Satellite image real-time segmentation method based on large model

Abstract

The invention relates to a satellite image real-time segmentation method based on a large model. The method comprises the following steps: establishing an image segmentation system based on the SAM large model; preprocessing the satellite image data; fine tuning is carried out on SAM large model network parameters, and model parameters are optimized to adapt to specific characteristics of satellite images; finely adjusting the structure of the SAM large model, and accessing an adversarial network to improve the image segmentation precision; performing real-time segmentation on the preprocessed satellite image by using the fine-tuned SAM large model to generate a segmentation mask; post-processing, including denoising and edge smoothing, is performed on the generated segmentation mask so as to improve the segmentation quality; according to the method, high resolution and complexity of the satellite image are considered, real-time segmentation and accurate optimization of the satellite image are realized through a deep learning algorithm, a system can continuously learn and optimize according to new data, segmentation efficiency and accuracy are ensured, and complex real-time application requirements are met.

Description

Technical Field

[0001] The present invention relates to the field of image processing and pattern recognition technology, and in particular, to a real-time satellite image segmentation method based on a large model. The method uses a large model, especially a Segment Anything (SAM) model, to quickly and accurately segment high-resolution satellite images. In the fields of remote sensing technology, geographic information system (GIS), environmental monitoring, agricultural monitoring, urban planning and disaster management, there is an increasing demand for real-time processing and analysis of satellite images. The present invention aims to provide an efficient and accurate satellite image segmentation solution to support decision-making and problem solving in the above fields. Background Art

[0002] With the rapid development of remote sensing technology, satellite images have become an important means of obtaining information about the earth's surface. In many fields such as agriculture, urban planning, and environmental monitoring, satellite images provide rich geographic and environmental data. Traditional image segmentation methods, such as those based on thresholds, edge detection, or region growing, often require a lot of manual intervention and have slow processing speeds, which cannot meet the needs of real-time processing. In addition, these methods are often limited by computing power and algorithm complexity when processing large-scale, high-resolution satellite images.

[0003] In recent years, deep learning technology has made significant progress in the field of image segmentation, especially the application of convolutional neural networks and recurrent neural networks, which has greatly improved the accuracy and efficiency of segmentation. However, existing deep learning methods still face challenges when processing satellite images, including how to process images acquired by different sensors, different resolutions and different conditions, and how to adapt to changing surface features and environmental changes.

[0004] As an advanced image segmentation model, the Segment Anything (SAM) model achieves the ability to transfer zero samples to new image distributions and tasks through prompt engineering technology. The design of the SAM model enables it to process various types of images and show excellent performance on multiple tasks. However, when the SAM model is directly applied to satellite image segmentation, it is also necessary to consider the high resolution, high spectrum and high dynamic range characteristics of satellite images, as well as the need for real-time processing.

[0005] The present invention aims to solve the limitations of the prior art by establishing an image segmentation system based on the SAM large model, preprocessing satellite image data, fine-tuning model parameters and structure, real-time segmentation, post-processing and dynamic optimization to achieve efficient and real-time segmentation of satellite images. This method pays special attention to the high resolution and complexity of satellite images, and achieves real-time segmentation and precise optimization of satellite images through deep learning algorithms. The system can continuously learn and optimize according to new data, ensuring the efficiency and accuracy of segmentation and meeting complex real-time application requirements. Summary of the invention

[0006] The present invention provides a real-time satellite image segmentation method based on a large model, aiming to solve the problems of low efficiency and insufficient accuracy of satellite image segmentation in the prior art, especially under the requirements of high resolution and real-time processing. The present invention realizes fast and high-precision segmentation of satellite images through deep learning technology, especially by utilizing the powerful processing capability of the Segment Anything (SAM) large model, and meets the requirements of real-time applications. The present invention can be widely used in the fields of military target detection, remote sensing technology, geographic information system, environmental monitoring, agricultural monitoring, urban planning and disaster management.

[0007] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a real-time segmentation method of satellite images based on a large model, comprising the following steps:

[0008] Step 1: Based on the SAM large model, adversarial network and semantic segmentation network, a satellite image segmentation large model is constructed;

[0009] Step 2: Preprocess the satellite image data;

[0010] Step 3: Adjust the network parameters of the SAM large model and optimize the parameters of the SAM large model to adapt to the specific characteristics of satellite images and improve the recognition and segmentation capabilities of the SAM large model for satellite images;

[0011] Step 4: Use the adjusted satellite image segmentation model to perform real-time segmentation on the preprocessed satellite image and generate a segmentation mask to keep synchronization with the satellite image data image sequence;

[0012] Step 5: Post-process the generated segmentation mask to obtain the real-time segmentation result.

[0013] The adversarial network generates a segmentation result according to the segmentation mask output by the mask decoder in the SAM large model and inputs it into the mask decoder. Through the learning of the SAM large model, the mask decoder outputs the final segmentation mask.

[0014] The semantic network adopts the Segformer model; the preprocessed image is semantically segmented by the Segformer model to obtain a mask, a bounding box and a target point, and input them into the prompt encoder of the SAM large model.

[0015] The image encoder in the SAM large model adds an Adapter module after the first feed-forward layer and the second feed-forward layer respectively;

[0016] For the preprocessed image, the feature map is obtained after multi-head attention, the first feedforward layer, and the first Adapter module in sequence, and the residual connection is performed with the preprocessed image. After passing through the first normalization layer, it is input into the second feedforward layer, and after passing through the second Adapter module, the residual connection is performed with the output of the first normalization layer, and the features are output through the second normalization layer.

[0017] The construction of the Adapter module includes the following steps:

[0018] The Adapter module consists of three parts: downward projection, ReLU activation function and upward projection;

[0019] Down-projection uses the first MLP layer to compress the given embedding to a lower dimension relative to the down-projected input image;

[0020] The ReLU activation function is used to perform nonlinear activation on the embedding after downward projection;

[0021] The up-projection uses a second MLP layer to expand the compressed embedding back to the original dimension of the down-projected input image.

[0022] The preprocessing of the satellite image data includes the following steps: removing noise, filling missing values, correcting outliers, resizing, normalizing, color correction, and contrast enhancement.

[0023] The network parameters of the SAM large model are adjusted and the parameters of the SAM large model are optimized as follows: when the network parameters of the SAM large model are trained, only the parameters of the Adapter module are trained, and the remaining parameters in the SAM large model are frozen.

[0024] The construction of the dynamic optimization mechanism includes the following steps:

[0025] Step (1) setting the initial parameters of the adversarial network in the satellite image segmentation model;

[0026] Step (2) using the constructed SAM large model, adversarial network, semantic network and each network parameter to perform simulation and calculate the error between the prediction result and the real data of the actual collected target characteristics;

[0027] Step (3) analyzing the distribution and characteristics of the error, using the gradient descent method to adjust the network parameters of the SAM large model, the adversarial network, and the semantic network, and updating the satellite image segmentation large model;

[0028] Step (4) returns to step (2) until the stopping condition is met, and uses an independent validation set to validate the modified satellite image segmentation model;

[0029] Step (5) During the application of the large satellite image segmentation model, new experimental data is collected and the data set is updated to achieve continuous optimization of the model.

[0030] A satellite image real-time segmentation system based on a large model, comprising:

[0031] A model building unit, used to build a large satellite image segmentation model based on the SAM large model, adversarial network and semantic segmentation network;

[0032] The model training unit is used to pre-process satellite image data; adjust the network parameters of the SAM large model, optimize the SAM large model parameters to adapt to the specific characteristics of satellite images, and improve the SAM large model's ability to recognize and segment satellite images;

[0033] The image segmentation unit is used to use the adjusted satellite image segmentation model to perform real-time segmentation on the pre-processed satellite image and generate a segmentation mask to keep synchronization with the satellite image data image sequence; post-process the generated segmentation mask to obtain a real-time segmentation result.

[0034] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the real-time segmentation method of satellite images based on a large model is implemented.

[0035] The present invention has the following beneficial effects and advantages:

[0036] The present invention comprehensively considers the high resolution and complexity of satellite images, provides more accurate image segmentation results, and realizes real-time segmentation of target images based on the powerful robustness, adaptability and generalization ability of the large model. The system can continuously learn and optimize according to new test data, ensuring the efficiency and accuracy of segmentation, and is suitable for complex real-time application requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is the SAM backbone network structure diagram;

[0038] Figure 2 is a structural diagram of the image encoder after fine-tuning in the present invention;

[0039] Figure 3It is a structural diagram after fine-tuning the large model in the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] like Figure 3 As shown, a real-time satellite image segmentation method based on a large model includes: establishing an image segmentation system based on the SegmentAnything (SAM) large model; preprocessing satellite image data, processing noise, missing values ​​and outliers in the data to ensure data quality as well as resizing and normalization to adapt to the input requirements of the SAM model; fine-tuning the network parameters of the SAM large model, optimizing the model parameters to adapt to the specific characteristics of satellite images, such as spectral characteristics and spatial resolution; fine-tuning the structure of the SAM large model, accessing the adversarial network to improve the image segmentation accuracy; using the fine-tuned SAM large model to perform real-time segmentation on the preprocessed satellite image to generate a segmentation mask; post-processing the generated segmentation mask, including denoising and edge smoothing, to improve the segmentation quality; designing a dynamic optimization mechanism to adjust the image segmentation system online according to the real-time segmentation results and user feedback to improve the segmentation accuracy. The present invention takes into account the high resolution and complexity of satellite images, realizes real-time segmentation and precise optimization of satellite images through deep learning algorithms, and the system can continuously learn and optimize according to new data, ensuring the efficiency and accuracy of segmentation, and meeting complex real-time application requirements.

[0042] The principle of image segmentation technology is based on dividing images into different areas or objects through algorithms for separate analysis or processing. With the help of deep learning, it trains neural networks to identify and locate specific objects in images, achieving accurate segmentation from pixel level to object level. It is widely used in medical imaging analysis, environmental perception of autonomous vehicles, robot navigation, and remote sensing image processing, greatly promoting the application of computer vision technology in practical scenarios.

[0043] like Figure 1As shown in the figure, Segment Anything Model (SAM) is an advanced image segmentation model that uses deep learning technology to achieve accurate recognition and segmentation of each object in the image. The SAM model uses powerful image encoders and prompt encoders to handle a variety of prompt types including points, boxes, and text, and quickly output high-quality segmentation masks. During the training process, SAM learns how to generate accurate segmentation results based on given prompts through iterative learning and a large amount of image data. The model can not only handle a single segmentation task, but also transfer to new image distributions and tasks through zero-sample learning, showing excellent generalization ability. Due to its efficient computing performance and excellent segmentation effect, the SAM model has broad application prospects in the field of satellite image analysis and processing. Its construction specifically includes the following steps:

[0044] Step 1: Establish an image segmentation system based on the Segment Anything (SAM) large model, where the SAM large model has the ability to process high-resolution satellite images.

[0045] Step 2: Preprocess the satellite image data, including but not limited to removing noise, filling missing values, correcting outliers to ensure data quality, and resizing and normalizing to adapt to the input requirements of the SAM large model. The preprocessing step further includes color correction and contrast enhancement of the satellite image to improve the accuracy of subsequent segmentation.

[0046] Use sliding average, median filtering and other methods to smooth the data to reduce the impact of noise; use the K nearest neighbor algorithm (KNN) to fill in missing values ​​based on similarity; and use the 3σ principle (triple standard deviation method) based on statistical methods to detect and filter outliers.

[0047] In statistics, the 68-95-99.7 principle is the percentage of values ​​within one standard deviation, two standard deviations, and three standard deviations from the mean in a normal distribution. More precise numbers are 68.27%, 95.45%, and 99.73%. This is the 3σ principle:

[0048] The probability that the value is distributed in (μ-σ,μ+σ) is 0.6827;

[0049] The probability that the value is distributed in (μ-2σ,μ+2σ) is 0.9545;

[0050] The probability that the value is distributed in (μ-3σ,μ+3σ) is 0.9973;

[0051] It can be considered that the values ​​of a normally distributed data set are almost all concentrated in the interval (μ-3σ,μ+3σ), and the possibility of exceeding this range is less than 0.3%. Therefore, the data that is not in this interval will be eliminated as outliers.

[0052] Step 3: Fine-tune the network parameters of the SAM large model and optimize the model parameters to adapt to the specific characteristics of satellite images, including spectral characteristics and spatial resolution, so as to improve the model's ability to recognize and segment satellite images.

[0053] The steps of fine-tuning the network parameters of the SAM large model include but are not limited to adjusting the weights of the image encoder self-attention module, the weights and biases of the fully connected layers, and adding or deleting network layers to adapt to specific satellite image features. The steps of fine-tuning the structure of the SAM large model include adding specific modules, such as the adversarial network module and the spatial pyramid pooling module, to improve the model's ability to recognize features of different regions in satellite images.

[0054] The Adapter mechanism is a technique for fine-tuning large models, which allows us to adapt to new tasks or domains by adding a small number of trainable parameters without changing the original model parameters. The core advantage of the Adapter mechanism lies in parameter efficiency. It fine-tunes the model by inserting additional small modules (i.e., Adapters) into the pre-trained model instead of updating the parameters of the entire model. During the fine-tuning process, the parameters of the original model are frozen, which means that the model retains the knowledge it has learned during the pre-training phase, while the newly added Adapter is responsible for learning specific knowledge for the new task.

[0055] like Figure 2As shown in the figure, SAM's image encoder first resizes the input image to 1024x1024, then downsamples it to 64x64 through a convolution with a convolution kernel of 16 and a stride of 16, and then adds the position code to the Transformer Block. Then design an Adapter module, which usually consists of three parts: down projection, ReLU activation function, and up projection. Down projection uses a simple MLP layer to compress the given embedding to a lower dimension, while up projection uses another MLP layer to expand the compressed embedding back to its original dimension. In SAM's image encoder, two Adapters are used for each ViT block. The first Adapter is located after the multi-head attention and before the residual connection, and the second Adapter is placed in the residual path of the MLP layer following the multi-head attention. All adapter parameters are sampled from a normal distribution with a mean of 0 and a standard deviation of 0.01. During the training phase, only the parameters of the newly added Adapter layer are trained, while the original ViT model parameters are kept frozen. This avoids large-scale perturbations to the pre-trained model while allowing the Adapter to learn specific knowledge of the new task. The ReLU activation function is used to perform nonlinear activation on the embedding after downward projection, increasing the nonlinear expression ability of the model and thus effectively learning the complex features in the input data.

[0056] Step 4: Fine-tune the structure of the SAM large model and connect the adversarial network to improve the image segmentation accuracy. The adversarial network is used to enhance the model's robustness to outliers and noise.

[0057] Generative Adversarial Networks (GAN) is a deep learning model. It learns the distribution of data through two adversarial neural networks, the generator and the discriminator. The goal of the generator is to generate fake data that is similar to real data, while the goal of the discriminator is to distinguish between real data and fake data generated by the generator. The two networks compete with each other during the training process, with the generator constantly learning how to generate more realistic data and the discriminator constantly learning how to more accurately identify real and fake data. In the end, the generator is able to generate high-quality fake data, so that the discriminator has difficulty distinguishing between real and fake. The principle of GAN is based on the zero-sum game idea in game theory, and reaches the Nash equilibrium point through a dynamic "game process", that is, the data distribution generated by the generator is as close to the real data distribution as possible. This mechanism has enabled GAN to achieve remarkable results in many fields such as image generation, image translation, video generation, and natural language processing, and has become a hot research direction in the field of artificial intelligence.

[0058] The discriminator D is actually a binary logistic regression problem. Optimizing the minimum cross entropy is equivalent to optimizing the maximum value of

[0059]

[0060] loss=crossEntryLoss(D(x),1)+crossEntryLoss(D(G(z)),0)

[0061] Among them, x represents the real data sample, Indicates the calculation of logarithmic loss for real samples. Indicates the calculation of logarithmic loss for generated samples, V(G,D) represents the value function in the generative adversarial network, and represents the game target of the generator and the discriminator. loss represents the loss function of the discriminator, and crossEntryLoss represents the cross entropy loss function.

[0062] The goal is to optimize the discriminator D. The generator G is fixed and not optimized. The generator G is the result of the last iteration optimization, so it can be abbreviated as

[0063]

[0064] The optimization function of the generator G is

[0065]

[0066] loss=crossEntryLoss(D(G(z)),1)

[0067] The goal is to optimize the generator G, and the discriminator D is fixed and not optimized. The discriminator D is the result of the last iteration optimization, so it can be abbreviated as

[0068]

[0069] Therefore, the objective function of GAN is

[0070]

[0071] Among them, V is the letter designated to represent the cross entropy, P data (x) represents the distribution of real data, z is the noise input to the generator G, P z (z) is the distribution of noise.

[0072] By constructing an adversarial network to form a game effect with SAM, the SAM segmentation effect is further improved, so that it can still maintain a high segmentation accuracy and rate when facing complex satellite images and small target segmentation.

[0073] Step 5: Use the fine-tuned SAM large model to perform real-time segmentation on the pre-processed satellite image to generate a segmentation mask, where real-time segmentation includes inputting the satellite image into the model and quickly outputting the segmentation result to keep synchronization with the image sequence. Input the pre-processed satellite image into the fine-tuned SAM large model and output the segmentation mask within a predetermined time, where the predetermined time is based on keeping synchronization with the image sequence.

[0074] Step 6: Post-process the generated segmentation mask. The post-processing step includes using morphological operations and image filtering techniques to further optimize the quality of the segmentation mask. This implementation includes denoising and edge smoothing to improve the segmentation quality, where the post-processing step is used to improve the continuity and clarity of the segmentation edge.

[0075] Image edge smoothing refers to the process of processing the edge of an image to make it look smoother and more natural, and to reduce or eliminate the jagged or discontinuous edge. This method is often used to improve the visual quality of an image, especially when the image is enlarged, reduced, or sharpened, and the edges may appear unnaturally sharp or burred. Edge smoothing technology can make the transition of an image softer, improving the overall aesthetics and visual effects of the image.

[0076] Gaussian blur is a commonly used image edge smoothing technique that achieves a blurring effect by replacing each pixel value in an image with the weighted average of its surrounding pixel values, where the weights are determined by a Gaussian distribution.

[0077]

[0078] Among them, I blurred (x, y) is the pixel value of the blurred image at position (x, y), I(x ′ ,y ′ ) is the original image at position (x ′ ,y ′ ) is the pixel value of the image, σ is the standard deviation of the Gaussian distribution, which controls the degree of blur. The larger σ is, the more obvious the blur effect is.

[0079] Post-processing the segmentation mask using Gaussian blur technique helps improve the performance of the model edge detection and image segmentation.

[0080] Step 7: Establish a dynamic optimization mechanism for the adversarial network, set the initial network model parameters according to the experimental data, use the current network model and parameters for simulation, and calculate the average error between the predicted results and the real data of the actual collected target characteristics; by analyzing the distribution and characteristics of the error, use the gradient descent method to adjust the network model parameters, update the network model, and ensure the generalization ability of the model; in the process of model application, collect new experimental data, and feed these data back to the correction mechanism to achieve continuous optimization of the model.

[0081] The construction of the dynamic optimization mechanism includes the following steps:

[0082] Step 1: Set the initial parameters of the adversarial network part of the large model system for satellite image segmentation;

[0083] Step 2: Use the constructed network model and parameters to perform simulations and calculate the error between the predicted results and the real data of the actual collected target characteristics, using the mean square error (MSE);

[0084] Step 3: Analyze the distribution and characteristics of the error, use the gradient descent method to adjust the network model parameters, and update the network model;

[0085] Step 4: Repeat steps 2 and 3 until the stopping condition is met, and use an independent validation set to validate the modified model to ensure the generalization ability of the model;

[0086] Step 5: During the model application process, new experimental data are collected and fed back into the correction mechanism to achieve continuous optimization of the model.

[0087] The present invention also includes a user interface that allows a user to input satellite image data, view segmentation results, and provide feedback to assist in dynamic optimization.

Claims

1. A real-time satellite image segmentation method based on a large model, characterized in that: The following steps are involved: Step 1: Based on the SAM large model, adversarial network and semantic segmentation network, a satellite image segmentation large model is constructed; Step 2: Preprocess the satellite image data; Step 3: Adjust the network parameters of the SAM large model and optimize the parameters of the SAM large model to adapt to the specific characteristics of satellite images and improve the recognition and segmentation capabilities of the SAM large model for satellite images; Step 4: Use the adjusted satellite image segmentation model to perform real-time segmentation on the preprocessed satellite image and generate a segmentation mask to keep synchronization with the satellite image data image sequence; Step 5: Post-process the generated segmentation mask to obtain the real-time segmentation result.

2. The method for real-time segmentation of satellite images based on a large model according to claim 1, characterized in that: The adversarial network generates a segmentation result according to the segmentation mask output by the mask decoder in the SAM large model and inputs it into the mask decoder. Through the learning of the SAM large model, the mask decoder outputs the final segmentation mask.

3. The method for real-time segmentation of satellite images based on a large model according to claim 1, characterized in that: The semantic network adopts the Segformer model; the preprocessed image is semantically segmented by the Segformer model to obtain a mask, a bounding box and a target point, and input them into the prompt encoder of the SAM large model.

4. The method for real-time segmentation of satellite images based on a large model according to claim 1, characterized in that: The image encoder in the SAM large model adds an Adapter module after the first feed-forward layer and the second feed-forward layer respectively; For the preprocessed image, the feature map is obtained after multi-head attention, the first feedforward layer, and the first Adapter module in sequence, and the residual connection is performed with the preprocessed image. After passing through the first normalization layer, it is input into the second feedforward layer, and after passing through the second Adapter module, the residual connection is performed with the output of the first normalization layer, and the features are output through the second normalization layer.

5. The method for real-time segmentation of satellite images based on a large model according to claim 4, characterized in that: The construction of the Adapter module includes the following steps: The Adapter module consists of three parts: downward projection, ReLU activation function and upward projection; Down-projection uses the first MLP layer to compress the given embedding to a lower dimension relative to the down-projected input image; The ReLU activation function is used to perform nonlinear activation on the embedding after downward projection; The up-projection uses a second MLP layer to expand the compressed embedding back to the original dimension of the down-projected input image.

6. The method for real-time segmentation of satellite images based on a large model according to claim 1, characterized in that: The preprocessing of the satellite image data includes the following steps: removing noise, filling missing values, correcting outliers, resizing, normalizing, color correction, and contrast enhancement.

7. The method for real-time segmentation of satellite images based on a large model according to claim 1, characterized in that: The network parameters of the SAM large model are adjusted and the parameters of the SAM large model are optimized as follows: when the network parameters of the SAM large model are trained, only the parameters of the Adapter module are trained, and the remaining parameters in the SAM large model are frozen.

8. The method for real-time segmentation of satellite images based on a large model according to claim 1, characterized in that: The construction of the dynamic optimization mechanism includes the following steps: Step (1) setting the initial parameters of the adversarial network in the satellite image segmentation model; Step (2) using the constructed SAM large model, adversarial network, semantic network and each network parameter to perform simulation and calculate the error between the prediction result and the real data of the actual collected target characteristics; Step (3) analyzing the distribution and characteristics of the error, using the gradient descent method to adjust the network parameters of the SAM large model, the adversarial network, and the semantic network, and updating the satellite image segmentation large model; Step (4) returns to step (2) until the stopping condition is met, and uses an independent validation set to validate the modified satellite image segmentation model; Step (5) During the application of the large satellite image segmentation model, new experimental data is collected and the data set is updated to achieve continuous optimization of the model.

9. A satellite image real-time segmentation system based on a large model, characterized in that: include: A model building unit, used to build a large satellite image segmentation model based on the SAM large model, adversarial network and semantic segmentation network; Model training unit, used to pre-process satellite image data; Adjust the network parameters of the SAM large model and optimize the parameters of the SAM large model to adapt to the specific characteristics of satellite images and improve the SAM large model's ability to recognize and segment satellite images; The image segmentation unit is used to use the adjusted satellite image segmentation model to perform real-time segmentation on the pre-processed satellite image and generate a segmentation mask to keep synchronization with the satellite image data image sequence; post-process the generated segmentation mask to obtain a real-time segmentation result.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the real-time satellite image segmentation method based on a large model as described in any one of claims 1 to 8 is implemented.

Images