A Remote Sensing Data-based Cultivated Land Use Change Detection Method Using an Improved Intelligent Algorithm

Abstract

The present invention discloses a method for detecting changes in cultivated land use of remote sensing data based on an improved intelligent algorithm, including S1, obtaining remote sensing images covering the target cultivated land area; S2, preprocessing the collected remote sensing data; S3, using the improved intelligent algorithm to extract features from the preprocessed remote sensing data to generate a multi-dimensional feature matrix; S4, using a dynamic spatio-temporal graph convolutional network to perform temporal analysis on the extracted feature information; S5, based on the results of the temporal analysis, identifying the expansion, reduction and changes in the use type of cultivated land; S6, generating a change detection result; S7, generating maps, charts and interactive visualization models; S8, generating a detailed change report and providing future change trend analysis in combination with historical data and prediction models. The present invention significantly improves the accuracy and efficiency of detecting changes in cultivated land use, and provides comprehensive result visualization and decision support functions.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing data processing and analysis, and particularly to a method for detecting changes in cultivated land use of remote sensing data based on an improved intelligent algorithm. Background Art

[0002] With the rapid development of remote sensing technology, remote sensing data has been more and more widely used in fields such as agriculture, environmental monitoring, and land use. Detection of changes in cultivated land use is one of the important applications of remote sensing data. By analyzing remote sensing data, the change situation of cultivated land use can be monitored, providing an important basis for agricultural management, land resource planning, and environmental protection. However, when dealing with large-scale remote sensing data, the existing methods for detecting changes in cultivated land use often face problems such as insufficient accuracy and low efficiency, which are mainly manifested in the following aspects.

[0003] Firstly, traditional methods for detecting changes in cultivated land use usually rely on manual annotation and simple image processing techniques. These methods are difficult to accurately identify and monitor subtle changes in cultivated land use when dealing with large-scale remote sensing data. For example, traditional image difference methods and classification methods are easily affected by factors such as illumination and season when facing complex cultivated land use change scenarios, resulting in low detection accuracy. In addition, manual annotation requires a large amount of manpower and time investment, making it difficult to meet the requirements of real-time monitoring and rapid response.

[0004] Secondly, when dealing with a large amount of data, the existing methods for detecting changes in cultivated land use have a slow processing speed and are difficult to cope with the rapidly changing cultivated land use situation. Traditional image processing methods often need to process large-scale remote sensing data frame by frame, with high computational complexity and long processing time. For example, methods based on pixel-level classification need to classify each pixel, with a huge amount of calculation and low processing efficiency. This is unacceptable in practical applications, especially in scenarios that require real-time monitoring and rapid response.

[0005] Thirdly, the existing methods for detecting changes in cultivated land use lack comprehensive result visualization and decision support functions. In the process of detecting changes in cultivated land use, only the detection results are not enough. Users (such as agricultural managers, land resource planners, and environmental protection agencies) need intuitive result displays and detailed change reports to understand the detection results and make scientific decisions based on them. However, traditional methods usually only provide simple detection results and lack powerful visualization and decision support functions, making it difficult to effectively utilize the detection results. For example, the change detection results generated by traditional methods are usually static images or tables, which cannot intuitively display information such as the specific location, change type, and change amount of the change area, and it is difficult for users to make timely and effective decisions based on these results.

[0006] In summary, existing cultivated land use change detection methods have many deficiencies in terms of accuracy, efficiency, and result visualization. With the development of remote sensing technology and the continuous increase in data volume, there is an urgent need for a cultivated land use change detection method based on improved intelligent algorithms to improve the accuracy and efficiency of detection, provide comprehensive result visualization and decision support functions, and meet the needs of practical applications. The present invention combines remote sensing technology, intelligent algorithms, and geographic information system technology to propose a new cultivated land use change detection method, aiming to solve the above technical problems, improve the accuracy and efficiency of detection, and provide comprehensive result visualization and decision support functions. Summary of the Invention

[0007] An object of the present invention is to propose a remote sensing data cultivated land use change detection method based on improved intelligent algorithms, which significantly improves the accuracy and efficiency of cultivated land use change detection and provides comprehensive result visualization and decision support functions.

[0008] A remote sensing data cultivated land use change detection method based on improved intelligent algorithms according to an embodiment of the present invention includes the following steps:

[0009] S1. Obtain remote sensing images covering the target cultivated land area through remote sensing satellites, unmanned aerial vehicles, or other remote sensing devices, and the remote sensing images include multispectral images and radar images;

[0010] S2. Perform image correction, noise removal, geometric correction, and image registration on the collected remote sensing data;

[0011] S3. Use improved intelligent algorithms to extract features from the preprocessed remote sensing data and generate a multi-dimensional feature matrix;

[0012] S4. Use a dynamic spatio-temporal graph convolutional network to perform temporal analysis on the extracted feature information, and detect the changes in cultivated land use by comparing remote sensing image data in different time periods;

[0013] S5. Based on the results of temporal analysis, use an integrated adversarial generation network and a multi-task learning model to detect and classify cultivated land use changes, and identify the expansion, reduction, and changes in the use type of cultivated land;

[0014] S6. Generate change detection results, including the specific location, change type, and change amount information of the change area;

[0015] S7. Use augmented reality technology and geographic information system technology to visually display the detection results and generate maps, charts, and interactive visualization models;

[0016] S8. Generate a detailed change report, including the specific location of the change area, the type of change, the amount of change, and time period information, and provide an analysis of future change trends by combining historical data and prediction models.

[0017] Optionally, the S1 includes:

[0018] S11. Obtain multispectral images covering the target cultivated area through optical remote sensing satellites and multispectral imaging satellites. The multispectral images include image data in the visible light band, near-infrared band, and mid-infrared band.

[0019] S12. Obtain multispectral images covering the target cultivated area through a multispectral sensor carried by a drone. The multispectral sensor has an optical resolution of 0.5 meters.

[0020] S13. Obtain synthetic aperture radar images covering the target cultivated area through remote sensing satellites.

[0021] S14. Obtain radar images covering the target cultivated area through a radar sensor carried by a drone. The radar sensor can penetrate clouds and vegetation to obtain information on the ground surface.

[0022] S15. The obtained multispectral images and radar images form remote sensing images, which are transmitted to a ground processing station through a satellite communication link and a wireless communication link.

[0023] S16. The data format of the remote sensing images includes standard formats such as GeoTIFF, HDF, and NetCDF.

[0024] Optionally, the S2 includes:

[0025] S21. Perform image correction on the collected multispectral images. The image correction includes radiometric correction and atmospheric correction. Radiometric correction is achieved by converting the radiation values recorded by the sensor into surface reflectance, and atmospheric correction is achieved by eliminating the effects of atmospheric scattering and absorption.

[0026] S22. Remove noise from the collected radar images, and process the radar images using Lee filters, Frost filters, and Gamma-Map filters.

[0027] S23. Perform correction of geometric distortion and conversion of geographic coordinate systems on the multispectral images and radar images. Use terrain correction and orthorectification methods to correct the remote sensing images to a unified geographic coordinate system.

[0028] S24. Perform image registration on the multispectral images and radar images using feature point matching and region-based matching methods, and perform spatial alignment on remote sensing images obtained at different time periods and by different sensors to ensure the pixel correspondence relationship between images.

[0029] S25. Perform data fusion on the calibrated and registered multi-spectral image I multispectral and the radar image I radar using a method based on weighted average and principal component analysis to generate a comprehensive remote sensing image I with more information fused :

[0030] I fused = α·I multispectral + β·I radar ;

[0031] Wherein, I fused is the fused remote sensing image, Ix ultispectral and I radar are the multi-spectral image and the radar image respectively, and α and β are weight coefficients.

[0032] Optionally, the S3 includes:

[0033] S31. Perform multi-level convolutional transformation on the preprocessed remote sensing image I fused data to construct an improved multi-layer convolutional neural network model. The convolutional kernel design of the improved multi-layer convolutional neural network model is optimized to an adaptive shape convolutional kernel, and the shape and size of the convolutional kernel are dynamically adjusted according to the content of the input remote sensing image I fused data feature map:

[0034] Y i，j，k = ∑ m，n X i+f(m)，j+f(n) ·W m，n，k + b k ;

[0035] Wherein, Y i，j，k is the output feature value of the kth convolutional kernel at the position (i, j), X i+f(m)，j+f(n) is the pixel value of the input feature map at the position (i + f(m), j + f(n)), W m,n，k is the weight of the kth convolutional kernel at the position (m, n), b k is the bias of the kth convolutional kernel, and the function f is used to dynamically adjust the shape and size of the convolutional kernel;

[0036] S32. After each convolutional operation, perform activation processing using an adaptive Swish activation function:

[0037] T(x) = x·sigmoid(βx);

[0038] Wherein, f(x) is the activation function, x is the input value, and β is a learnable parameter that can be automatically adjusted according to the training process;

[0039] S33. An improved pooling operation is performed using a multi-scale max pooling method to reduce the size of the feature map and retain the main feature information:

[0040]

[0041] where P i，j，k is the value of the pooled feature map at position (i, j), is the value of the convolutional output feature map within the multi-scale pooling window;

[0042] S34. After the convolutional transformation processing and the pooling operation, a self-attention mechanism is used to perform adaptive feature enhancement processing for the characteristics of the remote sensing image I fused data:

[0043]

[0044] where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector;

[0045] S35. The feature map after the adaptive feature enhancement processing is flattened to generate a multi-dimensional feature matrix. The dimension of the feature matrix is determined by the output feature map of the convolutional neural network and the result of the adaptive feature enhancement processing. The flattening operation converts the three-dimensional feature map into a two-dimensional matrix. The supporting formula for the flattening operation is:

[0046] F flattened =

[0047] Flatten(Attention(MultiScalePool(Swish(Conv(I fused , W, b)))));

[0048] where F flattened is the flattened multi-dimensional feature matrix, I fused is the fused remote sensing image, Conv(I fused , W, b) is the convolutional operation, Swish(·) is the adaptive Swish activation function, MultiScalePo0l(·) is the multi-scale max pooling operation, Attention(·) is the adaptive feature enhancement processing, and Flatten(·) is the flattening operation.

[0049] Optionally, the S4 includes:

[0050] S41. Using the multi-dimensional feature matrix F flattened as the input, a dynamic spatio-temporal graph convolutional network including a temporal convolutional layer and a graph convolutional layer is constructed;

[0051] S42. In the temporal convolutional layer, perform temporal convolutional operations on the input feature matrix using one-dimensional convolution to extract temporal features:

[0052]

[0053] Among them, T i，t is the value of the temporal convolution output at time step t and feature dimension i, F i，t+j is the value of the input feature matrix at time step t + j and feature dimension i, W j is the weight of the temporal convolution kernel, and k is the radius of the temporal convolution kernel;

[0054] S43. In the graph convolutional layer, perform spatial convolutional operations on the features after temporal convolution using graph convolution to extract spatial features:

[0055]

[0056] Among them, G i，t is the value of the graph convolution output at node i and time step t, N(i) is the set of neighbor nodes of node i, A ij is the adjacency matrix value between node i and node j, d i and d j are the degrees of node i and node j respectively, and T j，t is the feature value after temporal convolution;

[0057] S44. Perform activation processing on the output of the graph convolutional layer;

[0058] S45. Fuse the outputs of the temporal convolutional layer and the graph convolutional layer to construct a dynamic spatio-temporal feature matrix D st :

[0059] D st = concat(T, G);

[0060] Among them, D st is the dynamic spatio-temporal feature matrix, T is the output of the temporal convolutional layer, G is the output of the graph convolutional layer, and concat represents the concatenation operation in the feature dimension;

[0061] S46. Based on the dynamic spatio-temporal feature matrix D st , by comparing the remote sensing image data in different time periods, detect the change situation of cultivated land use, and calculate the change detection matrix C:

[0062]

[0063] Among them, C is the change detection matrix, and are the dynamic spatio-temporal feature matrices at time steps t1 and t2 respectively.

[0064] Optionally, S5 includes:

[0065] S51. Based on the change detection matrix C, construct an integrated adversarial generation network model and a multi-task learning model to detect and classify the changes in cultivated land use;

[0066] S52. In the integrated adversarial generation network model, the generator generates the change detection result, and the discriminator discriminates between the generated change detection result and the real data. The objective function of the generator is:

[0067]

[0068] where D(x) is the discrimination result of the discriminator for the real data x, G(z) is the change detection result generated by the generator, and z is the input noise of the generator;

[0069] S53. In the multi-task learning model, define multiple tasks, including cultivated land expansion detection task, cultivated land reduction detection task, and utilization type change detection task. The loss function is the weighted sum of the multi-task loss functions:

[0070] L MTL = αL expansion + βL reduction + γL type_change};

[0071] where L expansion is the loss of the cultivated land expansion detection task, L reduction is the loss of the cultivated land reduction detection task, L type_change is the loss of the utilization type change detection task, and α, β, and γ are the weight coefficients of the loss functions;

[0072] S54. During the training process, alternately optimize the generator and discriminator of the integrated adversarial generation network model, train the multi-task learning model, and improve the accuracy of detection and classification by minimizing the multi-task loss function:

[0073]

[0074] where θ G , θ D , θ MTL are the parameters of the generator, discriminator, and multi-task learning model respectively, and η is the learning rate;

[0075] S55. After the training is completed, use the optimized integrated adversarial generation network model and multi-task learning model to detect and classify the input change detection matrix C, identify the expansion, reduction, and utilization type changes of cultivated land, and output the detection and classification results.

[0076] Optionally, S6 includes:

[0077] S61. Based on the generated detection and classification results, construct a change detection result matrix R. The change detection result matrix R includes the specific location of the change area, the change type, and the change amount information.

[0078] S62. Project the change detection result matrix R onto the geographic coordinate system to generate the geographical location of the change area. The geographical location of the change area is jointly determined by the position coordinates in the change detection matrix C and the change type in the classification result.

[0079] S63. Precise position the specific location of the change area, and use the interpolation algorithm to smooth the boundary of the change area.

[0080] S64. Calculate the area and change amount of the change area. The change amount includes the cultivated land expansion amount, the cultivated land reduction amount, and the utilization type change amount. The calculation formula for the change amount is:

[0081] ΔA = A t2 - A t1 :

[0082] where ΔA is the change amount, and A t1 and A t2 are the areas of the change area at time steps t1 and t2 respectively.

[0083] S65. Generate a change detection report. The change detection report includes the specific location of the change area, the change type, the change amount, and the time period information. The format of the change detection report is:

[0084] Change detection report = {Location: (x, y), Change type: Type, Change amount: ΔA, Time period: [t1, t2]};

[0085] where (x, y) are the specific position coordinates of the change area, Type is the change type, and [t1, t2] is the time period.

[0086] S66. Store the generated change detection report in the database and provide a query interface for users to access and analyze.

[0087] The beneficial effects of the present invention are:

[0088] (1) The present invention uses an intelligent algorithm that combines multi-level convolutional transformation and adaptive feature enhancement to extract features from preprocessed remote sensing data, generating a multi-dimensional feature matrix. It performs temporal analysis using a dynamic spatio-temporal graph convolutional network. By comparing remote sensing image data from different time periods, it detects changes in cultivated land use. Based on the results of the temporal analysis, an integrated adversarial generation network and a multi-task learning model are used to detect and classify changes in cultivated land use, accurately identifying the expansion, reduction, and changes in the type of cultivated land use. The improvement measures significantly improve the accuracy of change detection and reduce false alarms and missed detections.

[0089] (2) The present invention optimizes data processing and preprocessing steps to ensure the high quality of input data, improving the reliability of detection results. The improved intelligent algorithm can quickly process large-scale remote sensing data, significantly enhancing the efficiency of data processing and analysis. In addition, by using a dynamic spatio-temporal graph convolutional network and an integrated adversarial generation network, it can complete the detection and classification of changes in cultivated land use in a relatively short time, meeting the requirements of real-time monitoring and rapid response. Description of the Drawings

[0090] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0091] Figure 1 It is a flowchart of a method for detecting changes in cultivated land use of remote sensing data based on an improved intelligent algorithm proposed by the present invention. Detailed Embodiments

[0092] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0093] Refer to Figure 1 , a method for detecting changes in cultivated land use of remote sensing data based on an improved intelligent algorithm, includes the following steps:

[0094] S1. Obtain remote sensing images covering the target cultivated land area through remote sensing satellites, unmanned aerial vehicles, or other remote sensing devices. The remote sensing images include multispectral images and radar images;

[0095] S2. Perform image correction, noise removal, geometric correction, and image registration on the collected remote sensing data;

[0096] S3. Use an improved intelligent algorithm to extract features from the preprocessed remote sensing data, generating a multi-dimensional feature matrix;

[0097] S4. Use a dynamic spatio-temporal graph convolutional network to perform temporal analysis on the extracted feature information, and detect the changes in cultivated land use by comparing the remote sensing image data in different time periods;

[0098] S5. Based on the results of the temporal analysis, use an integrated adversarial generation network and a multi-task learning model to detect and classify the changes in cultivated land use, and identify the expansion, reduction, and changes in the use type of cultivated land;

[0099] S6. Generate change detection results, including the specific location, change type, and change amount information of the change area;

[0100] S7. Use augmented reality technology and geographic information system technology to visually display the detection results, and generate maps, charts, and interactive visualization models;

[0101] S8. Generate a detailed change report, including the specific location, change type, change amount, and time period information of the change area, and provide future change trend analysis by combining historical data and prediction models.

[0102] In this embodiment, S1 includes:

[0103] S11. Obtain multi-spectral images covering the target cultivated land area through optical remote sensing satellites and multi-spectral imaging satellites. The multi-spectral images include image data in visible light bands, near-infrared bands, and mid-infrared bands;

[0104] S12. Obtain multi-spectral images covering the target cultivated land area through a multi-spectral sensor carried by a drone. The multi-spectral sensor has an optical resolution of 0.5 meters;

[0105] S13. Obtain synthetic aperture radar images covering the target cultivated land area through remote sensing satellites;

[0106] S14. Obtain radar images covering the target cultivated land area through a radar sensor carried by a drone. The radar sensor can penetrate clouds and vegetation to obtain surface information;

[0107] S15. The obtained multi-spectral images and radar images constitute remote sensing images, which are transmitted to the ground processing station through satellite communication links and wireless communication links;

[0108] S16. The data formats of the remote sensing images include GeoTIFF, HDF, and NetCDF standard formats.

[0109] In this embodiment, S2 includes:

[0110] S21. Perform image correction on the collected multi-spectral images. The image correction includes radiometric correction and atmospheric correction. The radiometric correction is achieved by converting the radiation values recorded by the sensor into surface reflectance, and the atmospheric correction is achieved by eliminating the effects of atmospheric scattering and absorption.

[0111] S22. Remove noise from the collected radar images. Use the Le e filter, Frost filter, and Gamma-Map filter to process the radar images.

[0112] S23. Perform geometric distortion correction and geographic coordinate system conversion on the multi-spectral images and radar images. Use terrain correction and orthorectification methods to correct the remote sensing images to a unified geographic coordinate system.

[0113] S24. Use feature point matching and region-based matching methods to register the multi-spectral images and radar images, and perform spatial alignment on the remote sensing images obtained at different times and by different sensors to ensure the pixel correspondence between the images.

[0114] S25. Perform data fusion on the corrected and registered multi-spectral image I multisprctral and radar image I radar using a method based on weighted average and principal component analysis to generate a comprehensive remote sensing image I fused :

[0115] I fused = α·I multispectral + β·I radar ;

[0116] where, I fused is the fused remote sensing image, I multisprctral and I radar are the multi-spectral image and the radar image respectively, and α and β are weight coefficients.

[0117] In this embodiment, S3 includes:

[0118] S31. Perform multi-level convolutional transformation processing on the preprocessed remote sensing image I fused data to construct an improved multi-layer convolutional neural network model. The convolutional kernel of the improved multi-layer convolutional neural network model is designed and optimized as an adaptive shape convolutional kernel, and the shape and size of the convolutional kernel are dynamically adjusted according to the content of the input remote sensing image I fused data feature map:

[0119] Y i，j，k = ∑ m，n X i+f(m)，j+f(n) ·W m，n，k + b k ;

[0120] Among them, Y i，j，k is the output eigenvalue of the k-th convolution kernel at position (i, j), and X i+f(m)，j+f(n) is the pixel value of the input feature map at position (i + f(m), j + f(n)), W m,n，k is the weight of the k-th convolution kernel at position (m, n), b k is the bias of the k-th convolution kernel, and the function f is used to dynamically adjust the shape and size of the convolution kernel;

[0121] S32. After each layer of convolution operation, an adaptive Swish activation function is used for activation processing:

[0122] T(x) = x · sigmoid(βx);

[0123] Among them, f(x) is the activation function, x is the input value, and β is a learnable parameter that can be automatically adjusted according to the training process;

[0124] S33. An improved pooling operation is performed using a multi-scale max pooling method to reduce the size of the feature map and retain the main feature information:

[0125]

[0126] Among them, P i，j，k is the value of the pooled feature map at position (i, j), is the value of the convolution output feature map within the multi-scale pooling window;

[0127] S34. After the convolution transformation processing and the pooling operation, a self-attention mechanism is used for adaptive feature enhancement processing for the remote sensing image I fused data characteristics:

[0128]

[0129] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector;

[0130] S35. The feature map after the adaptive feature enhancement processing is flattened to generate a multi-dimensional feature matrix. The dimension of the feature matrix is determined by the output feature map of the convolutional neural network and the result of the adaptive feature enhancement processing. The flattening operation converts the three-dimensional feature map into a two-dimensional matrix. The supporting formula for the flattening operation is:

[0131] F flattened =

[0132] Flatten(Attention(MultiScalePool(Swish(Conv(I fused, W, b)))))

[0133] Among them, F flattened is the flattened multi-dimensional feature matrix, I fused is the fused remote sensing image, Conv(I fused , W, b) is the convolution operation, Swish(·) is the adaptive Swish activation function, MultiScalePo0l(·) is the multi-scale max pooling operation, Attention(·) is the adaptive feature enhancement process, and Flatten(·) is the flattening operation.

[0134] In this embodiment, S4 includes:

[0135] S41. Using the multi-dimensional feature matrix F flattened as the input, construct a dynamic spatio-temporal graph convolutional network including a temporal convolutional layer and a graph convolutional layer;

[0136] S42. In the temporal convolutional layer, use one-dimensional convolution to perform temporal convolution operations on the input feature matrix to extract temporal features:

[0137]

[0138] Among them, T i，t is the value of the temporal convolution output at time step t and feature dimension i, F i，t+j is the value of the input feature matrix at time step t + j and feature dimension i, W j is the weight of the temporal convolution kernel, and k is the radius of the temporal convolution kernel;

[0139] S43. In the graph convolutional layer, use graph convolution to perform spatial convolution operations on the features after temporal convolution to extract spatial features:

[0140]

[0141] Among them, G i，t is the value of the graph convolution output at node i and time step t, N(i) is the set of neighbor nodes of node i, A ij is the adjacency matrix value between node i and node j, d i and d j are the degrees of node i and node j respectively, and T j，t is the feature value after temporal convolution;

[0142] S44. Perform activation processing on the output of the graph convolutional layer;

[0143] S45. Fuse the outputs of the temporal convolutional layer and the graph convolutional layer to construct a dynamic spatio-temporal feature matrix D st :

[0144] Dst = concat(T, G);

[0145] Among them, D st is the dynamic spatio-temporal feature matrix, T is the output of the time convolutional layer, G is the output of the graph convolutional layer, and concat represents the concatenation operation in the feature dimension;

[0146] S46. Based on the dynamic spatio-temporal feature matrix D st , by comparing the remote sensing image data in different time periods, detect the change situation of cultivated land use, and calculate the change detection matrix C:

[0147]

[0148] Among them, C is the change detection matrix, and are the dynamic spatio-temporal feature matrices at time steps t1 and t2 respectively.

[0149] In this embodiment, S5 includes:

[0150] S51. Based on the change detection matrix C, construct an integrated adversarial generation network model and a multi-task learning model to detect and classify the changes in cultivated land use;

[0151] S52. In the integrated adversarial generation network model, the generator generates the change detection result, and the discriminator discriminates the generated change detection result and the real data. The objective function of the generator is:

[0152]

[0153] Among them, D(x) is the discrimination result of the discriminator for the real data x, G(z) is the change detection result generated by the generator, and z is the input noise of the generator;

[0154] S53. In the multi-task learning model, define multiple tasks, including cultivated land expansion detection task, cultivated land reduction detection task, and utilization type change detection task. The loss function is the weighted sum of the multi-task loss functions:

[0155] L MTL = αL expansion + βL reduction + γL type_change};

[0156] Among them, L expansion is the loss of the cultivated land expansion detection task, L reduction is the loss of the cultivated land reduction detection task, L type_change is the loss of the utilization type change detection task, and α, β, and γ are the weight coefficients of the loss functions;

[0157] S54. During the training process, alternately optimize the generator and discriminator of the integrated adversarial generation network model, train the multi-task learning model, and improve the accuracy of detection and classification by minimizing the multi-task loss function:

[0158]

[0159] Among them, θ G , θ D , θ MTL are the parameters of the generator, discriminator, and multi-task learning model respectively, and η is the learning rate;

[0160] S55. After the training is completed, use the optimized integrated adversarial generation network model and multi-task learning model to detect and classify the input change detection matrix C, identify the expansion, reduction of cultivated land, and changes in utilization types, and output the detection and classification results.

[0161] In this embodiment, S6 includes:

[0162] S61. Based on the generated detection and classification results, construct a change detection result matrix R, and the change detection result matrix R includes the specific location, change type, and change amount information of the change area;

[0163] S62. Project the change detection result matrix R onto the geographic coordinate system to generate the geographical location of the change area, and the geographical location of the change area is jointly determined by the position coordinates in the change detection matrix C and the change type in the classification result;

[0164] S63. Precisely locate the specific location of the change area, and use the interpolation algorithm to smooth the boundary of the change area;

[0165] S64. Calculate the area and change amount of the change area. The change amount includes the cultivated land expansion amount, cultivated land reduction amount, and utilization type change amount. The calculation formula for the change amount is:

[0166] ΔA = A t2 -A t1 :

[0167] Among them, ΔA is the change amount, and A t1 and A t2 are the change area areas at time steps t1 and t2 respectively;

[0168] S65. Generate a change detection report. The change detection report includes the specific location, change type, change amount, and time period information of the change area. The format of the change detection report is:

[0169] Change detection report = {Location: (x, y), Change type: Type, Change amount: ΔA, Time period: [t1, t2]};

[0170] Among them, (x, y) are the specific position coordinates of the change area, Type is the change type, and [tl, t2] is the time period;

[0171] S66. Store the generated change detection report in the database and provide a query interface for users to access and analyze.

[0172] Embodiment 1:

[0173] In a typical agricultural management and land resource planning scenario, the agricultural administration of a certain city needs to monitor the utilization of cultivated land within the city to timely understand the changes in cultivated land utilization, optimize the cultivated land utilization plan, and ensure the sustainable development of agricultural production and the rational utilization of land resources. The city covers an area of approximately 5,000 square kilometers, and the cultivated land area accounts for about 60% of the total area of the city. To achieve this goal, the agricultural administration decides to adopt the method for detecting changes in cultivated land utilization from remote sensing data based on an improved intelligent algorithm of the present invention to conduct a detailed monitoring and analysis of the changes in cultivated land utilization in the city.

[0174] In this embodiment, the agricultural administration uses remote sensing satellites and unmanned aerial vehicles (UAVs) to obtain multi-spectral images and radar images covering the whole city. The specific steps are as follows:

[0175] Obtain multi-spectral images covering the cultivated land areas of the whole city through optical remote sensing satellites and multi-spectral imaging satellites, including image data in the visible light band, near-infrared band, and mid-infrared band. In addition, obtain high-resolution multi-spectral images through multi-spectral sensors carried by UAVs, and obtain synthetic aperture radar images through remote sensing satellites and UAVs to ensure the comprehensiveness and accuracy of the data.

[0176] Perform radiometric correction and atmospheric correction on the collected multi-spectral images to eliminate the influence of illumination and atmospheric scattering. Remove noise from the radar images, and use Lee filter, Frost filter, and Gamma-Map filter to process the speckle noise. Subsequently, perform geometric correction and image registration on the multi-spectral images and radar images, correct the remote sensing images to a unified geographic coordinate system, and perform data fusion to generate a comprehensive remote sensing image containing more information.

[0177] Adopt an intelligent algorithm combining multi-level convolutional transformation and adaptive feature enhancement to extract features from the preprocessed remote sensing data and generate a multi-dimensional feature matrix. The convolutional kernel of the convolutional neural network is designed and optimized as an adaptive shape convolutional kernel, which dynamically adjusts the shape and size of the convolutional kernel according to the content of the feature map of the input remote sensing image data, and uses an adaptive Swish activation function for activation processing. Reduce the size of the feature map through a multi-scale max pooling method, retain the main feature information, and finally generate a multi-dimensional feature matrix.

[0178] The time series analysis of the extracted feature information is carried out using a dynamic spatio-temporal graph convolutional network. By comparing the remote sensing image data in different time periods, the changes in cultivated land use are detected. The dynamic spatio-temporal graph convolutional network includes a temporal convolutional layer and a graph convolutional layer, which extract temporal features and spatial features respectively, and an adaptive Swish activation function is used for activation processing to construct a dynamic spatio-temporal feature matrix.

[0179] Based on the results of the time series analysis, an integrated adversarial generation network and a multi-task learning model are used to detect and classify the changes in cultivated land use. The generator generates the change detection results, and the discriminator discriminates between the generated change detection results and the real data. In the multi-task learning model, tasks such as cultivated land expansion detection, cultivated land reduction detection, and utilization type change detection are defined. During the training process, the generator and discriminator of the integrated adversarial generation network model are alternately optimized to minimize the multi-task loss function and improve the accuracy of detection and classification.

[0180] A change detection result matrix is generated, including information such as the specific location, change type, and change amount of the changed area. The change detection result matrix is projected onto the geographic coordinate system to generate the geographical location of the changed area, and precise positioning and interpolation processing are performed on the specific location of the changed area to calculate the area and change amount of the changed area. Using geographic information system technology, the detection results are visually displayed to generate maps, charts, and interactive visualization models for easy intuitive understanding by users.

[0181] To verify the effectiveness of the method of the present invention, the agricultural administration compared the performance of the method of the present invention and traditional methods in cultivated land use change detection. The specific data is shown in Table 1 below:

[0182] Table 1 Comparison data of the method of the present invention and traditional methods

[0183] Project Traditional method Method of the present invention Detection accuracy 85％ 95％ Processing efficiency (time required per 50,000 square kilometers) 72 hours 24 hours Accuracy of change area recognition 80％ 92％ Calculation error of change amount ±15％ ±5％ Visualization effect Static image Interactive visualization model

[0184] It can be seen from the data in Table 1 above that the method of the present invention is superior to traditional methods in terms of detection accuracy, processing efficiency, recognition accuracy of changed areas, and calculation error of change amounts. In addition, the method of the present invention provides an interactive visualization model, enabling users to more intuitively understand the detection results and providing a more comprehensive decision support function.

[0185] In summary, the method for detecting changes in cultivated land use from remote sensing data based on an improved intelligent algorithm of the present invention significantly improves the detection accuracy and efficiency in practical applications, provides comprehensive result visualization and decision support functions, solves the problems of insufficient accuracy, low efficiency, and lack of visualization in the prior art, and has important technical and economic value.

[0186] The present invention uses an intelligent algorithm that combines multi-level convolutional transformation and adaptive feature enhancement to extract features from preprocessed remote sensing data, generate a multi-dimensional feature matrix, and perform temporal analysis using a dynamic spatio-temporal graph convolutional network. By comparing remote sensing image data from different time periods, the change situation of cultivated land use is detected. Based on the results of the temporal analysis, an integrated adversarial generation network and a multi-task learning model are used to detect and classify the changes in cultivated land use, accurately identify the expansion, reduction, and changes in the use type of cultivated land, and the improvement measures significantly improve the accuracy of change detection and reduce false alarms and missed reports.

[0187] The present invention ensures the high quality of the input data by optimizing the data processing and preprocessing steps, improving the reliability of the detection results. The improved intelligent algorithm can quickly process large-scale remote sensing data, significantly enhancing the efficiency of data processing and analysis. In addition, by using a dynamic spatio-temporal graph convolutional network and an integrated adversarial generation network, the detection and classification of changes in cultivated land use can be completed in a relatively short time, meeting the requirements of real-time monitoring and rapid response.

[0188] As mentioned above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A method for detecting cultivated land use changes based on remote sensing data based on an improved intelligent algorithm, characterized in that: The following steps are involved: S1. Acquire remote sensing images covering the target cultivated land area through remote sensing satellites, drones or other remote sensing equipment, wherein the remote sensing images include multispectral images and radar images; S2, performing image correction, noise removal, geometric correction and image registration on the collected remote sensing data; S3, using improved intelligent algorithm to extract features from pre-processed remote sensing data and generate a multi-dimensional feature matrix; S4. Use the dynamic spatiotemporal graph convolutional network to perform time series analysis on the extracted feature information and detect the changes in cultivated land use by comparing remote sensing image data in different time periods; S5. Based on the results of time series analysis, an integrated adversarial generative network and a multi-task learning model are used to detect and classify changes in cultivated land use and identify changes in the expansion, contraction, and utilization type of cultivated land; S6. Generate change detection results, including specific location of the changed area, change type and change amount information; S7. Use augmented reality technology and geographic information system technology to visualize the test results and generate maps, charts and interactive visualization models; S8. Generate a detailed change report, including the specific location, change type, change amount and time period of the change area, and provide future change trend analysis by combining historical data and prediction models; The S3 includes: performing multi-level convolution transformation processing on the pre-processed remote sensing image data, constructing an improved multi-layer convolutional neural network model, optimizing the convolution kernel design to be an adaptive shape convolution kernel, and dynamically adjusting the shape and size of the convolution kernel according to the content of the input remote sensing image data feature map; The S4 includes: S41, taking the multi-dimensional feature matrix as input, constructing a dynamic spatiotemporal graph convolutional network including a temporal convolutional layer and a graph convolutional layer; S42, in the time convolution layer, using one-dimensional convolution to perform a time series convolution operation on the input feature matrix to extract time features; S43. In the graph convolution layer, the graph convolution is used to perform a spatial convolution operation on the features after the temporal convolution to extract the spatial features. S44, performing activation processing on the output of the graph convolutional layer; S45. Fuse the outputs of the temporal convolution layer and the graph convolution layer to construct a dynamic spatiotemporal feature matrix D st ; S46, based on the dynamic spatiotemporal feature matrix D st ,By comparing the remote sensing image data of different time periods, the changes in cultivated land use are detected and the change detection matrix C is calculated.

2. According to claim 1, a method for detecting cultivated land use changes based on remote sensing data and improved intelligent algorithm is characterized in that: The S1 includes: S11, acquiring a multispectral image covering the target cultivated land area through an optical remote sensing satellite and a multispectral imaging satellite, wherein the multispectral image includes image data of a visible light band, a near infrared band, and a mid-infrared band; S12, acquiring a multispectral image covering the target cultivated land area by a multispectral sensor carried by the UAV, wherein the multispectral sensor has an optical resolution of 0.5 meters; S13, acquiring a synthetic aperture radar image covering the target cultivated land area through a remote sensing satellite; S14, obtaining a radar image covering the target cultivated land area by using a radar sensor carried by the drone, wherein the radar sensor can penetrate clouds and vegetation to obtain information on the ground surface; S15, the acquired multispectral image and radar image constitute a remote sensing image, and the remote sensing image is transmitted to a ground processing station via a satellite communication link and a wireless communication link; S16. The data formats of the remote sensing images include GeoTIFF, HDF and NetCDF standard formats.

3. The method for detecting cultivated land utilization changes based on remote sensing data based on an improved intelligent algorithm according to claim 2, characterized in that: The S2 includes: S21, performing image correction on the collected multispectral image, wherein the image correction includes radiation correction and atmospheric correction, wherein the radiation correction is achieved by converting the radiation value recorded by the sensor into the reflectivity of the ground object, and the atmospheric correction is achieved by eliminating the influence of atmospheric scattering and absorption; S22, removing noise from the collected radar image, and processing the radar image using Lee filter, Frost filter and Gamma-Map filter; S23, correcting the geometric distortion of multispectral images and radar images and converting the geographic coordinate system, using terrain correction and orthorectification methods to correct the remote sensing images to a unified geographic coordinate system; S24, performing image registration on the multispectral image and the radar image using feature point matching and region-based matching methods, and spatially aligning remote sensing images acquired in different time periods and by different sensors to ensure pixel correspondence between images; S25, multispectral image I after correction and registration multispectral and radar image I radar Perform data fusion based on weighted average and principal component analysis to generate a comprehensive remote sensing image containing more information. fused : I fused =α·I multispectral +β·I radar ; Among them, I fused is the fused remote sensing image, I multispectral and I radar are multispectral images and radar images respectively, and α and β are weight coefficients.

4. The method for detecting cultivated land utilization changes based on remote sensing data based on an improved intelligent algorithm according to claim 3 is characterized in that: The S3 includes: S31, preprocessing the remote sensing image I fused The data is processed by multi-level convolution transformation, and an improved multi-layer convolutional neural network model is constructed. The convolution kernel design of the improved multi-layer convolutional neural network model is optimized to an adaptive shape convolution kernel. According to the input remote sensing image I fused The content of the data feature map dynamically adjusts the shape and size of the convolution kernel: Y i,j,k =∑ m,n X i+f(m),j+f(n) ·W m,n,k +b k ; Among them, Y i,j,k is the output eigenvalue of the kth convolution kernel at position (i, j), X i+f(m),j+f(n) is the pixel value of the input feature map at position (i+f(m),j+f(n)), W m,n,k is the weight of the kth convolution kernel at position (m,n), b k is the bias of the kth convolution kernel, and the function f is used to dynamically adjust the shape and size of the convolution kernel; S32. After each layer of convolution operation, an adaptive Swish activation function is used for activation processing: f(x) = x·sigmoid(βx); Among them, f(x) is the activation function, x is the input value, and β is a learnable parameter that can be automatically adjusted according to the training process; S33, use the multi-scale maximum pooling method to improve the pooling operation, reduce the size of the feature map, and retain the main feature information: Among them, P i,j,k is the value of the feature map at position (i, j) after pooling, The value of the convolution output feature map within the multi-scale pooling window; S34, after the convolution transformation and pooling operation, the self-attention mechanism is used to target the remote sensing image I fused Adaptive feature enhancement processing of data characteristics: Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector; S35, flattening the feature map after the adaptive feature enhancement process to generate a multi-dimensional feature matrix. The dimension of the feature matrix is ​​determined by the output feature map of the convolutional neural network and the result of the adaptive feature enhancement process. The flattening operation converts the three-dimensional feature map into a two-dimensional matrix. The supporting formula of the flattening operation is: F flattened =Flatten(Attention(MultiScalePool(Swish(Conv(I fused ,W,b))))); Among them, F flattened is the flattened multi-dimensional feature matrix, I fused is the fused remote sensing image, Conv(I fused ,W,b) is a convolution operation, Swish(·) is an adaptive Swish activation function, MultiScalePool(·) is a multi-scale maximum pooling operation, Attention(·) is an adaptive feature enhancement process, and Flatten(·) is a flattening operation.

5. The method for detecting cultivated land utilization changes based on remote sensing data and improved intelligent algorithm according to claim 4 is characterized in that: The S4 includes: S41, multi-dimensional feature matrix F flattened As input, a dynamic spatiotemporal graph convolutional network consisting of a temporal convolutional layer and a graph convolutional layer is constructed; S42. In the time convolution layer, a one-dimensional convolution is used to perform a time series convolution operation on the input feature matrix to extract the time features: Among them, T i,t is the value of the temporal convolution output at time step t and feature dimension i, F i,t+j is the value of the input feature matrix at time step t+j and feature dimension i, W j is the weight of the temporal convolution kernel, and k is the radius of the temporal convolution kernel; S43. In the graph convolution layer, graph convolution is used to perform spatial convolution operation on the features after temporal convolution to extract spatial features: Among them, G i,t is the value of the graph convolution output at node i and time step t, N(i) is the set of neighbor nodes of node i, and A ij is the adjacency matrix value between node i and node j, d i and d j are the degrees of node i and node j respectively, T j,t is the eigenvalue after time convolution; S44, performing activation processing on the output of the graph convolutional layer; S45. Fuse the outputs of the temporal convolution layer and the graph convolution layer to construct a dynamic spatiotemporal feature matrix D st : D st =concat(T,G); Among them, D st is the dynamic spatiotemporal feature matrix, T is the output of the temporal convolution layer, G is the output of the graph convolution layer, and concat represents the concatenation operation on the feature dimension; S46, based on the dynamic spatiotemporal feature matrix D st By comparing remote sensing image data of different time periods, the changes in cultivated land use are detected and the change detection matrix C is calculated: Where C is the change detection matrix, and are the dynamic spatiotemporal feature matrices of time steps t1 and t2 respectively.

6. The method for detecting cultivated land utilization changes based on remote sensing data based on improved intelligent algorithm according to claim 5, characterized in that: The S5 includes: S51. Based on the change detection matrix C, an integrated adversarial generative network model and a multi-task learning model are constructed to detect and classify changes in cultivated land use; S52. In the integrated adversarial generative network model, the generator generates change detection results, and the discriminator discriminates the generated change detection results and the real data. The objective function of the generator is: Where D(x) is the discriminator’s discriminant result on the real data x, G(z) is the change detection result generated by the generator, and z is the input noise of the generator; S53. In the multi-task learning model, multiple tasks are defined, including the cultivated land expansion detection task, the cultivated land reduction detection task and the utilization type change detection task. The loss function is the weighted sum of the multi-task loss functions: L MTL =αL expansion +βL reduction +γL type_change ; Among them, L expansion is the loss of farmland expansion detection task, L reduction To reduce the loss of detection tasks, L type_change To utilize the loss of type change detection task, α, β and γ are the weight coefficients of the loss function; S54. During the training process, the generator and discriminator of the integrated adversarial generative network model are alternately optimized to train the multi-task learning model and improve the accuracy of detection and classification by minimizing the multi-task loss function: Among them, θ G ,θ D ,θ MTL are the parameters of the generator, discriminator and multi-task learning model respectively, and η is the learning rate; S55. After the training is completed, the optimized integrated adversarial generative network model and multi-task learning model are used to detect and classify the input change detection matrix C, identify the expansion, reduction and changes in the utilization type of cultivated land, and output the detection and classification results.

7. The method for detecting cultivated land utilization changes based on remote sensing data and improved intelligent algorithm according to claim 6 is characterized in that: The S6 includes: S61, constructing a change detection result matrix R based on the generated detection and classification results, where the change detection result matrix R includes specific location, change type and change amount information of the change area; S62, projecting the change detection result matrix R into the geographic coordinate system to generate the geographic location of the change area, where the geographic location of the change area is determined by the position coordinates in the change detection matrix C and the change type in the classification result; S63, accurately locate the specific position of the changed area, and use an interpolation algorithm to smooth the boundary of the changed area; S64, calculating the area and change amount of the changed area, wherein the change amount includes the amount of cultivated land expansion, the amount of cultivated land reduction and the amount of change in the type of use, and the calculation formula of the change amount is: ΔA=A t2 -IN t1 ; Among them, ΔA is the change, A t1 and A t2 are the areas of the changing regions at time steps t1 and t2 respectively; S65: Generate a change detection report, which includes the specific location of the changed area, the change type, the change amount and the time period information. The format of the change detection report is: Change detection report = {position: (x, y), change type: Type, change amount: ΔA, time period: [t1, t2]}; Where (x, y) is the specific location coordinates of the change area, Type is the change type, and [t1, t2] is the time period; S66. Store the generated change detection report in a database and provide a query interface for user access and analysis.

Images