Remote sensing image description method based on priori and semantic alignment

Abstract

The application discloses a remote sensing image description method based on priori and semantic alignment, and aims at solving the problems of target illusion, inaccurate quantity statistics and spatial relation expression deviation of an airport remote sensing image description method. An image description model is constructed, the image description model comprises a visual encoder, a double-channel gate non-linear mapper, a panoramic segmentation network, a fusion layer and a text decoder, and an airport remote sensing image to be described is input into the image description model, and the image description model outputs description text of the airport remote sensing image to be described. The application belongs to the technical field of remote sensing image processing.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing, and more specifically to a remote sensing image description method based on prior and semantic alignment. Background Technology

[0002] With the rapid development of remote sensing Earth observation technology, high-resolution remote sensing images are playing an increasingly important role in areas such as regional planning and target monitoring. Airports, as typical complex remote sensing scenarios, contain various key targets including terminals, runways, taxiways, aprons, and aircraft. These targets are not only diverse in type and scale, but also exhibit complex spatial relationships and frequent changes in operational status. Therefore, how to automatically generate accurate, complete, and semantically expressive text descriptions from remote sensing images has become an important research direction for intelligent interpretation of remote sensing images.

[0003] Traditional remote sensing image description methods are mostly built on an encoder-decoder framework, typically using convolutional neural networks or visual Transformers to extract image features, and then combining them with recurrent neural networks or Transformer decoders to generate descriptive text. While these methods achieve a certain degree of mapping from visual information to natural language expression, they largely rely on global visual features to drive text generation, lacking the ability to explicitly model key targets, quantitative relationships, and spatial layouts within the scene. Consequently, in complex scenes, they are prone to problems such as overly coarse descriptions, incomplete semantic information, and inaccurate representation of spatial relationships.

[0004] In recent years, the development of multimodal large language models has provided a new technical path for image description. By mapping visual features to the semantic space of a language model, multimodal large language models can generate more natural, fluent, and richer descriptive text by leveraging the knowledge and reasoning capabilities gained through large-scale pre-training. However, existing multimodal large language models are mainly designed for natural scenes and still have significant limitations in remote sensing image description tasks. Especially in high-resolution airport scenes, due to the complexity of background elements, the dense distribution of targets, and large scale differences, multimodal large language models are easily affected by complex backgrounds such as terminals, boarding bridges, taxiways, and shadows of ground features, leading to distortion in the identification of key targets. This, in turn, results in problems such as target illusion, quantitative deviations, and inaccurate descriptions of spatial relationships during the description process.

[0005] Furthermore, most existing remote sensing image description studies rely on publicly available datasets for training. However, the descriptive texts in these datasets are typically brief, focusing on macroscopic scene overviews and lacking systematic characterization of key aspects such as target attributes, quantity, spatial relationships, and operational status. This makes it difficult to meet the practical needs of fine-grained, structured, and high-fidelity semantic interpretation in airport scenarios. Simultaneously, existing methods rely heavily on implicit alignment between visual features and language models during text generation, underutilizing structured scene information such as category, quantity, and location. Consequently, they struggle to establish stable constraints on core elements within airport scenes.

[0006] In summary, existing technologies for airport remote sensing image description still suffer from problems such as inconsistencies in the quantitative description of key targets, deviations in the expression of spatial relationships, and insufficient semantic generalization capabilities for complex scenes. Therefore, it is necessary to propose an airport remote sensing image description method that can fully utilize prior information about scene structure, enhance visual semantic alignment capabilities, and improve the realism and accuracy of descriptions. Summary of the Invention

[0007] To address the problems of target illusion, inaccurate quantity statistics, and spatial relationship expression in airport remote sensing image description methods, this invention proposes a remote sensing image description method based on prior and semantic alignment.

[0008] The technical solution adopted in this invention is:

[0009] Construct an image description model and acquire remote sensing images of the airport to be described. Input the remote sensing images of the airport to be described into the image description model, and the image description model outputs the description text of the remote sensing images of the airport to be described.

[0010] The image description model includes a visual encoder, a dual-channel gated nonlinear mapper, a panoramic segmentation network, a fusion layer, and a text decoder.

[0011] The beneficial effects of this invention are as follows:

[0012] This invention constructs an image description model comprising a visual encoder, a dual-channel gated nonlinear mapper, a panoramic segmentation network, a fusion layer, and a text decoder. The dual-channel gated nonlinear mapper enhances the joint modeling capability of local regions through a spatial merging strategy and utilizes parallel-generated gated branches and feature branches for dynamic content adjustment. This process uses adaptive weighting of the feature branches by the gated branches to specifically amplify the semantic responses of key targets such as terminals, runways, taxiways, and aircraft at the continuous feature level, while suppressing complex background interference and improving the effectiveness and stability of spatial projection of visual features into the language model. The panoramic segmentation network utilizes the fine-grained structure provided by the underlying panoramic segmentation mask to deconstruct complex visual scenes into discrete physical attributes through perceptual entity filtering, scene anchor point localization, and aircraft nearest neighbor clustering resolution. This hierarchical resolution approach aligns with the spatial pattern of densely packed local targets and strong overall structural correlation in airport scenes. By injecting structured cue text into the text decoder input sequence, it provides clear prior constraints such as infrastructure categories, quantities, and spatial locations for the text generation process, improving the problems of disordered entity relationships and missing key features in the generated text description. The fusion layer generates unified multimodal features that fuse visual mapping features and hierarchical structured cue text, and uses a text decoder to generate descriptive text for airport remote sensing images. This invention achieves complementary advantages through the continuous high-level visual semantics provided by a dual-channel gated nonlinear mapper and the discrete structured attribute priors extracted from panoramic segmentation results. Driven by the autoregressive language generation loss of the image description model, the model can simultaneously rely on visual features to perceive the global scene and use structured cue text to finely depict the distribution of local multi-target objects and the operational status of infrastructure. This enhances the semantic understanding capability of the image description model in complex airport remote sensing scenarios and improves the accuracy, realism, and hierarchical logic of airport remote sensing image description text. Attached Figure Description

[0013] Figure 1 This is a flowchart of an image description model;

[0014] Figure 2 This is a schematic diagram of step two;

[0015] Figure 3 This is a structural diagram of a dual-channel gated nonlinear mapper; Detailed Implementation

[0016] Specific implementation method one: Combining Figures 1-3 This embodiment describes a remote sensing image description method based on prior and semantic alignment, which includes the following steps:

[0017] Step 1: Acquire airport remote sensing images and preprocess them to obtain preprocessed airport remote sensing images. The specific process is as follows:

[0018] Step 1: Use the Gaofen-1 remote sensing satellite to collect remote sensing images of multiple airports. The spatial resolution of each airport remote sensing image is 4000×4000 pixels.

[0019] Steps 1 and 2: Preprocess each airport remote sensing image to obtain a set of airport remote sensing image patches. The specific process is as follows:

[0020] Each airport remote sensing image is formatted and resized to meet the input requirements of the image description model, resulting in an adjusted airport remote sensing image. Each adjusted airport remote sensing image is then segmented using an overlapping sliding cropping method, with a sliding step size of 600 pixels and a cropping window size of 800×800 pixels, resulting in a set of airport remote sensing image blocks.

[0021] Step 2, as follows Figure 2 As shown, a prompt word template is designed based on the preprocessed airport remote sensing image. The prompt word template and the preprocessed airport remote sensing image are input into a large language model. The large language model outputs a long text description of the preprocessed airport remote sensing image. The long text description is manually checked and semantically refined to obtain a short text description. The specific process is as follows:

[0022] A prompt word template is designed based on the overall logic between the infrastructure context, aircraft position, and airport operational status within each airport remote sensing image patch in the airport remote sensing image patch set. The infrastructure context refers to whether the airport remote sensing image patch represents an airport scene. The aircraft position refers to the aircraft's location on the airport remote sensing image patch, including upper left, lower left, upper right, and lower right. The airport operational status refers to whether the airport is in a congested or normal operating state. The specific content of the prompt word template is: determining whether the input airport remote sensing image patch represents an airport scene, describing the specific location of the aircraft within the airport remote sensing image patch, and comprehensively judging whether the current airport operational status is congested or normal.

[0023] The prompt word template and a set of airport remote sensing image patches are input into a large language model, which outputs a long text description for each airport remote sensing image patch. The large language model is either the Qwen3 model or the Qwen2.5 model. All long text descriptions of airport remote sensing image patches are manually verified, focusing on the accuracy of infrastructure identification such as terminals, runways, and taxiways, the rationality of aircraft numbers and group distribution, and the consistency of airport operational status descriptions. Long text descriptions with target illusions, missing key features, or semantic confusion are corrected or removed, ultimately forming a set of semantically complete and hierarchically clear long text descriptions. Further, while retaining key information such as the main airport scene, aircraft distribution, and airport operational status, each long text description in the long text description set is semantically refined to obtain a set of short text descriptions for visual-language feature alignment. Finally, the long text description set and the short text description set are mixed and used as a training set for training the image description model.

[0024] Step 3: Construct an image description model. The image description model includes a Rice-ViT visual encoder, a dual-channel gated nonlinear mapper, a panoramic segmentation network, a fusion layer, and a Qwen3-0.6B text decoder. The dual-channel gated nonlinear mapper includes, in sequence, a layer normalization processing layer, a spatial merging layer, a first linear transformation layer, a dimension cutting layer, an activation fusion layer, and a second linear transformation layer.

[0025] Step 4: Train the image description model using the preprocessed airport remote sensing image and the training set. Input the preprocessed airport remote sensing image, and the image description model outputs a descriptive text for the preprocessed airport remote sensing image, resulting in the trained image description model. For example... Figure 1 As shown. The specific process is as follows:

[0026] Step 41: Input the set of airport remote sensing image patches obtained in Steps 1 and 2 into the Rice-ViT visual encoder. The Rice-ViT visual encoder outputs the visual embedding features of each airport remote sensing image patch. ,in, For the real number field, For batches, The length of the visual token. The dimension of a single visual embedding feature.

[0027] Step 42, as follows Figure 3 As shown. The visual embedding features of each airport remote sensing image patch are input into a dual-channel gated nonlinear mapper, and the dual-channel gated nonlinear mapper outputs the mapping features of each airport remote sensing image patch. The specific process is as follows:

[0028] (1) Visual embedding features of each airport remote sensing image patch Within the input layer normalization processing layer, the output layer normalizes the visual embedding features of each airport remote sensing image patch.

[0029] (2) The processed visual embedding features of each airport remote sensing image patch are input into the spatial merging layer. The spatial merging strategy is used to reorganize the visual embedding features of adjacent spatial tokens in the channel dimension. The spatial merging layer outputs the intermediate features of each airport remote sensing image patch. Specifically:

[0030] Let the spatial merging scale be Based on the processed visual embedding features of each airport remote sensing image patch, each The visual embedding features of adjacent spatial tokens are aggregated into a new feature, and the feature dimension of each new feature is... The number of tokens after each merge is This allows us to obtain the intermediate features of each airport remote sensing image patch. .

[0031] (1)

[0032] (2)

[0033] in, This indicates that visual embedded features are rearranged and flattened according to spatial merging rules.

[0034] (3) Extract the intermediate features from each airport remote sensing image patch The intermediate features are input into the first linear transformation layer and output by the first linear transformation layer. corresponding 1-dimensional feature .

[0035] (4) The feature one Within the input dimension, slice the layers and then divide the features along the channel dimension. According to the above The last dimension of the dimension is equally divided, and the gated branch output features of each airport remote sensing image patch are output from the dimension-splitting layer. and gated branch output characteristics ,in, , The specific process is as follows:

[0036] (3)

[0037] (4)

[0038] in, The projection matrix within the first linear transformation layer. For the bias term of the first linear transformation layer, Indicates the relationship between feature one Perform segmentation along the channel dimension.

[0039] (5) Output features of the gated branch of each airport remote sensing image patch and gated branch output characteristics Input activation fusion layer, using The nonlinear activation function outputs features on the gated branch. Activate the gated branch and output its features. With gated branch output characteristics Element-wise multiplication is performed to activate the gated fusion features of each airport remote sensing image patch output by the fusion layer. This further enhances the selective modeling capability of visual representations. The process essentially utilizes gated branches. For feature branches Adaptive weighting is applied so that visual responses in different dimensions are dynamically amplified or attenuated based on task relevance.

[0040] (5)

[0041] in, This indicates element-wise multiplication.

[0042] (6) Gated fusion features of each airport remote sensing image patch The input is processed by the second linear transform layer, which outputs the mapping features of each airport remote sensing image patch. The expression is:

[0043] (6)

[0044] In the formula, This is the projection matrix of the second linear transformation layer. This is the bias term for the second linear transformation layer.

[0045] A dual-channel gated nonlinear mapper adaptively reshapes and maps visual embedding features to the language space. Through layer normalization, space merging strategy, nonlinear mapping and gated modulation, it suppresses redundant responses to complex backgrounds, enhances target-related semantic features, and maps visual embedding features to the language model embedding space, thereby improving the effectiveness and stability of visual features when projecting to the language space.

[0046] Step 43: Input the set of airport remote sensing image patches into the panoramic segmentation network to obtain the panoramic segmentation result for each airport remote sensing image patch. Extract the infrastructure information from the panoramic segmentation results and generate hierarchical prompt text for each airport remote sensing image patch based on the infrastructure information. The specific process is as follows:

[0047] (i) Input the set of airport remote sensing image patches obtained in steps one and two into the panoramic segmentation network (MFGS2Net) to obtain the panoramic segmentation result of each airport remote sensing image patch. ,in, The total number of segments. For the first Each segmented unit , For the first The infrastructure category labels within each segmented unit include terminals, runways, taxiways, and aircraft. For the first The bounding box of infrastructure in each segmented unit. , Let be the coordinates of any vertex of the bounding box. The width of the bounding box. The height of the bounding box. For the first The area of ​​each segmentation unit. Let the width and height of the panoramic segmentation result for each airport remote sensing image patch be respectively... and .

[0048] (ii) Set the minimum effective area threshold for the aircraft as follows: The minimum effective building area threshold is Based on the minimum effective area thresholds for aircraft and buildings, the panoramic segmentation results of each airport remote sensing image patch are constrained. Effective aircraft, effective buildings, and infrastructure are then selected to obtain the set of effective aircraft in the panoramic segmentation results of each airport remote sensing image patch. Effective building collection and infrastructure collection According to the infrastructure set Obtain the collection of infrastructure categories It covers various land features such as runways and taxiways.

[0049] (7)

[0050] (8)

[0051] (9)

[0052] (iii) The effective building set from each airport remote sensing image patch The building with the largest area was selected as the scene anchor point. Obtain the scene anchor point Coverage in the corresponding panoramic segmentation results Based on coverage Anchor points for scenes Scale classification is performed, specifically: if the coverage rate Then set the scene anchor point Divided into small; if the coverage Then set the scene anchor point Divided into large; if the coverage Then set the scene anchor point Classified as a large type. Scene anchor points are obtained based on the aforementioned classification criteria. The scale classification results.

[0053] (10)

[0054] (11)

[0055] in, As scene anchor point The area.

[0056] Synchronously extract the scene anchor points center coordinates , the center coordinates The position in the corresponding panoramic segmentation result is converted into specific textual location terms, such as center coordinates. If the text is located in the upper left corner of the corresponding panoramic segmentation result, the converted text location term will be "upper left". Then, the text location term is compared with the aforementioned scene anchor points. The scale classification results are concatenated to generate global anchor point text descriptions for the panoramic segmentation results, such as a giant anchor point located in the upper left corner of the panoramic segmentation results. When the effective building set... When the largest effective building is missing, the default general scene anchor point is used to generate the global anchor point text description of the corresponding panoramic segmentation result.

[0057] (iv) Set up the valid aircraft collection for each airport remote sensing image patch. Treat it as an undirected adjacency graph vertex set And set the cluster distance threshold as Then extract the effective aircraft set from each airport remote sensing image patch. Any two aircraft and center point and Obtain the center point and Two-dimensional Euclidean distance :

[0058] (12)

[0059] (13)

[0060] (14)

[0061] in, For airplane The coordinates of any vertex of the bounding box. For airplane The width of the bounding box, For airplane The height of the bounding box, For airplane The coordinates of any vertex of the bounding box. For airplane The width of the bounding box, For airplane The height of the bounding box. If Then determine the aircraft and airplane There are adjacency relationships between them and the construction of aircraft. and airplane If there is no undirected edge, then the aircraft is determined. and airplane There are no adjacency relationships between them. Therefore, a valid set of aircraft can be obtained. The relationships between all aircraft are determined. Next, an undirected adjacency graph is obtained based on all undirected edges. edge set Thus, a complete undirected adjacency graph was obtained.

[0062] (15)

[0063] Then, the complete undirected adjacency graph is traversed, connected components are extracted, and multiple aircraft connected by undirected edges are grouped into a local aircraft group. The total number of aircraft contained in each local aircraft group and the specific location of each local aircraft group in each airport remote sensing image block are extracted to obtain the information of each local aircraft group in each airport remote sensing image block.

[0064] (v) The infrastructure set of each airport remote sensing image patch When runways and taxiways are detected within the panoramic segmentation mask, they are merged to obtain the runway-taxiway area, and a text description of the runway-taxiway area is generated. This text description is then used as the background context for subsequent text decoders. If runways and taxiways are not detected, the apron area is directly extracted and a text description of the apron is generated. This text description is then used as the background context for subsequent text decoders.

[0065] (vi) Define the global anchor point text description for each airport remote sensing image patch extracted above as follows: Information on each local group of aircraft is Background information context is The complete structured semantic representation of an airport remote sensing image patch is denoted as: The hierarchical structure prompt generation mechanism generates preset fixed text tags, , and Merge into prompt text See formula (14), then display the prompt text. Encoded as a prompt embedding.

[0066] (16)

[0067] In the formula, These are pre-defined fixed text tags (guide words) that add a beginning to the generated natural language paragraph, indicating to the large language model that the following is an overview of the scene. This indicates a text concatenation operation. This represents a combined function that organizes information from multiple local groups of aircraft into a natural language description in a preset order.

[0068] Step 44: Map the features of each airport remote sensing image patch obtained in Step 42. And the hierarchical prompt text obtained in step four or three The input is fused within the fusion layer, and the fusion layer outputs the multimodal features of each airport remote sensing image patch.

[0069] The default task instruction text (e.g., describing this remote sensing image of the airport) is: ,Will The task instruction text is input into the Qwen3-0.6B model, where its built-in tokenizer and embedding layer are used to perform word segmentation and spatial mapping, ultimately yielding the text embedding sequence of the task instruction. Simultaneously, the built-in word segmenter and embedding layer of the Qwen3-0.6B model are used to process the hierarchical prompt text obtained in step four-three. Word segmentation and spatial mapping are performed to obtain the structured prompt text embedding sequence. .

[0070] (17)

[0071] (18)

[0072] In the formula, This indicates word segmentation operation. This indicates a space mapping operation.

[0073] The mapping features obtained in step four-two The text embedding sequence of the task instructions With the prompt text embedded sequence The images are concatenated along the sequence dimension to obtain the multimodal features of each airport remote sensing image patch. .

[0074] (19)

[0075] in, This indicates a tensor splicing operation.

[0076] Steps four and five: Let the text description sequence of each airport remote sensing image patch in the training set of Step two be... ,in, This represents the total number of tokens in the standard description text. Indicates the first The real token label corresponding to each time step.

[0077] Multimodal features of each airport remote sensing image patch The text descriptions of the corresponding airport remote sensing image patches in the training set from step two are input into the Qwen3-0.6B text decoder. The Qwen3-0.6B text decoder outputs the natural language description text for each airport remote sensing image patch. The specific process is as follows:

[0078] The multimodal input sequence of each airport remote sensing image patch is used. And the text description of the corresponding airport remote sensing image patch in the training set in step two. The input is processed by the Transformer structure within the Qwen3-0.6B text decoder, which processes the multimodal input sequence. Mapping features in The text embedding sequence of task instructions and Embedded sequence with prompt text Interactive computation is performed using an autoregressive decoding method based on the input multimodal sequence. and before The generated text sequence Calculate the first The probability distribution of candidate words at each time step is analyzed, and the token with the highest probability is selected as the current time step. The generated text sequence .

[0079] (20)

[0080] Text sequence Compared to the past The generated historical text sequence is then concatenated before proceeding. Text sequence generation at each time step. This process is repeated iteratively until the text sequence generated at the current time step is reached. The decoding process ends with a preset sequence terminator, and the natural language description text for each airport remote sensing image patch is output.

[0081] Step 46: Repeat steps 41 to 45 to train the image description model. Based on the natural language description text of each airport remote sensing image patch and the text description of the corresponding airport remote sensing image patch in the training set in step 2, calculate the autoregressive language generation loss using the standard cross-entropy function. The image description model is optimized using this autoregressive language generation loss to obtain a well-trained image description model.

[0082] (twenty one)

[0083] in, For the text decoder in the first Generate a real token at each time step. The conditional probability.

[0084] Step 5: Collect remote sensing images of the airport to be described, input the remote sensing images of the airport to be described into the trained image description model, and the trained image description model outputs natural language description text of the remote sensing images of the airport to be described.

[0085] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A remote sensing image description method based on prior and semantic alignment, characterized in that: It includes the following steps: Construct an image description model and acquire remote sensing images of the airport to be described. Input the remote sensing images of the airport to be described into the image description model, and the image description model outputs the descriptive text of the remote sensing images of the airport to be described. The image description model includes a visual encoder, a dual-channel gated nonlinear mapper, a panoramic segmentation network, a fusion layer, and a text decoder.

2. The remote sensing image description method based on prior and semantic alignment according to claim 1, characterized in that: The visual encoder is a Rice-ViT visual encoder.

3. The remote sensing image description method based on prior and semantic alignment according to claim 2, characterized in that: The dual-channel gated nonlinear mapper sequentially includes a layer normalization processing layer, a spatial merging layer, a first linear transformation layer, a dimension cutting layer, an activation fusion layer, and a second linear transformation layer.

4. The remote sensing image description method based on prior and semantic alignment according to claim 3, characterized in that: The text decoder is the Qwen3-0.6B text decoder.

5. The remote sensing image description method based on prior and semantic alignment according to claim 4, characterized in that: The process of inputting the remote sensing image of the airport to be described into the image description model, and the image description model outputting the descriptive text of the remote sensing image of the airport to be described, is as follows: S1. Input the airport remote sensing image to be described into the Rice-ViT visual encoder. The Rice-ViT visual encoder outputs the visual embedding features of the airport remote sensing image to be described. S2. Input the visual embedding features into a dual-channel gated nonlinear mapper, and the dual-channel gated nonlinear mapper outputs the mapping features of the airport remote sensing image to be described. S3. Input the remote sensing image of the airport to be described into the panoramic segmentation network to obtain the panoramic segmentation result of the remote sensing image of the airport to be described. Extract the infrastructure information from the panoramic segmentation result and generate the prompt text of the remote sensing image of the airport to be described based on the infrastructure information. S4. Input the mapping features of the airport remote sensing image to be described and the prompt text into the fusion layer for fusion. The fusion layer outputs the multimodal features of the airport remote sensing image to be described. S5. Input the multimodal features of the airport remote sensing image to be described into the Qwen3-0.6B text decoder, and the Qwen3-0.6B text decoder outputs the descriptive text of the airport remote sensing image to be described.

6. The remote sensing image description method based on prior and semantic alignment according to claim 5, characterized in that: The specific process of S2 is as follows: S21. Input the visual embedding features of the airport remote sensing image to be described into the normalization processing layer, and the normalization processing layer outputs the corresponding processed visual embedding features. S22. The processed visual embedding features are input into the spatial merging layer, and the spatial merging layer outputs the corresponding intermediate features. The specific process is as follows: Let the spatial merging scale of the spatial merging layer be . Based on the processed visual embedding features of the input, each The visual embedding features of adjacent spatial tokens are aggregated into a new feature, resulting in the intermediate feature of the airport remote sensing image to be described. S23. Input the intermediate features into the first linear transformation layer, and the first linear transformation layer outputs... One-dimensional feature; S24. Input the feature one into the dimension slicing layer, and along the channel dimension, slice the feature one according to... The last dimension is divided into equal parts, and the gated branches corresponding to the dimensional divisions are output as features. and gated branch output characteristics ; S25, Output the gated branch feature and the gated branch output features Input activation fusion layer, using The nonlinear activation function outputs features of the gated branch. Activate the gated branch and output its features. With gated branch output characteristics Perform element-wise multiplication to activate the gated fusion feature output of the fusion layer; S26. The gated fusion features are input into the second linear transformation layer, and the second linear transformation layer outputs the mapping features of the airport remote sensing image to be described.

7. The remote sensing image description method based on prior and semantic alignment according to claim 6, characterized in that: The specific process of S3 is as follows: S31. Input the remote sensing image of the airport to be described into the panoramic segmentation network to obtain the corresponding panoramic segmentation result. S32. Set the minimum effective area threshold for aircraft and the minimum effective area threshold for buildings. Based on the minimum effective area threshold for aircraft and the minimum effective area threshold for buildings, constrain the panoramic segmentation results to obtain the effective aircraft set, effective building set and infrastructure set in the panoramic segmentation results. S33. Select the building with the largest area from the effective building set as the scene anchor point, obtain the coverage rate of the scene anchor point on the panoramic segmentation result, determine the scale classification result of the scene anchor point based on the coverage rate, extract the center coordinates of the scene anchor point simultaneously, convert the position of the center coordinates on the panoramic segmentation result into text directional words, and concatenate the text directional words with the scale classification result of the scene anchor point to generate a global anchor point text description of the panoramic segmentation result. When the largest building is missing from the valid building set, the default general scene anchor point is used to generate the global anchor point text description of the panoramic segmentation result; S34. Treat the set of valid aircraft as the vertex set of an undirected adjacency graph, set a clustering distance threshold, and then calculate the two-dimensional Euclidean distance between any two aircraft in the set of valid aircraft. If the two-dimensional Euclidean distance is less than the clustering distance threshold, it is determined that there is an adjacency relationship between the two aircraft and an undirected edge between the two aircraft is constructed. Otherwise, it is determined that there is no adjacency relationship between the two aircraft. Obtain the relationships between all aircraft in the valid aircraft set; The edge set of the undirected adjacency graph is obtained from all undirected edges. The complete undirected adjacency graph is obtained from the vertex set and edge set. The complete undirected adjacency graph is traversed, and connected components are extracted. Multiple aircraft connected by undirected edges are classified into a local aircraft group. The total number of aircraft contained in each local aircraft group and the position of each local aircraft group in the remote sensing image of the airport to be described are extracted, and the information of each local aircraft group in the remote sensing image of the airport to be described is obtained. S35. When runways and taxiways are detected within the panoramic segmentation mask of the infrastructure set, the runways and taxiways are merged to obtain the runway-taxiway area, a text description of the runway-taxiway area is generated, and the text description of the runway-taxiway area is used as the background cue context of the text decoder. If runways and taxiways are not detected, the apron is directly extracted and a text description of the apron is generated, and the apron text description is used as the background cue context of the text decoder. S36. By using a hierarchical structure prompt generation mechanism, the preset fixed text labels, the global anchor text description, the information of each local group of aircraft, and the background prompt context are merged into prompt text to obtain the prompt text of the remote sensing image of the airport to be described.

8. The remote sensing image description method based on prior and semantic alignment according to claim 7, characterized in that: The process of determining the scale classification result of the scene anchor point based on the coverage in S33 is as follows: If coverage Then the scene anchor points are divided into small ones; If coverage Then the scene anchor points will be divided into large ones; If coverage Then the scene anchor point will be classified as huge.

9. The remote sensing image description method based on prior and semantic alignment according to claim 8, characterized in that: The prompt text in S36 is: （1） in, For preset fixed text labels, For global anchor text description, Information for each local group of aircraft, To provide background context, This indicates a text concatenation operation. This is a combined function that organizes information from multiple local groups of aircraft into a natural language description in a preset order.

10. The remote sensing image description method based on prior and semantic alignment according to claim 9, characterized in that: The specific process of S4 is as follows: Suppose the task instruction text describes this remote sensing image of the airport. The task instruction text is input into the Qwen3-0.6B model. The built-in word segmenter and embedding layer of the Qwen3-0.6B model are used to perform word segmentation and spatial mapping on the task instruction text, respectively, to obtain the text embedding sequence of the task instruction. Simultaneously, the built-in word segmenter and embedding layer of the Qwen3-0.6B model are used to process the prompt text. Perform word segmentation and spatial mapping to obtain the embedded sequence of prompt text; The mapping features, the text embedding sequence of the task instructions, and the prompt text embedding sequence are concatenated to obtain the multimodal features of the airport remote sensing image to be described.

Images