capture the time series pattern of land use type changes, focus on key time nodes, realize high-precision and automatic change detection, and further refine the change area in combination with object-oriented analysis method, improving the accuracy and reliability of the results. At the same time, spatial heterogeneity analysis can deeply understand the driving mechanism of land use change. By constructing a three-dimensional land resource model containing terrain, vegetation height, land use type and change information, the visualization and spatial analysis of land resources are realized, providing an intuitive tool for dynamic supervision and analysis of land resources. Model refinement and attribute assignment further improve the authenticity and information content of the model. Through dynamic monitoring of indicators, early warning triggering, report generation and information push, the automation, intelligence and dynamic supervision of land resources are realized, and abnormal changes in land resources can be discovered and warned in time, the supervision efficiency is improved, and support is provided for the decision-making of relevant departments. The data-driven early warning rule library is constructed to improve the accuracy and effectiveness of the early warning. Therefore, the present invention provides a method for dynamic supervision of land natural resources, which is based on multi-source data, deep learning, time series analysis and three-dimensional modeling. The method effectively solves the problems of low efficiency, high cost and difficulty in dynamic supervision of traditional land resource supervision methods, and realizes the automation, intelligence and dynamic supervision of land resources.

[0011] Preferably, step S1 comprises the following steps:

[0012] Step S11: acquiring remote sensing images of the land natural resource area to obtain original remote sensing image data;

[0013] Step S12: Collect auxiliary data on the land natural resource area to obtain an auxiliary data set;

[0014] Step S13: converting the data format of the original remote sensing image data and the auxiliary data set, and performing data registration to obtain a registered data set;

[0015] Step S14: performing atmospheric correction and radiation correction on the registration data set to obtain a corrected data set;

[0016] Step S15: performing data fusion and data trimming on the correction data set to obtain an initial data set.

[0017] The present invention provides a basic data source for subsequent land resource classification, change detection and three-dimensional modeling by acquiring high-resolution, multi-spectral remote sensing image data, ensuring data quality, thereby improving the accuracy and reliability of subsequent analysis. The application of Gaofen-2 satellite data ensures the high spatial resolution and rich spectral information of the image, and can more finely identify the characteristics of the ground objects. Collecting a variety of auxiliary data, including meteorological, topographic, soil and historical land use data, can supplement remote sensing image information, provide more comprehensive surface information, help improve land classification accuracy, and provide more abundant reference basis for change analysis and early warning. For example, meteorological data can be used for atmospheric correction, terrain data can be used for terrain correction and three-dimensional modeling, and soil data and historical land use data can be used to assist land classification and change analysis. Converting data from different sources and different formats into a unified format and coordinate system ensures the consistency and comparability of the data, laying the foundation for subsequent multi-source data fusion and analysis. Accurate registration ensures the spatial correspondence between different data and avoids analysis deviations caused by registration errors. Atmospheric correction eliminates the effects of atmospheric scattering and absorption on the reflectivity of ground objects, and radiation correction converts the DN values ​​recorded by the sensor into surface reflectivity, thereby improving the authenticity and reliability of image data and providing more accurate input data for subsequent land classification and change detection. This makes image data acquired at different times or by different sensors comparable. Through data fusion, the high spatial resolution of panchromatic images and the spectral information of multispectral images are combined to generate high-resolution multispectral images, which improves the accuracy of ground object identification. Data clipping removes redundant data outside the study area, reduces the amount of data processing, improves processing efficiency, and makes subsequent analysis more focused on the target area.

[0018] Preferably, step S2 comprises the following steps:

[0019] Step S21: construct a sample data set for the initial data set to obtain a training sample set and a verification sample set;

[0020] Step S22: extracting hyperspectral features from the training sample set using a convolutional neural network to obtain hyperspectral features; extracting LiDAR features from the training sample set using PointNet technology to obtain LiDAR features;

[0021] Step S23: performing feature fusion on the hyperspectral features and the LiDAR features to obtain fused features; constructing a classifier based on the fused features and the training sample set, and performing land classification model training to obtain a land classification model;

[0022] Step S24: extracting features and classifying the initial data set using the land classification model to obtain classification results;

[0023] Step S25: performing post-processing on the classification result to obtain a post-processed classification result;

[0024] Step S26: Use the verification sample set to evaluate the accuracy of the post-processing classification results to obtain a land classification map.

[0025] The present invention provides a reliable data basis for the training and verification of deep learning models by constructing high-quality, representative training sample sets and verification sample sets, ensuring that the model can effectively learn the features of land objects and improve the classification accuracy. The stratified random sampling method ensures the uniformity and representativeness of the samples and avoids the overfitting of the model to certain specific land object types. The spectral features of hyperspectral data are extracted by 3D-CNN, which can effectively capture the spatial and spectral dimension information in the hyperspectral data and improve the feature expression ability. At the same time, the three-dimensional structural features of LiDAR data are extracted by PointNet, which can effectively capture the geometric information in the point cloud data and avoid the information loss caused by the traditional method of rasterizing the point cloud data, thereby improving the accuracy of land object classification. Through feature fusion, the hyperspectral features and LiDAR features are combined, and the information of multi-source data is comprehensively utilized to improve the expression ability of land object features. The classifier is constructed using a fully connected neural network, and combined with the ReLU activation function and the Dropout layer, the nonlinear relationship between the fused features and the land use type can be effectively learned, and the generalization ability of the model can be improved to avoid overfitting. The application of the Adam optimizer and the cross entropy loss function can effectively train the model and improve the classification accuracy. The trained land classification model is used to extract features and classify the initial data set, which realizes the automation of land resource classification, greatly improves classification efficiency, and reduces the need for manual intervention. The deep learning model can automatically learn complex land features and has higher classification accuracy and efficiency than traditional methods. The conditional random field (CRF) is used to post-process the classification results. By considering the relationship between adjacent pixels, it can effectively remove noise and isolated points in the classification results, smooth the classification boundaries, and improve the spatial consistency of the classification results, thereby improving the accuracy and visual effect of the final classification map. The classification results are evaluated for accuracy using an independent validation sample set. By calculating indicators such as confusion matrix, overall accuracy, and Kappa coefficient, the performance of the classification model can be objectively evaluated, providing a reference basis for model optimization and improvement, and ensuring that the final generated land classification map has reliable accuracy.

[0026] Preferably, step S3 comprises the following steps:

[0027] Step S31: generating a multi-temporal land classification map according to the land classification map to obtain a multi-temporal land classification map; performing multi-temporal data alignment on the multi-temporal land classification map to obtain an aligned land classification map sequence;

[0028] Step S32: constructing time series features for the aligned land classification map sequence to obtain a time series feature data set;

[0029] Step S33: performing LSTM-based change detection according to the time series feature data set to obtain LSTM change detection results;

[0030] Step S34: performing object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection graph;

[0031] Step S35: Perform change analysis on the preliminary change detection graph to obtain a change analysis result; generate a change detection graph based on the change analysis result to obtain a change detection graph.

[0032] The present invention generates a multi-phase land classification map and performs accurate multi-phase data alignment to provide a consistent and comparable data basis for subsequent change detection analysis. Data alignment eliminates the geometric deviation between images of different phases and ensures the accuracy of change detection results. By constructing a time series feature data set and converting multi-phase land use information into time series data, the dynamic change process of land use types can be better captured, providing richer input information for LSTM-based change detection, and improving the accuracy and reliability of change detection. The extracted features such as the number of changes and duration can more comprehensively reflect the dynamic characteristics of land use changes. By analyzing the time series features using the LSTM network, the time series pattern of land use type changes can be effectively learned, the areas where changes have occurred can be identified, and the attention mechanism can be combined to enable the model to pay more attention to the moments that are important for change detection, thereby improving the accuracy of change detection. The bidirectional LSTM can consider past and future information at the same time, further improving the accuracy of detection. Combining the change detection results of LSTM with object-oriented analysis methods can effectively utilize the spatial information and object features of the image, refine the change area, eliminate the falsely detected pixels, make the boundary of the change area clearer, and improve the accuracy and spatial consistency of the change detection results. Through a detailed analysis of the change detection results, including statistical change area, analysis of change type and spatial distribution characteristics, and combined with GWR analysis of spatial heterogeneity, we can gain an in-depth understanding of the laws and driving factors of land use change. The final generated change detection map with legend can clearly and intuitively display land use change information, providing a scientific basis for land resource management and decision-making.

[0033] Preferably, step S33 includes the following steps:

[0034] Step S331: performing time series data preprocessing on the time series feature data set to obtain preprocessed time series data;

[0035] Step S332: performing forward LSTM feature extraction according to the preprocessed time series data to obtain forward LSTM features; performing reverse LSTM feature extraction according to the preprocessed time series data to obtain reverse LSTM features;

[0036] Step S333: performing bidirectional LSTM feature concatenation on the forward LSTM feature and the reverse LSTM feature to obtain a Bi-LSTM feature;

[0037] Step S334: Calculate the attention weight of the Bi-LSTM feature to obtain the attention weight; normalize the attention weight to obtain the normalized attention weight;

[0038] Step S335: performing attention weighting on the Bi-LSTM features and the normalized attention weights to obtain attention weighted features;

[0039] Step S336: performing time dimension feature aggregation on the attention weighted features to obtain aggregated features; performing full connection layer classification on the aggregated features to obtain an inactivated classification result;

[0040] Step S337: Perform Sigmoid activation on the inactivated classification results to obtain LSTM change detection results.

[0041] The present invention preprocesses the time series feature data set by the Z-score standardization method, eliminates the difference in dimensions between different features, improves the efficiency and stability of model training, and prevents the model from being biased towards features with larger values, so that the model can more effectively learn the change pattern in the time series data. Forward and reverse LSTM feature extraction are performed separately, and context information in the time series data can be captured from different angles. The forward LSTM focuses on the impact of information at past moments on the current moment, while the reverse LSTM focuses on the impact of information at future moments on the current moment, so as to more comprehensively understand the change law of time series data. The features extracted by the forward and reverse LSTMs are spliced, and the time series information in two directions is integrated to form a more expressive Bi-LSTM feature, which can more effectively express the complex change pattern in the time series data and improve the accuracy of change detection. By calculating the attention weight and normalizing it, different importance is given to features of different time steps, so that the model can pay more attention to the time steps that are critical to change detection, thereby improving the sensitivity and accuracy of the model and effectively filtering out irrelevant information. The normalized attention weights are weighted with the Bi-LSTM features to highlight the feature information of important time steps, so that the model can more effectively use key information for change detection and further improve the detection accuracy. The attention weighted features are aggregated in the time dimension, the time series information is integrated into a feature vector, and classified through the fully connected layer to achieve the discrimination of the changed and unchanged states. The Sigmoid activation function is used to convert the output of the fully connected layer into a probability value, making the results easier to interpret and analyze, facilitating the setting of thresholds for change detection, and can be directly used for subsequent processing and analysis.

[0042] Preferably, step S35 includes the following steps:

[0043] Step S351: determining the change type according to the preliminary change detection map and the multi-temporal land classification map to obtain a change type map;

[0044] Step S352: performing statistics on the change area according to the change type diagram to obtain change statistical data;

[0045] Step S353: performing spatial heterogeneity analysis on the change type map and the auxiliary data set to obtain spatial heterogeneity analysis results;

[0046] Step S354: generating a legend according to the change type map, the change statistical data and the spatial heterogeneity analysis result to obtain a legend;

[0047] Step S355: Add the legend to the preliminary change detection map to obtain a change detection map.

[0048] The present invention can accurately determine the specific change type of each change pixel by combining the preliminary change detection map and the multi-phase land classification map, such as the conversion of cultivated land to construction land, forest land to grassland, etc., so as to generate a more informative change type map, and provide a more detailed data basis for subsequent analysis and management. The area and proportion of each change type are counted, and quantitative information of land use change is provided, which can help users quickly understand the scale and importance of different change types, and provide data support for land resource management and decision-making. The spatial heterogeneity analysis of the change type map and auxiliary data set using the GWR model can reveal the influencing factors and spatial differences of land use change in different geographical locations. For example, the influence of factors such as terrain slope, distance from the road, and population density on different types of land use changes can be analyzed, so as to more deeply understand the driving mechanism of land use change and provide a scientific basis for formulating more effective land resource management strategies. A legend is generated according to the change type map, change statistics data and spatial heterogeneity analysis results, so that the change detection map is easier to understand and explain. A clear legend can help users quickly identify different change types, and understand their corresponding areas, proportions and main influencing factors, which improves the visualization effect of the results and the efficiency of information transmission. Adding a legend to the preliminary change detection map generates the final change detection map, making the map information more complete, facilitating users to intuitively understand and use the change detection results, and providing more complete data for subsequent report generation and information push.

[0049] Preferably, step S4 comprises the following steps:

[0050] Step S41: obtaining LiDAR data in the initial data set; performing point cloud processing on the LiDAR data to obtain point cloud data;

[0051] Step S42: constructing a digital elevation model according to the point cloud data to obtain a digital elevation model;

[0052] Step S43: extracting vegetation height from the point cloud data and the digital elevation model to obtain a vegetation height map;

[0053] Step S44: using the digital elevation model as the terrain basis and the land classification map as the texture map to construct a basic three-dimensional model to obtain a basic three-dimensional model;

[0054] Step S45: Refining the basic three-dimensional model according to the vegetation height map and the change detection map to obtain a refined model; assigning attributes to the refined model to obtain a three-dimensional land resource model.

[0055] The present invention obtains LiDAR data from the initial data set and performs point cloud processing, thereby providing a clean and reliable data basis for subsequent DEM construction and vegetation height extraction. The removal of noise points ensures the quality of point cloud data and improves the accuracy of subsequent processing. The point cloud data is used to construct a high-precision DEM, which provides accurate terrain information, lays a foundation for the construction of a three-dimensional model, and provides a reference datum for the extraction of vegetation height. The progressive encryption triangulation filtering algorithm can effectively distinguish between ground points and non-ground points, thereby improving the accuracy of DEM. By combining point cloud data and DEM to extract vegetation height, more accurate vegetation height information can be obtained, enriching the details of the three-dimensional land resource model, and providing important data for subsequent ecosystem analysis and land resource assessment. With DEM as the terrain basis and the land classification map as the texture map, a basic three-dimensional model is constructed, which realizes the visualization of land use types in three-dimensional space and provides users with intuitive land resource spatial distribution information. The basic three-dimensional model is refined using the vegetation height map and the change detection map, so that the model reflects the surface morphology more realistically and highlights the land use change area. By giving the model attribute information, such as land use type, area, vegetation height, change type, etc., the three-dimensional model not only has visualization functions, but also has data analysis and query functions, providing a powerful tool for dynamic supervision and analysis of land resources.

[0056] Preferably, step S5 comprises the following steps:

[0057] Step S51: extracting indicators from the three-dimensional land resource model and the change detection map, and performing indicator analysis to obtain a land resource indicator data set;

[0058] Step S52: Acquire historical land resource data and early warning target data; construct an early warning rule base according to the historical land resource data and the early warning target data to obtain an early warning rule base;

[0059] Step S53: dynamically monitor the land resource indicator data set to obtain dynamic monitoring results; perform early warning trigger processing on the dynamic monitoring results according to the early warning rule library to obtain early warning information;

[0060] Step S54: Generate a supervision report based on the early warning information and the land resource indicator data set to obtain a supervision report;

[0061] Step S55: Push information based on early warning information and supervision reports to achieve dynamic supervision of land and natural resources.

[0062] The present invention extracts key indicators from a three-dimensional land resource model and a change detection map, and performs statistical analysis to generate a land resource indicator data set, thereby providing a data basis for subsequent dynamic monitoring and early warning. By analyzing the change trend and spatial distribution characteristics of the indicators, the dynamic changes of land resources can be understood in a timely manner. Combining historical land resource data, early warning target data and expert knowledge, an early warning rule base is constructed to provide early warning thresholds and rules for dynamic supervision of land resources, thereby realizing the automation and intelligence of early warning. This enables the system to automatically identify potential land resource problems and issue early warnings in a timely manner. By regularly performing dynamic monitoring on the land resource indicator data set and comparing the monitoring results with the early warning rule base, abnormal changes in land resources can be discovered in a timely manner, and early warnings can be triggered to generate early warning information, thereby reminding relevant departments to take measures in a timely manner to effectively prevent illegal development and destruction of land resources. A supervision report is generated based on the early warning information and the land resource indicator data set, providing a comprehensive summary and analysis of the land resource status for relevant departments, making it easy to grasp the dynamic changes and early warning of land resources, and providing support for land resource management and decision-making. Pushing early warning information and supervision reports to relevant departments and personnel ensures timely communication of information, facilitates quick response and measures by relevant departments, improves the efficiency of land resource supervision, and promotes information sharing and openness and transparency. The application of the WebGIS platform further facilitates information acquisition and sharing, and improves public participation.

[0063] Preferably, step S52 includes the following steps:

[0064] Step S521: Acquire historical land resource data and early warning target data; perform expert knowledge labeling processing on the historical land resource data to obtain expert knowledge labeling data;

[0065] Step S522: extracting early warning features from historical land resource data and expert knowledge mark data to obtain early warning feature data;

[0066] Step S523: dividing the warning feature data into data sets to obtain a warning training data set;

[0067] Step S524: Select a machine learning model according to the warning target data to obtain a selected machine learning model; perform warning model training according to the selected machine learning model and the warning training data set to obtain a warning model;

[0068] Step S525: performing model evaluation and optimization on the early warning model according to the early warning training data set to obtain an optimized early warning model;

[0069] Step S526: extract warning rules using the optimized warning model, and construct a warning rule library to obtain a warning rule library.

[0070] The present invention obtains historical land resource data and early warning target data, and marks the historical data in combination with expert knowledge, so as to provide a marked data set with expert experience for subsequent early warning feature extraction and model training, so that the model can learn the expert's judgment standard on the land resource status, and improve the accuracy and reliability of the early warning model. Features related to early warning are extracted from historical land resource data, and are associated with expert knowledge marked data to construct early warning feature data, which provides effective input features for subsequent model training, so that the model can better learn the law of land resource changes and early warning mode. The early warning feature data is divided into a training data set and a test data set, which provides an independent data set for model training and evaluation, avoids overfitting, ensures the generalization ability of the model, and enables the model to be better applied to new data. According to the early warning target data and the early warning feature data, a suitable machine learning model is selected, and the training data set is used for model training, so that the model can learn the early warning mode in the historical data, and lays a foundation for the subsequent early warning rule extraction. The early warning model is evaluated and optimized using the test data set, which improves the prediction accuracy and stability of the model and ensures the reliability of the early warning rule. The optimized early warning model is used to extract early warning rules and build an early warning rule library, which realizes the generation of early warning rules based on data-driven, makes the early warning rules more in line with the actual situation, improves the accuracy and effectiveness of the early warning, and provides a basis for dynamic monitoring and early warning triggering.

[0071] Preferably, the present invention also provides a land natural resource dynamic supervision system for executing the land natural resource dynamic supervision method as described above, the land natural resource dynamic supervision system comprising:

[0072] The multi-source data acquisition module is used to collect multi-source data of land natural resource areas to obtain original remote sensing image data and auxiliary data sets; multi-source data preprocessing is performed based on the original remote sensing image data and auxiliary data sets to obtain an initial data set;

[0073] The intelligent land classification module is used to construct a land classification model based on a convolutional neural network according to the initial data set to obtain a land classification model; and to perform intelligent land classification processing using the land classification model to obtain a land classification map;

[0074] The map change detection module is used to perform LSTM-based change detection according to the land classification map to obtain the LSTM change detection result; perform object-oriented change detection and refinement on the LSTM change detection result to obtain a preliminary change detection map; perform change analysis on the preliminary change detection map to obtain the change analysis result; generate a change detection map according to the change analysis result to obtain a change detection map;

[0075] The three-dimensional land resource modeling module is used to construct a digital elevation model based on the initial data set to obtain a digital elevation model; construct a basic three-dimensional model based on the digital elevation model and the land classification map to obtain a basic three-dimensional model; refine the model and assign attributes to the basic three-dimensional model based on the change detection map to obtain a three-dimensional land resource model;

[0076] The dynamic supervision and early warning module is used to dynamically monitor the indicators of the three-dimensional land resource model and the change detection map to obtain dynamic monitoring results; to trigger early warning processing on the dynamic monitoring results according to the pre-built early warning rule library to obtain early warning information; and to push resource supervision information based on the early warning information to realize the dynamic supervision task of land natural resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A schematic diagram of the steps of a method for dynamic supervision of land natural resources;

[0078] Figure 2 Detailed implementation flow chart of step S3 in the present invention;

[0079] Figure 3 It is a schematic diagram of the detailed implementation steps of step S4 in the present invention.

[0080] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0081] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0082] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0083] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0084] To achieve this, please refer to Figures 1 to 3 , a dynamic supervision of land natural resources, including the following steps:

[0085] Step S1: collect multi-source data of land natural resource areas to obtain original remote sensing image data and auxiliary data sets; perform multi-source data preprocessing based on the original remote sensing image data and auxiliary data sets to obtain an initial data set;

[0086] Step S2: constructing a land classification model based on a convolutional neural network according to the initial data set to obtain a land classification model; performing intelligent land classification processing using the land classification model to obtain a land classification map;

[0087] Step S3: Perform LSTM-based change detection according to the land classification map to obtain LSTM change detection results; perform object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection map; perform change analysis on the preliminary change detection map to obtain a change analysis result; generate a change detection map according to the change analysis result to obtain a change detection map;

[0088] Step S4: constructing a digital elevation model according to the initial data set to obtain a digital elevation model; constructing a basic three-dimensional model according to the digital elevation model and the land classification map to obtain a basic three-dimensional model; refining and assigning attributes to the basic three-dimensional model according to the change detection map to obtain a three-dimensional land resource model;

[0089] Step S5: Dynamically monitor the indicators of the three-dimensional land resource model and the change detection map to obtain dynamic monitoring results; perform early warning trigger processing on the dynamic monitoring results according to the pre-built early warning rule library to obtain early warning information; push resource supervision information according to the early warning information to realize the dynamic supervision task of land natural resources.

[0090] In the embodiment of the present invention, reference Figure 1 As shown, it is a schematic diagram of the process flow of a dynamic supervision of land natural resources of the present invention. In this example, the dynamic supervision of land natural resources includes the following steps:

[0091] Step S1: collect multi-source data of land natural resource areas to obtain original remote sensing image data and auxiliary data sets; perform multi-source data preprocessing based on the original remote sensing image data and auxiliary data sets to obtain an initial data set;

[0092] In the embodiment of the present invention, multispectral remote sensing images are first acquired using the Gaofen-2 satellite, and meteorological, topographic (DEM), soil and historical land use data are collected as auxiliary data sets. Then, the data are format converted, registered, atmospheric corrected and radiometric corrected using GDAL and ENVI tools. Finally, the image fusion is performed using the Gram-Schmidt method, and all data are clipped to the study area to generate an initial data set containing multi-source information.

[0093] Step S2: constructing a land classification model based on a convolutional neural network according to the initial data set to obtain a land classification model; performing intelligent land classification processing using the land classification model to obtain a land classification map;

[0094] In the embodiment of the present invention, training and validation sample sets are constructed based on high-resolution images and field survey data. Then, 3D-CNN is used to extract hyperspectral features, PointNet is used to extract LiDAR features, and the two are spliced ​​for feature fusion. Next, a fully connected neural network classifier containing two hidden layers and a Dropout layer is constructed, and trained using cross entropy loss and Adam optimizer. The initial data set is classified using the trained model, and CRF is used for post-processing to finally generate a land classification map, and the validation set is used for accuracy evaluation.

[0095] Step S3: Perform LSTM-based change detection according to the land classification map to obtain LSTM change detection results; perform object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection map; perform change analysis on the preliminary change detection map to obtain a change analysis result; generate a change detection map according to the change analysis result to obtain a change detection map;

[0096] In an embodiment of the present invention, multi-phase land classification maps are registered and aligned, and then a time series feature data set is constructed, including land use type codes, number of changes, and duration. Next, a bidirectional LSTM network combined with an attention mechanism is used for change detection, and the features are input into the network and classified to obtain LSTM change detection results. Subsequently, eCognition software is used for object-oriented change detection and refinement to generate a preliminary change detection map. Finally, a change analysis is performed, including statistical change area, spatial distribution characteristics, and GWR analysis of spatial heterogeneity, and finally a final change detection map containing a legend is generated.

[0097] Step S4: constructing a digital elevation model according to the initial data set to obtain a digital elevation model; constructing a basic three-dimensional model according to the digital elevation model and the land classification map to obtain a basic three-dimensional model; refining and assigning attributes to the basic three-dimensional model according to the change detection map to obtain a three-dimensional land resource model;

[0098] In an embodiment of the present invention, PDAL is used to preprocess and point cloud the LiDAR data, and then based on the point cloud data, a DEM is generated using a progressively encrypted triangulated network filtering algorithm and an inverse distance weighted interpolation method. Next, the vegetation height map is extracted by calculating the elevation difference between the non-ground point and the DEM. Then, in Blender, based on the DEM, the land classification map is used as a texture to construct a basic three-dimensional model. Finally, the model is refined using the vegetation height map and the change detection map, the display of vegetation height and change area is added, and the model object is given attributes such as land use type, area, height, and change type to generate the final three-dimensional land resource model.

[0099] Step S5: Dynamically monitor the indicators of the three-dimensional land resource model and the change detection map to obtain dynamic monitoring results; perform early warning trigger processing on the dynamic monitoring results according to the pre-built early warning rule library to obtain early warning information; push resource supervision information according to the early warning information to achieve the dynamic supervision task of land natural resources;

[0100] In the embodiment of the present invention, vegetation coverage, building height, change area and other indicators are extracted from the three-dimensional model and the change detection map, and statistical analysis is performed to generate a land resource indicator data set. Then, historical land resource data and early warning target data are obtained, and an early warning rule base is constructed in combination with expert knowledge. Next, the indicator data set is dynamically monitored regularly, and the results are matched with the early warning rule base to trigger an early warning and generate early warning information. Finally, a supervision report is generated based on the early warning information and indicator data, and the early warning information and report are pushed to relevant departments and personnel via email, text messages and WebGIS platform to realize dynamic supervision of land natural resources.

[0101] Preferably, step S1 comprises the following steps:

[0102] Step S11: acquiring remote sensing images of the land natural resource area to obtain original remote sensing image data;

[0103] Step S12: Collect auxiliary data on the land natural resource area to obtain an auxiliary data set;

[0104] Step S13: converting the data format of the original remote sensing image data and the auxiliary data set, and performing data registration to obtain a registered data set;

[0105] Step S14: performing atmospheric correction and radiation correction on the registration data set to obtain a corrected data set;

[0106] Step S15: performing data fusion and data trimming on the correction data set to obtain an initial data set.

[0107] In an embodiment of the present invention, for a part of Beijing located at 116° to 117° east longitude and 39° to 40° north latitude, the Gaofen-2 satellite (GF-2) is used to obtain multispectral remote sensing image data covering the area. The image acquisition time is 10:30 am on September 15, 2023, when the weather is clear and the cloud cover is less than 5%. The panchromatic sensor and multispectral sensor carried by the GF-2 satellite obtain image data with spatial resolutions of 1 meter and 4 meters, respectively. The panchromatic band ranges from 450 to 900 nanometers, and the multispectral band includes blue light (450-520 nanometers), green light (520-590 nanometers), red light (630-690 nanometers) and near infrared (770-890 nanometers). The acquired original remote sensing image data is stored in GeoTIFF format and contains image metadata information, such as imaging time, satellite attitude angle, etc.

[0108] Obtain meteorological data corresponding to the study area of ​​Beijing at 10:30 am on September 15, 2023 from the National Meteorological Information Center, including temperature, air pressure, humidity, wind speed, wind direction, etc. Download the 30-meter resolution DEM data of the study area from the geospatial data cloud platform, which is generated based on the ASTER GDEM product. Obtain soil type data for the study area from the Resources and Environmental Science Data Center, which uses the classification system of the World Soil Resources Reference Library (WRB). Obtain the 2022 land use status data of the study area from the Ministry of Natural Resources, which is classified using the national land use classification standard. All auxiliary data are converted to GeoTIFF format and maintain the same geographic coordinate system and projection method as the remote sensing image data.

[0109] All original remote sensing image data and auxiliary data were converted into a unified GeoTIFF format using the GDAL library. The GF-2 panchromatic image was selected as the reference image, and the image registration tool in the ENVI software was used to perform geometric precision correction on the multispectral image. Image registration was achieved by control points and polynomial transformation, and the registration accuracy was controlled within 0.5 pixels. The auxiliary data was reprojected to the same WGS 84 coordinate system and UTM projection (Zone 50N) as the reference image. All registered data was stored in GeoTIFF format with the same pixel size and geographic range.

[0110] The Flaash model is used to perform atmospheric correction on the GF-2 multispectral image to eliminate the effects of atmospheric scattering and absorption on the reflectivity of the ground objects. The input parameters of the Flaash model include image acquisition time, sensor parameters, atmospheric model parameters, etc. After atmospheric correction, the DN value of the image is converted to surface reflectivity. The DN value of the multispectral image is converted to radiance value using the radiation calibration tool in the ENVI software, and then the radiance value is converted to surface reflectivity using the surface reflectance conversion formula. The corrected image data is stored in the form of surface reflectivity and the band information is retained.

[0111] The Gram-Schmidt method was used to fuse the GF-2 panchromatic image and the multispectral image to generate a multispectral image with a spatial resolution of 1 meter. This method uses the high spatial resolution information of the panchromatic band to improve the spatial resolution of the multispectral image while retaining the multispectral information. Using the vector boundary of the study area, the GDAL library was used to crop the fused multispectral image, DEM data, soil type data, and land use status data to remove data outside the study area. The cropped data constituted the initial data set and was stored in GeoTIFF format, providing a data basis for subsequent land resource classification and change detection.

[0112] Preferably, step S2 comprises the following steps:

[0113] Step S21: construct a sample data set for the initial data set to obtain a training sample set and a verification sample set;

[0114] Step S22: extracting hyperspectral features from the training sample set using a convolutional neural network to obtain hyperspectral features; extracting LiDAR features from the training sample set using PointNet technology to obtain LiDAR features;

[0115] Step S23: performing feature fusion on the hyperspectral features and the LiDAR features to obtain fused features; constructing a classifier based on the fused features and the training sample set, and performing land classification model training to obtain a land classification model;

[0116] Step S24: extracting features and classifying the initial data set using the land classification model to obtain classification results;

[0117] Step S25: performing post-processing on the classification result to obtain a post-processed classification result;

[0118] Step S26: Use the verification sample set to evaluate the accuracy of the post-processing classification results to obtain a land classification map.

[0119] In an embodiment of the present invention, based on the visual interpretation method, a representative sample area is selected from the high-resolution remote sensing image in the initial data set, and the sample area is manually annotated in combination with the field survey data and the high-precision map to determine its land use type. The study area is divided into six land use types: cultivated land, forest land, grassland, construction land, water body and unused land. To ensure the uniformity and representativeness of the sample, a stratified random sampling method is adopted to extract no less than 500 sample points from each land use type. Each sample point corresponds to a 3×3 pixel area, and the spectral value of the multispectral image in the area, the height value and density value of the LiDAR point cloud and other features are extracted to form a sample data set. The sample data set is randomly divided into a training sample set and a verification sample set in a ratio of 7:3 for training and verification of the deep learning model.

[0120] A three-dimensional convolutional neural network (3D-CNN) was constructed for hyperspectral feature extraction. The 3D-CNN contains three convolutional layers with kernel sizes of 3×3×7, 3×3×5, and 3×3×3, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The feature map is downsampled using a maximum pooling layer. The hyperspectral data in the training sample set are input into the 3D-CNN to obtain the hyperspectral features. The pre-trained PointNet model is used to extract features from the LiDAR data in the training sample set. The LiDAR point cloud data with a radius of 1 meter around each sample point is input into the PointNet network to obtain the feature vector of each point. Then, the feature vectors of all points are subjected to a maximum pooling operation to obtain the LiDAR feature vector representing the sample area.

[0121] Hyperspectral features and LiDAR features are concatenated to obtain fused features. A fully connected neural network is constructed as a classifier, which contains two hidden layers with 256 and 128 neurons respectively. Each hidden layer is followed by a ReLU activation function and a Dropout layer (Dropout rate = 0.5). The fused features are input into the classifier, and the output layer contains six neurons, corresponding to six land use types. The land classification model is trained using the cross entropy loss function and the Adam optimizer. The learning rate is set to 0.001, the batch size is set to 64, and the number of training rounds is set to 100. During the training process, the validation sample set is used to monitor the model performance, and the early stopping strategy is used to prevent overfitting.

[0122] The hyperspectral data and LiDAR data in the initial dataset are input into the trained 3D-CNN and PointNet models respectively to extract the hyperspectral features and LiDAR features. The extracted features are concatenated to obtain fused features. The fused features are input into the trained classifier to obtain the classification probability of each pixel. The class with the highest probability is selected as the land use type of the pixel to obtain the classification result.

[0123] The classification results are post-processed using conditional random fields (CRF) to improve the spatial consistency of the classification results. The CRF model takes into account the interaction between adjacent pixels and can effectively remove noise and isolated points in the classification results. The parameters of the CRF model are optimized using a grid search method.

[0124] The validation sample set is used to evaluate the accuracy of the post-processing classification results. The confusion matrix, overall accuracy, Kappa coefficient, user accuracy, and producer accuracy are calculated to evaluate the classification accuracy. The post-processing classification results and accuracy evaluation results are visualized to generate a land classification map. Each pixel in the land classification map is assigned a corresponding land use type label and represented by a different color.

[0125] Preferably, step S3 comprises the following steps:

[0126] Step S31: generating a multi-temporal land classification map according to the land classification map to obtain a multi-temporal land classification map; performing multi-temporal data alignment on the multi-temporal land classification map to obtain an aligned land classification map sequence;

[0127] Step S32: constructing time series features for the aligned land classification map sequence to obtain a time series feature data set;

[0128] Step S33: performing LSTM-based change detection according to the time series feature data set to obtain LSTM change detection results;

[0129] Step S34: performing object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection graph;

[0130] Step S35: Perform change analysis on the preliminary change detection graph to obtain a change analysis result; generate a change detection graph based on the change analysis result to obtain a change detection graph.

[0131] As an example of the present invention, refer to Figure 2 As shown, in this example, step S3 includes:

[0132] Step S31: generating a multi-temporal land classification map according to the land classification map to obtain a multi-temporal land classification map; performing multi-temporal data alignment on the multi-temporal land classification map to obtain an aligned land classification map sequence;

[0133] In an embodiment of the present invention, land classification maps for the three years of 2021, 2022 and 2023 are obtained. The land classification maps of these three years are generated using the same technical scheme and parameter settings as step S2 to ensure data consistency. The GDAL library is used to perform multi-phase data alignment on the land classification maps of the three years. Taking the land classification map of 2023 as the reference image, the land classification maps of 2021 and 2022 are aligned to the land classification map of 2023. A polynomial transformation model is used for geometric correction, and the nearest neighbor resampling method is used for pixel value interpolation. The alignment accuracy is controlled within 0.5 pixels. The aligned land classification map sequence is stored in GeoTIFF format and has the same geographic coordinate system, projection mode and pixel size.

[0134] Step S32: constructing time series features for the aligned land classification map sequence to obtain a time series feature data set;

[0135] In an embodiment of the present invention, a time series feature is constructed for the aligned land classification map sequence obtained in step S31. For each pixel, its land use type code in 2021, 2022 and 2023 is extracted to form a time series of length 3. In addition, the number of times the land use type of each pixel changes between the three years and the duration of each land use type are calculated. These features are combined into a feature vector to form a time series feature dataset. The time series feature dataset is stored in NumPy array format with dimensions [image height, image width, feature dimension], where the feature dimension is the number of land use type codes plus the number of changes and duration.

[0136] Step S33: performing LSTM-based change detection according to the time series feature data set to obtain LSTM change detection results;

[0137] In an embodiment of the present invention, a bidirectional LSTM (Bi-LSTM) network is constructed, and change detection is performed in combination with an attention mechanism. The Bi-LSTM network includes two LSTM layers, which process forward and reverse time series data respectively. The number of hidden units in each LSTM layer is set to 128. The attention mechanism is used to assign different weights to features at different time steps. The time series feature data set is input into the Bi-LSTM network to obtain a feature vector for each pixel. The feature vector is then input into a fully connected layer for classification, and the output layer includes two neurons, representing the two states of change and no change, respectively. The output value is converted into a probability using a Sigmoid activation function. Pixels with a probability value greater than 0.5 are marked as changed, and an LSTM change detection result is obtained.

[0138] Step S34: performing object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection graph;

[0139] In an embodiment of the present invention, eCognition software is used to perform object-oriented change detection and refinement on the LSTM change detection results. First, multi-scale segmentation is performed based on multispectral images to segment the image into several homogeneous objects. Then, the average spectral value, shape index, texture features, etc. of each object are calculated. According to the LSTM change detection results, the object where the pixels marked as changed are located is marked as a changed object. For the object where the pixels not marked as changed are located, if the similarity between the object and the changed object is higher than a preset threshold (for example, 0.8), the object is also marked as a changed object. Finally, all the changed objects are merged to generate a preliminary change detection map.

[0140] Step S35: performing change analysis on the preliminary change detection graph to obtain a change analysis result; generating a change detection graph according to the change analysis result to obtain a change detection graph;

[0141] In an embodiment of the present invention, a change analysis is performed on the preliminary change detection map. The area of ​​conversion between different land use types, the frequency of change occurrence, and the spatial distribution characteristics are counted. The geographically weighted regression (GWR) model is used to analyze the spatial heterogeneity of land use change and its influencing factors, such as terrain slope, distance from the road, population density, etc. A legend is generated based on the change analysis results, including the color, area, and description information of different change types. The legend is added to the preliminary change detection map to generate a final change detection map. The final change detection map is stored in GeoTIFF format and contains attribute information such as change type and change area.

[0142] Preferably, step S33 includes the following steps:

[0143] Step S331: performing time series data preprocessing on the time series feature data set to obtain preprocessed time series data;

[0144] Step S332: performing forward LSTM feature extraction according to the preprocessed time series data to obtain forward LSTM features; performing reverse LSTM feature extraction according to the preprocessed time series data to obtain reverse LSTM features;

[0145] Step S333: performing bidirectional LSTM feature concatenation on the forward LSTM feature and the reverse LSTM feature to obtain a Bi-LSTM feature;

[0146] Step S334: Calculate the attention weight of the Bi-LSTM feature to obtain the attention weight; normalize the attention weight to obtain the normalized attention weight;

[0147] Step S335: performing attention weighting on the Bi-LSTM features and the normalized attention weights to obtain attention weighted features;

[0148] Step S336: performing time dimension feature aggregation on the attention weighted features to obtain aggregated features; performing full connection layer classification on the aggregated features to obtain an inactivated classification result;

[0149] Step S337: Perform Sigmoid activation on the inactivated classification results to obtain LSTM change detection results.

[0150] In an embodiment of the present invention, the Z-score standardization method is used to preprocess the time series feature data set. For each feature, its mean and standard deviation are calculated on all samples. Then, the mean is subtracted from each eigenvalue and divided by the standard deviation to obtain the standardized eigenvalue. The preprocessed time series data is stored in NumPy array format, and its dimensions remain unchanged, still [image height, image width, feature dimension], but the range of eigenvalues ​​is standardized to a mean of 0 and a standard deviation of 1. This step eliminates the dimensional differences between different features, which is conducive to the training and convergence of the model.

[0151] Create two LSTM layers for forward and reverse LSTM feature extraction. The number of hidden units in each LSTM layer is set to 128, and the input feature dimension is the same as the feature dimension of the time series feature dataset. Input the preprocessed time series data into the forward and reverse LSTM layers respectively. The forward LSTM layer processes the data in the order of the time series, while the reverse LSTM layer processes the data in the reverse order of the time series. The outputs of both LSTM layers are sequence features with the dimensions [image height, image width, time step, number of hidden units].

[0152] The forward LSTM features and the reverse LSTM features are concatenated on the last dimension (the number of hidden units). The concatenated Bi-LSTM feature dimensions are [image height, image width, time step, 2*number of hidden units], where 2*number of hidden units is 256. This step fuses the feature information extracted by the forward and reverse LSTMs to better capture the contextual information in the time series data.

[0153] Create a linear layer with an input dimension of 2*number of hidden units (256) and an output dimension of 1. Input the Bi-LSTM features into the linear layer to get the attention weights for each time step with the dimension [image height, image width, time step, 1]. Use the Softmax function to normalize the attention weights to ensure that the sum of the attention weights for each pixel is 1. The dimension of the normalized attention weights is the same as the dimension of the attention weights.

[0154] Multiply the Bi-LSTM features and the normalized attention weights to get the attention-weighted features. The dimensions of the attention-weighted features are the same as the Bi-LSTM feature dimensions, i.e. [image height, image width, time step, 2*number of hidden units]. This step incorporates the attention mechanism into the Bi-LSTM network, allowing the model to pay more attention to the moments that are important for change detection.

[0155] Perform an average pooling operation on the attention weighted features in the time dimension to obtain an aggregated feature with a dimension of [image height, image width, 2*number of hidden units]. Create a fully connected layer with an input dimension of 2*number of hidden units (256) and an output dimension of 2, corresponding to the changed and unchanged states. Input the aggregated feature into the fully connected layer to obtain an inactive classification result with a dimension of [image height, image width, 2].

[0156] The Sigmoid activation function is used to activate the inactivated classification results to obtain the LSTM change detection results, whose dimensions are the same as the inactivated classification results, that is, [image height, image width, 2]. Each pixel has two probability values, representing the probability of change and no change. This step converts the output of the model into probability values ​​for subsequent discrimination and analysis.

[0157] Preferably, step S35 includes the following steps:

[0158] Step S351: determining the change type according to the preliminary change detection map and the multi-temporal land classification map to obtain a change type map;

[0159] Step S352: performing statistics on the change area according to the change type diagram to obtain change statistical data;

[0160] Step S353: performing spatial heterogeneity analysis on the change type map and the auxiliary data set to obtain spatial heterogeneity analysis results;

[0161] Step S354: generating a legend according to the change type map, the change statistical data and the spatial heterogeneity analysis result to obtain a legend;

[0162] Step S355: Add the legend to the preliminary change detection map to obtain a change detection map.

[0163] In an embodiment of the present invention, a preliminary change detection map and a multi-temporal land classification map are used to determine the specific change type of each pixel that has changed. The change type is determined by comparing the land use type code of each pixel in the land classification map of different years. For example, if a pixel is cultivated land (coded as 1) in 2021 and becomes construction land (coded as 4) in 2023, the change type code of the pixel is 14, indicating that the cultivated land is converted to construction land. A unique code is defined for all possible land use type conversion combinations. Each changed pixel is assigned a corresponding change type code to generate a change type map. The change type map is stored in GeoTIFF format, and each pixel value represents its change type code. Pixels that have not changed are assigned a value of 0.

[0164] Based on the change type map, count the number of pixels for each change type. Combined with the spatial resolution of the image (for example, the resolution of the GF-2 image is 4 meters), convert the number of pixels into actual area. For example, one pixel represents 4 meters * 4 meters = 16 square meters, so 100 pixels represent 1600 square meters. Store the area statistics of each change type in a table, including the change type code, change type description (for example, farmland converted to construction land), change area, and proportion of the total change area. The change statistics are saved in CSV file format for subsequent analysis and report generation.

[0165] The geographically weighted regression (GWR) model was used to analyze the spatial heterogeneity of the change type map and the auxiliary dataset. The occurrence of each change type in the change type map was used as the dependent variable, and the terrain slope, distance from the road, population density, etc. in the auxiliary dataset were used as independent variables. The GWR model can capture the differences in the impact of independent variables on the dependent variable in different geographical locations. The analysis results include the regression coefficient of each independent variable at different locations, the local R-square value, etc. These results are stored in raster data format, and each pixel contains a set of regression coefficients and statistical indicators. In addition, the global statistical results of the GWR model (for example, the global R-square value, AIC value) are also recorded to evaluate the overall goodness of fit of the model.

[0166] Each change type is assigned a unique color according to the change type code defined in the change type map. Combined with the change statistics, the description, area and proportion of each change type are displayed in the legend. In addition, based on the results of the spatial heterogeneity analysis, information about the main influencing factors can be added to the legend, for example, which factors have a greater impact on a specific type of land use change. The legend is stored in a vector graphics format (such as SVG) or embedded in an image file (such as PNG).

[0167] Add a legend to the preliminary change detection map to generate the final change detection map. Place the legend in an appropriate location on the image, such as the lower right corner or on the left. Make sure the legend is clear and easy to understand, with an appropriate font size and colors that correspond to the colors in the change type map. The final change detection map is stored in GeoTIFF or other image format, containing all the change information as well as the legend, making it easier for users to understand and interpret the change detection results.

[0168] Preferably, step S4 comprises the following steps:

[0169] Step S41: obtaining LiDAR data in the initial data set; performing point cloud processing on the LiDAR data to obtain point cloud data;

[0170] Step S42: constructing a digital elevation model according to the point cloud data to obtain a digital elevation model;

[0171] Step S43: extracting vegetation height from the point cloud data and the digital elevation model to obtain a vegetation height map;

[0172] Step S44: using the digital elevation model as the terrain basis and the land classification map as the texture map to construct a basic three-dimensional model to obtain a basic three-dimensional model;

[0173] Step S45: Refining the basic three-dimensional model according to the vegetation height map and the change detection map to obtain a refined model; assigning attributes to the refined model to obtain a three-dimensional land resource model.

[0174] As an example of the present invention, refer to Figure 3 As shown, in this example, step S4 includes:

[0175] Step S41: obtaining LiDAR data in the initial data set; performing point cloud processing on the LiDAR data to obtain point cloud data;

[0176] In an embodiment of the present invention, LiDAR data is read from an initial data set, which is stored in LAS format and includes the three-dimensional coordinates (x, y, z), intensity values ​​and other related attributes of each laser point. PDAL (Point Data Abstraction Library) is used to read and preprocess the LAS file. Noise points, such as isolated points and low-intensity points, are removed. Points with intensity values ​​below 100 and isolated points that are too far from other points are filtered out by setting a specific threshold. The preprocessed LiDAR data is converted into point cloud data, stored in XYZ format, and includes the three-dimensional coordinate information of each point. The point cloud data is saved in a text file or PCD (Point Cloud Data) format for subsequent digital elevation model construction and vegetation height extraction.

[0177] Step S42: constructing a digital elevation model according to the point cloud data to obtain a digital elevation model;

[0178] In an embodiment of the present invention, a digital elevation model (DEM) is constructed using point cloud data. A progressively encrypted triangulated network filtering algorithm is used to classify ground points in point cloud data, and the point cloud data is divided into ground points and non-ground points. DEM is generated using ground point data. The inverse distance weighted interpolation method is used to generate DEM with a resolution of 1 meter. During the interpolation process, the search radius is set to 5 meters, and the elevation value of each pixel is determined by the weighted average of the ground points within a 5-meter range around it. The generated DEM is stored in GeoTIFF format and contains elevation value information.

[0179] Step S43: extracting vegetation height from the point cloud data and the digital elevation model to obtain a vegetation height map;

[0180] In the embodiment of the present invention, point cloud data and DEM are used to extract vegetation height. For each pixel, the average elevation value of non-ground points within the pixel range is calculated. The average elevation value of the non-ground point is subtracted from the DEM elevation value of the corresponding pixel to obtain the vegetation height of the pixel. The vegetation height value is rasterized to generate a vegetation height map with the same resolution as the DEM, which is 1 meter. The vegetation height map is stored in GeoTIFF format, and the pixel value represents the vegetation height.

[0181] Step S44: using the digital elevation model as the terrain basis and the land classification map as the texture map to construct a basic three-dimensional model to obtain a basic three-dimensional model;

[0182] In an embodiment of the present invention, Blender software is used to construct a basic three-dimensional model. DEM is imported into Blender as a reference surface for the three-dimensional terrain. A land classification map is applied to the DEM surface as a texture map. Different colors or textures are assigned to different areas according to the land use type. For example, cultivated land is set to green, forest land is set to dark green, construction land is set to gray, water bodies are set to blue, and unused land is set to brown. The generated basic three-dimensional model is saved in OBJ or FBX format, including three-dimensional terrain and land use type information.

[0183] Step S45: Refining the basic three-dimensional model according to the vegetation height map and the change detection map to obtain a refined model; assigning attributes to the refined model to obtain a three-dimensional land resource model;

[0184] In an embodiment of the present invention, in Blender, the height of the vegetation area in the basic three-dimensional model is adjusted according to the vegetation height map. The vegetation height map is applied to the model as a displacement map to make the vegetation area present a three-dimensional effect. According to the change detection map, the changed area is highlighted. For example, the changed area is set to red, and labels are added to indicate the change type and area. Attribute information is assigned to each object in the model, including land use type, area, vegetation height, change type, etc. The attribute information is stored in a table form and associated with the three-dimensional model. The final three-dimensional land resource model is saved in OBJ or FBX format and contains detailed attribute information for subsequent dynamic supervision and analysis.

[0185] Preferably, step S5 comprises the following steps:

[0186] Step S51: extracting indicators from the three-dimensional land resource model and the change detection map, and performing indicator analysis to obtain a land resource indicator data set;

[0187] Step S52: Acquire historical land resource data and early warning target data; construct an early warning rule base according to the historical land resource data and the early warning target data to obtain an early warning rule base;

[0188] Step S53: dynamically monitor the land resource indicator data set to obtain dynamic monitoring results; perform early warning trigger processing on the dynamic monitoring results according to the early warning rule library to obtain early warning information;

[0189] Step S54: Generate a supervision report based on the early warning information and the land resource indicator data set to obtain a supervision report;

[0190] Step S55: Push information based on early warning information and supervision reports to achieve dynamic supervision of land and natural resources.

[0191] In an embodiment of the present invention, indicators such as vegetation coverage, building height, and building area are extracted from a three-dimensional land resource model. Each object in the model is traversed using Python scripts and Blender API to calculate its surface area and volume. According to the object attribute information, the area and proportion of different land use types are counted. Indicators such as land use change area and change type are extracted from the change detection map. The change detection map is read using the GDAL library, the number of pixels of each change type is counted, and it is converted into area. The extracted indicators are stored in a land resource indicator dataset, which uses a CSV format and contains information such as indicator name, indicator value, and timestamp. Statistical analysis is performed on the indicator data, such as calculating the change trend of the indicator, spatial distribution characteristics, etc.

[0192] Obtain historical land resource data such as land use change data, meteorological data, and socio-economic data for the past five years. Obtain this data from relevant departments or public data platforms and preprocess it to make it consistent with the current data format. The warning target data includes preset land use change thresholds, vegetation coverage thresholds, etc. For example, the threshold for cultivated land reduction is set to 100 hectares per year, and the threshold for vegetation coverage reduction is set to 5%. Build a warning rule library based on historical land resource data and warning target data. The warning rule library adopts the form of if-then rules. For example, if the cultivated land reduction area exceeds the threshold, the warning is triggered. The warning rule library is stored in JSON format and contains rule conditions, warning levels, warning information, etc.

[0193] Dynamically monitor the land resource indicator dataset regularly (for example, daily or weekly). Compare the latest indicator data with historical data to analyze the changing trend of the indicator. Use Python scripts to read the land resource indicator dataset and the early warning rule library. Match the dynamic monitoring results with the rules in the early warning rule library. If an indicator exceeds the preset threshold or meets the early warning rule, an early warning is triggered and an early warning message is generated. The early warning information contains information such as the warning type, warning level, warning location, time, and reason, and is stored in JSON format.

[0194] Generate a supervision report based on the early warning information and land resource indicator dataset. The supervision report contains information such as the overall status of land resources, change trends, and early warning conditions. Use Python's report generation library, such as ReportLab, to automatically generate supervision reports. The report contains visualization elements such as charts and maps, as well as detailed indicator data and analysis results. The supervision report is saved in PDF format and contains timestamp information.

[0195] The generated warning information and supervision reports are pushed to relevant departments and personnel via email, SMS, etc. Use Python's email sending library, such as smtplib, to automatically send emails. At the same time, the warning information and supervision reports are uploaded to the WebGIS-based platform for online viewing and downloading by relevant departments and the public. The platform is built using open source libraries such as Leaflet or OpenLayers, and provides data query, visualization, analysis and other functions to realize the online release and sharing of dynamic land and natural resource supervision information.

[0196] Preferably, step S52 includes the following steps:

[0197] Step S521: Acquire historical land resource data and early warning target data; perform expert knowledge labeling processing on the historical land resource data to obtain expert knowledge labeling data;

[0198] Step S522: extracting early warning features from historical land resource data and expert knowledge mark data to obtain early warning feature data;

[0199] Step S523: dividing the warning feature data into data sets to obtain a warning training data set;

[0200] Step S524: Select a machine learning model according to the warning target data to obtain a selected machine learning model; perform warning model training according to the selected machine learning model and the warning training data set to obtain a warning model;

[0201] Step S525: performing model evaluation and optimization on the early warning model according to the early warning training data set to obtain an optimized early warning model;

[0202] Step S526: extract warning rules using the optimized warning model, and construct a warning rule library to obtain a warning rule library.

[0203] In an embodiment of the present invention, historical land resource data such as land use change data, vegetation coverage data, meteorological data (such as temperature and precipitation), terrain data (such as slope and elevation) and socio-economic data (such as population density and GDP) of the study area in the past ten years are obtained. At the same time, early warning target data are obtained, such as early warning thresholds for changes in different land use types and early warning thresholds for changes in vegetation coverage. Domain experts are invited to mark the historical land resource data, and the land resource status of each time period in the historical data is marked as one of the three states of "normal", "warning" or "alarm". Based on their professional knowledge and experience, combined with the early warning target data, the experts judge whether the land resource status of each time period reaches the warning or alarm level. The expert marking results are stored as expert knowledge marking data, saved in CSV format, including timestamps, various indicator data and corresponding expert marking status.

[0204] Extract features related to early warning from historical land resource data, such as land use type change rate, vegetation coverage change rate, annual average temperature, annual precipitation, population density growth rate, etc. Calculate statistical indicators such as change value, mean, variance, etc. of these features in different time periods. Associate the extracted features with expert knowledge tag data to form early warning feature data. The early warning feature data is stored in CSV format, including timestamps, various early warning features, and corresponding expert tag status.

[0205] The warning feature data is divided into a training data set and a test data set in a ratio of 7:3. The training data set is used to train the machine learning model, and the test data set is used to evaluate the model performance. During the division process, ensure that the sample ratio of each category ("normal", "warning", "alarm") in the training data set and the test data set is the same as the sample ratio of each category in the original data set to avoid sample imbalance.

[0206] According to the characteristics of the warning target data and warning feature data, select a suitable machine learning model, such as support vector machine (SVM), random forest, or gradient boosting decision tree (GBDT). Use Python's machine learning library Scikit-learn to train the selected machine learning model. Use the warning training data set to train the model and adjust the model parameters (such as the kernel function and penalty coefficient of SVM, the number and depth of trees of random forest, and the learning rate and number of trees of GBDT) to obtain the best performance.

[0207] Use the test data set to evaluate the trained early warning model and calculate the model's accuracy, precision, recall, F1 value and other indicators. According to the evaluation results, optimize the model. For example, adjust the model parameters, use different feature combinations or adopt different model training methods. Repeat steps S524 and S525 until the model performance meets the preset requirements.

[0208] If you use a decision tree model (such as random forest or GBDT), you can extract warning rules directly from the trained model. Each rule corresponds to a path in the decision tree, and the rule conditions consist of the node conditions on the path. If you use other types of models (such as SVM), you need to use other methods to extract warning rules. For example, you can analyze the weight coefficients of the model to determine which features have a greater impact on the warning results, and build warning rules based on these features. The extracted warning rules are stored in the warning rule library, which is saved in JSON format and contains rule conditions, warning levels, warning information, etc.

[0209] Preferably, the present invention also provides a land natural resource dynamic supervision system for executing the land natural resource dynamic supervision method as described above, the land natural resource dynamic supervision system comprising:

[0210] The multi-source data acquisition module is used to collect multi-source data of land natural resource areas to obtain original remote sensing image data and auxiliary data sets; multi-source data preprocessing is performed based on the original remote sensing image data and auxiliary data sets to obtain an initial data set;

[0211] The intelligent land classification module is used to construct a land classification model based on a convolutional neural network according to the initial data set to obtain a land classification model; and to perform intelligent land classification processing using the land classification model to obtain a land classification map;

[0212] The map change detection module is used to perform LSTM-based change detection according to the land classification map to obtain the LSTM change detection result; perform object-oriented change detection and refinement on the LSTM change detection result to obtain a preliminary change detection map; perform change analysis on the preliminary change detection map to obtain the change analysis result; generate a change detection map according to the change analysis result to obtain a change detection map;

[0213] The three-dimensional land resource modeling module is used to construct a digital elevation model based on the initial data set to obtain a digital elevation model; construct a basic three-dimensional model based on the digital elevation model and the land classification map to obtain a basic three-dimensional model; refine the model and assign attributes to the basic three-dimensional model based on the change detection map to obtain a three-dimensional land resource model;

[0214] The dynamic supervision and early warning module is used to dynamically monitor the indicators of the three-dimensional land resource model and the change detection map to obtain dynamic monitoring results; to trigger early warning processing on the dynamic monitoring results according to the pre-built early warning rule library to obtain early warning information; and to push resource supervision information based on the early warning information to realize the dynamic supervision task of land natural resources.

[0215] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0216] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for dynamic supervision of land natural resources, characterized in that: The following steps are involved: Step S1: collect multi-source data of land natural resource areas to obtain original remote sensing image data and auxiliary data sets; perform multi-source data preprocessing based on the original remote sensing image data and auxiliary data sets to obtain an initial data set; Step S2: constructing a land classification model based on a convolutional neural network according to the initial data set to obtain a land classification model; performing intelligent land classification processing using the land classification model to obtain a land classification map; Step S3: Perform LSTM-based change detection according to the land classification map to obtain LSTM change detection results; perform object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection map; Perform change analysis on the preliminary change detection graph to obtain change analysis results; Generate a change detection map according to the change analysis result to obtain a change detection map; Step S4: constructing a digital elevation model according to the initial data set to obtain a digital elevation model; constructing a basic three-dimensional model according to the digital elevation model and the land classification map to obtain a basic three-dimensional model; refining and assigning attributes to the basic three-dimensional model according to the change detection map to obtain a three-dimensional land resource model; Step S5: dynamically monitor the indicators of the three-dimensional land resource model and the change detection map to obtain dynamic monitoring results; perform early warning trigger processing on the dynamic monitoring results according to the pre-built early warning rule library to obtain early warning information; Resource supervision information is pushed based on early warning information to achieve dynamic supervision tasks of land and natural resources.

2. The method for dynamic supervision of land natural resources according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: acquiring remote sensing images of the land natural resource area to obtain original remote sensing image data; Step S12: Collect auxiliary data on the land natural resource area to obtain an auxiliary data set; Step S13: converting the data format of the original remote sensing image data and the auxiliary data set, and performing data registration to obtain a registered data set; Step S14: performing atmospheric correction and radiation correction on the registration data set to obtain a corrected data set; Step S15: performing data fusion and data trimming on the correction data set to obtain an initial data set.

3. The method for dynamic supervision of land natural resources according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: construct a sample data set for the initial data set to obtain a training sample set and a verification sample set; Step S22: extracting hyperspectral features from the training sample set using a convolutional neural network to obtain hyperspectral features; extracting LiDAR features from the training sample set using PointNet technology to obtain LiDAR features; Step S23: performing feature fusion on the hyperspectral features and the LiDAR features to obtain fused features; constructing a classifier based on the fused features and the training sample set, and performing land classification model training to obtain a land classification model; Step S24: extracting features and classifying the initial data set using the land classification model to obtain classification results; Step S25: performing post-processing on the classification result to obtain a post-processed classification result; Step S26: Use the verification sample set to evaluate the accuracy of the post-processing classification results to obtain a land classification map.

4. The method for dynamic supervision of land natural resources according to claim 1 is characterized in that: Step S3 includes the following steps: Step S31: generating a multi-temporal land classification map according to the land classification map to obtain a multi-temporal land classification map; performing multi-temporal data alignment on the multi-temporal land classification map to obtain an aligned land classification map sequence; Step S32: constructing time series features for the aligned land classification map sequence to obtain a time series feature data set; Step S33: performing LSTM-based change detection according to the time series feature data set to obtain LSTM change detection results; Step S34: performing object-oriented change detection and refinement on the LSTM change detection results to obtain a preliminary change detection graph; Step S35: Perform change analysis on the preliminary change detection graph to obtain a change analysis result; generate a change detection graph based on the change analysis result to obtain a change detection graph.

5. The method for dynamic supervision of land natural resources according to claim 4 is characterized in that: Step S33 includes the following steps: Step S331: performing time series data preprocessing on the time series feature data set to obtain preprocessed time series data; Step S332: performing forward LSTM feature extraction according to the preprocessed time series data to obtain forward LSTM features; performing reverse LSTM feature extraction according to the preprocessed time series data to obtain reverse LSTM features; Step S333: performing bidirectional LSTM feature concatenation on the forward LSTM feature and the reverse LSTM feature to obtain a Bi-LSTM feature; Step S334: Calculate the attention weight of the Bi-LSTM feature to obtain the attention weight; normalize the attention weight to obtain the normalized attention weight; Step S335: performing attention weighting on the Bi-LSTM features and the normalized attention weights to obtain attention weighted features; Step S336: performing time dimension feature aggregation on the attention weighted features to obtain aggregated features; performing full connection layer classification on the aggregated features to obtain an inactivated classification result; Step S337: Perform Sigmoid activation on the inactivated classification results to obtain LSTM change detection results.

6. The method for dynamic supervision of land natural resources according to claim 4 is characterized in that: Step S35 includes the following steps: Step S351: determining the change type according to the preliminary change detection map and the multi-temporal land classification map to obtain a change type map; Step S352: performing statistics on the change area according to the change type diagram to obtain change statistical data; Step S353: performing spatial heterogeneity analysis on the change type map and the auxiliary data set to obtain spatial heterogeneity analysis results; Step S354: generating a legend according to the change type map, the change statistical data and the spatial heterogeneity analysis result to obtain a legend; Step S355: Add the legend to the preliminary change detection map to obtain a change detection map.

7. The method for dynamic supervision of land natural resources according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: obtaining LiDAR data in the initial data set; performing point cloud processing on the LiDAR data to obtain point cloud data; Step S42: constructing a digital elevation model according to the point cloud data to obtain a digital elevation model; Step S43: extracting vegetation height from the point cloud data and the digital elevation model to obtain a vegetation height map; Step S44: using the digital elevation model as the terrain basis and the land classification map as the texture map to construct a basic three-dimensional model to obtain a basic three-dimensional model; Step S45: Refining the basic three-dimensional model according to the vegetation height map and the change detection map to obtain a refined model; assigning attributes to the refined model to obtain a three-dimensional land resource model.

8. The method for dynamic supervision of land natural resources according to claim 1 is characterized in that: Step S5 includes the following steps: Step S51: extracting indicators from the three-dimensional land resource model and the change detection map, and performing indicator analysis to obtain a land resource indicator data set; Step S52: Acquire historical land resource data and early warning target data; construct an early warning rule base according to the historical land resource data and the early warning target data to obtain an early warning rule base; Step S53: dynamically monitor the land resource indicator data set to obtain dynamic monitoring results; perform early warning trigger processing on the dynamic monitoring results according to the early warning rule library to obtain early warning information; Step S54: Generate a supervision report based on the early warning information and the land resource indicator data set to obtain a supervision report; Step S55: Push information based on early warning information and supervision reports to achieve dynamic supervision of land and natural resources.

9. The method for dynamic supervision of land natural resources according to claim 8, characterized in that: Step S52 includes the following steps: Step S521: Acquire historical land resource data and early warning target data; perform expert knowledge labeling processing on the historical land resource data to obtain expert knowledge labeling data; Step S522: extracting early warning features from historical land resource data and expert knowledge mark data to obtain early warning feature data; Step S523: dividing the warning feature data into data sets to obtain a warning training data set; Step S524: Select a machine learning model according to the warning target data to obtain a selected machine learning model; perform warning model training according to the selected machine learning model and the warning training data set to obtain a warning model; Step S525: performing model evaluation and optimization on the early warning model according to the early warning training data set to obtain an optimized early warning model; Step S526: extract warning rules using the optimized warning model, and construct a warning rule library to obtain a warning rule library.

10. A land natural resource dynamic supervision system, characterized in that: Used to execute the land natural resources dynamic supervision method as claimed in claim 1, the land natural resources dynamic supervision system comprises: The multi-source data acquisition module is used to collect multi-source data of land natural resource areas to obtain original remote sensing image data and auxiliary data sets; multi-source data preprocessing is performed based on the original remote sensing image data and auxiliary data sets to obtain an initial data set; The intelligent land classification module is used to construct a land classification model based on a convolutional neural network according to the initial data set to obtain a land classification model; and to perform intelligent land classification processing using the land classification model to obtain a land classification map; The map change detection module is used to perform LSTM-based change detection according to the land classification map to obtain the LSTM change detection result; perform object-oriented change detection and refinement on the LSTM change detection result to obtain a preliminary change detection map; perform change analysis on the preliminary change detection map to obtain the change analysis result; generate a change detection map according to the change analysis result to obtain a change detection map; The three-dimensional land resource modeling module is used to construct a digital elevation model based on the initial data set to obtain a digital elevation model; construct a basic three-dimensional model based on the digital elevation model and the land classification map to obtain a basic three-dimensional model; refine the model and assign attributes to the basic three-dimensional model based on the change detection map to obtain a three-dimensional land resource model; The dynamic supervision and early warning module is used to dynamically monitor the indicators of the three-dimensional land resource model and the change detection map to obtain dynamic monitoring results; to trigger early warning processing on the dynamic monitoring results according to the pre-built early warning rule library to obtain early warning information; and to push resource supervision information based on the early warning information to realize the dynamic supervision task of land natural resources.

Images