Land change detection system based on generative adversarial network

Abstract

The application relates to the technical field of neural networks, and more particularly to a land change detection system based on a generative adversarial network. A multi-source data fusion and feature extraction part is used for high-dimension fusion of multi-source remote sensing data, considers time series and spatial distribution, ensures the spatiotemporal consistency and integrity of information, and obtains fused data; a spatiotemporal attention mechanism is introduced to adaptively adjust the importance of pyramid features and obtain attention weights; a generative adversarial network part is used for designing a conditional generative adversarial network, obtaining conditional features based on the pyramid features and the attention weights; and a land change detection part is used for considering the gradient change and Euclidean distance of generated images, performing change detection based on the generated image pairs, and obtaining a change detection score. The application provides an efficient and flexible solution for land change monitoring and environmental protection.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of neural networks, and particularly relates to a land change detection system based on a generative adversarial network. BACKGROUND

[0002] With the rapid development of remote sensing technology, geographic information systems (GIS) and remote sensing image analysis have been widely applied in the fields of land change monitoring, environmental protection, urban planning and resource management. Land change detection is a method of identifying and quantifying land surface changes using multi-temporal remote sensing data. By analyzing image data at different time points, it extracts land surface change information, which can help governments and enterprises better understand land use, monitor the impact of natural disasters and manage the ecological environment. However, with the explosive growth of remote sensing data and the complexity of land surface changes, traditional land change detection techniques are not up to the task when faced with large-scale data and complex land surface features.

[0003] Currently, the mainstream methods of land change detection mainly fall into two categories: pixel-based change detection methods and object-based change detection methods. Pixel-based change detection methods are one of the most traditional change detection methods. They identify land surface changes by directly comparing the pixel value changes of the same location in multi-temporal remote sensing images. Common pixel-level change detection algorithms include image difference method, image ratio method, change vector analysis (CVA) and principal component analysis (PCA), etc. The advantages of these methods are simple calculation, easy implementation, and suitable for areas with obvious land surface changes. However, the disadvantages of this type of method are also very obvious: pixel-level detection requires high geometric correction and spectral consistency of the image, and is easily affected by image noise, shadows and seasonal changes, etc., resulting in inaccurate detection results. In addition, pixel-level methods cannot fully utilize the spatial information of the image, and are prone to "salt and pepper effect", i.e. a large number of noise points in the detection results. In order to overcome the shortcomings of pixel-level methods, object-based change detection methods have emerged. This method first segments the remote sensing image, dividing the image into multiple objects or regions with physical meaning, and then performs change detection at the object level. Object-level methods can fully utilize the spectral, spatial and texture information of the image, reduce the influence of noise, and improve the detection accuracy. Common object-level methods include object-oriented image analysis (OBIA) and multi-scale segmentation methods, etc. However, one of the main problems of object-level methods is that the segmentation accuracy depends on the resolution of the image and the parameter settings of the algorithm, and different parameter settings may lead to completely different segmentation results. In addition, object-level methods still have problems of insufficient detection accuracy and low computational efficiency when facing complex land surfaces and multi-scale changes. SUMMARY

[0004] The main purpose of the present application is to provide a land change detection system based on a generative adversarial network, which can accurately extract multi-scale features of surface changes, generate images similar to real changes, and evaluate the degree of surface changes through comprehensive change detection scores. The system significantly improves the accuracy and robustness of change detection, adapts to various complex surface environments and data types, solves the problems of insufficient multi-source data fusion, poor sensitivity to non-stationary spatio-temporal features, and insufficient multi-scale change detection capability in the prior art, and provides an efficient and flexible solution for land change monitoring and environmental protection.

[0005] To solve the above problems, the technical scheme of the present application is as follows:

[0006] The land change detection system based on a generative adversarial network comprises: a multi-source data fusion and feature extraction part for high-dimensional fusion of multi-source remote sensing data, considering time series and spatial distribution to ensure the spatio-temporal consistency and integrity of information, obtaining fused data; using an adaptive convolution kernel design to capture non-stationary spatio-temporal features in the fused data; based on the non-stationary spatio-temporal features, using a multi-scale feature pyramid construction method to ensure accurate detection of changes at different scales, obtaining pyramid features at different scales through multi-layer feature downsampling and upsampling operations; introducing a spatio-temporal attention mechanism to adaptively adjust the importance of the pyramid features to obtain attention weights; a generative adversarial network part for designing a conditional generative adversarial network, obtaining conditional features based on the pyramid features and the attention weights, the generator generating simulated changed images, and the discriminator evaluating the authenticity of these images to guide the generator to improve; a land change detection part for considering the gradient change and Euclidean distance of the generated images, performing change detection based on the generated image pairs to obtain change detection scores.

[0007] Further, the multi-source data fusion and feature extraction part performs high-dimensional fusion on multi-source remote sensing data, considers time series and spatial distribution to ensure the spatio-temporal consistency and integrity of information, and obtains fused data, the process specifically comprising: collecting original data from n different remote sensing data sources; using interpolation and resampling to align all original data in time and space to ensure that all original data are on the same spatio-temporal grid, giving each data source equal weight, and the obtained fused data is represented by the following formula:

[0008]

[0009] Wherein, is the fused data set, which is a function of time t and spatial position (x, y); x is the X-axis coordinate; y is the Y-axis coordinate; T is the total time span; and Ω is the spatial domain. is the signal strength of the ith data source, τ is the time offset, u is the X-axis offset, v is the Y-axis offset, λ i is the time attenuation coefficient of the ith data source, σ i is the spatial distribution standard deviation of the ith data source, Δt i is the time resolution of the ith data source, P i is the periodic change period of the ith data source, α i is the periodic change intensity coefficient of the ith data source; is the Gaussian kernel used to simulate spatial correlation, σ i is the correlation range parameter, which controls the range of correlation; is used to handle the difference in time resolution of different data sources.

[0010] Further, the multi-source data fusion and feature extraction part captures the non-stationary spatio-temporal features F in the fused data by the following formula using adaptive convolution kernel design:

[0011]

[0012] where Z is the normalization constant to ensure the integral of the convolution kernel is 1, σ t is the preset adaptive standard deviation in time dimension, σ s is the preset adaptive standard deviation in space dimension.

[0013] Further, the multi-source data fusion and feature extraction part obtains pyramid features of different scales by the following formula:

[0014]

[0015] where P l is the pyramid feature corresponding to the lth layer scale, DownSample l represents the lth downsampling operation; is the Laplacian operator; represents the kth upsampling operation, P k is the pyramid feature corresponding to the kth layer scale.

[0016] Further, the multi-source data fusion and feature extraction part introduces a spatio-temporal attention mechanism to adaptively adjust the importance of pyramid features, and the specific process of obtaining attention weights includes: calculating the spatio-temporal correlation matrix of each layer of pyramid features according to the following formula:

[0017]

[0018] where C ij is the element in the ith row and jth column of the spatio-temporal correlation matrix, t i, t j respectively the time of feature i and feature j; using the spatio-temporal correlation matrix, the attention weight of each layer of pyramid feature is calculated:

[0019]

[0020] wherein A l is the attention weight of the lth layer scale pyramid feature at time t and spatial position (x, y); F i , F j is the feature of feature i and feature j at time t and spatial position (x, y); σ F is an important parameter for controlling the similarity of feature values, and is a set value.

[0021] Further, the formula representation of the generator of the conditional generative adversarial network in the generative adversarial network part design is as follows:

[0022]

[0023] wherein z is a random noise vector, obeying a normal distribution P′ l is the conditional feature of the lth layer, which is calculated by the following formula:

[0024] P l ′=A l ·P l ;

[0025] A l is the attention weight of the lth layer scale pyramid feature; G(z, P′, A) is the simulated changed image generated by the generator; W G and b G are the weight matrix and bias vector of the generator respectively; φ(·) is a nonlinear activation function, which is a ReLU function or a LeakyReLU function; σ(·) is an output layer activation function, which is a tanh function or a sigmoid function; U is the weight matrix of the conditional input of the generator, which is used to fuse the conditional feature and the noise; A is the attention weight matrix; z is the noise; the output of the generator is a simulated image, which is as similar as possible to the real image in high-dimensional space.

[0026] Further, the formula representation of the discriminator is as follows:

[0027]

[0028] wherein I is the input image, which is a real image I real or a simulated image G(z, P′, A) generated by the generator; D(I, P′, A) is the probability output by the discriminator, indicating the authenticity of the input image I; W Dand b D and b are the weight matrix and bias vector of the discriminator respectively; V is the weight matrix of the conditional input of the discriminator; D(I, P', A) is the probability output by the discriminator, indicating the authenticity of the input image I.

[0029] Further, the loss function of the generator is as follows:

[0030]

[0031] where p z (z) is the probability distribution of the random noise z; is the expectation operator; P' is the conditional feature set; is the basic generator loss, used to maximize the probability that the generated image is identified as a real image by the discriminator; is a gradient normalization regularization term, used to balance the learning speed of the generator on different conditional features; and are the gradients with respect to the generator and the conditional features respectively, a k is the adaptive weight of each feature layer, and ∈ is a constant to prevent division by zero; is a reconstruction loss term, which ensures the consistency of the generated image with the real image in the conditional feature space, and uses the L2 norm to measure the difference between the generated image and the real image; |||2 represents the L2 norm operation; represents the loss part corresponding to the conditional feature of the k-th layer; ||| represents the operation of calculating the determinant of a matrix or the modulus of a vector; p P′ (P') is the probability distribution of the conditional feature P'; I real (P l ) is the conditional feature of the real image at the l-th layer scale;

[0032] Further, the loss function of the discriminator is as follows:

[0033]

[0034] where p data (I real ) is the probability distribution of the real image; 1-D(G(z, P', A), P', A) represents the authenticity discrimination output of the generated image by the discriminator, and the probability that the generated image is false is obtained;

[0035] is the first discriminator loss, used to maximize the probability that the discriminator correctly identifies the real image; ] is the second discriminator loss, used to maximize the probability that the discriminator correctly identifies the generated image; is a multi-scale contrastive loss term, which strengthens the sensitivity of the discriminator to subtle changes by the difference of the discriminator output at different scales;p and q are both norms of the contrast measure.

[0036] Further, the land change detection part detects the change based on the generated image pairs by the following formula to obtain a change detection score:

[0037]

[0038] wherein is the final change detection score, which quantifies the change between the two generated images, the higher the score, the more significant the change;N represents the total number of pixels;G1(x i , y i ) and G2(x i , y i ): respectively represent the pixel values of the first and second generated images at position (x i , y i ) at different times; and respectively represent the gradient vectors of the first and second generated images at position (x i , y i );γ: a weight parameter, the value range is between [0, 1], used to balance the importance of gradient change and Euclidean distance in change detection.

[0039] The land change detection system based on the generative adversarial network has the following beneficial effects:

[0040] Firstly, the present application has significant advantages in multi-source data fusion and feature extraction. Traditional land change detection methods often rely on a single data source or simple data fusion technology, and cannot fully utilize the diversity and richness of multi-source remote sensing data. By introducing adaptive convolution kernel design, the present application can dynamically adjust the shape and size of the convolution kernel, so as to better capture the non-stationary spatio-temporal features in the fused data. The use of adaptive convolution kernel enables the system to extract more delicate change features at different time and spatial scales, enhancing the sensitivity and adaptability of the system to complex land surface changes. This adaptive feature extraction method not only improves the utilization efficiency of data, but also significantly improves the accuracy of change detection, especially in the case of multi-source data fusion, which can fully utilize the advantages of different data sources.

[0041] Secondly, the application adopts a spatio-temporal attention mechanism to adaptively adjust the importance of pyramid features. The introduction of this mechanism effectively solves the problem of lack of flexibility in weight distribution for different changing areas in the prior art. By calculating the spatio-temporal correlation matrix of each layer of pyramid features, the system can adaptively adjust the weight of each feature in change detection according to the temporal and spatial correlation of the features. This attention mechanism ensures that the system can pay more attention to areas with significant changes in change detection, while reducing attention to unimportant areas, thereby improving the accuracy and efficiency of detection. Especially when dealing with large-scale remote sensing data and complex surface environments, the introduction of the spatio-temporal attention mechanism enables the system to dynamically optimize the distribution of feature weights, ensuring accurate identification and detection of key changing areas.

[0042] In addition, the multi-scale feature pyramid construction method of the application further enhances the system's ability to detect changes at different scales. Surface changes often exhibit multi-scale, multi-level feature changes, and traditional methods are difficult to comprehensively detect changes at multiple scales. The application constructs a multi-scale feature pyramid through multi-layer feature downsampling and upsampling operations, enabling the system to effectively capture change features at different scales. In particular, by combining the Laplacian operator for edge enhancement, the system performs well in capturing both detailed and global features. This multi-scale feature pyramid construction method enables the system to achieve significant detection results when detecting large-scale land use changes and small-scale local topographic changes, improving the system's adaptability and reliability in different application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The system structure diagram of the land change detection system based on the generative adversarial network provided by the embodiment of the application. DETAILED DESCRIPTION

[0044] In order to enable personnel in the art to better understand the application scheme, the technical solutions in the embodiments of the application will be described clearly and completely below in conjunction with the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the application.

[0045] Embodiment 1: Reference Figure 1, a land change detection system based on a generative adversarial network, the system comprising: a multi-source data fusion and feature extraction part for high-dimensional fusion of multi-source remote sensing data, considering time series and spatial distribution, ensuring the spatio-temporal consistency and integrity of the information, obtaining fused data; using a self-adaptive convolution kernel design to capture non-stationary spatio-temporal features in the fused data; based on the non-stationary spatio-temporal features, using a multi-scale feature pyramid construction method to ensure accurate detection of changes at different scales, obtaining pyramid features at different scales through multi-layer feature downsampling and upsampling operations; introducing a spatio-temporal attention mechanism to adaptively adjust the importance of the pyramid features to obtain attention weights; a generative adversarial network part for designing a conditional generative adversarial network, obtaining conditional features based on the pyramid features and the attention weights, a generator generating simulated changed images, and a discriminator evaluating the authenticity of the images to guide the generator to improve; a land change detection part for considering the gradient change and Euclidean distance of the generated images, performing change detection based on the generated image pairs to obtain a change detection score.

[0046] Specifically, the application relates to a land change detection system based on a generative adversarial network (GAN), which is characterized in that the system realizes accurate detection of land changes by fusing multi-source remote sensing data, using adaptive convolution kernels to capture non-stationary spatio-temporal features, constructing a multi-scale feature pyramid, introducing a spatio-temporal attention mechanism, and designing a conditional generative adversarial network. The system first starts from the multi-source data fusion and feature extraction part, and the purpose of the part is to high-dimensionally fuse information from different remote sensing data sources to ensure the consistency and integrity of data in time series and spatial distribution. The multi-source fusion technology compensates for the deficiencies of a single data source in spatial resolution, time resolution, spectral characteristics, etc. by integrating data from different sensors and time points, so that the system can more comprehensively and accurately capture the information of land surface changes. In actual operation, the multi-source data fusion not only considers the spatial feature distribution, but also fully considers the time series features, so that the fused data can better reflect the true situation of land surface changes by combining the change patterns of the two dimensions of time and space. After multi-source data fusion, the system enters the adaptive convolution kernel design stage, and the key of this step is to capture the non-stationary spatio-temporal features in the fused data. Traditional convolutional neural networks usually assume that the features of data are stationary when processing remote sensing data, but in actual situations, remote sensing data often has complex non-stationary characteristics, which reflect the changes and inconsistencies of data in different time and spatial scales. In order to solve this problem, the application introduces adaptive convolution kernel design, which dynamically adjusts the size and shape of the convolution kernel, so that the system can better adapt to the non-stationarity of data and capture more fine spatio-temporal features. For example, in areas with significant changes, the convolution kernel may need to be smaller and more concentrated in shape to capture the details of the changes; while in areas with gentle changes, the convolution kernel can be larger to obtain more extensive context information. Such design enables the system to more accurately capture the features of land changes when processing complex and varied remote sensing data.

[0047] After acquiring the non-stationary spatio-temporal features, the system adopts a multi-scale feature pyramid construction method. This method is introduced to ensure that the system can accurately detect land changes at different scales. The types and scales of land changes often vary, some changes may be small-scale and subtle, such as seasonal changes in vegetation, while some changes may be large-scale, such as large-scale land development or natural disasters causing land destruction. The multi-scale feature pyramid can effectively capture and express these changes at different scales by combining multi-layer feature downsampling and upsampling operations. Downsampling operation helps to extract low-frequency features, which is suitable for large-scale land change detection, while upsampling operation helps to restore high-frequency features, which is suitable for small-scale change detection. Through this multi-scale feature extraction method, the system can process large-scale and small-scale changes simultaneously in the same network framework, ensuring the comprehensiveness and accuracy of the detection. To further improve the flexibility and accuracy of feature selection, the system introduces a spatio-temporal attention mechanism. The main function of this mechanism is to adaptively adjust the importance of pyramid features, so that the system can focus more on those areas or time periods that are most critical to land change detection. The attention mechanism calculates the importance weight of each feature in the current detection task, and then weights the different features. This means that when the change features in a certain area or time period are particularly significant, the system can automatically increase the feature weight of that area or time period, making the detection result more accurate. For example, in some images after earthquakes or floods and other natural disasters, the attention mechanism can help the system pay more attention to the areas severely affected by the disaster, and ignore the areas with little change. This adaptive attention mechanism effectively improves the detection efficiency and accuracy of the system, especially when dealing with large-scale remote sensing data.

[0048] After the feature extraction and selection process described above, the system enters the generative adversarial network part. The core idea of the generative adversarial network is to make the generator generate images that are increasingly close to the real change images through adversarial training, while the discriminator guides the improvement of the generator by continuously evaluating the authenticity of the generated images. In the present invention, a conditional generative adversarial network (CGAN) is designed for land change detection tasks, and the input of the generator is the multi-scale pyramid features and attention weights extracted through the previous steps, which provide the generator with rich contextual information about land change. The generator generates simulated changed images based on these conditional features, and the discriminator compares these generated images with real images, gradually optimizes the performance of the generator by feeding back the generation error of the generator. In this way, through the adversarial training between the generator and the discriminator, the system can generate more realistic and close-to-real change images, which provides a more reliable basis for subsequent change detection. Finally, the system enters the land change detection part, the main task of which is to perform change detection based on the generated image pairs. In this process, the system not only considers the gradient change of the generated images, but also combines the Euclidean distance for change detection. Gradient change is mainly used to capture the change of image details, which is particularly effective for identifying small features such as edge changes; while the Euclidean distance is used to measure the overall difference between two images, which can reflect larger scale changes. By combining these two methods, the system can more comprehensively and accurately evaluate the change of land. The final change detection score is based on the gradient change and Euclidean distance calculation results of the image pair, combining a comprehensive scoring mechanism, which can provide a quantitative evaluation of the degree of change, providing a scientific basis for land change management and decision-making.

[0049] In embodiment 2, the multi-source data fusion and feature extraction part, which fuses multi-source remote sensing data in high dimension, considers time series and spatial distribution, ensures the spatio-temporal consistency and integrity of information, and obtains the process of fused data, specifically includes: collecting original data from n different remote sensing data sources; using interpolation and resampling to align all original data in time and space to ensure that all original data are on the same spatio-temporal grid, giving each data source equal weight, and the obtained fused data is represented by the following formula:

[0050]

[0051] Wherein, is the fused data set, which is a function of time t and spatial position (x, y); x is the X-axis coordinate; y is the Y-axis coordinate; T is the total time span; Ω is the spatial domain; is the signal intensity of the i-th data source, τ is the time offset; u is the X-axis offset; v is the Y-axis offset; λ iis the time decay coefficient for the i-th data source; σ i is the spatial distribution standard deviation for the i-th data source; Δt i is the time resolution for the i-th data source; P i is the periodic variation period for the i-th data source; α i is the periodic variation intensity coefficient for the i-th data source; is the Gaussian kernel used to simulate spatial correlation, σ i is the correlation range parameter, which controls the range of correlation; is used to handle the differences in time resolution of different data sources.

[0052] Specifically, this fusion process first collects raw data from n different remote sensing data sources, which may include satellite images, unmanned aerial vehicle images, and ground radar data, etc. Due to differences in acquisition time, spatial resolution, and spectral characteristics, these data often have large differences. Therefore, before fusing these data, they need to be preprocessed. In the formula, all raw data are aligned to a unified time and spatial grid through interpolation and resampling, which ensures that the information of all data sources can be represented and utilized at the same time point and spatial location. After data alignment, the formula further introduces a time decay factor This factor is used to simulate the decay characteristics of the information of each data source over time. Here, λ i represents the time decay coefficient of the i-th data source, reflecting the effectiveness of the information of this data source over time. As time goes on, the representativeness and accuracy of remote sensing data may gradually decrease, so by introducing a time decay factor, more weight can be given to newer data in the fusion process, thereby improving the timeliness and accuracy of the fusion result. At the same time, in order to handle the spatial distribution characteristics of the data, a Gaussian kernel function is introduced in the formula. The Gaussian kernel function is a common smoothing function used to simulate spatial correlation, which controls the spatial influence range of data through a standard deviation parameter σ i In remote sensing data processing, the spatial resolution and accuracy of different data sources are different, and the introduction of the Gaussian kernel function can effectively smooth and integrate data from different sources, ensuring that they can reasonably reflect the actual ground features in the fusion process. By controlling the parameter σ i , the influence range of each data source in the fused data can be adjusted to ensure the spatial consistency of the fusion result. In addition, remote sensing data often have different time resolutions, which leads to inconsistent sampling frequencies of different data sources in time. In order to effectively handle this difference, the formula uses function. The sinc function is a commonly used tool in signal processing and data analysis, which is used to maintain the bandwidth of the signal during interpolation. In the context of the present invention, the use of the sinc function can effectively handle the differences in time resolution of different data sources, ensuring that the fusion results in the time dimension are not distorted or biased. Finally, the periodic variation characteristics of the data sources are also considered in the formula The periodic variation of remote sensing data is usually caused by natural phenomena (such as seasonal changes, day-night changes) or human activities (such as regular land cultivation). By modeling these periodic variations, the system can more accurately identify and separate the differences between the signal changes caused by periodic variations and the actual land change signals. Here, α i and P i represent the periodic variation intensity and period of the i-th data source, respectively. By adjusting these parameters, the system can more flexibly adapt to the characteristics of different data sources, improving the detection accuracy of land changes.

[0053] In embodiment 3, the multi-source data fusion and feature extraction part uses an adaptive convolution kernel design to capture non-stationary spatio-temporal features F in the fused data through the following formula:

[0054]

[0055] where Z is a normalization constant to ensure that the integral of the convolution kernel is 1; σ t is a preset adaptive standard deviation in the time dimension; and σ s is a preset adaptive standard deviation in the spatial dimension.

[0056] Specifically, this formula realizes the extraction of spatio-temporal features by performing triple integration on the fused data and combining the form of the Gaussian kernel function. Here, the core of the convolution kernel design is to perform weighted processing through the Gaussian distribution in the time and space dimensions, so that the convolution operation can not only capture local changes in the data, but also adapt to feature changes in different regions and time periods. This design adjusts the shape and range of the convolution kernel by using the time dimension adaptive standard deviation σ t and the spatial dimension adaptive standard deviation σ s , so as to provide different sensitivities at different time points and spatial positions. Specifically, the form of the Gaussian kernel function reflects the weighting strategy of the data in time and space, where τ, u and v represent the time and spatial offsets, respectively. This adaptive convolution kernel design effectively solves the limitations of traditional convolution operations. Traditional convolution kernels are usually fixed in a certain size and shape, which is not suitable for handling non-stationary features, as it cannot adapt to changes in data at different locations and times. The present invention introduces adaptive standard deviations σt and σ s , so that the size and shape of the convolution kernel can be dynamically adjusted according to the changes of the actual data. For example, in the time dimension, σ t can adjust the "time window" size of the convolution kernel according to the speed of data change; similarly, in the spatial dimension, σ s can adjust the "spatial range" of the convolution kernel according to the spatial distribution of data features. This adaptive mechanism makes the convolution operation more flexible and adaptive, so that it can more accurately capture non-stationary features in the data. In addition, in order to ensure that the weighted calculation result of the convolution kernel can reasonably reflect the distribution of the real data, a normalization constant Z is introduced in the formula, which ensures that the integral result of the convolution kernel is 1, thereby avoiding data distortion caused by non-normalized kernel function. The normalization process ensures that the intensity of the output feature only depends on the characteristics of the data itself when the data is convolved, and will not be unnecessarily amplified or reduced due to the change in the size of the convolution kernel. This design is particularly important in the framework of the generative adversarial network, because the generative adversarial network needs to be iteratively optimized between the generator and the discriminator, and any unreasonable feature extraction will directly affect the accuracy of the final change detection result. Through this adaptive convolution kernel design, the system can capture non-stationary spatio-temporal features after multi-source data fusion, thereby providing more accurate and detailed feature input for the generator of the generative adversarial network. In the framework of the generative adversarial network, the generator uses these extracted spatio-temporal features to simulate the image of land change, while the discriminator evaluates the authenticity of these simulated images to guide the optimization of the generator. The adversarial training process of the generator and the discriminator depends on the quality and expressiveness of the input features, so the effectiveness of the adaptive convolution kernel in the feature extraction stage directly determines the pros and cons of the overall performance of the generative adversarial network. The adaptive convolution kernel design not only improves the flexibility and accuracy of feature extraction, but also enhances the robustness of the system in dealing with complex land surface changes, so that the land change detection system can more accurately identify and evaluate various types of land changes.

[0057] In an embodiment, the multi-source data fusion and feature extraction part obtains pyramid features of different scales through the following formula:

[0058]

[0059] wherein P l is the pyramid feature corresponding to the lth scale; DownSample l represents the lth downsampling operation; is the Laplacian operator; represents the kth upsampling operation; P k is the pyramid feature corresponding to the kth scale.

[0060] Specifically, remote sensing data, by its very nature, involves multi-scale and multi-level observations of the Earth's surface. Different types of surface changes (such as vegetation cover, building changes, and soil erosion) may appear more pronounced or more subtle at different scales. Therefore, single-scale feature extraction methods often fail to fully reflect complex surface changes, while the introduction of multi-scale feature pyramids provides an effective way to address this problem. In this invention, the feature pyramid is constructed by performing multiple downsampling and upsampling operations on the feature map. The downsampling operation aims to reduce the resolution of the feature map, thereby capturing large-scale surface changes; the upsampling operation aims to recover detailed information in the lower-resolution feature map, enabling the system to simultaneously focus on small-scale changes. In this way, the system obtains multi-level feature representations at different scales, with each layer of the feature map representing a different spatial scale. To further enrich and refine these feature representations, the Laplace operator is introduced into the formula. The Laplacian operator, a second-order differential operator, is commonly used in image processing to enhance edge features, effectively capturing changing regions and edge information in images. In the construction of multi-scale feature pyramids, the application of the Laplacian operator further highlights and amplifies the change information in the feature maps, enabling generative adversarial networks to more accurately capture change details when generating simulated change images. Furthermore, to ensure effective fusion of feature information across different levels of the multi-scale pyramid, an accumulation and upsampling strategy is employed in the formula. Specifically, the feature map of the l-th layer of the pyramid depends not only on the downsampling operation of that layer... l (F)) and also by upsampling the feature maps of the first l-1 layers (U k) and Laplace transform, and the accumulation makes the high-resolution features of lower levels effectively feedback to the low-resolution features of higher levels. This bottom-up feature fusion strategy can effectively integrate feature information of different scales, ensuring that both macro information of large-scale features and detailed information of small-scale features are retained in each layer of the feature pyramid. In this way, the system can be more comprehensive and accurate in land change detection, especially in handling complex scenes of multi-scale surface changes. At the same time, this multi-scale feature pyramid construction method has unique advantages in the application of generative adversarial networks. Generative adversarial networks rely on the game between the generator and the discriminator to optimize the model performance. In the context of land change detection, the generator needs to generate simulated images similar to the real changes, while the discriminator evaluates the authenticity of these images and optimizes the generator in the opposite direction. In this process, the rich feature representation provided by the multi-scale feature pyramid can provide more detailed and comprehensive input information for the generator, thereby improving the quality and authenticity of the generated images. Especially through the combination of Laplace operator and up-sampling operation, the generator can more accurately reproduce the change features of different scales when generating simulated images, making it more difficult for the discriminator to distinguish between real images and generated images, and ultimately improving the performance of the entire generative adversarial network.

[0061] In embodiment 5, a spatio-temporal attention mechanism is introduced in the multi-source data fusion and feature extraction part, and the importance of pyramid features is adaptively adjusted. The specific process of obtaining the attention weight includes: calculating the spatio-temporal correlation matrix of each layer of pyramid features according to the following formula:

[0062]

[0063] Wherein, C ij is the element in the i-th row and j-th column of the spatio-temporal correlation matrix; t i , t j are the time of feature i and feature j, respectively; using the spatio-temporal correlation matrix, the attention weight of each layer of pyramid features is calculated:

[0064]

[0065] Wherein, A l is the attention weight of the l-th layer scale pyramid feature at time t and spatial position (x, y); F i , F j are the features of feature i and feature j at time t and spatial position (x, y); σ F is an importance parameter that controls the similarity of feature values, and is a set value.

[0066] Specifically, the correlation matrix quantifies the spatio-temporal relationship between different features through the form of a Gaussian kernel function, i.e., considering the differences in the temporal dimension (t i -t j ) and the spatial dimension (x i -x j ) and (y i -y j ). This design utilizes the characteristics of Gaussian distribution to measure similarity: when the temporal and spatial distances of two features are close, the correlation is stronger, which means that these two features may have similar behavior on certain change patterns. Therefore, the formula attenuates the temporal and spatial differences through an exponential function, so that the correlation value can reasonably reflect the real spatio-temporal relationship between features. This process is crucial because it provides the basis data for subsequent attention weight calculation, enabling the system to dynamically adjust according to the relative importance between features. After obtaining the spatio-temporal correlation matrix, the system further calculates the attention weight of each layer of pyramid features. The calculation of attention weight is based on a weighted feature similarity measure, which involves the Euclidean distance between each pair of features, i.e., |F i -F j | 2 . This measure method fully considers the spatial and temporal changes of features, and maps these changes to a similarity measure through an exponential function. Specifically, this method introduces a parameter σ F to control the influence range of feature similarity, the larger the σ F value, the smaller the influence of feature similarity on attention weight, and vice versa. In this way, the system can adaptively adjust the attention weight of features under different change intensities and change patterns, so that those more important features in change detection get more attention, while the influence of secondary features is appropriately weakened. This adaptive adjustment mechanism of weight makes the system able to make more accurate judgments in the feature extraction stage, optimizing the performance of the entire detection process.

[0067] In addition, the spatio-temporal attention mechanism of the present application also pays special attention to the balance between local and global features of the characteristics. In the context of multi-source data fusion, different data sources may have different temporal and spatial resolutions, which requires the system to not only capture large-scale global change features, but also identify and analyze small-scale local change details. Through the above calculation method of spatio-temporal correlation matrix and attention weight, the present application can effectively balance the contribution of global and local features, so that the generative adversarial network can maintain global consistency while not losing local details when simulating and detecting land surface changes. This ability is crucial for land change detection, because land surface changes often have both large-scale significant changes (such as deforestation, urban expansion) and subtle local changes (such as small-scale building renovation or land erosion). Only by balancing global and local features, the system can accurately identify and locate the change area in a complex and variable land surface environment. Further, the spatio-temporal attention mechanism plays a crucial role in the framework of the generative adversarial network. The generative adversarial network optimizes the detection ability of the system through the game between the generator and the discriminator, where the generator is responsible for simulating the real land surface change, and the discriminator is responsible for judging the difference between the generated image and the real image. In this adversarial training process, the generator relies on the quality of the input features to generate change images as close to reality as possible, while the discriminator optimizes the feedback according to the authenticity of these images. By introducing the spatio-temporal attention mechanism, the system can provide more accurate and targeted feature information for the generator, making the generated change image more realistic and detailed. This improvement not only improves the training efficiency of the generative adversarial network, but also significantly enhances the generalization ability of the system, making it perform better in processing different types and complexities of remote sensing data. The introduction of the spatio-temporal attention mechanism also helps the generative adversarial network overcome some inherent problems in traditional methods. For example, traditional land change detection methods usually rely on fixed feature extraction and matching strategies, which often appear insufficient when dealing with non-stationary spatio-temporal data. By adaptively adjusting the feature weights, the present application enables the system to dynamically adjust the detection strategy according to the actual situation of the change area. This flexibility greatly improves the accuracy and robustness of the detection, enabling the system to better adapt to various complex situations in practical applications. Especially when faced with large-scale remote sensing data, the spatio-temporal attention mechanism can significantly reduce computational complexity, as it can selectively focus on areas with significant changes, thereby reducing unnecessary computational overhead and improving overall detection efficiency.

[0068] Example 6: Formula representation of the generator of the conditional generative adversarial network

[0069]

[0070] where z is a random noise vector, following a normal distribution P′ l is the conditional feature of the l-th layer, which is calculated by the following equation:

[0071] P′ l = A l · P l ;

[0072] A l is the attention weight of the l-th layer scale pyramid feature; G(z, P′, A) is the simulated changed image generated by the generator; W G and b G are the weight matrix and bias vector of the generator, respectively; φ(·) is a nonlinear activation function, which is a ReLU function or a LeakyReLU function; σ(·) is an output layer activation function, which is a tanh function or a sigmoid function; U is the weight matrix of the conditional input of the generator, which is used to fuse the conditional features and the noise; A is the attention weight matrix; z is the noise; the output of the generator is a simulated image, which is as similar as possible to the real image in high-dimensional space.

[0073] Specifically, the input of the generator first includes a random noise vector z, which follows a standard normal distribution N(0, 1). The introduction of random noise is a common strategy in generative adversarial networks (GAN), which aims to provide initial randomness to the generation process, ensuring that the generated images have diversity and unpredictability. The diversity of the noise vector can effectively prevent the generator from falling into the trap of “mode collapse”, i.e., the generator only learns to produce a limited set of images without diversity. Combined with the noise vector is the conditional feature P′ l which is obtained by weighting the pyramid features P l of each layer, and the weight is provided by the spatiotemporal attention mechanism, i.e., P′ l = A l · P l . In this way, the conditional feature effectively integrates the information of multi-scale features and attention mechanism, ensuring that the generator can better generate images according to the spatiotemporal correlation of the input data. In order to effectively fuse the conditional feature and the noise vector, a fusion operation is introduced in the design of the generator, which is completed by the weight matrix U, which is specifically represented as the sum of the random noise vector z and all layers of the conditional feature The fusion is performed. This fusion step ensures that the input of the generator can comprehensively consider randomness and the characteristics of the input data to form an input vector with rich information. Subsequently, this fusion vector is mapped through a nonlinear activation function φ(·), usually a ReLU or LeakyReLU function, to further enhance the expression ability of the generator. The use of a nonlinear activation function can effectively capture complex input data patterns, enabling the generator to better model the complexity of land change in high-dimensional space. The output layer of the generator uses an activation function σ(·), which can be a tanh function or a sigmoid function. These activation functions are used to limit the output of the generator within a certain range, ensuring that the generated image values are physically reasonable. For example, the tanh function limits the output value to [-1, 1], which is very effective for image data normalization; while the sigmoid function limits the output value to [0, 1], which is very useful for generating grayscale images or probability images. The weight matrix W G and the bias vector b G of the generator are then continuously optimized through adversarial training, with the ultimate goal of generating a simulated change image that is as close as possible to the real image in high-dimensional feature space.

[0074] In Example 7, the formula of the discriminator is as follows:

[0075]

[0076] where I is the input image, which is either a real image I real or a simulated image G(z, P', A) generated by the generator; D(I, P', A) is the probability output by the discriminator, indicating the authenticity of the input image I; W D and b D are the weight matrix and bias vector of the discriminator, respectively; V is the weight matrix of the conditional input of the discriminator; and D(I, P', A) is the probability output by the discriminator, indicating the authenticity of the input image I.

[0077] Specifically, the main task of the discriminator is to distinguish whether the input image is from real data or simulated data generated by the generator, so as to perform adversarial training with the generator under the framework of a generative adversarial network (GAN) and optimize the performance and detection accuracy of the system. The input of the discriminator not only includes the image itself (which can be a real image I real or a simulated image G(z, P', A) generated by the generator), but also includes multi-scale feature pyramid conditional features P' land the corresponding attention weight A. This design idea aims to ensure that the discriminator can combine multi-dimensional information when evaluating the authenticity of the image, and make more comprehensive and accurate judgments. The discriminator receives an image input I, which can be a real image or a generated image, in addition to a feature vector containing pyramid features P' l and attention weights. In the formula, the input vector is composed of the image I and the sum of all layer condition features This combination of inputs ensures that the discriminator can consider both the visual features and structural features of the image. The weight matrix V is responsible for weighted fusion of the input vector, followed by nonlinear transformation through the nonlinear activation function φ(·) (usually ReLU or LeakyReLU). The application of the nonlinear activation function aims to increase the ability of the discriminator to capture complex feature patterns, enabling it to distinguish more subtle changes and feature differences. After nonlinear transformation, the result is dot multiplied with the discriminator's weight matrix W D and the bias vector b DThe step of this linear transformation is to adjust the feature space so that the output of the discriminator can more accurately reflect the authenticity of the input image. Finally, the discriminator maps the output value to the interval [0, 1] through the output layer activation function σ(·) (usually sigmoid function), and the output result can be interpreted as the probability value of the image being real. The choice of sigmoid function here is because it can well represent the characteristics of probability value, limiting the output of the discriminator within a reasonable range, making the result easier to interpret and optimize. The core of the design of the discriminator lies in its comprehensive utilization of multi-source information. By combining the features of the input image and the pyramid conditional features, the discriminator can make more detailed and accurate judgments on the authenticity of the image. Especially in the context of land change detection of remote sensing data, different types of surface changes will exhibit different visual and structural features in the image, and the discriminator needs to be able to identify the subtle differences between these features. Through the introduction of conditional features and attention mechanisms, the discriminator not only focuses on the global features of the image, but also deeply understands the feature patterns of local changes, thereby improving its ability to distinguish between real and generated images. This design enables the generator and the discriminator to constantly promote each other in the process of adversarial training, and the generator will continuously optimize its generation strategy according to the feedback of the discriminator, while the discriminator will improve its discrimination ability with the improvement of the generator. Ultimately, the adversarial learning process of this generative adversarial network can greatly improve the detection accuracy and robustness of the system for land change. Through this design based on multi-scale features and attention mechanisms, the discriminator becomes a key component in the entire system, not only as the optimization target of the generator, but also as one of the core components to ensure the effective training of the generative adversarial network. Through this dual input and conditional discrimination strategy, the invention can demonstrate stronger detection capability and higher efficiency in complex remote sensing data analysis tasks, especially in the processing of high-dimensional and multi-scale features. In this way, the discriminator is not just a simple true or false judgment, but through deep mining and complex analysis of features, it ensures that each result output by the system has high reliability and accuracy.

[0078] Example 8: Loss function of the generator As follows:

[0079]

[0080] where p z (z) is the probability distribution of random noise z; is the expected operator; P' is the conditional feature set; is the basic generator loss, which is used to maximize the probability that the generated image is identified as a real image by the discriminator; is a gradient normalization regularization term, which is used to balance the learning speed of the generator under different conditional features; and are the gradients with respect to the generator and the conditional features, respectively, and k is the adaptive weight for each feature layer, and ∈ is a constant to prevent division by zero; is a reconstruction loss term that ensures the generated image is consistent with the real image in the conditional feature space, using the L2 norm to measure the difference between the generated image and the real image; |||2 denotes the L2 norm operation; denotes the loss part corresponding to the conditional features of the k-th layer; ||| denotes the operation of calculating the determinant of a matrix or the modulus of a vector; p P′ (P') is the probability distribution of the conditional features P'; I real (P') is the conditional feature of the real image at the l-th scale. l

[0081] Specifically, the loss function of the generator includes a basic generator loss part, which is used to maximize the probability of the generated image being identified as a real image by the discriminator. This part of the loss is achieved through an adversarial learning mechanism, which uses the expectation operator to optimize the probability distribution of the generated image. Specifically, the generator tries to minimize the performance of the discriminator in distinguishing real images from generated images, i.e., by minimizing to maximize the probability of the discriminator identifying the generated image as a real image. In this adversarial framework, the generator continuously improves its generation strategy, making it increasingly difficult for the discriminator to distinguish between the generated image and the real image. This process of mutual game gradually improves the authenticity and detail performance of the generated image. However, relying solely on the basic generator loss may result in uneven learning of the generator in different conditional features, so the present invention introduces a gradient normalization regularization term to balance the learning speed and quality of the generator in different conditional features. The purpose of this regularization term is to adjust the gradient contribution of different feature layers to prevent the generator from learning too fast or too slow in certain feature layers. The formula expresses this balancing strategy, by calculating the gradient norm ratio of the generator in different feature layers, the generator can automatically adjust the importance weight k of each feature layer, so as to ensure that all feature layers can be properly learned during the training process. Here, ∈ is a small constant to prevent division by zero errors and ensure the stability of gradient calculation. Through this gradient normalization method, the generator can achieve balanced learning in different feature scales, making the model's response to various surface change patterns more consistent and stable. In addition, to ensure that the generated image is highly consistent with the real image in the feature space, the loss function of the generator also includes a reconstruction loss term. This reconstruction loss term uses the L2 norm to measure the difference between the generated image and the real image in the multi-scale conditional feature space. The formula described in ​This is to optimize this goal. By minimizing the Euclidean distance between generated images and real images in the feature space, this loss term ensures that the generator can not only generate realistic visual effects, but also approximate the distribution of real data in a higher-dimensional feature space. In other words, this reconstruction loss effectively forces the generator to consider not only the overall visual similarity when generating images, but also to ensure consistency with real images in terms of detailed features. This is particularly important for land change detection systems, as subtle changes in the land surface are often the key to detection and early warning systems.

[0082] Embodiment 9: Loss function of discriminator As follows:

[0083]

[0084] where p data (I real ) is the probability distribution of real images; 1-D(G(z, P', A), P', A) represents the discriminant output of the discriminator for the authenticity of the generated image, and the probability of the generated image being false is obtained;

[0085] is the first discriminator loss, which is used to maximize the probability of the discriminator correctly identifying real images; is the second discriminator loss, which is used to maximize the probability of the discriminator correctly identifying generated images; is a multi-scale contrastive loss term that enhances the sensitivity of the discriminator to subtle changes by the difference between the outputs of discriminators of different scales; p and q are both norms of contrast measures.

[0086] Specifically, the core part of the generator loss function is the basic generator loss, which realizes the game between the generator and the discriminator through an adversarial learning mechanism. Specifically, the goal of the generator is to generate images as close to real as possible, so that the discriminator is difficult to distinguish their true and false. To this end, the generator attempts to minimize the misjudgment rate of the discriminator for its generated images, that is, by minimizing the expression To achieve this, the generator is trained to maximize the probability that the generated image is identified as real by the discriminator. In this way, the generator adjusts its parameters during training to try to "fool" the discriminator, thereby generating more realistic images. Under the framework of GAN, this adversarial learning mechanism promotes the generator and discriminator to compete and co-evolve during training, ultimately enabling the generator to simulate high-precision land surface changes. However, relying solely on adversarial loss is not enough to ensure the comprehensive learning effect of the generator. In the application of GAN, different feature layers may contribute differently to the learning of the generator, so a mechanism is needed to balance the learning speed and effect of the generator in different feature layers. For this purpose, the present invention introduces a gradient normalization regularization term to adjust the learning intensity of the generator in different conditional features. The core idea of this regularization term is to balance the learning process of the generator by measuring the gradient contribution of different feature layers. The gradient normalization part in the formula is designed to control the learning rate of each feature layer and prevent some feature layers from over-learning while others from under-learning. In this way, the generator can learn evenly in the multi-scale feature space, thereby better capturing the overall picture of the land surface changes. Here, ∈ is a constant to prevent division by zero, ensuring the stability and numerical safety of the calculation. The introduction of this regularization term effectively avoids the problem of gradient explosion or disappearance, making the learning process of the generator more stable and efficient. In addition, to further enhance the consistency of the generated image and the real image in the feature space, the generator loss function also includes a reconstruction loss term. This reconstruction loss term measures the difference between the generated image and the real image in the multi-scale conditional feature space through the L2 norm. Specifically, the reconstruction loss term The purpose of the reconstruction loss is to minimize the Euclidean distance between the generated image and the real image in the feature space. This distance measurement method ensures that the generator not only generates realistic images in terms of visual effects, but also maintains similarity with real images in high-dimensional feature space. By optimizing this reconstruction loss, the generator is forced to generate images that not only mimic the overall visual features, but also capture subtle surface change features. This is crucial for land change detection systems, as changes in the ground often exhibit different features at different scales, and small changes in these features may be the focus of detection systems for identification and analysis. Through the joint action of the above three parts, the loss function of the generator can comprehensively guide the learning direction and strategy of the generator during the training process. The basic generator loss ensures that the generator always seeks the optimal solution in the confrontation with the discriminator, thereby generating realistic images; the gradient normalization regularization term balances the learning rate of different feature layers, avoiding overfitting or underfitting problems; the reconstruction loss term further ensures that the generated image maintains high consistency with the real image in the feature space. The design of this multiple loss strategy enables the generator to effectively learn and simulate various land surface change conditions when processing complex remote sensing data, while maintaining an efficient learning process. Finally, the innovation of this loss function design lies in its ability to solve the challenges faced by traditional generative adversarial networks when dealing with high-dimensional, multi-scale data, thereby improving the overall performance and accuracy of the land change detection system.

[0087] In an embodiment, the land change detection part detects land change based on the generated image pairs by the following formula to obtain a land change detection score:

[0088]

[0089] wherein is the final land change detection score, which quantifies the change between the two generated images, and the higher the score, the more significant the change; N represents the total number of pixels; G1(x i , y i ) and G2(x i , y i ): represent the pixel values of the first and second generated images at position (x i , y i ) at different times, respectively; and represent the gradient vectors of the first and second generated images at position (x i , y i ), respectively; γ: a weight parameter, with a value range of [0, 1], used to balance the importance of gradient change and Euclidean distance in land change detection.

[0090] Specifically, two main comparison indicators are image gradient difference and pixel value difference. These two indicators represent different dimensions of image changes in detail and overall. Gradient difference is mainly used to capture the changes in the edge and detail part of the image. Gradient is a measure of image gray scale change, which can sensitively reflect the changes of edges and textures in the image. Therefore, by comparing the gradient difference of two generated images, the system can better detect subtle but significant edge information in surface changes, such as changes in building outlines, road widening or narrowing, etc. This sensitivity to edge changes is particularly important for capturing high-precision land change information, because edge changes are often the most intuitive manifestation of surface changes. On the other hand, pixel value difference ||G1(x i , y i -G2(x i , y i )||2is used to evaluate the overall brightness or color changes of the image, which may reflect larger-scale surface changes, such as reduction or increase in vegetation coverage, large-scale land development or flood inundation, etc. By calculating the Euclidean distance difference of generated images at each pixel position, the system can quantify the change amplitude of overall visual features between two images. The introduction of pixel value difference enables the detection process to capture large-scale macroscopic change features, ensuring the effectiveness of the system in large-scale change detection. The weight parameter γ in the formula is used to balance the gradient change and pixel value change. The value of parameter γ ranges between 0 and 1, determining the relative importance of the two differences in the final change detection score. By adjusting γ, the system can find the best balance between sensitivity and robustness in detection according to actual application requirements. For example, in some tasks, if the features of surface changes are mainly reflected in the changes of edges and details, a higher γ value can be selected to enhance the weight of gradient change; while in another case, if the focus is on the changes of overall surface features, the γ value can be reduced to increase the proportion of pixel value difference in change detection. The flexibility of this weight adjustment makes the change detection method of the present application adaptable to different types of remote sensing data and detection tasks, thus providing a highly customizable detection scheme. In addition, the normalization factor in the formula is used to average the change difference at all pixel positions to obtain a global change detection score. This score provides a unified quantitative indicator for changes at each pixel position, enabling the system to make overall evaluation of the degree of change in the entire image. In this way, the CDS score can not only reflect the details of local changes, but also summarize the overall changes of the surface, providing a powerful decision support tool for land management and environmental monitoring.

[0091] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalent replacements; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A land change detection system based on generative adversarial networks, characterized in that, The system includes: a multi-source data fusion and feature extraction part, used to perform high-dimensional fusion of multi-source remote sensing data, considering time series and spatial distribution to ensure the spatiotemporal consistency and integrity of information, and obtain fused data; an adaptive convolutional kernel design is used to capture non-stationary spatiotemporal features in the fused data; based on the non-stationary spatiotemporal features, a multi-scale feature pyramid construction method is used to ensure accurate detection of changes at different scales, and pyramid features at different scales are obtained through downsampling and upsampling operations of multi-layer features; a spatiotemporal attention mechanism is introduced to adaptively adjust the importance of pyramid features to obtain attention weights; a generative adversarial network part, used to design a conditional generative adversarial network, based on pyramid features and attention weights to obtain conditional features, the generator generates images simulating changes, and the discriminator guides the generator to improve by evaluating the authenticity of these images; a land change detection part, used to consider the gradient changes and Euclidean distance of the generated images, and to perform change detection based on the generated image pairs to obtain change detection scores; the multi-source data fusion and feature extraction part, which performs high-dimensional fusion of multi-source remote sensing data, considering time series and spatial distribution to ensure the spatiotemporal consistency and integrity of information, and the process of obtaining fused data specifically includes: from Raw data were collected from several different remote sensing data sources. Interpolation and resampling were used to align all the raw data in time and space to ensure they were on the same spatiotemporal grid. Each data source was assigned equal weight, and the resulting fused data was represented using the following formula: ； in, The merged dataset is time. and spatial location The function; X-axis coordinate; Y-axis coordinate; Total time span; For spatial domain; For the first Signal strength of each data source This is the time offset; This is the X-axis offset; This is the Y-axis offset; For the first Time decay coefficient of each data source; For the first Spatial distribution standard deviation of each data source; For the first The time resolution of each data source; For the first The periodic change cycle of each data source; For the first The periodic variation intensity coefficient of each data source; A Gaussian kernel is used to simulate spatial correlation; Used to handle differences in time resolution between different data sources.

2. The land change detection system based on generative adversarial networks as described in claim 1, characterized in that, In the multi-source data fusion and feature extraction section, an adaptive convolutional kernel design is used to capture non-stationary spatiotemporal features in the fused data through the following formula. : ； in, This is a normalization constant to ensure that the integral of the convolution kernel is 1; The standard deviation is set to the preset time dimension. The standard deviation is an adaptive value for the preset spatial dimension.

3. The land change detection system based on generative adversarial networks as described in claim 2, characterized in that, In the multi-source data fusion and feature extraction section, pyramid features at different scales are obtained using the following formula: ； in, For the first Pyramid features corresponding to layer scales; Indicates the first Sub-sampling operation; For the Laplace operator; Indicates the first The next upsampling operation; For the first Pyramid features corresponding to layer scales.

4. The land change detection system based on generative adversarial networks as described in claim 3, characterized in that, In the multi-source data fusion and feature extraction section, a spatiotemporal attention mechanism is introduced to adaptively adjust the importance of pyramid features. The specific process for obtaining attention weights includes: calculating the spatiotemporal correlation matrix of each pyramid feature using the following formula: ； in, The first in the spatiotemporal correlation matrix Okay, number Column elements; Features and characteristics The time frame is used; the attention weights of each pyramid feature are calculated using the spatiotemporal correlation matrix. ； in, In time and spatial location Upper Attention weights for hierarchical pyramid features; Features and characteristics In time and spatial location Features above; The importance parameter for controlling feature similarity is set to a specific value.

5. The land change detection system based on generative adversarial networks as described in claim 4, characterized in that, The conditions for designing a Generative Adversarial Network (GAN) are expressed in the following formula: ； in, The vector is a random noise vector that follows a normal distribution. ; For the first The conditional characteristics of a layer are calculated using the following formula: ； For the first Attention weights for hierarchical pyramid features; The simulated image generated by the generator; and These are the weight matrix and bias vector of the generator, respectively; It is a non-linear activation function, either a ReLU function or a LeakyReLU function; is the activation function for the output layer, which can be either the tanh function or the sigmoid function; The weight matrix is ​​the conditional input to the generator, used to fuse conditional features and noise; is the attention weight matrix; the generator output is a simulated image that is as similar as possible to the real image in a high-dimensional space.

6. The land change detection system based on generative adversarial networks as described in claim 5, characterized in that, The formula for the discriminator is expressed as follows: ； in, The input image is a real image. Or simulated images generated by the generator ; The probability output by the discriminator represents the input image. The authenticity; and These are the weight matrix and bias vector of the discriminator, respectively; The weight matrix is ​​the conditional input to the discriminator.

7. The land change detection system based on generative adversarial networks as described in claim 6, characterized in that, Loss function of generator As shown below: ； in, For random noise The probability distribution; For expected operators; For the set of conditional features; It is the basic generator loss, used to maximize the probability that the generated image is recognized as a real image by the discriminator; It is a gradient normalization regularization term used to balance the learning speed of the generator on different conditional features; and These are the gradients with respect to the generator and the conditional features, respectively. These are the adaptive weights for each feature layer. It is a constant that prevents division by zero; It is a reconstruction loss term that ensures the consistency between the generated image and the real image in the conditional feature space. The L2 norm is used to measure the difference between the generated image and the real image. Represents L2 norm operations; Indicates the first The loss component corresponding to the conditional features of the layer; This represents the operation of finding the determinant of a matrix or the modulus of a vector; Conditional features The probability distribution; For real images in the first Conditional features at the layer scale.

8. The land change detection system based on generative adversarial networks as described in claim 7, characterized in that, loss function of discriminator As shown below: ； in, Probability distribution of real images This represents the output of the discriminator's judgment on the authenticity of the generated image, yielding the probability that the generated image is fake; It is a multi-scale contrast loss term that enhances the discriminator's sensitivity to subtle changes by using the differences in discriminator outputs at different scales; and Both are norms of the contrast quantity.

9. The land change detection system based on generative adversarial networks as described in claim 8, characterized in that, The land change detection section uses the following formula to perform change detection based on the generated image pairs, and obtains the change detection score: ； in, This is the final change detection score, used to quantify the changes between the two generated images. The higher the score, the more significant the change. Indicates the total number of pixels; and : These represent different times and locations. The pixel values ​​of the first and second generated images; and These represent the positions of the first and second generated images, respectively. The gradient vector on; A weight parameter, with a value range of... Between these two factors, the importance of gradient change and Euclidean distance in change detection is balanced.

Images