Land resource dynamic monitoring and early warning method and system based on multi-source remote sensing data fusion

Abstract

The application relates to the technical field of land resource monitoring, in particular to a land resource dynamic monitoring and early warning method and system based on multi-source remote sensing data fusion, which comprises the following steps: periodically collecting optical images, SAR echoes and LiDAR point clouds by using an unmanned aerial vehicle, constructing a three-dimensional digital model of the ground and carrying out land division; fusing to form a multi-dimensional feature vector, establishing an LSTM land feature evolution model, predicting the change rate interval of each feature in the current period based on a historical sequence; constructing a time series difference change detection algorithm, calculating the land change rate and taking the predicted interval as an abnormality judgment threshold, and screening potential abnormal land blocks; adopting a double-branch convolutional neural network to identify crop states, growth stages and illegal construction behaviors, determining and determining abnormalities and giving confidence; performing spatial clustering on the determined abnormal land blocks, obtaining accurate boundaries in combination with the three-dimensional model, generating multi-level early warning information, and improving the decision efficiency and response speed of land resource monitoring.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of land resource monitoring, in particular to a land resource dynamic monitoring and early warning method and system based on multi-source remote sensing data fusion. BACKGROUND

[0002] Traditional land resource monitoring mainly relies on manual patrol and satellite remote sensing image interpretation, and has the disadvantages of long monitoring cycle, poor timeliness, limited coverage range and the like. Existing remote sensing monitoring technologies mostly use a single data source, such as using only optical images for land cover classification or relying only on radar data for change detection. Such a single data source monitoring method has many limitations in practical application: optical images are easily affected by cloud and fog weather, and cannot realize all-weather monitoring; although synthetic aperture radar can penetrate clouds, image interpretation is difficult, and it is difficult to accurately identify ground object types; laser radar can obtain accurate three-dimensional information, but lacks spectral characteristics and cannot effectively distinguish different land use types.

[0003] Existing methods mostly use fixed threshold values to determine abnormal changes, and cannot adapt to the land use change rules of different regions and different seasons, which easily misjudges normal agricultural production activities as abnormal changes or misses real illegal land use behaviors. In addition, the current land resource monitoring has deficiencies in abnormal identification and early warning response. The monitoring result can only give the position information of the changed area, and cannot accurately determine the specific type and severity of the change, lacking fine classification and identification capability. The generation of early warning information is also relatively rough, and cannot perform hierarchical early warning according to the type, size and confidence of the abnormality, affecting the decision efficiency and response speed of land resource monitoring.

[0004] In view of this, the application provides a land resource dynamic monitoring and early warning method based on multi-source remote sensing data fusion. SUMMARY

[0005] To achieve the above-mentioned purpose, the application provides a land resource dynamic monitoring and early warning method based on multi-source remote sensing data fusion and a system, and the specific technical solutions are as follows:

[0006] The land resource dynamic monitoring and early warning method based on multi-source remote sensing data fusion comprises the following steps:

[0007] Periodically collecting optical images, SAR echoes and LiDAR point cloud data of the ground by an unmanned aerial vehicle, generating a three-dimensional digital model of the ground based on the LiDAR point cloud data and performing land parcel division, mapping the ground optical images and SAR echoes obtained in the same period to the corresponding land parcels by constructing a space-time registration algorithm, and generating a three-dimensional digital model of the ground;

[0008] For the divided plots, the texture feature parameters of the optical image, the backscattering coefficient and polarization features of the SAR echo, and the elevation statistics and surface roughness of the LiDAR point cloud are extracted, and a multi-dimensional feature vector of the plot is generated through a feature fusion algorithm;

[0009] A plot feature evolution model based on a long short-term memory network is constructed, the multi-dimensional feature vector of the plot based on the historical period is used to deduce the change rule of the plot in the current period, and a plot feature change rate interval is generated;

[0010] A change detection algorithm based on time series difference is established, the plot change rate in the current period is calculated according to the multi-dimensional feature vector of the plot, the plot feature change rate interval is used as a judgment threshold range of the potential abnormal plot, and the potential abnormal plot is marked according to the judgment threshold range of the potential abnormal plot;

[0011] A plot state recognition model based on a double-branch convolutional neural network is constructed, the optical image and SAR echo of the potential abnormal plot are input into the plot state recognition model, the ground crop planting state, growth stage and ground illegal construction behavior are recognized, and it is determined whether the potential abnormal plot is a determined abnormal plot, and the classification result and confidence score of the determined abnormal plot are output;

[0012] The determined abnormal plots are subjected to spatial clustering analysis to form a ground abnormal area, the geographical boundary of the ground abnormal area is demarcated according to the ground three-dimensional digital model, and multi-level warning information is generated according to the classification result and confidence score of the determined abnormal plot.

[0013] Preferably, an integrated optical camera, synthetic aperture radar and laser radar scanner is carried by a UAV to collect optical image, SAR echo and LiDAR point cloud data of the monitoring area;

[0014] Based on the collected LiDAR point cloud data, after noise removal and point cloud thinning preprocessing, and separation of ground points and non-ground points, an irregular triangle network interpolation method is used to generate a ground three-dimensional digital model;

[0015] A watershed algorithm based on slope and curvature is used to divide and code the plots of the ground three-dimensional digital model by fusing the edge detection results of the optical image;

[0016] A space-time registration algorithm is constructed to geometrically correct and register the optical image and SAR echo and map them to the corresponding plots.

[0017] Preferably, the optical image is converted into a gray image, a gray level co-occurrence matrix method is used to calculate and extract energy, contrast, correlation and entropy in a set distance and multiple directions, and the feature values in multiple directions are averaged to obtain the feature parameters;

[0018] The SAR echo is radiometrically calibrated and speckle noise is suppressed, the backscattering coefficient is calculated, and a scattering matrix is constructed from the SAR echo data, and a polarization decomposition method is used to extract components representing surface scattering, dihedral scattering and volume scattering characteristics, and polarization entropy and average scattering angle are calculated.

[0019] Preferably, for the point cloud data in each plot, calculate the elevation statistical parameters, including the average elevation, the standard deviation of the elevation, the maximum elevation and the minimum elevation;

[0020] By grid division of the point cloud in the plot and using the local plane fitting method, the surface roughness is calculated.

[0021] After extracting the feature parameters of each data source, a feature fusion algorithm is used to normalize all the extracted feature parameters, and the normalized optical texture features, SAR polarization features and LiDAR elevation features are spliced in a predetermined order to generate a plot multi-dimensional feature vector.

[0022] Preferably, a plot feature evolution model is constructed based on a long short-term memory network.

[0023] The plot multi-dimensional feature vectors of the historical periods are organized into a time series, the feature change rate sequence between adjacent periods is calculated and standardized to serve as the input of the plot feature evolution model;

[0024] The architecture of the long short-term memory network is designed, including an input layer, a fully connected layer and an output layer; the input layer includes an LSTM layer of a forget gate, an input gate and an output gate, and the output layer is used to output the feature change rate prediction value of the next period and the uncertainty estimate for quantifying the prediction confidence.

[0025] Preferably, based on the feature change rate prediction value and the uncertainty estimate predicted by the plot feature evolution model, a change rate confidence interval is constructed for each dimension of the plot, forming a multi-dimensional feature change rate interval of the plot.

[0026] Preferably, the plot multi-dimensional feature vectors of the current and historical monitoring periods are obtained, a feature difference vector is calculated, and the plot change rate of the current period is calculated.

[0027] Based on the coefficient of variation of each feature dimension in the historical data, the weight is determined, and the weighted Euclidean distance is used to calculate the comprehensive change rate of the plot.

[0028] The comprehensive relative change rate of the plot is calculated, and the positive change rate and the negative change rate representing the degree of increase and decrease of the feature value are calculated.

[0029] Preferably, a dynamic decision threshold range is set based on the feature change rate interval of the plot, and when the calculated feature change rate of the plot exceeds the dynamic decision threshold range, the abnormality degree of the plot feature deviating from the prediction interval is calculated and summarized as the comprehensive change rate abnormality degree of the plot;

[0030] An abnormality decision mechanism based on multi-criteria fusion is constructed, and when the plot meets the conditions of any abnormality decision mechanism, the plot is marked as a potential abnormal plot:

[0031] The abnormality decision mechanism includes: the change rate of a preset number of feature dimensions exceeding the prediction interval, the comprehensive change rate abnormality degree of the plot being greater than a preset threshold, and the comprehensive change rate of the plot exceeding the comprehensive change rate interval derived from the boundaries of each dimension of the prediction interval.

[0032] Preferably, a plot state recognition model based on a double-branch convolutional neural network is constructed to identify whether a potential abnormal plot is a confirmed abnormal plot.

[0033] The double-branch convolutional neural network processes optical images and SAR data respectively, the optical image branch adopts a residual connection structure and inserts a multi-scale feature extraction layer composed of multiple parallel hole convolutions to capture features of different receptive fields;

[0034] The SAR data branch adds polarization feature encoding at the input end of the optical image branch network structure.

[0035] Preferably, a feature fusion mechanism based on an attention mechanism is constructed to adaptively calculate the weights of the optical and SAR two-branch features and perform weighted fusion.

[0036] A multi-task learning structure is designed at the end of the double-branch convolutional neural network, and parallel classification heads are set to simultaneously recognize crop planting states, growth stages, and ground illegal construction behaviors.

[0037] According to the prediction results of the plot state recognition model on the recognition task, a comprehensive confidence score is calculated, and combined with a preset illegal construction decision threshold and a crop abnormality confidence threshold, it is determined whether a potential abnormal plot is a confirmed abnormal plot, and the classification result is output.

[0038] Preferably, the identified confirmed abnormal plots are subjected to spatial clustering analysis, and a density-based clustering algorithm is used to merge plots that are spatially adjacent and have the same abnormal class into continuous ground abnormal areas.

[0039] Using a three-dimensional digital model of the ground, the geographic boundary of the abnormal area is calibrated and the geometric properties of the ground abnormal area are calculated by a convex hull algorithm and an Alpha-shape algorithm.

[0040] Preferably, a multi-level early warning mechanism is constructed, a risk score is calculated according to the severity of the abnormal type, the recognition confidence and the area of the region to determine the early warning level, and a structured early warning report containing the abnormal position and type information is generated.

[0041] The land resource dynamic monitoring and early warning system based on multi-source remote sensing data fusion is used to realize the land resource dynamic monitoring and early warning method based on multi-source remote sensing data fusion, and comprises a data acquisition module, a feature extraction module, a feature change rate prediction module, a potential abnormal land block identification module, an abnormal land block identification and classification module, and a hierarchical early warning module.

[0042] The data acquisition module periodically collects optical images, SAR echoes and LiDAR point cloud data of the ground by using a drone, generates a three-dimensional digital model of the ground based on the LiDAR point cloud data and performs land division, maps the optical images and SAR echoes of the same period obtained from the ground to the corresponding land by using a time-space registration algorithm, and generates a three-dimensional digital model of the ground.

[0043] The feature extraction module extracts the texture feature parameters of the optical images, the backscattering coefficient and polarization features of the SAR echoes, and the elevation statistics and surface roughness of the LiDAR point cloud for the divided land, and generates a multi-dimensional feature vector of the land by using a feature fusion algorithm.

[0044] The feature change rate prediction module constructs a land feature evolution model based on a long short-term memory network, deduces the change rule of the land in the current period based on the multi-dimensional feature vector of the land in the historical period, and generates a land feature change rate interval.

[0045] The potential abnormal land block identification module establishes a change detection algorithm based on time series difference, calculates the change rate of the land in the current period according to the multi-dimensional feature vector of the land, takes the land feature change rate interval as a threshold range for judging potential abnormal land blocks, and marks the potential abnormal land blocks according to the threshold range for judging potential abnormal land blocks.

[0046] The abnormal land block identification and classification module constructs a land state identification model based on a double-branch convolutional neural network, inputs the optical images and SAR echoes of the potential abnormal land blocks into the land state identification model, identifies the ground crop planting state, growth stage and ground illegal construction behavior, determines whether the potential abnormal land blocks are determined abnormal land blocks, and simultaneously outputs the classification results and confidence score of the determined abnormal land blocks.

[0047] The hierarchical early warning module performs spatial clustering analysis on the determined abnormal land blocks to form a ground abnormal region, calibrates the geographical boundary of the ground abnormal region according to the three-dimensional digital model of the ground, and generates multi-level early warning information according to the classification results and confidence score of the determined abnormal land blocks.

[0048] The application has the beneficial effects that: through the cooperative collection and space-time registration of multi-source remote sensing data, the application realizes the all-around stereoscopic acquisition of ground information. The three-dimensional digital model generated by the LiDAR point cloud provides a high-precision terrain basis, and provides a unified geographic reference framework for subsequent accurate division of plots and spatial matching of multi-source data.

[0049] The application extracts complementary feature parameters of different remote sensing data sources to construct a multi-dimensional description system of plots. The multi-dimensional feature vector fully utilizes the respective advantages of optics, microwaves and laser radars, and realizes the comprehensive representation of surface cover types, physical structures and geometric shapes.

[0050] The application adopts a long short-term memory network to establish a plot feature evolution model, which can effectively capture the time sequence dependence and long-term change trend of plot features; and by generating a plot feature change rate interval, a dynamic benchmark is provided for anomaly detection.

[0051] The change detection algorithm based on time series difference of the application can accurately quantify the actual change degree of the plot, and realizes adaptive screening of abnormal plots in combination with a dynamic threshold range. The false positive rate caused by seasonal changes and normal agricultural activities can be effectively reduced, and the accuracy of land resource monitoring is significantly improved.

[0052] The double-branch convolutional neural network constructed by the application fully utilizes the complementary advantages of optical and SAR data, realizes accurate identification of plot states in all-weather and all-day, the model can accurately determine the authenticity of abnormal plots, and outputs specific abnormal types and confidence scores, thereby improving the accuracy of illegal land use identification.

[0053] The application integrates discrete abnormal plots into continuous abnormal areas through spatial clustering analysis, improves the practicability of the monitoring results, and provides accurate spatial positioning information and differentiated response strategies for regulatory units in combination with the accurate geographic positioning of the three-dimensional digital model and the multi-level early warning mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A land resource dynamic monitoring and early warning method flowchart based on multi-source remote sensing data fusion is provided for the application;

[0055] Figure 2 A multi-dimensional feature extraction and fusion flowchart of multi-remote sensing data is provided for the application;

[0056] Figure 3 A plot feature evolution modeling and prediction flowchart is provided for the application;

[0057] Figure 4 A change detection and potential anomaly identification flowchart is provided for the application;

[0058] Figure 5A land plot state recognition and abnormality discrimination flowchart is provided for the present application.

[0059] Figure 6 A land resource dynamic monitoring and early warning system structure diagram based on multi-source remote sensing data fusion is provided for the present application. DETAILED DESCRIPTION

[0060] In order to make the above objectives, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0061] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from the description, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited by the specific embodiments disclosed below.

[0062] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is independent of or selectively excludes other embodiments.

[0063] Embodiment 1

[0064] Reference Figures 1 to 4 For the first embodiment of the present application, as shown in Figure 1 , a land resource dynamic monitoring and early warning method based on multi-source remote sensing data fusion is provided.

[0065] Step 1: Periodically collect optical images, SAR echoes and LiDAR point cloud data of the ground by the unmanned aerial vehicle, generate a three-dimensional digital model of the ground based on the LiDAR point cloud data and perform land plot division, map the optical images and SAR echoes of the ground obtained in the same period to the corresponding land plot by constructing a space-time registration algorithm, and generate a three-dimensional digital model of the ground.

[0066] The ground data is collected by a multi-sensor platform carried by a UAV. The UAV flies along a preset route at a fixed height and speed. The optical camera, synthetic aperture radar and laser radar scanner carried by the UAV work synchronously to collect optical images, SAR echo and LiDAR point cloud data of the monitoring area. The optical camera collects high-resolution images in the visible and near-infrared wave bands. The synthetic aperture radar sensor transmits microwave signals and receives SAR echo data reflected by the ground. The laser radar scanner transmits laser pulses and records the time and intensity information of the returned signals to form LiDAR point cloud data. By integrating a high-precision GNSS receiver and an inertial measurement unit on the UAV, the position and attitude parameters of the flight platform are recorded in real time to provide accurate geographic reference information for subsequent data processing.

[0067] A three-dimensional digital model of the ground is constructed based on the collected LiDAR point cloud data. The original point cloud data is preprocessed, including noise point elimination and point cloud thinning. Statistical filtering algorithm is used to eliminate outlier noise points, and voxel grid downsampling method is used for point cloud thinning. The three-dimensional space is divided into cubic voxels with a side length of 0.5 m. All points in each voxel are replaced by its centroid point, which reduces data redundancy while maintaining terrain features.

[0068] The ground points are separated from the processed point cloud data. The improved progressive morphological filter algorithm is used to separate the ground points from the non-ground points. Based on the separated ground points, the irregular triangle network (TIN) interpolation method is used to generate a continuous digital elevation model with an interpolation resolution of 0.5 m, forming a complete three-dimensional digital model of the ground.

[0069] The generated three-dimensional digital model of the ground is divided into plots. The watershed algorithm based on slope and curvature is used to identify the terrain feature lines. The slope and curvature of each grid point are calculated, where z is the elevation value, x and y are the plane coordinates. When or , it is marked as a terrain feature point, and are the slope threshold and curvature threshold, respectively. Exemplarily, is set to 15°, is set to 0.1. The feature points are connected to form plot boundary lines. The boundary is optimized in combination with the edge detection results of the optical image. The Canny edge detection operator is used to extract the image edges, which are fused with the terrain feature lines to generate accurate plot boundaries. The divided plots are encoded to give them unique identifiers, recording the spatial range, center coordinates and area attributes of the plots.

[0070] ​A spatio-temporal registration algorithm is constructed for the precise mapping of optical images and SAR echoes to the parcels. The sensors are geometrically corrected based on the position and attitude information of the UAV obtained from the differential GPS / INS integrated navigation system mounted on the UAV to establish the conversion relationship between the sensor coordinate system and the geographic coordinate system. For optical image data, the sensor coordinates of the image data collected by the UAV are mapped into geographic coordinates. For SAR echo data, the geometric positioning is performed using the range-Doppler model to establish the mapping relationship between the slant range image coordinates and the ground coordinates.

[0071] Based on the completion of geometric correction, a fine registration method based on feature points is used to improve the registration accuracy. SIFT feature points of optical images and SAR images are extracted, and the same name point pairs are obtained through feature descriptor matching. The RANSAC algorithm is used to remove the mis-matching points and estimate the affine transformation parameters between the images. The registered optical images and SAR echo data are cropped and resampled according to the parcel boundaries, and each parcel corresponds to a set of multi-source remote sensing data slices. A parcel data index table is established to record the storage location and metadata information of each parcel in different sensors at different times, including collection time, sensor parameters, and weather conditions and other auxiliary information.

[0072] This step realizes multi-aspect monitoring of the monitored ground through multi-source remote sensing technology, and efficiently manages the collected multi-source remote sensing data, providing a standardized data basis for subsequent feature extraction and change detection, and improving the automation degree and processing efficiency of land resource monitoring.

[0073] Step 2: For the divided parcels, the texture feature parameters of the optical images, the backscattering coefficient and polarization features of the SAR echoes, and the elevation statistics and surface roughness of the LiDAR point cloud are extracted, and a multi-dimensional feature vector of the parcel is generated through a feature fusion algorithm. Referring to Figure 2 The flow chart of multi-dimensional feature extraction and fusion of multi-remote sensing data for this step.

[0074] For each parcel divided in step 1, multi-dimensional feature parameters are extracted from optical images, SAR echoes and LiDAR point cloud data respectively. For texture feature extraction of optical images, the RGB image corresponding to the parcel is first converted into a grayscale image, and the texture features are calculated using the gray level co-occurrence matrix (GLCM) method. When calculating the gray level co-occurrence matrix, the spatial relationship parameters of the pixel pairs are set, including distance d and direction θ, where the direction θ takes 0°, 45°, 90° and 135°, and the distance d is determined according to the image resolution. For example, when the image resolution is 0.1 meters / pixel, d takes values of 1, 3 and 5 pixels. The normalized gray level co-occurrence matrix is constructed as follows where i and j represent the gray values of the pixel pairs, and the texture feature parameters are calculated based on the matrix.

[0075] Four main texture features—energy, contrast, correlation, and entropy—are extracted from the gray-level co-occurrence matrix of optical images. Energy features... The formula reflects the coarseness of image texture and the uniformity of grayscale distribution. Where L is the number of gray levels. Contrast characteristics characterize the sharpness of the texture, and are calculated using the formula: Contrast Characteristics Correlation characteristics The linear correlation between pixel gray levels is described by the following formula: ,in , These are the mean values ​​in the row and column directions, respectively. , This represents the corresponding standard deviation. Entropy feature. The formula reflects the complexity of the texture. The final value of each texture feature is obtained by averaging the feature values ​​in the four directions. This multi-directional averaging process improves the rotation invariance of the features.

[0076] For feature extraction from SAR echo data, radiometric calibration and speckle noise suppression are performed first. A Lee filter is used to filter the SAR image, with a filter window size of 7×7 pixels to effectively suppress speckle noise while preserving edge information. The backscattering coefficients for each polarization channel are calculated, and the digital numerical values ​​(DN) are converted to backscattering coefficients. The conversion formula is: Where K is a calibration constant determined by sensor parameters, and DN is the original integer value stored for each pixel after sensor imaging and analog-to-digital conversion of the remote sensing image, used to represent the relative magnitude of the electromagnetic signal strength received by that pixel. For SAR data, a scattering matrix S is constructed, and polarization feature parameters are extracted.

[0077] Based on the scattering matrix S, the polarization decomposition characteristics are calculated. The Pauli decomposition method is used to decompose the scattering matrix into three basic scattering mechanism components: single scattering component. Characterizing surface scattering properties, secondary scattering component Reflecting dihedral scattering characteristics, volume scattering component Visible scattering characteristics and These are the diagonal elements of the scattering matrix S of a fully polarimetric SAR. Representing the horizontal polarization of the transmitter and the horizontal polarization of the receiver, the resulting complex scattering coefficients are: , This represents the complex scattering coefficients obtained from both the transmit and receive vertical polarizations. The polarization entropy is then calculated. and average scattering angle , the polarization entropy calculation formula is , wherein is the normalized probability of the eigenvalue of the coherence matrix, is the gray level index parameter of the pixel pair, and the average scattering angle , is the scattering angle corresponding to each eigenvector; the polarization feature can effectively represent the scattering mechanism and physical characteristics of the ground object, and improve the accuracy of ground object classification.

[0078] For LiDAR point cloud data, the elevation statistics and surface roughness features of the land parcel are extracted. The elevation statistical parameters of the point cloud data in each land parcel are calculated, including the average elevation , the standard deviation of the elevation , the maximum elevation and the minimum elevation , wherein N is the number of point clouds in the land parcel, is the elevation value of the kth point. The elevation range and the elevation coefficient of variation are calculated, which reflect the vertical structure characteristics of the land parcel.

[0079] The surface roughness is calculated by using the local plane fitting method. The point cloud in the land parcel is divided into grids, for example, the grid size is set to 1 meter x 1 meter, and the least square method is used to fit the local plane in each grid; the distance of each point to the fitted plane is calculated, and the surface roughness is defined as , wherein M is the number of grids, is the number of points in the mth grid. The terrain slope and the terrain relief are calculated, is the average elevation of the nth grid, and X and Y correspond to the actual distance axes of the row and column directions of the grid on the ground.

[0080] After extracting the features of each data source, a feature fusion algorithm is used to generate a multi-dimensional feature vector of the land parcel. First, all features are normalized, and the minimum-maximum normalization method is used to map the feature values to the [0, 1] interval. The optical texture feature, SAR polarization feature and LiDAR elevation feature are spliced in order to form a feature vector with a dimension of D , is the Dth normalized feature component.

[0081] ​The step fully utilizes the complementary advantages of different sensor data by a multi-source data feature extraction and fusion method. The optical image provides the spectral and texture information of the ground, the SAR data reflects the microwave scattering characteristics of the ground, and the LiDAR data accurately describes the three-dimensional structure of the terrain. The multi-dimensional feature vector after fusion comprehensively describes the comprehensive properties of the land parcel, provides a rich and reliable feature basis for subsequent change detection and anomaly recognition, and can significantly improve the accuracy and reliability of land resource monitoring.

[0082] Step 3: Constructing a land parcel feature evolution model based on a long short-term memory network, deducing the change rule of the land parcel in the current period based on the multi-dimensional feature vector of the land parcel in the historical period, and generating a land parcel feature change rate interval. Referring to Figure 3 , the flow chart of land parcel feature evolution modeling and prediction for this step.

[0083] A long short-term memory network (LSTM) based land parcel feature evolution model is constructed to learn and predict the time series change rule of the multi-dimensional feature vector of the land parcel. The historical time series data is preprocessed, and the multi-dimensional feature vector of each land parcel in different monitoring periods is organized into a time series. Let the feature vector of land parcel a in the tth monitoring period be , the historical feature sequence of this land parcel is constructed as , where T is the total number of historical monitoring periods. The historical feature change rate vector between adjacent periods is calculated as , where is the monitoring period interval, and the change rate sequence is constructed as .

[0084] In order to eliminate the influence of different feature dimensions and numerical ranges, the change rate data is standardized, and the mean value of the change rate of each feature dimension is calculated as and the standard deviation is , where represents the bth dimensional feature change rate of land parcel a in the tth period, and the standardized change rate is .

[0085] The architecture of the LSTM network is designed, including the input layer, the LSTM layer, the fully connected layer and the output layer. The input layer receives the feature change rate sequence within the sliding time window; the LSTM layer contains multiple LSTM units, each of which contains three gating mechanisms: the forget gate, the input gate and the output gate. The forget gate determines which information in the cell state at the previous moment needs to be forgotten, the input gate controls the reception degree of the current input information, the update of the cell state is realized by combining the outputs of the forget gate and the input gate, and the output gate controls the output of the cell state to the hidden state.

[0086] A fully connected layer is connected after the LSTM layer for feature mapping, and the ReLU activation function is used in the fully connected layer Increase non-linear expressive power. The output layer produces two outputs, including: a predicted value of the feature rate of change. and uncertainty estimation The predicted rate of change represents the expected rate of change vector for the land parcel in the next period, while the uncertainty estimate represents the variance of the prediction, used to construct the confidence interval for the rate of change. The output layer uses a linear activation function with the same dimension as the input feature vector.

[0087] The land parcel feature evolution model is trained using historical data to construct a training set and employs the mean squared error loss function. Optimize prediction accuracy, among which The number of training samples. The true rate of change is used, while introducing an uncertainty loss term. This allows the model to learn about the uncertainty of predictions. Total loss function. ,in The balancing coefficient is used. The Adam optimizer is used for model training, and an early stopping strategy is employed to prevent overfitting.

[0088] After training, the feature change rate interval is generated for each plot using the plot feature evolution model. For plot a, the feature change rate sequence of its most recent w periods is input, and the model outputs the predicted change rate for the next period. and uncertainty estimation .

[0089] Based on the prediction results, a confidence interval for the rate of change of the features is constructed. For the b-th dimension feature, the interval for its rate of change is: ,in, The confidence coefficient is determined based on the required confidence level. This indicates that the land parcel feature evolution model applies to land parcel a in the next period t+1. Point prediction values ​​of dimensional features, This represents the variance estimate of the predicted values ​​given by the land parcel feature evolution model. For example, at a 95% confidence level... The value is set to 1.96. Combining the change rate intervals of each dimension forms the multidimensional characteristic change rate interval of the land parcel. ,in and Let be the lower and upper bounds of the rate of change of the b-th dimension feature, respectively. The number of feature dimensions.

[0090] This step constructs a land parcel feature evolution model based on an LSTM network, enabling effective learning and prediction of the temporal patterns of multidimensional feature change rates for land parcels. The constructed model fully utilizes the temporal dependence information of change rates contained in historical monitoring data, capturing both long-term trends and short-term fluctuation patterns of land parcel feature changes. The generated feature change rate intervals not only provide the expected range of normal change rates for land parcels but also quantify the uncertainty of predictions, providing a scientifically sound benchmark for subsequent anomaly detection. Compared to traditional statistical methods, this deep learning-based temporal modeling method exhibits stronger nonlinear fitting capabilities and adaptability, handling complex multidimensional feature change rate evolution patterns and significantly improving the intelligence level and early warning accuracy of dynamic land resource monitoring.

[0091] Step 4: Establish a change detection algorithm based on temporal difference. Calculate the land parcel change rate for the current period based on the multi-dimensional feature vector of the land parcels. Use the range of land parcel feature change rates as the threshold range for identifying potentially anomalous land parcels. Mark potentially anomalous land parcels according to this threshold range. (See also...) Figure 4 This is a flowchart for change detection and potential anomaly identification in this step.

[0092] In step 4, a change detection algorithm based on temporal difference is established to identify potential abnormal plots by calculating the degree of change of the multidimensional feature vector of the plot during adjacent monitoring weeks.

[0093] Obtain the multidimensional feature vectors of the land parcels in the current monitoring period t and the previous monitoring period t-1. For land parcel a, let its feature vector in the current period be... The eigenvector of the previous period is ,in Indicates land parcel In the current monitoring cycle The 3D eigenvalues Indicates land parcel In the previous monitoring cycle The 3D eigenvalues Let be the dimension of the feature vector. Calculate the feature difference vector between two cycles. ,in This represents the change in the b-th dimension feature. Indicates land parcel The The change in dimensional features between two adjacent monitoring weeks is used to calculate the land parcel change rate for the current period. ,in For the monitoring period time interval, Let be the rate of change of the b-th dimension feature.

[0094] To comprehensively evaluate the overall change degree of the plot, the weighted Euclidean distance is used to calculate the plot comprehensive change rate. First, the weight coefficient of each feature dimension is determined, and the weight is determined based on the variation coefficient of the feature in the historical data. The historical variation coefficient of the b-th feature is calculated wherein is the standard deviation of the feature in the historical data, is the historical mean; the feature weight is defined as The features with smaller variation coefficients obtain larger weights, because these features are relatively stable under normal circumstances, and their changes are more likely to reflect abnormal situations. The comprehensive change rate of plot a is The calculation formula is This change rate comprehensively reflects the change intensity of the plot in multiple feature dimensions.

[0095] On the basis of calculating the change rate, the relative change rate concept is introduced to eliminate the influence of the feature dimension. For each feature dimension, the relative change rate is calculated wherein is a small constant to prevent division by zero. The exemplary value can be 0.001.

[0096] The comprehensive relative change rate of the plot is : ;

[0097] Meanwhile, the directional change index is calculated, and the positive change rate is defined as :

[0098] ;

[0099] and the negative change rate is ;

[0100] The positive change rate and the negative change rate respectively represent the degree of increase and decrease of the feature value, and this separation processing is helpful for identifying different types of plot change patterns.

[0101] The plot feature change rate interval generated in step 3 is used to construct a dynamic threshold range for abnormality judgment. For the b-th feature of plot a, its predicted change rate interval is obtained from step 3 wherein is the lower bound of the change rate, is the upper bound of the change rate. The feature change rate abnormality degree is defined as when , when , when , . To prevent small constants from being zero. An exemplary value can be 0.01. The integrated change rate anomaly degree of the plot reflects the degree of deviation of the change rate of the plot from the expected range.

[0102] An abnormality determination mechanism based on multi-criteria fusion is constructed, and based on the change rate interval provided in step 3, an integrated abnormality determination criterion is set. When plot a meets any of the following conditions, plot a is marked as a potential abnormal plot:

[0103] (1) There are at least characteristic dimensions whose change rates exceed the prediction interval, that is , where is an indicator function; (2) the integrated change rate anomaly degree , where is an anomaly degree threshold; (3) the integrated change rate exceeds the integrated interval derived from all feature change rate intervals , where , .

[0104] This step realizes accurate quantification and abnormal identification of plot dynamic changes by establishing a change detection algorithm based on time series difference. The algorithm constructed in this step fully utilizes the feature change rate interval generated in step 3 as a dynamic threshold, so that the abnormality determination criterion can adapt to the historical change pattern and seasonal regularity of the plot. The multi-criteria fusion strategy considers different situations such as single-dimensional anomaly, multi-dimensional anomaly and integrated anomaly, improving the comprehensiveness of the detection. The dynamic threshold method based on prediction interval in this step has higher adaptability and lower false positive rate than the traditional fixed threshold method, providing reliable candidate targets for subsequent fine identification of plot state, and effectively improving the accuracy and timeliness of land resource monitoring and early warning.

[0105] Step 5: Construct a plot state recognition model based on a double-branch convolutional neural network, input the optical image and SAR echo of the potential abnormal plot into the plot state recognition model, identify the ground crop planting state, growth stage and ground illegal construction behavior, and determine whether the potential abnormal plot is a confirmed abnormal plot. The classification results and confidence scores of the confirmed abnormal plot are output. Referring to Figure 5 , the plot state recognition and abnormality discrimination flowchart for this step.

[0106] A plot state recognition model based on a double-branch convolutional neural network is constructed to fine identify and classify the potential abnormal plot marked in step 4. The optical image and SAR echo data of the potential abnormal plot are preprocessed, and the image area corresponding to the plot is cropped and adjusted to a uniform size.

[0107] ​Further, for the optical image, RGB three-channel data is extracted and normalized, mapping the pixel value from [0, 255] to [0, 1] interval, and the normalization formula is wherein is the original pixel value, is the normalized value. For the SAR echo data, the complex form of the echo signal is converted into an amplitude image , and the calculation formula is wherein I is the in-phase component, and Q is the quadrature component, and the amplitude value of the amplitude image is logarithmically transformed , is a small constant to prevent infinity in the logarithmic operation. Exemplarily, the value is 1e-10.

[0108] A double-branch convolutional neural network architecture is designed, and a double-branch structure convolutional neural network is used to process the optical image and the SAR data respectively. The double-branch structure convolutional neural network includes an optical image branch and a SAR data branch.

[0109] The optical image branch of the double-branch structure convolutional neural network includes multiple convolution modules, each module being composed of a convolution layer, a normalization layer and an activation layer. Exemplarily, the first convolution module uses 64 convolution kernels with a size of 7x7 and a step of 2 to extract low-level features. The subsequent convolution modules adopt a residual connection structure, including two 3x3 convolution layers, and the number of convolution kernels is 128, 256 and 512 in turn. The calculation formula of the residual connection is wherein is the input feature map, is the convolution operation, is the convolution weight, is the output feature map. The residual structure adopted can effectively alleviate the gradient disappearance problem of the deep network and improve the feature learning ability.

[0110] A multi-scale feature extraction layer is inserted between the convolution modules, and a dilated convolution is used to realize feature capture of different receptive fields. Three parallel dilated convolution branches are set, and exemplarily, the dilated rates , , are , , respectively, and the convolution kernel size is 3x3. The output feature maps of the three branches are spliced to form a multi-scale feature representation wherein represents the convolution output with a dilated rate of . Feature fusion is performed through a 1x1 convolution, which reduces the channel dimension while maintaining the expression ability of the features.

[0111] The SAR data branch of the double-branch convolutional neural network adopts a similar network structure to the optical image branch, but is optimized for the characteristics of SAR images. A polarization feature encoding module is added before the first convolutional layer to encode SAR data of different polarization modes into a feature tensor. For full-polarimetric SAR data, the covariance matrix is constructed where is the scattering vector of the th pixel, is the conjugate transpose, and the superscript indicates the conjugate transpose, is the total number of pixels. The eigenvalue decomposition of the covariance matrix is performed where U is the eigenvector matrix, is the conjugate transpose of the matrix , and is the eigenvalue diagonal matrix. The eigenvalues and eigenvectors are used as input channels to enhance the network's ability to utilize polarization information.

[0112] A feature fusion mechanism is constructed to adaptively fuse the features of the optical image branch and the SAR data branch. An attention mechanism is used to calculate the weights of the features of the two branches:

[0113] The optical feature weight is:

[0114]

[0115] The SAR feature weight is:

[0116]

[0117] where and are the feature vectors of the two branches, and are learnable weight parameters.

[0118] The fused feature is calculated by The adaptive fusion mechanism can dynamically adjust the contribution of each data source according to the characteristics of different plots.

[0119] A multi-task learning structure is designed in the classification layer of the network to simultaneously identify crop planting status, growth stage, and illegal building behavior. Three parallel classification heads are set up, each containing two fully connected layers and a softmax output layer.

[0120] The crop planting status classification head outputs the probability distribution of each crop class ​​wherein , is the logit value of the th crop, is the logit value of the th crop.

[0121] Growth stage classification head outputs probabilities of , is the predicted probability of the th class of the growth stage classification head. Violation detection head outputs binary classification results , representing probabilities of normal and violation, and are two class probabilities of the violation detection head, representing predicted probabilities of normal or violation, respectively, exemplarily, are set to 10 main crops, are set to 5 growth stages.

[0122] The training of the field state recognition model adopts a multi-task joint loss function, and the total loss wherein , , are cross-entropy losses of the three tasks, respectively, is an L2 regularization term, , , , are loss weight coefficients. The network is trained using a stochastic gradient descent optimizer with momentum.

[0123] For each potential abnormal field, its optical image and SAR data are input to obtain the prediction results of the three tasks. The comprehensive confidence score , is the information entropy of the prediction result of the th potential abnormal field, reflecting the uncertainty of the prediction, wherein is the th potential abnormal field. When the violation probability , it is directly determined as a certain abnormal field, wherein is the violation determination threshold. For crop-related abnormalities, when the predicted crop type does not match the historical record or the growth stage deviates from the normal range, and the confidence score , it is determined as a certain abnormal field, wherein is the confidence threshold.

[0124] The step realizes accurate identification and fine-grained classification of potential abnormal land blocks by constructing a land block state recognition model based on a double-branch convolutional neural network. The land block state recognition model fully utilizes the spectral texture information of optical images and the backscattering characteristics of SAR data, captures different levels of feature of ground objects through multi-scale feature extraction, and effectively integrates the complementary advantages of multi-source information through an adaptive feature fusion mechanism. The multi-task learning framework constructed can complete multiple recognition tasks at the same time, improving the efficiency and practicality of the model. The land block state recognition model constructed in this step can accurately distinguish between normal changes and real abnormalities, significantly reducing the false positive rate, and providing reliable technical support for accurate monitoring and timely warning of land resources.

[0125] Step 6: Spatial clustering analysis is performed on the determined abnormal land blocks to form ground abnormal regions, the accurate geographic boundaries of the ground abnormal regions are marked according to the three-dimensional digital model of the ground, and multi-level warning information is generated according to the classification results and confidence scores of the determined abnormal land blocks.

[0126] Spatial clustering analysis is performed on the determined abnormal land blocks identified in step 5, and the DBSCAN algorithm based on density is used to merge spatially adjacent and same-class abnormal land blocks into continuous abnormal regions. Set the spatial neighborhood radius and the minimum number of land blocks Two parameters, for any determined abnormal land block, calculate the set of land blocks in its neighborhood , where represents the Euclidean distance between the land block and the center of mass, is the radius distance representing the spatial neighborhood. When , mark as a core land block, and all land blocks in its neighborhood with the same abnormal class are included in the same cluster. By recursively expanding the neighborhood of the core land block, a spatially connected abnormal region is formed. For example, for farmland illegal construction detection, set to 50 meters, to 3 land blocks, to ensure that the merged region has a meaningful spatial scale. The density-based clustering method used in this step can identify abnormal regions of any shape, avoiding the shape restrictions of traditional methods.

[0127] The geographical boundary of the abnormal area is accurately calibrated by using the ground three-dimensional digital model. For each cluster formed abnormal area, the boundary point set of all land blocks contained is extracted. The convex hull algorithm is used to calculate the outer contour of the abnormal area, and then the Alpha-shape algorithm is used to optimize the boundary shape to generate a boundary line that is more consistent with the actual feature contour. The two-dimensional boundary line is projected onto the three-dimensional digital model, and the elevation information of the boundary points is obtained by interpolation to construct a three-dimensional boundary model of the abnormal area.

[0128] The geometric properties of the abnormal area are calculated, including the area , the perimeter , and the shape complexity index , where is the boundary point coordinate, is the number of boundary points, represents the serial number of the th point on the boundary of the abnormal area. These geometric properties are stored together with the geographical coordinates of the abnormal area to form a complete spatial information description of the abnormal area.

[0129] According to the classification results and confidence scores of the abnormal land blocks, a multi-level early warning mechanism is constructed, and a three-level early warning system is set: the first level of warning corresponds to high-risk abnormalities, the second level of warning corresponds to medium-risk abnormalities, and the third level of warning corresponds to low-risk abnormalities. The determination of the warning level considers the severity of the abnormal type and the recognition confidence, and defines the risk score , where is the severity score of the abnormal type, is the average confidence score, is the normalized area of the abnormal area, , , are the corresponding weight coefficients. For example, the of illegal construction behavior can be set to 0.9, the of not planting according to regulations can be set to 0.6, and the first level of warning is triggered when . For each warning level, a structured warning report is generated containing the abnormal location, type, area, confidence, and recommended disposal measures.

[0130] The multi-level early warning information is sorted according to priority, and high-level early warning is preferentially pushed to the regulatory unit. An early warning information data package is constructed, including vector boundary data of the abnormal area, key frame image data, abnormal attribute table and early warning report document. The abnormal area is superimposed and displayed on the basic map by using a geographic information system platform, and different colors are used to identify different early warning levels, for example, red represents first-level early warning, orange represents second-level early warning, and yellow represents third-level early warning. A time series change comparison chart is generated to show the evolution process of the abnormal area in different monitoring periods. At the same time, an early warning feedback mechanism is established to record the disposal results and actual checking conditions of each early warning, which is used for subsequent model optimization and threshold adjustment.

[0131] This step integrates discrete abnormal land blocks into abnormal areas with practical significance through spatial clustering and boundary demarcation, improving the practicality of the monitoring results; the construction of the three-dimensional boundary model fully utilizes the elevation information of the LiDAR data, realizing the stereoscopic expression of the abnormal area. The multi-level early warning mechanism constructed in this step differentiates management according to the risk level, so that the limited regulatory resources are reasonably allocated. The early warning process realizes a complete closed loop from data acquisition, intelligent analysis to decision support, providing scientific, efficient and precise technical support for dynamic monitoring of land resources, and significantly improving the informatization and intelligentization level of land resource management.

[0132] Embodiment 2

[0133] Reference Figure 6 For the second embodiment of the present application, a land resource dynamic monitoring and early warning system based on multi-source remote sensing data fusion is provided.

[0134] The system comprises a data acquisition module, a feature extraction module, a feature change rate prediction module, a potential abnormal land block identification module, an abnormal land block identification and classification module, and a hierarchical early warning module.

[0135] The data acquisition module periodically acquires optical images, SAR echoes and LiDAR point cloud data of the ground by using a UAV, generates a three-dimensional digital model of the ground based on the LiDAR point cloud data and divides the land blocks, and constructs a space-time registration algorithm to map the optical images and SAR echoes of the same period to the corresponding land blocks to generate a three-dimensional digital model of the ground.

[0136] The feature extraction module extracts the texture feature parameters of the optical images, the backscattering coefficient and polarization features of the SAR echoes, and the elevation statistics and surface roughness of the LiDAR point cloud for the divided land blocks, and generates a multi-dimensional feature vector of the land block by using a feature fusion algorithm.

[0137] The feature change rate prediction module constructs a plot feature evolution model based on a long short-term memory network, deduces a plot change rule in a current period based on a plot multi-dimensional feature vector in a historical period, and generates a plot feature change rate interval.

[0138] The potential abnormal plot identification module establishes a change detection algorithm based on time series difference, calculates a plot change rate in a current period according to a plot multi-dimensional feature vector, takes a plot feature change rate interval as a determination threshold range of a potential abnormal plot, and marks the potential abnormal plot according to the determination threshold range of the potential abnormal plot.

[0139] The abnormal plot identification and classification module constructs a plot state identification model based on a double-branch convolutional neural network, inputs optical images and SAR echoes of the potential abnormal plot into the plot state identification model, identifies a ground crop planting state, a growth stage and a ground illegal building behavior, determines whether the potential abnormal plot is a determined abnormal plot, and simultaneously outputs a classification result and a confidence score of the determined abnormal plot.

[0140] The hierarchical early warning module performs spatial clustering analysis on the determined abnormal plot to form a ground abnormal area, calibrates an accurate geographic boundary of the ground abnormal area according to a ground three-dimensional digital model, and generates multi-level early warning information according to the classification result and the confidence score of the determined abnormal plot.

[0141] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. The described device embodiments are merely illustrative. For example, the division of the units is only a logical function division, and other division manners can be used in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0142] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above specific embodiments, and the above specific embodiments are merely illustrative, rather than limiting. A person of ordinary skill in the art can make changes, modifications, replacements and variations to the above embodiments under the inspiration of the present application without departing from the scope of the present application and the protection scope of the claims, and these are all within the protection scope of the present application.

Claims

1. A method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion, characterized in that, include: By periodically collecting ground optical images, SAR echoes and LiDAR point cloud data by UAVs, a ground 3D digital model is generated based on the LiDAR point cloud data and the land parcels are divided. A spatiotemporal registration algorithm is constructed to map the ground optical images and SAR echoes acquired in the same period to the corresponding land parcels, thereby generating a ground 3D digital model. For the divided plots, texture feature parameters of optical images are extracted, backscattering coefficient and polarization features of SAR echoes are extracted, and elevation statistics and surface roughness of LiDAR point clouds are extracted. Multidimensional feature vectors of plots are generated through feature fusion algorithms. Construct a land parcel feature evolution model based on a long short-term memory network, infer the current period's land parcel change pattern based on historical periodic multidimensional feature vectors of land parcels, and generate a range of land parcel feature change rates; A change detection algorithm based on time-series difference is established. The change rate of land parcels in the current period is calculated based on the multi-dimensional feature vector of land parcels. The range of land parcel feature change rate is used as the judgment threshold range of potential abnormal land parcels. Potential abnormal land parcels are marked according to the judgment threshold range of potential abnormal land parcels. A land parcel status identification model based on a dual-branch convolutional neural network is constructed. The optical images and SAR echoes of potential anomalous land parcels are input into the land parcel status identification model to identify the planting status and growth stage of crops on the ground and illegal construction behavior on the ground. The model determines whether potential anomalous land parcels are confirmed anomalous land parcels and outputs the classification results and confidence scores of confirmed anomalous land parcels. A land parcel status recognition model based on a dual-branch convolutional neural network is constructed to identify whether potential abnormal land parcels are confirmed abnormal land parcels. The dual-branch convolutional neural network processes optical images and SAR data separately. The optical image branch adopts a residual connection structure and inserts a multi-scale feature extraction layer composed of multiple parallel dilated convolutions to capture features of different receptive fields. The SAR data branch adds polarization feature coding to the input end of the optical image branch network structure; A feature fusion mechanism based on attention mechanism is constructed to adaptively calculate the weights of optical and SAR features and perform weighted fusion. A multi-task learning structure is designed at the end of a dual-branch convolutional neural network, and parallel classification heads are set up to simultaneously identify crop planting status, growth stage and illegal construction behavior on the ground. Based on the prediction results of the land parcel status identification model for the identification task, a comprehensive confidence score is calculated. Combined with the preset threshold for illegal construction and the confidence threshold for crop anomalies, the potential abnormal land parcels are determined to be abnormal land parcels, and the classification results are output. Spatial clustering analysis is performed on identified anomalous land parcels to form ground anomaly regions. The geographical boundaries of these regions are then determined using a 3D digital model of the ground. Finally, multi-level early warning information is generated based on the classification results and confidence scores of the identified anomalous land parcels.

2. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 1, characterized in that, The drone is equipped with an integrated optical camera, synthetic aperture radar and lidar scanner to collect optical images, SAR echo and LiDAR point cloud data of the monitored area. Based on the collected LiDAR point cloud data, after noise removal and point cloud thinning preprocessing, and after separating ground points from non-ground points, an irregular triangular mesh interpolation method is used to generate a ground 3D digital model. A watershed algorithm based on slope and curvature, combined with edge detection results from optical imagery, is used to divide and encode the 3D digital model of the ground. A spatiotemporal registration algorithm is constructed to perform geometric correction and registration of optical images and SAR echoes and map them to the corresponding land parcels.

3. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 2, characterized in that, Optical images are converted into grayscale images. The grayscale co-occurrence matrix method is used to calculate and extract energy, contrast, correlation and entropy at a set distance and in multiple directions. The feature values ​​in multiple directions are averaged to obtain feature parameters. Radiometric calibration and speckle noise suppression were performed on the SAR echoes. The backscattering coefficient was calculated, and a scattering matrix was constructed from the SAR echo data. The polarization decomposition method was used to extract the components characterizing surface scattering, dihedral scattering, and volume scattering. At the same time, the polarization entropy and the average scattering angle were calculated.

4. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 3, characterized in that, For the point cloud data within each plot, calculate the elevation statistics parameters, including the average elevation, elevation standard deviation, maximum elevation, and minimum elevation; Surface roughness is calculated by dividing the point cloud within the plot into a grid and using a local plane fitting method. After extracting the feature parameters from each data source, a feature fusion algorithm is used to normalize all the extracted feature parameters. The normalized optical texture features, SAR polarization features, and LiDAR elevation features are then spliced ​​together in a predetermined order to generate a multi-dimensional feature vector of the land parcel.

5. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 4, characterized in that, A land parcel feature evolution model was constructed based on a long short-term memory network; The multidimensional feature vectors of land parcels in historical cycles are organized into a time series, and the feature change rate series between adjacent weeks are calculated and standardized as input to the land parcel feature evolution model. The architecture of a Long Short-Term Memory (LSTM) network is designed, including an input layer, a fully connected layer, and an output layer. The input layer includes an LSTM layer with a forget gate, an input gate, and an output gate. The output layer is used to output the predicted value of the feature change rate for the next cycle and an uncertainty estimate for quantifying the prediction confidence.

6. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 5, characterized in that, Based on the predicted value of the feature change rate and the uncertainty estimate of the plot feature evolution model, a change rate confidence interval is constructed for each dimension feature of the plot, forming a multidimensional feature change rate interval for the plot.

7. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 6, characterized in that, Obtain the multidimensional feature vectors of land parcels for the current and historical monitoring periods, calculate the feature difference vectors, and calculate the land parcel change rate for the current period; The weights are determined based on the coefficients of variation of each feature dimension in historical data, and the overall change rate of the land parcel is calculated using weighted Euclidean distance. Calculate the overall relative change rate of the land parcel, as well as the positive and negative change rates that characterize the degree of increase and decrease in the feature value.

8. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 7, characterized in that, Based on the range of land parcel feature change rate, a dynamic judgment threshold range is set. When the calculated land parcel feature change rate exceeds the dynamic judgment threshold range, the abnormality of the land parcel feature deviating from the prediction range is calculated and summarized into the comprehensive change rate abnormality of the land parcel. A multi-criteria anomaly detection mechanism is constructed. When a plot of land meets any of the conditions of the anomaly detection mechanism, the plot is marked as a potential anomaly plot. The anomaly detection mechanism includes: the existence of a preset number of feature dimensions whose rate of change exceeds the prediction interval; the anomaly degree of the comprehensive rate of change of the land parcel being greater than a preset threshold; and the comprehensive rate of change of the land parcel exceeding the comprehensive rate of change range derived from the boundaries of each dimension of the prediction interval.

9. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 8, characterized in that, Spatial clustering analysis was performed on the identified anomalous land parcels. A density-based clustering algorithm was used to merge spatially adjacent land parcels with the same anomalous category into continuous ground anomalous areas. Using a 3D digital model of the ground, the geographic boundaries of the anomaly region are identified and the geometric properties of the anomaly region are calculated through the convex hull algorithm and the Alpha-shape algorithm.

10. The method for dynamic monitoring and early warning of land resources based on multi-source remote sensing data fusion according to claim 9, characterized in that, A multi-level early warning mechanism is constructed. Risk scores are calculated based on the severity of the anomaly type, the confidence level of identification, and the area to determine the early warning level, and a structured early warning report containing information on the location and type of the anomaly is generated.

11. A land resource dynamic monitoring and early warning system based on multi-source remote sensing data fusion, used to implement the land resource dynamic monitoring and early warning method based on multi-source remote sensing data fusion as described in any one of claims 1-10, characterized in that, include: The system includes a data acquisition module, a feature extraction module, a feature change rate prediction module, a potential abnormal land parcel identification module, an abnormal land parcel identification and classification module, and a graded early warning module. The data acquisition module periodically collects ground optical images, SAR echoes, and LiDAR point cloud data through UAVs. Based on the LiDAR point cloud data, it generates a ground three-dimensional digital model and divides the land into plots. It constructs a spatiotemporal registration algorithm to map the ground optical images and SAR echoes acquired in the same period to the corresponding plots, thereby generating a ground three-dimensional digital model. The feature extraction module extracts texture feature parameters from optical images, backscattering coefficients and polarization features from SAR echoes, and elevation statistics and surface roughness from LiDAR point clouds for the divided land parcels, and generates multidimensional feature vectors for the land parcels through feature fusion algorithms. The feature change rate prediction module constructs a land parcel feature evolution model based on a long short-term memory network, infers the current period's land parcel change pattern based on historical periodic multi-dimensional feature vectors of land parcels, and generates a land parcel feature change rate interval. The potential abnormal land parcel identification module establishes a change detection algorithm based on time-series difference, calculates the land parcel change rate of the current period based on the multi-dimensional feature vector of the land parcel, uses the range of land parcel feature change rate as the judgment threshold range of potential abnormal land parcels, and marks potential abnormal land parcels according to the judgment threshold range of potential abnormal land parcels. The abnormal land parcel identification and classification module constructs a land parcel status identification model based on a dual-branch convolutional neural network. It inputs the optical images and SAR echoes of potential abnormal land parcels into the land parcel status identification model to identify the planting status, growth stage of ground crops, and illegal construction behavior on the ground. It determines whether potential abnormal land parcels are confirmed abnormal land parcels and outputs the classification results and confidence scores of confirmed abnormal land parcels. The hierarchical early warning module performs spatial clustering analysis on identified abnormal land parcels to form ground anomaly areas, marks the geographical boundaries of the ground anomaly areas based on the ground three-dimensional digital model, and generates multi-level early warning information based on the classification results and confidence scores of the identified abnormal land parcels.

Images