Method, system and medium for generating land use change image

Abstract

The present application relates to a method, system, and medium for generating land use change images. The generation method includes: inputting first noise information and first feature information into a diffusion model, so that the diffusion model outputs predicted second noise information; the first feature information includes a change mask, a diffusion time step, a time label, and a ground feature category label; generating a target image based on the second noise information; the target image indicates an image before or after the land change. This method can utilize the diffusion model to construct an association representation between active and passive features through a collaborative mechanism using noise information and feature information, thereby generating before-after image sample pairs with strong realism and significant differences, effectively solving the overfitting and underfitting problems caused by sample scarcity in the training stage of deep learning models, and thus improving the quality and quantity of training samples.

Description

Technical Field

[0001] The present application relates to the technical field of land change detection, and in particular to a method, system, and medium for generating a land use change image. Background Art

[0002] Land-use change monitoring is widely used in various fields, including urban planning, ecological and environmental protection, and resource management. Traditional land-use change detection methods primarily rely on time-series analysis of remote sensing data. The most common approach involves identifying changes in land-use types by monitoring remote sensing data over different time periods. However, a single data source often fails to fully capture the diversity and complexity of land-use change. To address this deficiency, multi-source data fusion techniques have emerged. However, these methods often face two challenges: First, differences in spatial resolution and scale between active and passive data can lead to information redundancy or loss due to direct fusion. Second, traditional fusion methods often rely on artificially designed features or shallow models, making it difficult to deeply explore the deep semantics within the data. The introduction of convolutional neural networks (CNNs) and attention mechanisms has further enhanced the models' ability to extract heterogeneous features and focus on key areas. However, these methods rely on a large number of high-quality training samples. In the context of multi-source remote sensing change detection, high-quality, paired pre- and post-change annotated data is extremely scarce, which has become a key bottleneck restricting model performance. Summary of the Invention

[0003] Based on this, a method, system and medium for generating land use change images are provided to solve the problem of lack of high-quality training samples in the prior art.

[0004] In a first aspect, a method for generating a land use change image is provided, the method comprising:

[0005] Inputting first noise information and first feature information into a diffusion model so that the diffusion model outputs predicted second noise information; the first feature information includes a change mask, a diffusion time step, a time label, and a feature category label; the change mask indicates an area in an image corresponding to the first noise information that needs to change, the time label indicates a time phase before and after the land change, the feature category label indicates a feature type in the image, and the diffusion time step indicates the current step number in the diffusion process;

[0006] Generate a target image according to the second noise information; the target image indicates an image before the land change or after the land change;

[0007] The diffusion model is trained according to the following steps:

[0008] Acquire multiple groups of noise images as training samples; the noise images are obtained by adding third noise information to remote sensing images, and the remote sensing images include spectral images and radar images;

[0009] Inputting the noise image and the second feature information into the model to be trained, so that the conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, and the prediction sub-model of the model to be trained generates fourth noise information according to the conditional feature vector and the noise image;

[0010] The value of the loss function is calculated according to the third noise information and the fourth noise information, and the model parameters of the diffusion model are adjusted according to the value of the loss function until a preset iteration stop condition is met to obtain the diffusion model.

[0011] Optionally, inputting the noise image and the second feature information into a model to be trained, so that a conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, includes:

[0012] The noise image and the second feature information are input into the model to be trained, so that the conditional encoder converts the time label, the feature category label and the diffusion time step into embedding vectors through an embedding layer and / or a linear transformation, and the embedding vectors are fused to generate the conditional feature vector. The change mask encoder of the conditional encoder extracts features from the change mask to generate a spatial feature map.

[0013] Optionally, the variation mask encoder includes two layers of convolutional networks, each layer of the convolutional network includes a 3x3 convolution, a ReLU activation function and BatchNorm.

[0014] Optionally, the loss function is a mean square error between the third noise information and the fourth noise information.

[0015] Optionally, generating a target image according to the second noise information includes:

[0016] A target image is generated according to a sampling formula and the second noise information.

[0017] Optionally, after generating a target image according to the second noise information, the method further includes:

[0018] The target image is used as the first noise information to be input to the diffusion model for the next time, and is iterated a preset number of times to generate the target image after iteration; the preset number of times is determined according to the diffusion time step.

[0019] Optionally, after generating the target image according to the sampling formula and the second noise information, the method further includes:

[0020] Inputting a first image and a second image into a multi-scale fusion network, so that the multi-scale fusion network generates a first feature corresponding to the first image and a second feature corresponding to the second image; the target image includes the first image and the second image, the first image indicates an image before land change, and the second image indicates an image after land change;

[0021] The first feature and the second feature are input into a land use change prediction network to generate a land use change result based on the first feature and the second feature; the land use change result represents whether the land use has changed and / or the type of land use change.

[0022] Optionally, the multi-scale fusion network includes a half-scale active-passive convolutional attention fusion network and a quarter-scale active-passive convolutional attention fusion network.

[0023] In a second aspect, a system for generating a land use change image is provided, the system comprising:

[0024] An output module, configured to input first noise information and first feature information into a diffusion model, so that the diffusion model outputs predicted second noise information; the first feature information includes a change mask, a diffusion time step, a time label, and a feature category label; the change mask indicates an area in an image corresponding to the first noise information that needs to change, the time label indicates a time phase before and after a land change, the feature category label indicates a feature type in the image, and the diffusion time step indicates a current step number in a diffusion process;

[0025] A generating module, configured to generate a target image according to the second noise information; the target image indicates an image before or after land change;

[0026] Model training module, used to:

[0027] Acquire multiple groups of noise images as training samples; the noise images are obtained by adding third noise information to remote sensing images, and the remote sensing images include spectral images and radar images;

[0028] Inputting the noise image and the second feature information into the model to be trained, so that the conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, and the prediction sub-model of the model to be trained generates fourth noise information according to the conditional feature vector and the noise image;

[0029] The value of the loss function is calculated according to the third noise information and the fourth noise information, and the model parameters of the diffusion model are adjusted according to the value of the loss function until a preset iteration stop condition is met to obtain the diffusion model.

[0030] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the method described in the first aspect is implemented.

[0031] On the basis of conforming to the common sense in this field, the above-mentioned preferred conditions can be arbitrarily combined to obtain the preferred embodiments of the present application.

[0032] The above-mentioned land use change image generation method, system, and medium utilize a diffusion model to construct a correlation representation between active and passive features through a collaborative mechanism using noise information and feature information, thereby generating realistic and significantly different before-after image sample pairs. This application combines multiple features in remote sensing image data, including spectral images and radar images, and uses a collaborative mechanism to construct correlations between the features. This can obtain multiple sample pairs, which can effectively solve the overfitting and underfitting problems caused by sample scarcity during the training phase of deep learning models, thereby improving the quality and quantity of training samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flowchart of a method for generating a land use change image in one embodiment;

[0034] Figure 2 A schematic diagram of a model input for a method for generating a land use change image in one embodiment;

[0035] Figure 3 A schematic diagram of a conditional encoder for a method for generating a land use change image in one embodiment;

[0036] Figure 4 is a second schematic diagram of a diffusion model of a method for generating a land use change image in one embodiment;

[0037] Figure 5 Schematic diagram of the structure of a multi-scale fusion network of a method for generating a land use change image in one embodiment;

[0038] Figure 6 1 is a schematic structural diagram of a system for generating a land use change image in one embodiment;

[0039] Figure 7 FIG. 1 is a schematic structural diagram of an electronic device in an embodiment. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0041] Figure 1 A method for generating a land use change image provided in an embodiment of the present application includes:

[0042] S11 . Input the first noise information and the first feature information into a diffusion model, so that the diffusion model outputs predicted second noise information.

[0043] The first noise information may be presented in the form of an image, and may include Gaussian noise, noise with a specific distribution, multi-source noise, and the like.

[0044] The second noise information may be presented in the form of an image. It can be understood that the diffusion model identifies the noise information in the first noise information and generates the second noise information. The first noise information and the second noise information may be presented in the form of an image.

[0045] The first feature information includes change mask, diffusion time step, time label, and feature category label.

[0046] The diffusion model can be a conditional UNet (U-shaped network) or a conditional Transformer (conditional converter).

[0047] The change mask indicates the area of ​​the image corresponding to the first noise information that needs to change. If the first noise information is a noisy image, the change mask indicates the area of ​​the image with the first noise information that needs to change. The time tag indicates the time phase before and after the land change, such as the time phase before the land change and the time phase after the land change. The feature category tag indicates the type of feature in the image, such as farmland, urban area, or water body. The diffusion time step indicates the current step number of the diffusion model in the diffusion process.

[0048] The first noise information and the first feature information are input into a diffusion model. The diffusion model can be used to simulate or generate images of before and after phases in a remote sensing scene for data enhancement and sample expansion.

[0049] The first feature information guides the model to generate a specific remote sensing image phase. The change mask indicates areas of spatially predicted change (e.g., from a pre-detection algorithm, user annotation, or simulation strategy) and can be binary or probabilistic. The feature category label guides the diffusion model to generate semantically consistent scene content. The time label indicates whether the model is generating an image before or after the change. If all other conditions of the first feature information remain the same, changing the time label can switch the phase of the generated target image.

[0050] S12. Generate a target image according to the second noise information.

[0051] The target image indicates an image before land change or after land change.

[0052] After the diffusion model outputs the predicted second noise information, a target image can be generated based on the second noise information. The target image can indicate images before and after the land change.

[0053] This application constructs a diffusion model that introduces controllable conditional vectors (primary feature information) to guide the diffusion process, gradually restoring remote sensing images that meet specific semantic requirements in a high-dimensional noisy space. First, before the image generation task begins, a series of input conditions (primary feature information) can be constructed to guide the model's generation direction. These conditions include a change mask, a time tag, a feature class label, a polarization mode (e.g., VV (Vertical Transmit and Vertical Receive), VH (Vertical Transmit and Horizontal Receive)), and a spatial location code (e.g., region ID or latitude and longitude). These conditions are typically input as images or embedded vectors and fed into the diffusion model's "conditional encoder" module for unified encoding. Masked conditions, such as the change region map, are directly converted into image feature maps through convolutional layers, while category or label conditions (e.g., time, polarization, feature type) are converted into vector form through embedding layers or linear transformations. All these conditional information is integrated to form a unified "conditional feature representation" that guides image generation throughout the diffusion denoising process.

[0054] In the diffusion model, conditional UNet or conditional Transformer can be used as the backbone structure of the diffusion model. In each step of back diffusion, the model receives three key inputs: the current noise image x t, the diffusion time step t, and the "conditional feature vector" constructed in the previous step. These conditions will be injected into the backbone network, specifically by concatenating with image channels, weighted fusion with intermediate features, or feature modulation through the attention mechanism, thereby guiding the model to understand "what kind of image should be generated in what area and under what attributes". For example, the change mask will tell the model "only perform image difference synthesis in these areas"; the time label will allow the model to learn the generation features of the front-phase and back-phase images respectively; the polarization mode label will activate the corresponding SAR channel output; and the ground object category label will help the model synthesize remote sensing features that are more in line with specific types of semantics. This conditional guidance mechanism allows the model to no longer blindly generate images, but to explicitly synthesize target images according to the structural rules set by humans.

[0055] Throughout the denoising sampling process, starting with a pure noise image, the diffusion model gradually reduces noise at each step, approaching a high-resolution, multi-channel remote sensing image. Throughout this process, all conditional information is continuously integrated with the model's feature channels, ensuring that the final output image not only possesses high-quality spatial and textural structure but also meets specific task requirements, such as generating SAR dual-polarization images or images with specific variation patterns.

[0056] Ultimately, the model output can be a complete image pair, including a multispectral image before and after the change, and the corresponding SAR image. The number of channels can be set according to the task requirements (for example, RGB+NIR for a total of 4 channels, SAR VV+VH for a total of 2 channels). These output images are typically sized at the standard 256×256 remote sensing tile size and can be used for subsequent change detection training, data augmentation, or simulation analysis.

[0057] Overall, by tightly integrating the conditional embedding mechanism with the diffusion denoising network, this application achieves collaborative modeling and generation of multi-source remote sensing imagery. This allows for the synthesis of high-resolution, multi-channel remote sensing data pairs with reasonable variation logic, even in the absence of real samples. This approach balances flexibility with precise control, possessing broad engineering applicability and scientific research value.

[0058] For example, the method for generating a land use change image can be as follows: setting the target phase, determining whether the phase to be generated by the diffusion model is the front phase (t1) or the back phase (t2), and setting the corresponding time tag, inputting the first feature information including the time tag into the diffusion model, and the diffusion model internally processes the first feature information through a conditional encoder and an MLP (Multilayer Perceptron) to guide diffusion generation. Then, the first noise information is input into the diffusion model. The first noise information is a pre-initialized Gaussian noise image with the same size as the target image. Alternatively, a fixed random seed can be used to generate the first noise information as the starting point of the diffusion process. The diffusion model performs T-step back diffusion sampling. In the diffusion model, the image generated in each step will be used as the input of the next step until it gradually approaches the target image. It can be understood that in each step of the back diffusion process, the current image will be input. And the first feature information, the first feature information includes the time step s and all condition vectors, the diffusion model will be based on the current first noise information and the first feature information to generate an image closer to the target . Generated Will be used as input for the next step , continue denoising, repeating the above steps until the time step reaches 0, generating the second noise information. Then, based on this second noise information, generate the target image x0. To generate an image of another time phase, modify the time tag to that time phase and repeat the above steps. This generates an image pair of the preceding and following time phases. Both the multispectral image and the SAR (Synthetic Aperture Radar) image can be stitched or paired for use.

[0059] In one embodiment, after generating the target image according to the second noise information, the method further includes:

[0060] The target image is used as the first noise information for the next input and is input into the diffusion model until a preset number of iterations are performed to generate the target image after iteration; the preset number of iterations is determined according to the diffusion time step.

[0061] In the diffusion model, the image generated at each step will be used as the input of the next step until it gradually approaches the target image. It can be understood that in each step of the back diffusion process, the current image will be input. And the first feature information, the first feature information includes the time step s and all condition vectors, the diffusion model will be based on the current first noise information and the first feature information to generate an image closer to the target . Generated Will be used as input for the next step , continue denoising, repeat the above steps until the time step is 0, generate the second noise information, and then generate the target image x0 based on the second noise information.

[0062] When generating the before and after phases, although the space, change, and ground object conditions are consistent, the time labels are different, and the semantics of the generation tasks are different, one is "before the change" and the other is "after the change". Single diffusion sampling focuses on one target, which can avoid the mixing of erroneous information in the generation process, and finally splices the two sampling results together to form a complete image pair. The diffusion model of this application uses two independent diffusion reverse samplings to gradually generate high-quality active and passive remote sensing images from pure noise under the conditions of the before and after phases respectively. The model receives a unified change mask, ground object category, polarization label and spatial position encoding in each diffusion process to ensure the spatial and semantic consistency between the before and after phases, and distinguish the before and after features through time labels. Finally, the model output includes the before-phase multispectral image, the before-phase SAR image, the after-phase multispectral image and the after-phase SAR image, which together form a pair of active and passive remote sensing image pairs for subsequent change detection, data enhancement and simulation analysis tasks.

[0063] This application synthesizes high-quality before-after temporal remote sensing data pairs in the absence of real-world change samples. This method utilizes a conditional diffusion mechanism to model the change region and jointly generate remote sensing data including multispectral (primary source) and SAR (surface source) channels, achieving joint modeling and generation enhancement of multi-source remote sensing data.

[0064] Through the conditional diffusion mechanism, this application achieves: multi-source integrated joint generation to ensure the coordination and consistency of active and passive images in multiple levels such as space, change, and category; multi-dimensional condition-driven, with precise generation control capabilities, and the ability to specify complex conditions such as change areas, polarization patterns, and time attributes; modal complementarity enhancement, through feature intercommunication and collaboration, using the penetration characteristics of SAR to supplement optical blind spots, and using optical texture details to enhance SAR performance, thereby achieving an overall improvement in generation quality; directly synthesizing change samples without real samples, providing sufficient simulation data for remote sensing change detection and algorithm training, and significantly reducing data acquisition costs.

[0065] In one embodiment, generating the target image according to the second noise information includes: generating the target image according to a sampling formula and the second noise information.

[0066] The target image that is closer to the target can be generated through the sampling formula.

[0067] The diffusion model is trained according to the following steps:

[0068] Acquire multiple groups of noise images as training samples; the noise images are obtained by adding third noise information to remote sensing images, and the remote sensing images include spectral images and radar images;

[0069] Inputting the noise image and the second feature information into the model to be trained, so that the conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, and the prediction sub-model of the model to be trained generates fourth noise information according to the conditional feature vector and the noise image;

[0070] The value of the loss function is calculated according to the third noise information and the fourth noise information, and the model parameters of the diffusion model are adjusted according to the value of the loss function until a preset iteration stop condition is met to obtain the diffusion model.

[0071] The purpose of the diffusion model is to learn to gradually restore the active and passive remote sensing images of the target phase from the noisy image at any time step under specific conditions. The input includes the noisy image, which is obtained by adding third noise information to the remote sensing image, and also includes second feature information. The second feature information includes condition information, diffusion time step, time label, change mask, ground feature category label, etc. First, the third noise information is added to the remote sensing image to simulate the t-th step of the diffusion process to obtain the noisy image x. Then, the conditional feature is constructed, and the change mask is used to generate a spatial feature map through the change mask encoder; the global conditions (such as time label, category, step size, etc.) are mapped to feature modulation parameters through the conditional encoder. ; Re-inject features, change mask encoding and The network is concatenated in the channel dimension and used as the backbone network input. Global conditioning is applied to the FiLM modulation (Feature-wise Linear Modulation) within each ResBlock. Finally, the model predicts the output of the fourth noise information, which is then reconstructed using the third and fourth noise information for supervised learning. Parameter update: Backpropagation is performed using the Adam optimizer to update the network weights. The goal is for the model to accurately reproduce the noise it has added.

[0072] Understandably, if Figure 2 As shown, the input of the diffusion model includes multispectral image X, such as a 4-channel RGB+NIR image with a resolution of N×M; SAR image Y, such as a 2-channel VV / VH polarization image with a resolution of N×M; and a change mask M. C, a binary image indicating the area of ​​change; the previous and next time phase labels S, such as the previous or next time phase; the feature category label L, such as city, farmland, or water body; and the time step t, the current step number t∈{0,1,...,T} in the diffusion process. These conditional inputs are fed into the conditional encoder module. The conditional encoder module receives multidimensional conditional information and encodes different types of input (image and label) into high-dimensional conditional features in a unified format. The final encoded conditional features are provided to the main diffusion denoising network, serving as guidance signals for the entire generation process. This information can come from existing annotations, sensor parameters, or user settings.

[0073] like Figure 3 As shown, the input phase label S can be 1 or 0, and the phase label is mapped to the embedding vector , D = 128, D represents the dimension of the embedding vector, the feature category label L can be input as an integer category label, one-hot encoding (such as 10 categories → length 10 vector), linear projection to the embedding vector , the time step t indicates the current step number t∈{0,1,…,T} in the diffusion process, which is encoded into a vector using sinusoidal embedding: ,D=128, , This embedding method can provide the diffusion model with periodic characteristics of the diffusion stage, which helps to adjust the generation strategy of each step. The final output generates the concatenated label vector part. , the global condition vector Modulation and attention guidance for backbone networks. Variation mask for The input of the binary matrix is ​​transformed into a mask encoder: a two-layer small convolutional network (3x3 convolution + ReLU + BatchNorm) to extract local spatial features and output feature maps. , take C m =8, change mask feature For and input noise image x t (The dimensions are consistent with the active and passive images to be generated) Directly stitching in the channel dimension: .

[0074] like Figure 4 As shown, the prediction sub-model receives (noise + variation mask) as input channel, receiving Global label conditions) are used to modulate intermediate features (such as FiLM, conditional normalization) in each ResBlock, ultimately ensuring that the diffusion process can be accurately guided by conditions in space, time, and categories. Noise image at the current time step , including multispectral channel (main source) + SAR channel (source), spatial feature map generated by conditional encoder , N×M is the resolution, which usually comes from spatial conditions such as change mask, and the stitching method is: This part is directly used as multi-channel input in the input layer of the active-passive collaborative conditional diffusion backbone network and enters the encoder. The spliced ​​input ensures that the diffusion process perceives the changing area and spatial conditions from the beginning, and the spatial guidance is injected from the input end. Global conditions (including the previous and next phase labels, time steps, categories, etc.) are not directly spliced ​​into the image, but the intermediate features are modulated in ResBlock (such as FiLM, conditional normalization). The active and passive collaborative conditional diffusion backbone network consists of multiple layers of ResBlock. The intermediate features processed by each ResBlock are ,use Generate modulation parameters: Through a small MLP, Mapping to scaling factor , translation coefficient Then use the FiLM layer (Feature-wise Linear Modulation): , conditional modulation is injected layer by layer into the active-passive collaborative conditional diffusion backbone network ResBlock to ensure that information such as time and category accurately guides the entire generation process from the high-dimensional feature layer.

[0075] In this application, the active-passive collaborative conditional diffusion backbone network adopts a multi-layer ResBlock structure based on UNet. In each ResBlock, the global condition vector from the conditional encoder is Mapping to scaling factors through a small MLP network and offset coefficient , and linearly modulated (FiLM) with the intermediate feature stream along the channel dimension. This layer-by-layer conditional injection ensures that multi-dimensional conditional information, such as time and category, is accurately guided throughout the diffusion denoising process, effectively improving the semantic consistency and rationality of the generated image.

[0076] The characteristic of a trained diffusion model is that it does not directly receive the previous or next phase image to generate the result. Instead, it receives random noise and a set of conditions. In the reverse diffusion process (gradual denoising from pure noise to clean images), the previous and next phase images are generated. This application does not directly input the actual previous and next phase data into the network, but instead tells it "now generate the previous phase or the next phase", allowing the model to learn how to generate a pair of matching images based on the conditions. The "active and passive collaborative diffusion enhancement" mechanism proposed in this application uses a diffusion model to simultaneously model the change process of optical images and SAR images, and learns the difference patterns between them through a collaborative mechanism to generate realistic and representative before and after change samples, effectively compensating for the problem caused by the lack of samples.

[0077] In one embodiment, the noise image and the second feature information are input into the model to be trained, so that the conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, including:

[0078] The noise image and the second feature information are input into the model to be trained, and the conditional encoder converts the time label, the feature category label and the diffusion time step into an embedding vector through an embedding layer and / or a linear transformation respectively. The embedding vectors are fused to generate a conditional feature vector. The change mask encoder of the conditional encoder extracts features from the change mask to generate a spatial feature map.

[0079] Before the image generation task begins, a series of input conditions must be constructed to guide the model's generation. These conditions include a change mask (indicating the image regions to be changed), a time tag (distinguishing previous and subsequent time phases), feature category labels (such as urban, farmland, or water bodies), polarization patterns (e.g., VV / VH), and spatial location codes (e.g., region ID or latitude and longitude). These conditions are typically input as images or embedded vectors and fed into the model's "conditional encoder" module for unified encoding. Masked conditions, such as the change region map, are directly converted into image feature maps through convolutional layers, while category or label conditions (such as time, polarization, and feature type) are converted into vector form through embedding layers or linear transformations. All of this conditional information is integrated to form a unified "conditional feature representation," which guides image generation throughout the diffusion denoising process.

[0080] In one embodiment, the variation mask encoder includes two layers of convolutional networks, each of which includes 3x3 convolution, ReLU activation function and BatchNorm.

[0081] In one embodiment, the loss function is a mean square error between the third noise information and the fourth noise information.

[0082] In one embodiment, after generating the target image according to the sampling formula and the second noise information, the method further includes:

[0083] Inputting the first image and the second image into a multi-scale fusion network, so that the multi-scale fusion network generates a first feature corresponding to the first image and a second feature corresponding to the second image; the target image includes the first image and the second image, the first image indicates an image before land change, and the second image indicates an image after land change;

[0084] The first feature and the second feature are input into a land use change prediction network to generate a land use change result according to the first feature and the second feature; the land use change result represents whether the land use has changed and / or the type of land use change.

[0085] The input data of the multi-scale fusion network includes 18 channels, of which multispectral data and SAR data provide rich spatial information. Vegetation indices (NDVI (Normalized Difference Vegetation Index) and SAVI (Soil-Adjusted Vegetation Index)) are used to reflect the vegetation cover and soil conditions in the region. Water body index (NDWI (Normalized Difference Water Index)) identifies the distribution of water bodies, such as rivers, lakes, and reservoirs. Texture features (GLCM (Gray-Level Co-occurrence Matrix), LBP (Local Binary Patterns), Gabor filter) extract spatial structural information such as building density and transportation networks. Total feature channels: Vegetation index: 2 channels (NDVI, SAVI), Vegetation indices (NDVI, SAVI): These two indices reflect the vegetation cover and soil moisture changes in the region, respectively. NDVI is mainly used to describe vegetation growth, while SAVI can better extract vegetation information based on soil influence.

[0086] Nationally Determined Water Index (NDWI): This index is used to identify the distribution of water bodies, such as rivers, lakes, and reservoirs, and helps extract information on water body changes.

[0087] Texture features (GLCM, LBP, Gabor filter): These texture features are used to extract spatial structural features such as building density and transportation networks. GLCM is used to calculate the texture co-occurrence matrix of the image, LBP extracts local binary patterns, and Gabor filter is used for multi-scale texture analysis.

[0088] The above indicators constitute two vegetation index channels, one water index channel, and 11 texture feature channels. Together with the four channels extracted from polarimetric SAR data (GLCM texture, LBP local binary pattern, Gabor filter response, and Sobel edge map), the total input channels are 18. The data size of each channel is N × M, which means the spatial resolution is N × M, so the input data dimension is 18 × (N × M).

[0089] In order to process data features at different scales, this technical solution proposes a multi-scale active and passive data fusion convolutional attention neural network, which is divided into a multi-scale active and passive convolutional fusion network and a multi-scale active and passive convolutional attention fusion network. The multi-scale active and passive convolutional fusion network consists of three parts: the original scale active and passive convolutional fusion network, the half-scale active and passive convolutional fusion network, and the quarter-scale active and passive convolutional fusion network.

[0090] The original-scale active and passive convolutional fusion network input data is sized 18 × (N × M). Feature extraction is performed using eight convolutional layers with kernels of 18 × 3 × 3, using the same padding method. After the convolution operation, eight original-scale active and passive fusion features are obtained, with a size of 8 × (N × M).

[0091] The half-scale active and passive convolutional fusion network has an input size of 18×(N×M). Feature extraction is performed by using 16 convolutional layers with kernels of 18×2×2, a stride of 2, and no padding. The 18 active and passive features are fused to generate 16 half-scale active and passive fusion features with a dimension of 16×(N / 2×M / 2). This is followed by a half-scale deconvolution layer with a kernel of 16×16×2×2, a stride of 2, and valid padding (no padding). The 16×(N / 2×M / 2) half-scale active and passive fusion features are spatially upscaled to generate 16 upscaled half-scale active and passive fusion features with a dimension of 16×(N×M).

[0092] The input size of the quarter-scale active and passive convolutional fusion network is 18×M(N×M). Feature extraction is performed by using a convolutional layer with 24 convolution kernels of 18×4×4, a stride of 4, and no padding. The 18 active and passive features are fused to generate 24 quarter-scale active and passive fusion features with a dimension of 24×(N / 4×M / 4). This is followed by a quarter-scale deconvolution layer with a deconvolution kernel of 24×24×4×4, a stride of 4, and valid padding (no padding). The 24×(N / 4×M / 4) quarter-scale active and passive fusion features are spatially upgraded to generate 24 upgraded quarter-scale active and passive fusion features with a dimension of 24×(N×M).

[0093] The multi-scale active and passive feature fusion layer combines the original-scale active and passive fusion features, the upgraded half-scale active and passive fusion features, and the upgraded quarter-scale active and passive fusion features to form a multi-scale active and passive fusion feature with a size of 48 × (N × M). This feature layer contains rich information at different scales, which can enhance the model's ability to perceive complex land use changes.

[0094] To further optimize the feature extraction process and increase the model's focus on key features, this application introduces an attention mechanism based on a multi-scale convolutional network. This network consists of a half-scale active and passive convolutional attention fusion network and a quarter-scale active and passive convolutional attention fusion network.

[0095] like Figure 5 As shown in the figure, the half-scale active and passive convolutional attention fusion network takes a 48×(N×M) feature input and extracts it through a convolutional layer. The convolution kernel is 48×2×2, with a stride of 2 and no padding. This convolutional layer generates eight half-scale active and passive fusion features of size 8×(N / 2×M / 2). Next, a deconvolution operation is performed using a kernel of 8×8×2×2 with a stride of 2, outputting a dimensionally increased feature map of size 8×(N×M). The attention mechanism layer splits each pixel block, converting each channel into a sub-block of size 1, and calculates the query matrix Q, key matrix K, and value matrix V. By calculating self-attention weights, attention is applied to the V matrix to obtain weighted output features. After this layer, the output feature map of the half-scale active and passive convolutional attention features is 8×(N×M).

[0096] The Quartet-Scale Active and Passive Convolutional Attention Fusion Network extracts features through convolutional layers with a kernel size of 4×48×4×4, a stride of 4, and no padding. This convolution operation generates four quarter-scale active and passive fusion features of size 4×(N / 4×M / 4). The features are then upscaled through two deconvolution layers, ultimately outputting a 4×(N×M) feature map of quarter-scale active and passive convolutional attention features. Similarly, an attention mechanism is applied to this network to further enhance the model's feature extraction capabilities. The final output feature map is 4×(N×M).

[0097] The multi-scale active and passive convolutional attention fusion layer superimposes the half-scale active and passive convolutional attention features and the quarter-scale active and passive convolutional attention features to form a multi-scale active and passive convolutional attention fusion feature with a size of 12 × (N × M). This feature layer contains rich information at different scales and can enhance the model's ability to perceive complex land use changes.

[0098] This application proposes a land use change detection method based on active and passive multi-scale data fusion. By combining multiple features from remote sensing imagery, including multispectral and synthetic aperture radar (SAR) data, and employing a multi-scale convolutional neural network and attention mechanism, this technical solution can accurately identify and predict land use changes, providing a new approach for land use change monitoring.

[0099] The first and second features generated by the multi-scale fusion network are input into the land-use change prediction network. The previous phase land-use prediction network accepts the 12×(N×M) multi-scale active and passive convolutional attention fusion features obtained from the previous phase as input. This is processed through a fully connected network to predict the land-use category of the previous phase. The cross-entropy loss function L1 is used as the loss function to calculate the error between the predicted result and the true label for each pixel.

[0100] The post-phase land use prediction network is similar to the pre-phase network. It takes as input a 12×(N×M) feature map of the post-phase, processes it through a fully connected network, and outputs the land use category for the post-phase. The loss function is the L2 cross-entropy loss function, which calculates the error between the predicted and true labels.

[0101] The land-use change prediction network uses a 2×12×(N×M) feature map from the previous and next phases to predict the type of land-use change. Through a fully connected layer, the network maps the 2×12-dimensional input features to a 2-dimensional output, indicating whether a change has occurred. The loss function uses the cross-entropy loss function.

[0102] To enhance model consistency, the semantic-land-use change consistency loss (LC) is proposed, considering the relationship between the semantic information of the previous and next phases and land-use change. By comparing the semantic judgment of the previous and next phases with the land-use change prediction results, the accuracy of the prediction results is further improved.

[0103] By constructing training samples of the previous and next phases, the gradient descent method is used to optimize the following joint loss function: L = (N × M) × (L1 + L2 + LB + LC).

[0104] Among them, L1, L2, LB and LC are the front-phase cross entropy loss, the back-phase cross entropy loss, the land use change loss and the semantic-land use change consistency loss respectively. Finally, by optimizing the loss function, the land use change detection task is achieved.

[0105] The embodiments provided in this application introduce for the first time a "active-passive collaborative diffusion generation" mechanism. This mechanism uses a diffusion generation model to model the changing characteristics of optical remote sensing and SAR data, respectively, and utilizes conditional diffusion to generate pairs of before-and-after samples. This effectively alleviates the problem of insufficient training samples while also improving sample diversity and authenticity. Compared to traditional data enhancement methods such as rotation, cropping, and noise superposition, this method can simulate the differences between active and passive images at the physical level, more closely resembling actual remote sensing scenarios, and improving the generalization capability and training efficiency of deep learning models.

[0106] This application also proposes a multi-level active and passive convolutional fusion network structure combined with an attention mechanism at the original scale, half-scale, and quarter-scale levels, overcoming the limitations of existing methods in modeling at a fixed scale. By using a deconvolutional structure to increase the dimensionality of low-scale features and then concatenating and fusing them with original-scale features, the network can capture change information at multiple spatial scales, improving the ability to detect fine-grained local changes (such as urban sprawl and water degradation). This structure is particularly suitable for remote sensing analysis tasks in diverse geographic environments, surpassing traditional approaches that rely solely on a single scale or shallow feature fusion.

[0107] Based on the multi-scale active and passive fusion, this application further integrates the spatial attention mechanism to construct a multi-scale active and passive convolutional attention fusion network, which enables the network to automatically focus on the areas with the most significant changes, ignore background noise interference, and improve the accuracy of change detection. Especially in the presence of scattered noise in SAR images or cloud obscuration in optical images, the attention mechanism can dynamically adjust the model's perception intensity for different areas, greatly enhancing the model's robustness to environmental interference. This is particularly critical in complex terrains such as densely populated urban areas and mountainous landforms, and is a capability lacking in current mainstream change detection methods.

[0108] Unlike traditional strategies that predict classification and change separately, this paper proposes a semantic-land use change collaborative prediction network. By constructing a semantic consistency loss function for the preceding and subsequent time phases, it jointly models and trains the land use classification and change detection tasks. This approach not only improves classification accuracy but also enhances the model's ability to discern change boundaries and subtle differences. This consistency mechanism is particularly effective in areas with difficult-to-determine boundary transitions (such as the transition from farmland to construction land), effectively reducing the probability of misjudgment and providing more coherent and practically interpretable change results.

[0109] This application fully considers the data heterogeneity (differences between active and passive sources), annotation scarcity (few variation samples), and environmental complexity (such as shadows, terrain fluctuations, and weather interference) inherent in real-world remote sensing scenarios. It incorporates systematic innovations across modeling mechanisms, feature design, enhancement methods, and loss functions. Experiments demonstrate that this method surpasses existing technologies in accuracy and stability across multi-region, multi-scenario, and multi-variation datasets. It also possesses excellent cross-regional transfer capabilities, making it particularly suitable for deployment in regional land resource monitoring and urban planning support systems, and holds broad engineering application prospects.

[0110] In summary, the technical solution of this application has made systematic breakthroughs and innovations in active and passive remote sensing data fusion, data enhancement modeling, multi-scale feature extraction, attention mechanism guidance, and semantic consistency modeling, significantly improving the accuracy, stability, and scalability of land use change detection, and has clear technological advancement and significant application value.

[0111] The following is an embodiment showing how to use the present invention to conduct an experiment on land use change detection in Yiyang City, covering necessary implementation details such as data, parameters, equipment, tools, etc.

[0112] Experimental purpose: To verify the effectiveness of the multi-scale land use change detection method enhanced by active and passive collaborative diffusion in land use change detection, and compare it with single-modal and single-scale methods.

[0113] This experiment uses remote sensing data from the European Space Agency (ESA) Copernicus program, including Sentinel-1 (SAR data) and Sentinel-2 (optical multispectral data). Sentinel-1 provides dual-polarization SAR imagery (VV and VH channels) with a resolution of 10 meters; Sentinel-2 provides multispectral data in 13 bands, with four primary bands, B2, B3, B4, and B8, used as input features. This data, also at a resolution of 10 meters, facilitates alignment with the SAR imagery. Data from 2019 and 2021, with a resolution of 256×256, were selected for the before-and-after analysis.

[0114] The preprocessing process includes data cropping, registration, resampling, cloud removal, and normalization. First, Sentinel-1 and Sentinel-2 images are geometrically corrected and spatially registered using the geographic coordinate system (WGS84) to ensure that the same area has a consistent spatial location in the two data sources. Second, the optical imagery is atmospherically corrected using the Sen2Cor algorithm, and cloud masking is identified and removed using the Fmask algorithm. SAR data is radiometrically and terrain-corrected using the Radiometric Terrain Correction (RTC) method to eliminate the effects of mountains on radar echoes. Finally, all bands and indices are normalized to the range [0,1] for depth model input.

[0115] According to the invention, three spectral indices—NDVI (Normalized Difference Vegetation Index), SAVI (Soil Adjusted Vegetation Index), and NDWI (Normalized Difference Water Index)—are extracted from multispectral imagery. Regarding texture features, four channels—contrast, entropy, uniformity, and correlation—are calculated based on the gray-level co-occurrence matrix (GLCM). This is combined with the local binary pattern (LBP) image descriptor and Gabor filters (with six directional scales) to further enhance the spatial structure of the image. Furthermore, texture features are extracted from SAR images using the three aforementioned texture operators. A total of 18 input feature maps are generated, including 14 optical and texture channels and 4 SAR channels, all with an N×M dimension.

[0116] To address the scarcity of labeled data, a sample augmentation mechanism based on the conditional diffusion model was introduced. A small number of labeled change sample pairs were used as a training set to train a diffusion model for modeling optical and SAR data. By introducing semantic guidance conditions within the change regions, the diffusion model was able to synthesize realistic before-and-after sample pairs in unlabeled areas, effectively expanding the training set. In this experiment, 500 sample pairs were synthesized for each typical change type (such as urban expansion, water body reduction, and vegetation degradation) to improve the diversity and balance of the training data.

[0117] The network consists of a multi-scale convolutional feature extraction module and an attention mechanism module. The multi-scale module uses three spatial levels: original scale (1×), half scale (1 / 2×), and quarter scale (1 / 4×), constructing hierarchical features through convolution and deconvolution. The attention mechanism layer applies channel attention and spatial attention mechanisms to the dimensionally upgraded feature maps, improving the model's responsiveness to salient regions. The fused 48-channel multi-scale active and passive feature maps serve as input to the semantic prediction module. The network uses the Adam optimizer with an initial learning rate of 0.001 and 150 training epochs. Mini-Batch SGD is used for gradient optimization with a batch size of 16.

[0118] Separate semantic classification subnetworks (for predicting land use types) and change prediction networks were constructed for the preceding and subsequent time phases. Each subnetwork consists of two fully connected layers and a softmax output layer, enabling pixel-level classification or change detection prediction. By incorporating a semantic consistency loss function, the network simultaneously optimizes the logical consistency between the preceding and subsequent semantic predictions and the change map during training. The overall loss function is L = L1 (previous time phase semantic loss) + L2 (latter time phase semantic loss) + LB (land use change loss) + LC (semantic consistency loss), improving the model's overall performance.

[0119] Experimental results and performance evaluation

[0120] In the experiments, we compared the following three methods:

[0121] 1. This invention (Multi-scale active and passive data fusion convolutional attention neural network)

[0122] 2. This invention removes the collaborative diffusion enhancement and uses an active and passive multi-scale land use change detection method.

[0123] 3. Single-modal method (optical remote sensing data)

[0124] 4. Single-modal method (SAR data)

[0125] 5. Single-Scale Convolutional Network (Optical Remote Sensing)

[0126] 6. Single-Scale Convolutional Network (SAR Data)

[0127] The experimental results are shown in Table 1:

[0128] Table 1

[0129]

[0130] In summary, this application demonstrated high stability and generalization capabilities in experiments. Through active and passive collaborative diffusion enhancement techniques, the model can be applied to data-scarce areas for remote sensing monitoring tasks. Through a deep fusion design of multi-scale convolution and attention mechanisms, it can be widely deployed in geographic information systems such as land surveys, urban expansion analysis, and ecological degradation early warning, demonstrating significant engineering utility.

[0131] It should be understood that although Figure 1-4 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1-4At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0132] like Figure 6 As shown, the present application also provides a system for generating a land use change image, the generation system comprising:

[0133] Output module 61, configured to input first noise information and first feature information into a diffusion model, so that the diffusion model outputs predicted second noise information; the first feature information includes a change mask, a diffusion time step, a time label, and a feature category label; the change mask indicates an area in the image corresponding to the first noise information that needs to change, the time label indicates a time phase before and after the land change, the feature category label indicates the feature type in the image, and the diffusion time step indicates the current step number in the diffusion process;

[0134] A generating module 62 is configured to generate a target image according to the second noise information; the target image indicates an image before or after land change;

[0135] The model training module 63 is used to:

[0136] Acquire multiple groups of noise images as training samples; the noise images are obtained by adding third noise information to remote sensing images, and the remote sensing images include spectral images and radar images;

[0137] Inputting the noise image and the second feature information into the model to be trained, so that the conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, and the prediction sub-model of the model to be trained generates fourth noise information according to the conditional feature vector and the noise image;

[0138] The value of the loss function is calculated according to the third noise information and the fourth noise information, and the model parameters of the diffusion model are adjusted according to the value of the loss function until a preset iteration stop condition is met to obtain the diffusion model.

[0139] Optionally, the model training module is used to: input the noise image and the second feature information into the model to be trained, so that the conditional encoder converts the time label, the ground feature category label and the diffusion time step into embedding vectors through an embedding layer and / or a linear transformation, fuses the embedding vectors to generate the conditional feature vector, and the change mask encoder of the conditional encoder extracts features from the change mask to generate a spatial feature map.

[0140] Optionally, the variation mask encoder includes two layers of convolutional networks, each layer of convolutional network including 3x3 convolution, ReLU activation function and BatchNorm.

[0141] Optionally, the loss function is a mean square error between the third noise information and the fourth noise information.

[0142] Optionally, the generating module is used to generate a target image according to a sampling formula and the second noise information.

[0143] Optionally, the generation module is further configured to: input the target image as the first noise information for the next input into the diffusion model, until a preset number of iterations are performed to generate the target image after iteration; the preset number of iterations is determined according to the diffusion time step.

[0144] Optionally, the generation module is further configured to:

[0145] Inputting a first image and a second image into a multi-scale fusion network, so that the multi-scale fusion network generates a first feature corresponding to the first image and a second feature corresponding to the second image; the target image includes the first image and the second image, the first image indicates an image before land change, and the second image indicates an image after land change;

[0146] The first feature and the second feature are input into a land use change prediction network to generate a land use change result based on the first feature and the second feature; the land use change result represents whether the land use has changed and / or the type of land use change.

[0147] Optionally, the multi-scale fusion network includes a half-scale active-passive convolutional attention fusion network and a quarter-scale active-passive convolutional attention fusion network.

[0148] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. The system embodiment described above is only illustrative, in which the units described as separate components may or may not be physically separated, and the components of the units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present application solution.

[0149] Figure 7This is a structural diagram of an electronic device shown in an example embodiment of the present application. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, the method described in any of the above embodiments is implemented. Figure 7 The electronic device 70 shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0150] like Figure 7 As shown, electronic device 70 may be implemented as a general-purpose computing device, such as a server device. Components of electronic device 70 may include, but are not limited to, at least one processor 71, at least one memory 72, and a bus 73 connecting various system components (including memory 72 and processor 71).

[0151] The bus 73 includes a data bus, an address bus, and a control bus.

[0152] The memory 72 may include a volatile memory, such as a random access memory (RAM) 721 and / or a cache memory 722 , and may further include a read-only memory (ROM) 723 .

[0153] The memory 72 may also include a program tool 725 (or utility) having a set (at least one) of program modules 724, such program modules 724 including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include an implementation of a network environment.

[0154] The processor 71 executes various functional applications and data processing by running computer programs stored in the memory 72, such as the method provided in any of the above embodiments.

[0155] The electronic device 70 can also communicate with one or more external devices 74 (e.g., a keyboard, pointing device, etc.). This communication can occur via an input / output (I / O) interface 75. Furthermore, the electronic device 70 can communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 76. As shown, the network adapter 76 communicates with other modules of the electronic device 70 via a bus 73. It should be understood that, although not shown, other hardware and / or software modules can be used in conjunction with the electronic device 70, including but not limited to microcode, device drivers, redundant processors, external disk drive arrays, RAID (RAID) systems, tape drives, and data backup storage systems.

[0156] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, depending on the embodiment of the present application, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0157] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the method provided in any of the above embodiments when the program is executed by a processor.

[0158] Specifically, the readable storage medium may include but is not limited to: a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0159] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0160] An embodiment of the present application further provides a computer program product, including a computer program, which implements any of the above methods when executed by a processor.

[0161] The program code for executing the computer program product of the present application may be written in any combination of one or more programming languages, and the program code may be executed entirely on the user device, partially on the user device, as an independent software package, partially on the user device and partially on a remote device, or entirely on the remote device.

[0162] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0163] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for generating a land use change image, characterized in that: The generation method comprises: Inputting first noise information and first feature information into a diffusion model so that the diffusion model outputs predicted second noise information; the first feature information includes a change mask, a diffusion time step, a time label, and a feature category label; the change mask indicates an area in an image corresponding to the first noise information that needs to change, the time label indicates a time phase before and after the land change, the feature category label indicates a feature type in the image, and the diffusion time step indicates the current step number in the diffusion process; Generate a target image according to the second noise information; the target image indicates an image before the land change or after the land change; The diffusion model is trained according to the following steps: Acquire multiple groups of noise images as training samples; the noise images are obtained by adding third noise information to remote sensing images, and the remote sensing images include spectral images and radar images; Inputting the noise image and the second feature information into the model to be trained, so that the conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, and the prediction sub-model of the model to be trained generates fourth noise information according to the conditional feature vector and the noise image, wherein the second feature information includes a diffusion time step, a time label, a change mask, and a ground object category label; The value of the loss function is calculated according to the third noise information and the fourth noise information, and the model parameters of the diffusion model are adjusted according to the value of the loss function until a preset iteration stop condition is met to obtain the diffusion model.

2. The generation method according to claim 1, wherein: The step of inputting the noise image and the second feature information into a model to be trained, so that a conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, includes: The noise image and the second feature information are input into the model to be trained, so that the conditional encoder converts the time label, the feature category label and the diffusion time step in the second feature information into embedding vectors through an embedding layer and / or a linear transformation, and the embedding vectors are fused to generate the conditional feature vector. The change mask encoder of the conditional encoder extracts features from the change mask in the second feature information to generate a spatial feature map.

3. The generation method according to claim 2, wherein: The variation mask encoder includes two layers of convolutional networks, each of which includes 3x3 convolution, ReLU activation function and BatchNorm.

4. The generation method according to claim 1, wherein: The loss function is a mean square error between the third noise information and the fourth noise information.

5. The generation method according to claim 1, wherein: Generating a target image according to the second noise information includes: A target image is generated according to the sampling formula and the second noise information.

6. The generation method according to claim 1, wherein: After generating the target image according to the second noise information, the method further includes: The target image is used as the first noise information to be input to the diffusion model for the next time, and is iterated a preset number of times to generate the target image after iteration; the preset number of times is determined according to the diffusion time step.

7. The generation method according to claim 5, wherein: After generating the target image according to the sampling formula and the second noise information, the method further includes: Inputting a first image and a second image into a multi-scale fusion network, so that the multi-scale fusion network generates a first feature corresponding to the first image and a second feature corresponding to the second image; the target image includes the first image and the second image, the first image indicates an image before land change, and the second image indicates an image after land change; The first feature and the second feature are input into a land use change prediction network to generate a land use change result based on the first feature and the second feature; the land use change result represents whether the land use has changed and / or the type of land use change.

8. The generation method according to claim 7, wherein: The multi-scale fusion network includes a half-scale active and passive convolutional attention fusion network and a quarter-scale active and passive convolutional attention fusion network.

9. A system for generating land use change images, characterized in that: The generation system comprises: An output module, configured to input first noise information and first feature information into a diffusion model, so that the diffusion model outputs predicted second noise information; the first feature information includes a change mask, a diffusion time step, a time label, and a feature category label; the change mask indicates an area in an image corresponding to the first noise information that needs to change, the time label indicates a time phase before and after a land change, the feature category label indicates a feature type in the image, and the diffusion time step indicates a current step number in a diffusion process; A generating module, configured to generate a target image according to the second noise information; the target image indicates an image before or after land change; Model training module, used to: Acquire multiple groups of noise images as training samples; the noise images are obtained by adding third noise information to remote sensing images, and the remote sensing images include spectral images and radar images; Inputting the noise image and the second feature information into the model to be trained, so that the conditional encoder of the model to be trained generates a conditional feature vector according to the second feature information, and the prediction sub-model of the model to be trained generates fourth noise information according to the conditional feature vector and the noise image, wherein the second feature information includes a diffusion time step, a time label, a change mask, and a ground object category label; The value of the loss function is calculated according to the third noise information and the fourth noise information, and the model parameters of the diffusion model are adjusted according to the value of the loss function until a preset iteration stop condition is met to obtain the diffusion model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images