A remote sensing image description evaluation method based on cycle consistency
By generating and reconstructing images using a large multimodal remote sensing model, and combining this with a multimodal quality assessment model for cyclic consistency evaluation, the problems of reliance on manual annotation and semantic loss in remote sensing image description evaluation are solved, achieving efficient and accurate evaluation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-23
- Publication Date
- 2026-06-26
AI Technical Summary
Existing remote sensing image description and evaluation methods rely on costly manual annotation for fine-grained long text tasks. Traditional indicators have weak semantic expression capabilities, semantic loss occurs during text-to-image reconstruction, and reliable quantitative evaluation standards are lacking.
A remote sensing image description evaluation method based on cycle consistency is adopted. Fine-grained long text descriptions are generated by a remote sensing multimodal large model, and images are reconstructed by combining a weighted basis image strategy. A 0-5 level scoring system is designed to compare cycle consistency using a multimodal quality assessment model.
It achieves reference-free, objective, and accurate remote sensing image description and evaluation, reduces evaluation costs, preserves image spatial structure, improves the accuracy and reliability of evaluation, conforms to human visual perception, and provides fine-grained quantitative standards.
Smart Images
Figure CN122289763A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a remote sensing image description and evaluation method based on cycle consistency, belonging to the field of remote sensing image description and evaluation technology. Background Technology
[0002] With the deep integration of remote sensing technology and artificial intelligence, Remote Sensing Multimodal Large Language Model (RSMLLM) has shown great potential in remote sensing image captioning tasks. High-quality remote sensing image captioning not only requires the model to identify land cover categories, but also to accurately describe the spatial layout, texture features, and interrelationships of land cover. Currently, remote sensing image captioning datasets are evolving from traditional short texts to fine-grained long texts, but this also brings new challenges to the automatic evaluation of caption quality.
[0003] In existing technologies, the evaluation of remote sensing image descriptions mainly relies on traditional natural language processing metrics such as BLEU, METEOR, ROUGE, and CIDEr. These metrics are typically calculated based on the word frequency overlap between the test text and manually annotated reference text. However, traditional metrics have significant limitations when processing fine-grained long texts: First, high-quality manually annotated long texts are extremely costly and scarce, limiting the evaluation benchmark; second, traditional metrics focus too much on word alignment and struggle to measure the text's expressive quality in remote sensing semantics, spatial logic, and information sufficiency, often resulting in scores that do not align with human visual perception.
[0004] Furthermore, some existing evaluation methods attempt to conduct no-reference evaluation using a cyclical consistency approach of "image-text-image," that is, using a text-to-image model to restore the generated description to an image, and then qualitatively evaluating the consistency between the original image and the reconstructed image. However, in practical applications, due to the severe semantic gap (information loss) during the text-to-image conversion process, the reconstructed image often loses the original spatial details and layout features, leading to unreliable evaluation results. At the same time, there is no unified quantitative evaluation and grading standard for complex feature distributions in remote sensing scenarios, and a closed-loop evaluation system that comprehensively considers information sufficiency, minimum redundancy, and human understandability is lacking.
[0005] Therefore, a cyclic consistency evaluation method that can compensate for text-to-image conversion loss and has a reliable quantitative grading standard is needed to achieve an objective and accurate evaluation of the generative description of remote sensing multimodal large models. Summary of the Invention
[0006] The technical problem to be solved by this invention is to provide a remote sensing image description evaluation method based on cycle consistency, which solves the problems of existing remote sensing image description evaluation indicators failing in fine-grained long text tasks, over-reliance on manual reference annotation, and semantic loss in the process of text-to-image reconstruction, and achieves reference-free, highly reliable, and fine-grained automatic evaluation.
[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A remote sensing image description evaluation method based on cycle consistency includes the following steps: Step 1: Use a remote sensing multimodal large model to infer the original remote sensing image and generate a corresponding fine-grained long text description; Step 2: Extract the structural prior information of the original remote sensing image, construct a weighted basis image, and simultaneously input the fine-grained long text description and the weighted basis image into the preset text-based image model to fuse them and obtain the reconstructed remote sensing image; Step 3: Using a pre-defined multimodal quality assessment model as an evaluator, the original remote sensing image and the reconstructed remote sensing image are compared cyclically for consistency. Reasoning is performed from three dimensions: information sufficiency, redundancy, and understandability to obtain a consistency score level. Based on a pre-defined hierarchical mapping function, the consistency score level is mapped to a quantitative score. Step 4: Normalize the quantified scores to obtain the final remote sensing image description quality evaluation results.
[0008] As a preferred embodiment of the present invention, in step 1, the original remote sensing image is... Inputting a remote sensing multimodal large model RSMLLM, the system generates fine-grained long-text descriptions reflecting land cover categories, spatial layout, texture features, and topological relationships between land covers through an inference process. .
[0009] As a preferred embodiment of the present invention, the specific process of step 2 is as follows: Define feature extraction operators By analyzing the original remote sensing images Perform Gaussian noise restoration, image blur restoration, or edge feature extraction. Prior structural information: , in, For weighted basis images, To control the raw remote sensing images Hyperparameters for the degree to which pixel information is preserved; Fine-grained long text description With weighted basis image Input a pre-defined texture image model FLUX.2-dev-bnb-4bit, and obtain the reconstructed remote sensing image through weighted fusion. : , in, Operators representing the text-based graph model; The fusion weights are the weighted base images used to control the degree to which the reconstructed remote sensing image inherits the spatial structure of the original remote sensing image.
[0010] As a preferred embodiment of the present invention, the specific process of step 3 is as follows: Original remote sensing images With reconstructed remote sensing images Simultaneously, the pre-set multimodal quality assessment model Qwen3-VL-Plus is input. The assessment model performs reasoning analysis based on the cyclic consistency evaluation function from three dimensions: information sufficiency, redundancy, and understandability, and outputs a consistency score level. , Characterization and The degree of alignment in scene type, feature distribution, and spatial structure is scored based on a preset hierarchical mapping function to determine the consistency level. Mapped to a quantitative score; The cycle consistency evaluation function is expressed as: , in, For consistency rating level, For multimodal quality assessment models, This serves as a task prompt template to guide model evaluation; information sufficiency is used to assess fine-grained long text descriptions. Is it possible to reconstruct the original remote sensing image using a Wensheng image model? Images with consistent key features; redundancy used to evaluate fine-grained long text descriptions. Does it contain redundant or erroneous information unrelated to the reconstruction task? Understandability is used to assess the logical coherence of the remote sensing image description. Hierarchical mapping function The mapping criteria are as follows: 1) The matching degree of the main land cover types is less than 10%. ; 2) The matching degree of the main land cover types is between 10% and 30%. ; 3) The matching degree of major land cover types is between 30% and 50%, while the matching degree of spatial distribution characteristics is less than 50%. ; 4) The matching degree of major land cover types is between 50% and 70%, and the matching degree of spatial distribution characteristics is between 50% and 70%. ; 5) The main land cover types and spatial distribution characteristics have a 70% to 90% matching consistency, and the relative deformation error of the outlines and boundaries of the main land cover types is less than 20%. ; 6) The matching degree of the main land cover types is greater than 90%, and the detail error is less than 10%. .
[0011] As a preferred embodiment of the present invention, the specific process of step 4 is as follows: , in, To determine the final quality assessment results for remote sensing image descriptions, To quantify the score, and These are the preset highest and lowest grade scores, respectively.
[0012] A computer device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of the circular consistency-based remote sensing image description and evaluation method.
[0013] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the circular consistency-based remote sensing image description and evaluation method.
[0014] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: 1. This invention eliminates the reliance on high-quality manually annotated reference text by adopting a circular consistency path of "image-text-image", which greatly reduces the time and economic cost of model evaluation on large-scale remote sensing datasets and achieves objective evaluation without reference.
[0015] 2. This invention introduces a weighted base image strategy that includes Gaussian noise restoration, fuzzy restoration, or edge features. This preserves the spatial structure constraints of the original image during image reconstruction, effectively solves the spatial layout misalignment problem that easily occurs when generating pure text, compensates for the semantic gap in cross-modal conversion, and improves the accuracy of evaluation.
[0016] 3. The 0-5 level scoring system designed in this invention deeply integrates the characteristics of remote sensing images (such as land cover type, object distribution density, etc.). Compared with traditional word frequency matching indicators such as BLEU and CIDEr, it can better reflect the fine-grained semantic accuracy in long text descriptions, and has a fine-grained and highly reliable quantitative standard. The evaluation results are more in line with human visual perception.
[0017] 4. The evaluation framework proposed in this invention does not depend on a specific multimodal large model structure and can be widely applied to the performance evaluation of various remote sensing multimodal large models, the screening of training data quality, and geographic information perception tasks. It provides a standardized and automated quality assessment scheme for long text description tasks in the field of remote sensing, and is highly versatile and easy to extend. Attached Figure Description
[0018] Figure 1 This is a flowchart of the remote sensing image description and evaluation method based on cycle consistency according to the present invention; Figure 2 This is an architecture diagram of the remote sensing image description and evaluation method proposed in this invention. Detailed Implementation
[0019] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0020] To address the problems of high manual annotation costs, weak semantic expression capabilities of traditional metrics (such as BLEU and CIDEr), and semantic information loss (gap) during text-to-image reconstruction in existing remote sensing image captioning evaluation metrics for fine-grained long text tasks, this invention proposes a remote sensing image captioning evaluation method based on cycle consistency. This method utilizes "image..." text The iterative verification logic of the "image" introduces a weighted basis image strategy to compensate for reconstruction loss, and uses a multimodal large language model (MLLM) as an evaluator to achieve referenceless quantitative evaluation.
[0021] like Figure 1 and Figure 2 As shown, this invention proposes a remote sensing image description evaluation method based on cycle consistency, comprising the following steps: Step 1: Use a multimodal remote sensing model to infer information from the original remote sensing images, generating fine-grained text descriptions that reflect the types, attributes, and spatial layout of ground features. Original remote sensing images Input a Remote Sensing Multimodal Model (RSMLLM), and through a fine-tuned inference process, generate corresponding fine-grained long text descriptions. .describe It needs to cover the types of land features, spatial layout, texture features, and topological relationships between land features in the image.
[0022] Step 2: Construct a weighted basis image based on the original remote sensing image, and input it along with the text description into a preset text-based image model to reconstruct the image. To compensate for the spatial structure shift and semantic loss that may occur when generating images using pure text-driven methods, this invention introduces a weighted basis image strategy. 1) Define the feature extraction operator Through the Perform Gaussian noise restoration, image blur restoration, or edge detection processing to extract structural prior information from the image and construct a weighted basis image. ;
[0023] in, Hyperparameters that control the degree of information retention; Gaussian noise recovery: Preliminary restoration is performed after applying Gaussian noise, preserving pixel distribution characteristics; Image blur restoration: For Gaussian blurring is applied to preserve macroscopic semantic information; Edge feature extraction: using operators to extract The edge contours are used to construct a structured base image.
[0024] 2) Describe the text With weighted basis image Input a preset raw image model (such as FLUX.2-dev-bnb-4bit), and reconstruct the image through weighted fusion. The reconstruction process is represented as follows: , in, Represents the textual graph operator, The fusion weights are the weighted base images used to control the degree to which the reconstructed image inherits the original spatial structure.
[0025] Step 3: Perform cycle consistency comparison using a multimodal quality assessment model. Original image With reconstructed images Simultaneously input a multimodal quality assessment model (such as Qwen3-VL-Plus). The assessment model is based on Poca's definition of quality description and performs inference analysis from the following three dimensions based on the cycle consistency evaluation function: 1) Information adequacy: assessment Is it sufficient for the model to reconstruct the same... Images with consistent key features; 2) Minimal Redundancy: Evaluation Does it contain redundant or erroneous information unrelated to the reconstruction task? 3) Understandability: Assess the logical coherence of the description.
[0026] The evaluation model outputs a qualitative evaluation text. It represents the degree of alignment between two images in terms of scene type, ground feature distribution, and spatial structure.
[0027] Its cycle consistency evaluation function is defined as: , in, For the pre-defined multimodal quality assessment model, by comparison right The degree of reconstruction in terms of land cover categories, spatial locations, and attribute characteristics is used to verify the text description in reverse. The quality.
[0028] Based on hierarchical mapping function A quantitative score is generated and mapped to a preset 0-5 level rating standard. The specific grading logic is as follows: Level 0 (Completely Irrelevant): The scene's global features do not match at all. The generated image has a less than 10% match with the original image's main land cover types (land cover categories), failing to achieve semantic alignment with the original description. ; Level 1 (Extremely Low Consistency): Only occasional similarity exists in local pixels or textures; the matching degree of core descriptive land cover types is between 10% and 30%; the proportion of missing key objects exceeds 70%; or there are serious category determination errors. ; Level 2 (Low Consistency): Possesses 30% to 50% scene feature type matching, but the spatial distribution matching of major structures or objects is below 50%, with significant deviations in scale, outline shape, or orientation features. ; Level 3 (Medium Consistency): The macro-level scene type (e.g., city, farmland) is consistent, the main objects described in the description exist, and the land cover type matching degree reaches 50% to 70%; however, the error range of the spatial topology or relative position layout of the core objects is between 30% and 50%. ; Level 4 (High Consistency): The main land cover categories and spatial distribution characteristics have a 70% to 90% matching consistency, the spatial topological relationships roughly match the original map, and the relative deformation error of object outlines and boundaries is less than 20%. ; Level 5 (Near-Perfect Consistency): Global scene, local object details, and spatial structure are highly aligned; overall land cover type matching is greater than 90%; the difference in detail is less than 10, which is within the reasonable noise range of the model generation and does not affect downstream image semantic interpretation tasks. .
[0029] Step 4: Calculate the final evaluation index. To eliminate quantization bias between different models or tasks, the quantization scores are normalized to obtain the final remote sensing image description quality evaluation result. The calculation formula is: , in, , . The closer the value is to 1, the higher the quality of the description generated by the remote sensing multimodal large model.
[0030] This implementation method introduces... The constrained circular consistency architecture effectively solves the problem of difficult evaluation of long text descriptions in remote sensing. By leveraging the powerful expressiveness of the text-to-image model as an intermediary, it transforms abstract text semantic comparisons into intuitive image content alignment, thereby achieving more accurate performance evaluation than traditional metrics.
[0031] To verify the effectiveness of the remote sensing image description evaluation method based on cycle consistency proposed in this invention, this embodiment selects several representative remote sensing multimodal large models and conducts comparative experiments on public datasets of different scales, as detailed below: This embodiment selects two representative remote sensing description datasets, covering two task scenarios: short text and long text. 1) Sydney-captions (short text dataset): contains 613 high-resolution remote sensing images, each with 5 short, manually annotated descriptions. It mainly examines the model's ability to identify core land cover categories.
[0032] 2) RSIEval (Fine-grained long text dataset): Contains 100 high-quality remote sensing images and fine-grained long text annotations. The descriptions cover land cover attributes, spatial relationships and scene details, and are used to verify the parsing ability of evaluation methods under complex semantics.
[0033] Model Selection: This embodiment selects several mainstream types of remote sensing multimodal large models as evaluation objects, including: 1) Falcon-0.7B (Caption Fine-tuning Model): A model trained on a remote sensing command dataset using a dedicated Caption task.
[0034] 2) RemoteReasoner (non-caption fine-tuning model): It has only been fine-tuned for perception tasks such as object detection and land cover classification, and has not been optimized for long text descriptions.
[0035] 3) Qwen3-VL-8B (General Multimodal Base Model): A general large model that has not been fine-tuned in the field of remote sensing.
[0036] Evaluation indicators and experimental setup: Benchmark indicators: Traditional reference indicators BLEU-4, METEOR, and human scoring were selected as benchmarks.
[0037] The metric of this invention: the normalized cycle consistency score proposed in this invention. Experimental setup: The Wensheng image model used FLUX.2-dev-bnb-4bit, and the multimodal quality assessment model used Qwen3-VL-Plus. During the evaluation process, the fusion weights of the weighted base images were... Set to 0.3 to balance text guidance with spatial structure constraints.
[0038] On the two datasets mentioned above, the scores obtained by the method of this invention will be used. Correlation analysis was performed on the HumanCorrelation score and traditional indicators. The experimental results are shown in Table 1:
[0039] Table 1
[0040] Based on the comprehensive experimental results, the remote sensing image description evaluation method based on cycle consistency proposed in this invention has achieved significant success, with its evaluation performance comprehensively surpassing traditional reference indicators. Specifically, this method demonstrates stronger semantic capture capabilities and higher consistency with human perception across text evaluations of varying lengths.
[0041] On the short text dataset Sydney-captions, the correlation coefficient (HC) between the present invention's metric and human perception reached 82, an improvement of 21 percentage points compared to the BLEU-4 metric (61). On the more challenging fine-grained long text dataset RSIEval, the superiority of the present invention is even more pronounced, with a correlation coefficient of 89, while the traditional metric METEOR is only 41, and BLEU-4 is even lower at 38. This indicates that as the length and fine-grained information of the descriptive text increase, the evaluation efficiency of traditional metrics decreases significantly due to over-reliance on word matching, while the present invention, through the logical closed loop of "image reconstruction," can accurately identify semantic biases and spatial logic errors in the description. Furthermore, in the evaluation of high-performance models such as Qwen3-VL-8B, traditional metrics gave extremely low scores due to differences in model generation style and reference summary (e.g., BLEU-4 only 0.25), but the present invention still gave a logically consistent high score (66.2), clearly demonstrating the universality and reliability of the present invention's method in solving the problem of "long text evaluation failure," making it an effective remote sensing multimodal large model evaluation technique.
[0042] Based on the same inventive concept, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned remote sensing image description and evaluation method based on cycle consistency.
[0043] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned remote sensing image description and evaluation method based on cycle consistency.
[0044] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0045] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0048] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.
Claims
1. A method for evaluating remote sensing image description based on cycle consistency, characterized in that, Includes the following steps: Step 1: Use a remote sensing multimodal large model to infer the original remote sensing image and generate a corresponding fine-grained long text description. Step 2: Extract the structural prior information of the original remote sensing image, construct a weighted basis image, and simultaneously input the fine-grained long text description and the weighted basis image into the preset text-based image model to fuse them and obtain the reconstructed remote sensing image; Step 3: Using a pre-defined multimodal quality assessment model as an evaluator, the original remote sensing image and the reconstructed remote sensing image are compared cyclically for consistency. Reasoning is performed from three dimensions: information sufficiency, redundancy, and understandability to obtain a consistency score level. Based on a pre-defined hierarchical mapping function, the consistency score level is mapped to a quantitative score. Step 4: Normalize the quantified scores to obtain the final remote sensing image description quality evaluation results.
2. The remote sensing image description and evaluation method based on cycle consistency according to claim 1, characterized in that, In step 1, the original remote sensing image Inputting a remote sensing multimodal large model RSMLLM, the system generates fine-grained long-text descriptions reflecting land cover categories, spatial layout, texture features, and topological relationships between land cover features through an inference process. .
3. The remote sensing image description and evaluation method based on cycle consistency according to claim 1, characterized in that, The specific process of step 2 is as follows: Define feature extraction operators By analyzing the original remote sensing images Perform Gaussian noise restoration, image blur restoration, or edge feature extraction. Prior structural information: ， in, For weighted basis images, To control the raw remote sensing images Hyperparameters for the degree to which pixel information is preserved; Fine-grained long text description With weighted basis image Input a pre-defined texture image model FLUX.2-dev-bnb-4bit, and obtain the reconstructed remote sensing image through weighted fusion. : ， in, Operators representing the text-based graph model; The fusion weights are the weighted base images used to control the degree to which the reconstructed remote sensing image inherits the spatial structure of the original remote sensing image.
4. The remote sensing image description and evaluation method based on cycle consistency according to claim 1, characterized in that, The specific process of step 3 is as follows: Original remote sensing images With reconstructed remote sensing images Simultaneously, the pre-set multimodal quality assessment model Qwen3-VL-Plus is input. The assessment model performs reasoning analysis based on the cyclic consistency evaluation function from three dimensions: information sufficiency, redundancy, and understandability, and outputs a consistency score level. , Characterization and The degree of alignment in scene type, feature distribution, and spatial structure is scored based on a preset hierarchical mapping function to determine the consistency level. Mapped to a quantitative score; The cycle consistency evaluation function is expressed as: ， in, For consistency rating level, For multimodal quality assessment models, This serves as a task prompt template to guide model evaluation; information sufficiency is used to assess fine-grained long text descriptions. Is it possible to reconstruct the original remote sensing image using a Wensheng image model? Images with consistent key features; redundancy used to evaluate fine-grained long text descriptions. Does it contain redundant or erroneous information unrelated to the reconstruction task? Understandability is used to assess the logical coherence of the remote sensing image description. Hierarchical mapping function The mapping criteria are as follows: 1) The matching degree of the main land cover types is less than 10%. ; 2) The matching degree of the main land cover types is between 10% and 30%. ; 3) The matching degree of major land cover types is between 30% and 50%, while the matching degree of spatial distribution characteristics is less than 50%. ; 4) The matching degree of major land cover types is between 50% and 70%, and the matching degree of spatial distribution characteristics is between 50% and 70%. ; 5) The main land cover types and spatial distribution characteristics have a 70% to 90% matching consistency, and the relative deformation error of the outlines and boundaries of the main land cover types is less than 20%. ; 6) The matching degree of the main land cover types is greater than 90%, and the detail error is less than 10%. .
5. The remote sensing image description and evaluation method based on cycle consistency according to claim 1, characterized in that, The specific process of step 4 is as follows: ， in, To determine the final quality assessment results for remote sensing image descriptions, To quantify the score, and These are the preset highest and lowest grade scores, respectively.
6. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the remote sensing image description and evaluation method based on cycle consistency as described in any one of claims 1 to 5.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the remote sensing image description and evaluation method based on cycle consistency as described in any one of claims 1 to 5.