A remote sensing cloud simulation method and device of a classification guided diffusion model

Abstract

This invention discloses a remote sensing cloud simulation method and apparatus based on a classification-guided diffusion model, belonging to the field of remote sensing image cloud and fog simulation. The method includes: acquiring and preprocessing multi-temporal remote sensing images to construct a dataset containing registered pairs of cloudless and cloudy images; pre-training a feature extraction network based on cloud mask data and performing unsupervised clustering using the cloud features extracted by the network to achieve automated cloud classification; training a pre-constructed cloud simulation diffusion model using the cloudless image and cloud classification results as input, and the corresponding cloudy image as the ground truth supervision signal, to learn the mapping relationship from cloudless to cloudy states; and inputting the target cloudless image and a specified cloud category into the trained cloud simulation diffusion model to generate a simulated cloud-containing remote sensing image. This invention effectively improves the realism of generated clouds and fog, the naturalness of integration with ground features, and the controllability of generating specific cloud categories, providing sufficient and realistic training samples for cloud-related data-driven models.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image cloud and fog simulation technology, and in particular to the field of multi-temporal data-driven, classification-guided remote sensing image cloud simulation technology, and especially to a remote sensing cloud simulation method and apparatus with a classification-guided diffusion model. Background Technology

[0002] Remote sensing images, as the core data carrier for Earth observation, play an irreplaceable role in fields such as agricultural monitoring, ecological assessment, disaster emergency response, and land spatial planning. However, cloud and fog obstruction are common interference factors in the optical remote sensing imaging process, leading to blurred ground feature information and spectral distortion, severely reducing the usability of remote sensing data and posing significant challenges to subsequent interpretation and analysis.

[0003] To address the problem of cloud and fog obscuration, cloud removal technology for remote sensing images has become a research hotspot, and high-quality cloud and fog simulation datasets are a core prerequisite for training high-performance cloud removal models. Meanwhile, in scenarios such as remote sensing data augmentation, sensor performance testing, and cloud removal algorithm verification, there is an urgent need for cloud and fog simulation images with realistic, multi-scale, and multi-temporal characteristics.

[0004] Existing cloud and fog simulation technologies are mainly divided into two categories: one is a physical modeling method, which generates cloud and fog effects by simulating atmospheric scattering and cloud particle transport processes. However, this type of method relies on complex parameter tuning, has poor adaptability to different ground scene, and the generated cloud shape differs from the real remote sensing cloud. The other type is a traditional deep learning method, which can generate cloud and fog images with a certain degree of realism, but is prone to problems such as cloud shape distortion, unnatural integration with ground objects, and weak generalization ability. It is also difficult to accurately capture the dynamic change characteristics of clouds in multiple time phases.

[0005] Furthermore, existing technologies lack targeted adaptation to the multi-scale characteristics of remote sensing clouds and cannot achieve simulation requirements for specific cloud and fog types through cloud category guidance. Therefore, there is an urgent need for a cloud and fog simulation technology that can balance realism, generalization, and multi-scale adaptability to meet the diverse application needs in the field of remote sensing data processing. Diffusion models, with their powerful generation capabilities, stable training process, and precise guidance from cloud classification categories, provide a feasible path to solve the above problems. Summary of the Invention

[0006] To address the problems of insufficient cloud morphology realism, weak multi-scale cloud feature capture capability, unnatural integration with ground objects, and poor generalization in existing remote sensing cloud and fog simulation technologies, this invention provides a remote sensing cloud simulation method and device based on a classification-guided diffusion model. By constructing a classification-guided diffusion model driven by multi-temporal data, high-fidelity simulation of different cloud and fog types can be achieved, providing sufficient and realistic training samples for cloud-related data-driven models.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a remote sensing cloud simulation method using a classification-guided diffusion model, comprising the following steps: S10. Acquire and preprocess multi-temporal remote sensing images to construct a dataset containing registered cloudless and cloudy image pairs; S20. A feature extraction network is pre-trained based on cloud mask data, and the cloud features extracted by the network are used for unsupervised clustering to achieve automated cloud classification. S30. Using cloudless images and cloud classification results as input, and corresponding cloudy images as ground truth supervision signals, train a pre-built cloud simulation diffusion model to learn the mapping relationship from cloudless state to cloudy state. S40. Input the target cloudless image and the specified cloud category into the trained cloud simulation diffusion model to generate a high-fidelity simulated cloud-containing remote sensing image.

[0009] In one embodiment, step S10 includes: S101. Select images from multi-temporal remote sensing images where the cloud cover is below a preset threshold and the ground features are in a stable state, and use them as cloudless truth maps. S102. Based on the imaging time of the cloudless truth map, acquire multiple cloud-covered images taken by the same sensor in adjacent periods, covering different cloud cover ranges. S103. Based on the cloudless truth map, perform feature matching and geometric registration on the cloudy image; S104. Perform data cleaning on the registered remote sensing images and construct a multi-temporal dataset that meets the requirements.

[0010] In one embodiment, step S20 includes: S201. Construct an autoencoder structure based on VGG19 as a feature extraction network; S202. Extract cloud masks from the dataset using a cloud detection algorithm, and pre-train the feature extraction network using the cloud mask data; S203. Input the cloud mask data of the dataset into the pre-trained VGG19 module feature extraction network to obtain the output high-level feature map. S204. After dimensionality reduction and vectorization of the high-level feature map, a clustering algorithm is used to cluster cloud features to obtain cloud classification results.

[0011] In one embodiment, the VGG19 autoencoder uses the hierarchical structure, convolution kernel configuration, and downsampling logic of the VGG19 convolutional encoder as a reference to construct a decoder module that is mirror-symmetric to the encoder.

[0012] In one embodiment, the clustering algorithm is the KMeans algorithm, and the number of clusters is determined by the elbow rule.

[0013] In one embodiment, step S30 includes: S301. Construct a cloud simulation diffusion model with Transformer as the backbone network; the cloud simulation diffusion model includes an encoder, a bottleneck layer and a decoder, and uses a cosine noise scheduling strategy to generate Gaussian noise for different diffusion steps; S302. Continue preprocessing the input image data and convert the cloud classification results into text prompts; S303. The cloud simulation diffusion model is trained under supervision using cloudless images and text prompts as input and cloud-containing images as ground truth, and learns the mapping relationship from cloudless state to cloud-containing state.

[0014] Secondly, embodiments of the present invention provide a remote sensing cloud simulation device for a classification-guided diffusion model, comprising: The data preparation module is used to acquire and preprocess multi-temporal remote sensing images and construct a dataset containing registered cloudless and cloudy image pairs. The cloud feature learning and classification module is based on a feature extraction network pre-trained on cloud mask data, and uses the cloud features extracted by the network to perform unsupervised clustering to achieve automated cloud classification. The cloud simulation model training module takes cloudless images and cloud classification results as inputs and corresponding cloudy images as ground truth supervision signals to train a pre-built cloud simulation diffusion model and learn the mapping relationship from cloudless state to cloudy state. The cloud simulation inference module is used to input the target cloudless image and the specified cloud category into the trained cloud simulation diffusion model to generate a high-fidelity simulated cloud-containing remote sensing image.

[0015] Thirdly, embodiments of the present invention provide an apparatus, comprising: Image acquisition equipment used to collect multi-temporal remote sensing image data; Memory, which stores computer instructions; The processor, which is data-connected to the image acquisition device and the memory, executes the computer instructions to perform the remote sensing cloud simulation method of the classification-guided diffusion model as described in any embodiment of the first aspect, thereby generating a high-fidelity simulated cloud image.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the remote sensing cloud simulation method of the classification-guided diffusion model as described in any embodiment of the first aspect.

[0017] The descriptions of the second and third aspects of this invention can be referred to the detailed description of the first aspect; and the beneficial effects described in the second and third aspects can be referred to the analysis of the beneficial effects of the first aspect, which will not be repeated here.

[0018] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical effects: By using multi-temporal data-driven approaches, the spatiotemporal consistency between simulation data and real-world scenarios is ensured. Cloud feature clustering enables automated and refined classification of cloud and fog types. Finally, leveraging the powerful generation capabilities of the cloud simulation diffusion model, and guided by precise cloud category information, high-fidelity and controllable cloud and fog simulation is achieved. Compared to existing technologies, this invention effectively enhances the realism of generated clouds and fog, the naturalness of their integration with ground features, and the controllability of generating specific cloud types. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 A flowchart of a remote sensing cloud simulation method for a classification-guided diffusion model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the VGG19 autoencoder model structure provided in an embodiment of the present invention; Figure 3 The KMeans elbow method results are shown in the figure for the embodiments of the present invention; Figure 4 This is a schematic diagram of a remote sensing cloud simulation model for a classification-guided diffusion model provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of the classification-guided diffusion model remote sensing cloud simulation device provided in an embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram of the structure of a classification-guided diffusion model remote sensing cloud simulation device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: This invention discloses a remote sensing cloud simulation method based on a classification-guided diffusion model, referring to... Figure 1 As shown, it includes the following steps: S10. Acquire and preprocess multi-temporal remote sensing images to construct a dataset containing registered pairs of cloudless and cloudy images. For example, collect satellite images of the same area taken at different times, including completely cloudless "clean" images and images with various clouds and fog. Data preprocessing includes geometric radiometric registration, data cleaning, etc., to achieve accurate alignment, ensuring that they correspond one-to-one and guaranteeing the accuracy and quality of data alignment.

[0024] S20. A feature extraction network is pre-trained based on cloud mask data, and the cloud features extracted by this network are used for unsupervised clustering to achieve automated cloud classification. For example, a pre-trained VGG19 autoencoder is used to automatically learn the features of clouds in the image, and the KMeans clustering algorithm is used to classify the clouds into different types, such as thin clouds, thick clouds, and cirrus clouds.

[0025] S30. Using cloudless images and cloud classification results as input, and corresponding cloudy images as ground truth supervision signals, train a pre-built cloud simulation diffusion model to learn the mapping relationship from cloudless to cloudy states. In this step, a more powerful generative cloud simulation diffusion model is built, using cloudless images and specified cloud types as instructions to learn how to generate highly realistic cloudy images. During training, real cloudy images are used as the "standard answer" to guide the model.

[0026] S40. Input the target cloudless image and the specified cloud type into the trained cloud simulation diffusion model to generate a high-fidelity simulated cloud-containing remote sensing image. Finally, once the model is trained, any cloudless satellite image and the desired cloud type can be input, and the model can quickly generate a high-resolution simulation image with the specified type of cloud and fog.

[0027] In this embodiment, geometric registration preprocessing and data cleaning are performed on the acquired multi-temporal remote sensing data to complete the dataset construction. Cloud masks are extracted using the Fmask method to pre-train the VGG19 autoencoder. The autoencoder is then used to extract cloud features and perform cloud classification using the KMeans algorithm to obtain cloud classification results. A cloud simulation diffusion model is constructed and trained under supervision using cloudless images, cloud classification results as inputs, and images with clouds as ground truth. This completes the construction and training of the remote sensing image cloud simulation diffusion model. Based on this model, cloud simulation can be performed on given cloudless images and specified cloud types, generating high-fidelity remote sensing images and providing sufficient and realistic training samples for cloud-related data-driven models.

[0028] The following is a detailed description of each of the above steps: Step 1: Acquire and preprocess multi-temporal remote sensing images to construct a dataset containing registered pairs of cloudless and cloudy images. Geometric radiometric registration preprocessing and data cleaning are performed on the acquired multi-temporal remote sensing data. This step specifically includes: S101. Select images from multi-temporal remote sensing images where the cloud cover is below a preset threshold and the ground features are in a stable state, and use them as cloudless truth maps. S102. Based on the imaging time of the cloudless truth map, acquire multiple cloud-covered images taken by the same sensor in adjacent periods, covering different cloud cover ranges. S103. Based on the cloudless truth map, perform feature matching and geometric registration on the cloudy image; S104. Perform data cleaning on the registered remote sensing images and construct a multi-temporal dataset that meets the requirements.

[0029] For example, images of stable land cover periods with cloud cover <5% and NDVI change rate ≤10% are selected as cloudless ground truth images. NDVI is the Normalized Difference Vegetation Index, which can reflect crop growth and nutrient information. Subsequently, multi-phase cloud images with cloud cover of 10%-80% from the same sensor within 3-6 months are collected, using the ground truth image imaging time as the baseline. Using the cloudless ground truth image as the baseline, the ORB algorithm is used to perform geometric registration on the cloud images. By extracting stable land cover feature points, the root mean square error (RMSE) after registration is ensured to meet the following: Pixels Where N is the number of feature points. The coordinates of the feature points in the truth map. To obtain the coordinates of feature points in the cloud image after registration, the data is cleaned after registration to construct a multi-temporal remote sensing image dataset.

[0030] Although the registration algorithm has controlled the root mean square error (RMSE) to within 0.5 pixels, the human eye still needs to make a final confirmation. Operators will conduct random checks to ensure that the edges of ground features, such as houses, roads, and field ridges, completely overlap in both cloudless and clouded images, without any obvious "ghosting" or misalignment.

[0031] It can also remove unqualified image pairs. Even if the registration is perfect, some image pairs may have problems and need to be removed.

[0032] Furthermore, clouds in cloud-covered images not only obscure the view themselves, but their shadows can also cover important features. If the shaded area happens to be the focus of the study, this data set needs to be excluded.

[0033] If there are quality issues such as stripe noise, bad lines caused by sensor malfunction, color distortion, or overall blurring due to poor atmospheric conditions, the image also needs to be cleared.

[0034] The key point of this embodiment is to screen for periods when "ground features are in a stable state," but caution is still needed during the cleaning process. For example, after registration, if it is found that a building has suddenly been built on a vacant lot between cloudless and cloud-covered maps, or crops have been harvested, such real changes in the ground surface may be mistaken by the model as being caused by "clouds" and must be removed.

[0035] After this series of rigorous manual screenings, all the retained high-quality "cloudless-cloudy" paired data were systematically organized and numbered to construct a well-structured, clean, and reliable "multi-temporal remote sensing image dataset." This high-quality dataset forms a solid foundation for the subsequent training of a high-performance cloud simulation diffusion model.

[0036] Step 2: Pre-train a feature extraction network based on cloud mask data, and use the cloud features extracted by this network to perform unsupervised clustering to achieve automated cloud classification. For example, pre-train a VGG19 autoencoder using the dataset constructed in Step 1, extract latent features from the model, and combine them with KMeans for classification. This step specifically includes: S201. Construct an autoencoder structure based on VGG19 as a feature extraction network; S202. Extract cloud masks from the dataset using a cloud detection algorithm, and pre-train the feature extraction network using the cloud mask data; S203. Input the cloud mask data of the dataset into the pre-trained VGG19 module feature extraction network to obtain the output high-level feature map. S204. After dimensionality reduction and vectorization of the high-level feature map, a clustering algorithm is used to cluster cloud features to obtain cloud classification results.

[0037] The cloud mask for the dataset constructed in step 1 is extracted using the Fmask cloud detection algorithm, and an autoencoder model based on VGG19 is constructed. The network structure is as follows: Figure 2 As shown, these layers follow the original VGG19 hierarchical structure, with convolutional kernel configurations such as 3x3 convolutions and downsampling logic implemented through pooling layers. They are stacked to progressively reduce the spatial size of the feature maps while increasing the number of channels, capturing cloud features from local to global perspectives. The decoder is mirror-symmetric to the encoder, reconstructing the low-dimensional latent features obtained from the encoder into a cloud mask image of the same size as the original input. The decoder is constructed based on the encoder, using a reverse mirror-symmetric design. Each layer in the decoder exhibits the opposite trend in size and number of channels to the corresponding layer in the encoder.

[0038] The downsampling pooling layers in the encoder are replaced with corresponding upsampling layers in the decoder to gradually restore the spatial dimensions of the feature maps; the convolutional layers in the encoder have corresponding deconvolutional or convolutional layers in the decoder for fine reconstruction of image details; this symmetry ensures that the information flow can be effectively restored from the compressed latent features to the pixel space.

[0039] During the pre-training phase, the cloud mask image is input into the network. The encoder compresses it into latent features, and the decoder then attempts to reconstruct the original image from these features. By minimizing the error between the reconstructed image and the original input, the entire autoencoder is trained to learn how to retain the most essential and representative information of the cloud mask.

[0040] The extracted cloud masks were used to pre-train the network. After pre-training, the latent features of each image in the VGG19 model were extracted, and KMeans clustering was used to classify the latent features of the cloud mask images. The number of KMeans clusters was determined by the elbow method. Figure 3 As shown in the figure, the horizontal axis represents the number of KMeans cloud and fog types, and the vertical axis represents the sum of squares within a cluster, which represents the sum of the squares of the distances from a sample point to the centroid of its cluster. Combining the elbow method with the information in the figure, it can be seen that when the number of cloud and fog types exceeds 5, the rate of decrease of the sum of squares within a cluster slows down significantly. Therefore, the number of cloud and fog types in this embodiment is determined to be 5.

[0041] Step 3: Using cloudless images and cloud classification results as input, and the corresponding cloudy images as ground truth supervision signals, train a pre-built cloud simulation diffusion model to learn the mapping relationship from cloudless to cloudy states. Specifically, this includes: S301. Construct a cloud simulation diffusion model with Transformer as the backbone network; the cloud simulation diffusion model includes an encoder, a bottleneck layer and a decoder, and uses a cosine noise scheduling strategy to generate Gaussian noise for different diffusion steps; S302. Continue preprocessing the input image data and convert the cloud classification results into text prompts; S303. The cloud simulation diffusion model is trained under supervision using cloudless images and text prompts as input and cloud-containing images as ground truth, and learns the mapping relationship from cloudless state to cloud-containing state.

[0042] Step 3 is based on the multi-temporal remote sensing image data constructed in Step 1 and the cloud classification results obtained in Step 2. A cloud simulation diffusion model is established and trained. Multi-temporal data-driven classification-guided diffusion model remote sensing cloud simulation structure, such as Figure 4 As shown, this model uses a ControlNet-guided diffusion generation model as its core, combining clear, cloudless remote sensing images from the multi-temporal remote sensing image dataset constructed in step 1 with cloud category information obtained from KMeans to achieve controllable simulation generation of different cloud types. Clear, cloudless remote sensing images serve as the source of structural constraint information for ControlNet and are used as input to the diffusion model's control module. Furthermore, cloud category information is injected into the diffusion model in the form of prompts, thereby guiding the model to generate simulation results with characteristic cloud morphology and distribution features.

[0043] In this embodiment, the cloud simulation diffusion model uses a Transformer as the backbone network to construct the generative network of the diffusion model. The generative network includes an encoder, a bottleneck layer, and a decoder. The backbone parameters of the diffusion model are not updated during training; only the weights at the control module are updated. A cosine noise scheduling strategy is used to generate Gaussian noise for different diffusion steps. Paired cloud-covered images from the dataset are used as ground truth for supervised training. The network loss function is:

[0044] in, and Represent the truth image and the first Noisy image after one diffusion step, Indicates Gaussian noise. This represents network prediction noise. This indicates the control structure conditions.

[0045] Step 4: Input the target cloudless image and the specified cloud category into the trained cloud simulation diffusion model to generate a high-fidelity simulated cloud-containing remote sensing image.

[0046] Based on the cloud simulation diffusion model obtained in step 3, cloud simulation is performed on a given cloudless image and a specified cloud type to obtain a high-fidelity remote sensing image.

[0047] This invention employs a multi-temporal data-driven classification-guided diffusion model for remote sensing cloud simulation. It combines VGG19 to extract latent features and performs KMeans clustering to obtain cloud categories. These cloud categories are then used as input to the diffusion model to guide simulations of specific cloud types. By using the diffusion model for cloud simulation, the realism of cloud and fog simulations is improved, providing sufficient and realistic training samples for cloud-related data-driven models.

[0048] Example 2: This invention also provides a classification-guided diffusion model remote sensing cloud simulation device, referring to... Figure 5 As shown, it includes: The data preparation module is used to acquire and preprocess multi-temporal remote sensing images and construct a dataset containing registered cloudless and cloudy image pairs. The cloud feature learning and classification module is based on a feature extraction network pre-trained on cloud mask data, and uses the cloud features extracted by the network to perform unsupervised clustering to achieve automated cloud classification. The cloud simulation model training module takes cloudless images and cloud classification results as inputs and corresponding cloudy images as ground truth supervision signals to train a pre-built cloud simulation diffusion model and learn the mapping relationship from cloudless state to cloudy state. The cloud simulation inference module is used to input the target cloudless image and the specified cloud category into the trained cloud simulation diffusion model to generate a high-fidelity simulated cloud-containing remote sensing image.

[0049] This device can be integrated into satellite ground processing systems or cloud computing platforms to achieve operational use.

[0050] Example 3: This invention also provides a device capable of high-fidelity cloud and fog simulation of remote sensing images, which can be implemented by software and / or hardware. For example... Figure 6 As shown, the device includes: Image acquisition equipment used to collect multi-temporal remote sensing image data; Memory, which stores computer instructions; The processor, connected to the running data acquisition device and the memory, executes the computer instructions to perform the multi-temporal data-driven classification-guided diffusion model remote sensing cloud simulation method as described in any of the above embodiments, generating a high-fidelity simulated cloud image.

[0051] Specifically, the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0052] Example 4: The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a remote sensing cloud simulation method for a classification-guided diffusion model as described in Embodiment 1 above.

[0053] As a non-transitory computer-readable storage medium, memory can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program or instructions corresponding to the remote sensing cloud simulation method of the classification-guided diffusion model in Embodiment 1 above.

[0054] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function. The data storage area may store data created by the processor, etc.

[0055] In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device.

[0056] In some aspects, the memory may optionally include memory remotely located relative to the processor, which may be connected to the processor via a network.

[0057] Optionally, the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0058] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0059] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing cloud simulation method of a classification-guided diffusion model, characterized in that, Includes the following steps: S10. Acquire and preprocess multi-temporal remote sensing images to construct a dataset containing registered cloudless and cloudy image pairs; S20. A feature extraction network is pre-trained based on cloud mask data, and the cloud features extracted by the network are used for unsupervised clustering to achieve automated cloud classification. S30. Using cloudless images and cloud classification results as input, and corresponding cloudy images as ground truth supervision signals, train a pre-built cloud simulation diffusion model to learn the mapping relationship from cloudless state to cloudy state. S40. Input the target cloudless image and the specified cloud category into the trained cloud simulation diffusion model to generate a high-fidelity simulated cloud-containing remote sensing image.

2. The method of claim 1, wherein, Step S10 includes: S101. Select images from multi-temporal remote sensing images where the cloud cover is below a preset threshold and the ground features are in a stable state, and use them as cloudless truth maps. S102. Based on the imaging time of the cloudless truth map, acquire multiple cloud-covered images taken by the same sensor in adjacent periods, covering different cloud cover ranges. S103. Based on the cloudless truth map, perform feature matching and geometric registration on the cloudy image; S104. Perform data cleaning on the registered remote sensing images and construct a multi-temporal dataset that meets the requirements.

3. The method of claim 1, wherein, Step S20 includes: S201. Construct an autoencoder structure based on VGG19 as a feature extraction network; S202. Extract cloud masks from the dataset using a cloud detection algorithm, and pre-train the feature extraction network using the cloud mask data; S203. Input the cloud mask data of the dataset into the pre-trained VGG19 module feature extraction network to obtain the output high-level feature map. S204. After dimensionality reduction and vectorization of the high-level feature map, a clustering algorithm is used to cluster cloud features to obtain cloud classification results.

4. The method of claim 3, wherein, The VGG19 autoencoder is based on the hierarchical structure, convolution kernel configuration, and downsampling logic of the VGG19 convolutional encoder, and constructs a decoder module that is mirror-symmetric to the encoder.

5. The method of claim 3, wherein, The clustering algorithm is the KMeans algorithm, and the number of clusters is determined by the elbow rule.

6. The method of claim 1, wherein, Step S30 includes: S301. Construct a cloud simulation diffusion model with Transformer as the backbone network; the cloud simulation diffusion model includes an encoder, a bottleneck layer and a decoder, and uses a cosine noise scheduling strategy to generate Gaussian noise for different diffusion steps; S302. Continue preprocessing the input image data and convert the cloud classification results into text prompts; S303. The cloud simulation diffusion model is trained under supervision using cloudless images and text prompts as input and cloud-containing images as ground truth, and learns the mapping relationship from cloudless state to cloud-containing state.

7. A remote sensing cloud simulation device for classifying a guided diffusion model, characterized in that, include: The data preparation module is used to acquire and preprocess multi-temporal remote sensing images and construct a dataset containing registered cloudless and cloudy image pairs. The cloud feature learning and classification module is based on a feature extraction network pre-trained on cloud mask data, and uses the cloud features extracted by the network to perform unsupervised clustering to achieve automated cloud classification. The cloud simulation model training module takes cloudless images and cloud classification results as inputs and corresponding cloudy images as ground truth supervision signals to train a pre-built cloud simulation diffusion model and learn the mapping relationship from cloudless state to cloudy state. The cloud simulation inference module is used to input the target cloudless image and the specified cloud category into the trained cloud simulation diffusion model to generate a high-fidelity simulated cloud-containing remote sensing image.

8. An apparatus, comprising: include: Image acquisition equipment used to collect multi-temporal remote sensing image data; Memory, which stores computer instructions; The processor, which is data-connected to the image acquisition device and the memory, executes the computer instructions to perform the remote sensing cloud simulation method of the classification-guided diffusion model as described in any one of claims 1-7, thereby generating a high-fidelity simulated cloud image.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the remote sensing cloud simulation method of the classification-guided diffusion model as described in claims 1-7.

Images