Diffusion generation driven remote sensing image segmentation, height estimation and edge detection joint interpretation method

Abstract

The invention discloses a diffusion generation driven remote sensing image segmentation, height estimation and edge detection joint interpretation method, and belongs to the field of remote sensing image analysis and processing. According to the method, a unified diffusion framework is deployed through a proposed model DiffRSMT, and the unified diffusion framework is composed of two main components: a label unified coding module and a task interactive decoding module. The label unified coding module aggregates heterogeneous task labels into a unified potential representation, so that the network can effectively capture task interaction features. The task interaction decoding module adopts a time condition attention mechanism to enhance interaction between tasks in a denoising stage, so that the multi-task learning capability of the model can be improved. In addition, a time condition convolution module is introduced to further refine the denoising process. A wide range of experiments carried out on a plurality of remote sensing data sets show that when remote sensing images are combined to carry out segmentation, height estimation and boundary detection, DiffRSMT is superior to several most advanced multi-task interpretation methods.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image analysis and processing, specifically relating to a diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method. Background Technology

[0002] Remote sensing image interpretation aims to extract meaningful information from remote sensing data, such as semantic segmentation, boundary detection, and height estimation. These techniques play a crucial role in Earth observation and are widely used in environmental monitoring, urban planning, land use analysis, and disaster response. Single-task learning models are typically trained for a specific task (such as segmentation or detection). While they may achieve good performance on a single task, they are prone to wasting computational resources and have insufficient model generalization ability. In contrast, multi-task learning methods, by combining multiple related tasks within a unified model, can achieve feature sharing and mutual reinforcement, thereby improving the overall performance and computational efficiency of the model.

[0003] Current multi-task learning methods are generally categorized into encoder-based and decoder-based approaches based on the network location of shared features. Encoder-based methods capture common semantic features of multiple tasks in the early stages of the model through shared feature extraction networks; while decoder-based methods focus on achieving high-level feature interactions between tasks during the feature decoding stage, using attention mechanisms or feature fusion modules to achieve task feature complementarity and improve the model's multi-task joint modeling capability. With the deepening of multi-task learning research, a new paradigm, the base model, has emerged. These models are equipped with large task backbone networks, handling multiple remote sensing interpretation tasks simultaneously through a unified modeling framework. However, these models often rely on the encoder portion for feature sharing, neglecting explicit task interaction modeling in the task decoding stage, thus limiting the expressive power of semantic associations between tasks. In recent years, generative models have received widespread attention in remote sensing image interpretation. By probabilistically modeling the distribution of underlying data, generative models can demonstrate powerful generative capabilities in tasks such as reconstruction, enhancement, and completion. Among them, diffusion models, as an emerging generative framework, have achieved significant results in the field of image generation and understanding.

[0004] However, existing remote sensing methods based on diffusion models are mostly limited to single-task scenarios and have not fully explored their potential in multi-task learning. Current approaches typically train a separate diffusion model for each task and execute multiple diffusion processes independently during the inference phase. This approach has several drawbacks. First, because the diffusion processes for each task are independent, the model struggles to capture semantic relevance and complementary features between different tasks. Second, multiple independent diffusion processes significantly increase the computational cost of inference. Furthermore, the diffusion modeling of heterogeneous task labels (such as segmentation masks and depth maps) lacks a unified framework, making it difficult to achieve an integrated design for multi-task generation.

[0005] In summary, although existing remote sensing image interpretation methods based on deep learning and diffusion models have made some progress in single-task scenarios, they still have key problems in multi-task learning scenarios, such as insufficient task interaction modeling, high computational cost, and difficulty in uniformly handling heterogeneous tasks. Summary of the Invention

[0006] This invention addresses the problems existing in the prior art by providing a diffusion-driven remote sensing image segmentation, height estimation, and edge detection joint interpretation method. This method enables multi-task feature sharing and task interaction modeling within a unified diffusion model, thus balancing computational efficiency and generation consistency.

[0007] To address the above technical problems, this invention provides the following technical solution: a diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method, comprising the following steps:

[0008] S1. Use a shared visual encoder to extract multi-scale shared features from the preprocessed remote sensing image, and then use a lightweight task head to obtain the initial multi-task prediction sequence.

[0009] S2. Construct a unified label encoding module to map heterogeneous truth labels from several tasks to a unified latent representation. In performing lightweight mapping and Same-dimensional context prior representation ; S3, In Unified Latent Representation Add noise and combine it with contextual prior representations Generate noisy features ;

[0010] S4. Construct a temporally conditional task interactive attention module, inputting multi-scale shared features and initial multi-task prediction sequences, combined with noisy features. The system outputs task-specific features; a denoising decoder module is constructed, which takes task-specific features as input, performs upsampling operations and several temporal conditional residual convolutions, and performs task-specific decoding to output remote sensing image segmentation, height estimation, and edge detection results.

[0011] S5. Based on the unified label encoding module constructed in step S2, the time-conditional task interaction attention module and the denoising decoder module in step S4, construct a remote sensing image interpretation neural network driven by diffusion generation.

[0012] The neural network is trained using multi-scale shared features and an initial multi-task prediction sequence as input, and remote sensing image segmentation, height estimation, and edge detection results as output, to obtain a diffusion-driven remote sensing image interpretation model. During training, a joint loss function and an optimizer are used for supervised learning. After training, the model is used for inference on a test dataset. The model improves upon noise features... Denoising is performed to generate task output.

[0013] Furthermore, in step S4 above, the time-conditional task interaction attention module includes:

[0014] TC Fusion Unit: Input multi-scale shared features and initial multi-task prediction sequence Then, perform time-weighted fusion and output the fused features. ;

[0015] Attention Unit: Input Fusion Features and noise characteristics A linear mapping is performed on the query Q, key K, and value V matrices, and time-based embedding is added. Relevance weights between tasks are calculated using scaled dot product attention, enabling efficient interaction of multi-task features and outputting interactive enhancement features. ;

[0016] TC Routing Unit: Input interaction enhancement features, which are connected by a set of shared expert networks, task embeddings, and time steps, and output task-specific features.

[0017] Furthermore, the aforementioned step S1 includes the following sub-steps:

[0018] S1.1. Extract multi-scale shared features from the preprocessed remote sensing image using a shared visual encoder, as shown in the following formula: ;

[0019] S1.2, Based on multi-scale shared features Initial multi-task predictions are obtained using lightweight task headers: ;

[0020] S1.3, Concatenate or fuse the initial multi-task predictions into a sequence: .

[0021] Furthermore, the aforementioned step S2 includes the following sub-steps:

[0022] S2.1, For each task truth labels Embedded using a task-specific embedder:

[0023] ,

[0024] in The embedding function for task k outputs... t is the time step;

[0025] S2.2 Embed and concatenate all tasks and map them to a unified latent representation using a small MLP, as follows:

[0026] ,

[0027] in, Cat indicates concatenating k elements. feature;

[0028] S2.3, Constructing the Task Context Prior , is represented as:

[0029] ,

[0030] in, Indicates to Pooling / aggregation operations, A lightweight mapping MLP used to generate and Context of the same dimension.

[0031] Furthermore, in step S3 mentioned above, the noise time is selected. Then, predefined Gaussian noise Add to the unified latent representation Above, generate the corresponding latent noise representation. :

[0032]

[0033] in, As a time-gating factor, it is used for mixing when injecting noise. and , As input for subsequent modules, This represents the cumulative noise attenuation coefficient.

[0034] Furthermore, the aforementioned TC fusion unit is configured to perform the following steps:

[0035] S4A-1, Introduction of Time Step The signal-to-noise ratio (SNR) is used to guide the fusion process, and the SNR is defined as:

[0036] ,

[0037] S4A-2, weighting coefficients obtained through Sigmoid transformation. :

[0038] ,

[0039] S4A-3, in the Step, sharing features across multiple scales With the initial multi-task prediction sequence Perform time-weighted fusion and fusion features Calculate as follows:

[0040] .

[0041] Furthermore, the aforementioned attention unit is configured to perform the following action: feature fusion. With noise characteristics In the common input attention structure, a linear mapping is performed on the query (Q), key (K), and value (V) matrices, and temporal parameters are incorporated, such as embedding. get:

[0042] ,

[0043] in, , , Features after linear mapping of Q, K, and V , , These represent the features after temporal embedding of Q, K, and V, respectively.

[0044] Furthermore, the aforementioned TC routing unit is configured to perform the following actions: [regarding time steps] Task A small route enhances the interactive features. With task embedding and time step embedding Connect these together and generate a weighted combination of expert data, using task-specific features as an example, as follows:

[0045]

[0046] in, It is the expert provided by the routing Normalized weights, for A lightweight team of experts above.

[0047] Furthermore, the aforementioned construction of the joint loss function is as follows:

[0048] ,

[0049] in, These represent the weight coefficients for the three types of tasks, For semantic segmentation loss, To overestimate the loss, For boundary detection loss.

[0050] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0051] This invention proposes a multi-task interpretation framework based on a diffusion-generative model. By introducing a unified label encoding module and combining it with a task interaction decoding module, it avoids the limitation of traditional methods that only share features at a shallow level, thereby improving the model's representational power and task collaboration effect. Simultaneously, by introducing a time-conditional task interaction attention mechanism, the intensity of information fusion between tasks can be dynamically adjusted, effectively improving the model's accuracy in complex scenarios and its robustness in multi-task modeling. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the model structure of the present invention.

[0053] Figure 2 The diagram shows the structure of the Time Condition Task Interaction Attention Module. In the diagram, (a) is the overall structure of the Time Condition Task Interaction Attention Module, (b) is the structure of the TC Fusion Module, and (c) is the structure of the TC Routing Module.

[0054] Figure 3 This is a schematic diagram of the temporal conditional residual convolution module.

[0055] Figure 4 The diagram shows the training and inference process of the model. In the diagram, (a) is a schematic diagram of the training phase and (b) is a schematic diagram of the inference phase.

[0056] Figure 5 This is a multi-task interpretation result diagram of remote sensing image driven by diffusion generation according to the present invention. In the diagram, (a) is the input remote sensing image, (b) is the segmentation result interpretation diagram, (c) is the height estimation result interpretation diagram, and (d) is the edge detection result interpretation diagram.

[0057] Figure 6 Yes, the flowchart of this invention. Detailed Implementation

[0058] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0059] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0060] like Figure 1 As shown, this invention provides a diffusion-driven remote sensing image segmentation, height estimation, and edge detection joint interpretation method, which fully utilizes the correlation features between multiple tasks to improve feature sharing and semantic collaboration capabilities among tasks. Secondly, by designing a unified label encoding module, unified encoding of features from different heterogeneous tasks is achieved. Furthermore, by constructing a temporal conditional task interaction attention module, an attention interaction mechanism between tasks is introduced during the diffusion denoising process, enhancing information fusion and representation learning capabilities among multiple tasks. Through these improvements, this invention can efficiently and accurately fulfill the multi-task requirements of remote sensing images, meeting the practical needs of large-scale remote sensing data processing. The specific steps are as follows:

[0061] S1. Use a shared visual encoder to extract multi-scale shared features from the preprocessed remote sensing image, and then use a lightweight task head to obtain the initial multi-task prediction sequence.

[0062] S2. Construct a unified label encoding module to map heterogeneous truth labels from several tasks to a unified latent representation. In performing lightweight mapping and Same-dimensional context prior representation ; S3, In Unified Latent Representation Add noise and combine it with contextual prior representations Generate noisy features ;

[0063] S4. Construct a temporally conditional task interactive attention module, inputting multi-scale shared features and initial multi-task prediction sequences, combined with noisy features. The system outputs task-specific features; a denoising decoder module is constructed, which takes task-specific features as input, performs upsampling operations and several temporal conditional residual convolutions, and performs task-specific decoding to output remote sensing image segmentation, height estimation, and edge detection results.

[0064] S5. Based on the unified label encoding module constructed in step S2, the time-conditional task interaction attention module and the denoising decoder module in step S4, construct a remote sensing image interpretation neural network driven by diffusion generation.

[0065] The neural network is trained using multi-scale shared features and initial multi-task predictions as inputs, and remote sensing image segmentation, height estimation, and edge detection results as outputs, to obtain a diffusion-driven remote sensing image interpretation model. During training, a joint loss function and optimizer are used for supervised learning. After training, the model is used for inference on a test dataset. The model improves upon noise features... Denoising is performed to generate task output.

[0066] Before performing step S1, the input remote sensing image dataset is preprocessed. These images contain four channels: RGB and near-infrared, along with corresponding semantic segmentation and elevation maps. The data format must be a standard image format (such as TIFF or JPEG), with each image having four channels (RGB and NIR) and high resolution. Then, the input images undergo pixel value normalization, scaling the value of each pixel to between 0 and 1, using the following formula: ,in, These are the original pixel values. and These represent the minimum and maximum pixel values ​​in the image, respectively. Finally, the images are resized to a uniform size (e.g., 512×512 pixels) to ensure consistent input image dimensions and facilitate model training.

[0067] In a preferred embodiment of the present invention, step S1 includes the following sub-steps:

[0068] S1.1. Extract multi-scale shared features from the preprocessed remote sensing image using a shared visual encoder (such as ResNet), as shown in the following formula: ;

[0069] S1.2 To enrich the conditional information of the diffusion process, based on multi-scale shared features... Initial multi-task predictions are obtained using a lightweight task header: ;

[0070] S1.3, Concatenate or fuse the initial multi-task predictions into a sequence: The initial prediction sequence here serves both as early task information and as a conditional input for subsequent temporal conditional fusion.

[0071] In a preferred embodiment of the present invention, step S2 includes the following sub-steps:

[0072] S2.1, For each task truth labels Embedded using a task-specific embedder:

[0073] ,

[0074] in The embedding function for task k outputs... t is the time step;

[0075] S2.2 Embed and concatenate all tasks and map them to a unified latent representation using a small MLP, as follows:

[0076] ,

[0077] in, Cat indicates concatenating k elements. feature;

[0078] S2.3, Constructing the Task Context Prior , is represented as:

[0079] ,

[0080] in, Indicates to Pooling / aggregation operations, A lightweight mapping MLP used to generate and Context of the same dimension.

[0081] In a preferred embodiment of the present invention, step S3 specifically involves performing forward diffusion and forming training samples by adding noise: selecting the noise time. Then, predefined Gaussian noise Add to the unified latent representation Above, generate the corresponding latent noise representation. :

[0082]

[0083] in, As a time-gating factor, it is used for mixing when injecting noise. and , As input for subsequent modules, This represents the cumulative noise attenuation coefficient. Unlike previous methods that required a separate diffusion process for each task, the proposed label unified coding module allows the module to learn a shared latent representation across all tasks. This unified representation captures task-specific features in a consistent latent space, facilitates joint learning, and further enhances the model's generalization ability.

[0084] In a preferred embodiment of the present invention, a remote sensing image interpretation neural network based on diffusion generation is constructed for backdiffusion, such as... Figure 2As shown in Figure (a), the temporal conditional task interaction attention module in the network undertakes the core functions of progressive denoising, information recovery, and feature interaction between tasks.

[0085] TC fusion unit: such as Figure 2 As shown in (b), the input multi-scale shared features and initial multi-task prediction sequence Then, perform time-weighted fusion and output the fused features. The TC fusion unit shown is configured to perform the following steps:

[0086] S4A-1, Introduction of Time Step The signal-to-noise ratio (SNR) is used to guide the fusion process, and the SNR is defined as:

[0087] ,

[0088] S4A-2, weighting coefficients obtained through Sigmoid transformation. :

[0089] ,

[0090] S4A-3, in the Step, sharing features across multiple scales With the initial multi-task prediction sequence Perform time-weighted fusion and fusion features Calculate as follows:

[0091] .

[0092] In the early stages of diffusion, the model relies more on task feature prediction, while in the later stages it gradually strengthens the use of global semantic features, thereby achieving a balance of time-dependent feature weights.

[0093] Attention unit: such as Figure 2 As shown in (a), the input fusion features and noise characteristics A linear mapping is performed on the query Q, key K, and value V matrices, and time-based embedding is added. Relevance weights between tasks are calculated using scaled dot product attention, enabling efficient interaction of multi-task features and outputting interactive enhancement features. The attention unit is configured to perform the following action: fuse features. With noise characteristics In the common input attention structure, a linear mapping is performed on the query (Q), key (K), and value (V) matrices, and temporal parameters are incorporated, such as embedding. get:

[0094] ,

[0095] in, , , Features after linear mapping of Q, K, and V , , These represent the features after temporal embedding of Q, K, and V, respectively.

[0096] By calculating the relevance weights between tasks using scaled dot product attention, efficient interaction of multi-task features is achieved. The output features are then processed through residual connections to obtain interactive enhanced features.

[0097]

[0098] Generated interactive enhancement features Interpretation for a specific task.

[0099] TC routing unit: such as Figure 2 As shown in (c), the input interaction enhancement features are utilized by the TC routing unit, which leverages a shared set of expert networks to simultaneously extract features for a specific task. Specifically, it allows... become A lightweight team of experts above. Regarding time step... Task A small route will feature With task embedding and time step embedding Connect these together and generate a weighted combination of expert insights, incorporating task-specific features:

[0100]

[0101] in, It is the expert provided by the routing Normalized weights. The process is as follows: Figure 2 As shown in (c), then task-specific features They are sent to their respective noise reduction decoders for further refinement.

[0102] To further recover clean, task-specific outputs from latent noise features, a denoising decoder module is applied to the neural network. The denoising decoder module consists of a temporal conditional residual convolution module, an upsampling operation, and a prediction head. The temporal conditional residual convolution module incorporates time step information to guide the denoising process. Let... It is through learnable MLPs from the embedded time step The derived scale and shift vectors. The output for each block is defined as:

[0103]

[0104] in This represents a standard convolutional transformation (e.g., a 3x3 convolution followed by normalization and ReLU activation). The temporal conditional residual convolution module structure is as follows: Figure 3 As shown, the upsampling operation uses bilinear interpolation. After passing through several temporal conditional residual convolutional modules, the final feature map is obtained and input into the prediction head for a specific task to generate the final result.

[0105] In step S5, to ensure balanced performance across different tasks in multi-task learning during network training, task-specific supervised losses, including semantic segmentation loss, are designed. Highly estimating the loss and boundary detection loss The semantic segmentation task uses pixel-level cross-entropy loss, the height estimation task uses mean squared error (MSE) loss, and the boundary detection task uses binary cross-entropy (BCE) loss. These three losses jointly constrain the output predictions of each task, enabling the model to simultaneously learn classification accuracy, depth continuity, and boundary consistency. The combined joint loss is defined as:

[0106]

[0107] in, These represent the weight coefficients for the three task categories. The model optimization employs an adaptive weight update mechanism, automatically adjusting the relative weights of each loss term based on changes in the task gradient, thereby preventing any single task from excessively dominating the training process. Ultimately, through end-to-end joint training, the model achieves unified optimization of information flow between diffusion generation and task interpretation, significantly improving the semantic consistency, boundary accuracy, and structural integrity of the generated multi-task results.

[0108] The model was optimized using the AdamW optimizer with a learning rate, resulting in weight decay of 1e-4 and 1e-5. The batch size was set to 8, and the model was trained for 50 iterations. After training, inference was performed on the test dataset. The semantic segmentation results were output as images, with each pixel corresponding to a category. The height estimation results were output as grayscale images, where grayscale values ​​represent height. The boundary detection results were output as binary images, with white pixels representing boundary locations.

[0109] Step 8: Model training and inference process as follows Figure 4 As shown, the encoding processes differ between the training and inference phases. For example... Figure 4 As shown in (a), during the training phase, Gaussian noise is added to the unified latent representation obtained from the ground truth labels. To generate noisy latent representations In contrast, such as Figure 4As shown in (b), during the inference phase, truth tags are replaced with noise, and the model works by processing noisy latent representations. Denoising is performed to generate the task output. The joint interpretation results of remote sensing image segmentation, height estimation, and edge detection based on the diffusion generation model are as follows: Figure 5 As shown in the figure, (a) shows the input remote sensing image, (b) shows the segmentation result, (c) shows the height estimation result interpretation, and (d) shows the edge detection result interpretation.

[0110] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method, characterized in that, The steps are as follows: S1. Use a shared visual encoder to extract multi-scale shared features from the preprocessed remote sensing image, and then use a lightweight task head to obtain the initial multi-task prediction sequence. S2. Construct a unified label encoding module to map heterogeneous truth labels from several tasks to a unified latent representation. In performing lightweight mapping and Same-dimensional context prior representation ; S3, In Unified Latent Representation Add noise and combine it with contextual prior representations Generate noisy features ; S4. Construct a temporally conditional task interactive attention module, inputting multi-scale shared features and initial multi-task prediction sequences, combined with noisy features. Output task-specific features; A denoising decoder module is constructed. The input features are task-specific, and after upsampling, several temporal conditional residual convolutions, and task-specific decoding, the output results are remote sensing image segmentation, height estimation, and edge detection. S5. Based on the unified label encoding module constructed in step S2, the time-conditional task interaction attention module and the denoising decoder module in step S4, construct a remote sensing image interpretation neural network driven by diffusion generation. The neural network is trained using multi-scale shared features and an initial multi-task prediction sequence as input, and remote sensing image segmentation, height estimation, and edge detection results as output, to obtain a diffusion-driven remote sensing image interpretation model. During training, a joint loss function and an optimizer are used for supervised learning. After training, the model is used for inference on a test dataset. The model improves upon noise features... Denoising is performed to generate task output.

2. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 1, characterized in that, In step S4, the time-conditional task interaction attention module includes: TC Fusion Unit: Input multi-scale shared features and initial multi-task prediction sequence Then, perform time-weighted fusion and output the fused features. ; Attention Unit: Input Fusion Features and noise characteristics A linear mapping is performed on the query Q, key K, and value V matrices, and time-based embedding is added. Relevance weights between tasks are calculated using scaled dot product attention, enabling efficient interaction of multi-task features and outputting interactive enhancement features. ; TC Routing Unit: Input interaction enhancement features, which are connected by a set of shared expert networks, task embeddings, and time steps, and output task-specific features.

3. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Extract multi-scale shared features from the preprocessed remote sensing image using a shared visual encoder, as shown in the following formula: ; S1.2, Based on multi-scale shared features Initial multi-task predictions are obtained using lightweight task headers: ; S1.3, Concatenate or fuse the initial multi-task predictions into a sequence: .

4. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.1, For each task truth labels Embedded using a task-specific embedder: ， in The embedding function for task k outputs... t is the time step; S2.2 Embed and concatenate all tasks and map them to a unified latent representation using a small MLP, as follows: ， in, Cat indicates concatenating k elements. feature; S2.3, Constructing the Task Context Prior , is represented as: ， in, Indicates to Pooling / aggregation operations, A lightweight mapping MLP used to generate and Context of the same dimension.

5. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 1, characterized in that, In step S3, select the noise time. Then, predefined Gaussian noise Add to the unified latent representation Generate noisy features. : ， in, As a time-gating factor, it is used for mixing when injecting noise. and , As input for subsequent modules, This represents the cumulative noise attenuation coefficient.

6. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 2, characterized in that, The TC fusion unit is configured to perform the following steps: S4A-1, Introduction of Time Step The signal-to-noise ratio (SNR) is used to guide the fusion process, and the SNR is defined as: ， S4A-2, weighting coefficients obtained through Sigmoid transformation. : ， S4A-3, in the Step, sharing features across multiple scales With the initial multi-task prediction sequence Perform time-weighted fusion and fusion features Calculate as follows: 。 7. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 6, characterized in that, The attention unit is configured to perform the following actions: S4B-1, Receiver Fusion Features and noise characteristics Perform a linear mapping on the query Q, key K, and value V matrix and incorporate time as an embedding. get: ， in, , , Features after linear mapping of Q, K, and V , , These represent the features after temporal embedding of Q, K, and V, respectively. S4B-2, Output Interaction Enhancement Feature As shown in the following formula: .

8. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 7, characterized in that, The TC routing unit is configured to perform the following actions: [Regarding time steps] Task A small route enhances the interactive features. Embedded with tasks and time step embedding Connect these together and generate a weighted combination of expert data, using task-specific features as an example, as follows: ， in, It is the expert provided by the routing Normalized weights, for A lightweight team of experts above.

9. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 1, characterized in that, The output of the noise reduction decoder module is: , in Indicates the standard convolution variation, Embedded for time step.

10. The diffusion-generated remote sensing image segmentation, height estimation, and edge detection joint interpretation method according to claim 1, characterized in that, The joint loss function is constructed as follows: ， in, These represent the weight coefficients for the three types of tasks, For semantic segmentation loss, To overestimate the loss, For boundary detection loss.

Images