Public emotion-driven urban park landscape updating intelligent generation method

Abstract

The invention discloses a public emotion-driven urban park landscape update intelligent generation method, which comprises the following steps of: acquiring a user public comment text related to an urban park based on a social media platform, and taking the user public comment text as original data for acquiring public emotion text-image association; screening and cleaning the user comment text data; quantizing expression emotion of the user comments based on natural language processing; establishing a supervision fine tuning training set composed of user comments with positive emotions; inputting the established training set into a Stable Diffusion model subjected to low-rank adaptation LoRA fine tuning, and performing landscape fine tuning and generation; evaluating a fine-tuned landscape optimization generation result; and the improvement effect of the landscape optimization scheme on public emotion is checked. Urban park landscape space scene updating schemes can be automatically generated in batches, the matching degree of the generated schemes and public emotion demands is further improved, and the design period is effectively shortened.

Description

Technical Field

[0001] This invention relates to landscape architecture planning and design and urban planning methods, and in particular to an intelligent generation method for urban park landscape renewal driven by public sentiment. Background Technology

[0002] Urban parks, as core public spaces that enhance residents' well-being and promote social cohesion, are highly valued in modern urban planning and landscape architecture. However, in the era of stock planning, existing urban park landscape optimization and renewal technologies still largely rely on experience-driven planning models. Public participation has long remained at a superficial level, such as "information disclosure and opinion collection," making it difficult to accurately respond to the public's emotional needs and hindering the high-quality development of urban parks.

[0003] While generative AI technology offers a new paradigm for design innovation, it still faces key bottlenecks in urban park landscape optimization, particularly in core dimensions such as accurately responding to public emotional needs and adapting to human-scale experiences. Specifically: 1. Superficial public participation: Traditional technologies have long remained at a superficial level, lacking the ability to deeply analyze multimodal content on social platforms, such as comment sentiment and image semantics, making it difficult to accurately capture public emotional preferences; 2. Lack of emotion-space correlation: While existing generative models can achieve cross-modal generation of "text-image," they have not established a precise mapping logic between "emotion and space," making it difficult to transform public emotional needs into optimization instructions for landscape spatial features; 3. Misalignment of perspective and scale: Existing intelligent planning and design often generates spatial schemes based on a "top-down perspective" (such as land cover and plot boundaries), while the core experience of urban park landscapes originates from the "human scale." This misalignment of perspective leads to a low degree of matching between the generated landscape details and residents' actual perceptual needs, making it difficult to achieve "emotion-driven" refined optimization. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method for intelligently generating urban park landscape updates that is driven by public sentiment and can utilize big data from social media to achieve efficient quantitative expression of public opinion and generate content that matches the public's emotional needs.

[0005] Technical solution: The present invention provides an intelligent generation method for urban park landscape renewal driven by public sentiment, comprising the following steps:

[0006] (1) Obtain user comments related to urban parks from social media platforms, including text, images and time / location metadata, as raw data for obtaining the "text-image" association of public sentiment;

[0007] (2) Clean the publicly available user comment data and exclude data that does not meet the training criteria;

[0008] (3) Quantify the sentiment of user comment text based on natural language processing to provide a basis for sentiment linking landscape images in the comments;

[0009] (4) Establish the “text-image” association of user comments and create a supervised fine-tuning training set consisting of user comments with positive sentiments;

[0010] (5) The established training set is used for the LoRA-tuned Stable Diffusion model. A low-rank adaptation matrix is ​​added to the attention layer of the U-Net architecture, the original weights are frozen, and only the newly added low-rank parameters are trained to perform landscape fine-tuning and generation.

[0011] (6) Evaluate the results of the fine-tuned landscape optimization;

[0012] (7) Examine the emotional enhancement effect of the landscape optimization scheme.

[0013] Furthermore, in step (1), publicly available comment data of users is obtained from social media platforms, including text and images. Each data is stored in a separate folder, which contains a comment text and its corresponding image. User comments containing only a single data type are not included in the acquisition scope.

[0014] Furthermore, in step (2), cleaning the publicly available user comment data includes:

[0015] Text data cleaning: Remove comments posted too early; remove text containing errors, duplicate comments, Baidu Encyclopedia citations, dialects, irrelevant content, or text shorter than the specified word count; exclude advertisements and duplicate descriptions of scenic spots;

[0016] Image data cleaning: Images are focused on city parks, excluding portraits, close-ups of objects, animals, plants, large crowds, or any content unrelated to city parks; images reflect the human eye's eye level, excluding images taken from above, below, by drone, or with excessively large or small focal lengths; images with poor lighting conditions are also excluded.

[0017] Furthermore, in step (3), sentiment analysis tools are used to perform sentiment quantification analysis on user comments related to the urban integrated park.

[0018] Furthermore, in step (4), the steps for establishing the "text-image" association of user comments and creating the landscape optimization training set are as follows:

[0019] (41) Use a tagger to identify the tags of each landscape image associated with the user's comments. Generate a corresponding document for each image, which contains all the identified visual tags and preset delimiters. The visual features and key information of the constituent elements of the image are effectively marked.

[0020] (42) Set a label confidence threshold or perform further manual verification to filter out low-confidence non-landscape-related labels, and merge or delete semantically repetitive or highly similar labels;

[0021] (43) Based on the sentiment analysis results, the training dataset is grouped according to the sentiment score, and the Booru dataset label manager is used to complete the label management;

[0022] (44) Assign a specific trigger word to each training set. During the training phase, the trigger word is used as an emotion-oriented target, and the emotion preference score is incorporated into the model as an additional high-order feature. When generating images using the fine-tuned model, the trigger word with a specific emotion score prompts the model to generate landscape images that correspond to the corresponding emotion tendency based on the learned emotion-feature correspondence.

[0023] Further, in step (5), inputting the training set into the StableDiffusion model fine-tuned by low-rank adaptive LoRA includes:

[0024] In a PyTorch 2.0.1 environment, LoRA v1.8.3 was used as the training framework to perform four fine-tuning tasks, each corresponding to a sentiment score.

[0025] Fine-tuning based on the open-source model Stable-Diffusion-v1-5.

[0026] Furthermore, in step (6), the evaluation of the fine-tuned landscape optimization generation results includes:

[0027] Quantitative evaluation includes Inception Score (IS) and Fraser Inception Distance (FID). Inception Score (IS) measures the entropy of the classification probability distribution of the generated image at the output of the Inception model and the KL divergence between the distribution of each image and the overall distribution. Fraser Inception Distance (FID) measures the difference between the generated image and the real dataset.

[0028] Qualitative evaluation: The model's ability to generate images with consistent or diverse features under similar or different input conditions; the model's ability to generate images that meet specific requirements based on text prompts.

[0029] Furthermore, the Inception Score IS averages the entropy of the classification probability distribution at the model output and the KL divergence between the distribution of each image and the overall distribution by weighted scores, capturing the concentration of the class distribution and its global diversity, and evaluating the quality and diversity of the generated images.

[0030] Furthermore, the Fraser Inception Distance (FID) is calculated as the Fraser distance between two multivariate Gaussian distributions of features extracted from the intermediate layers of the Inception network, assuming that the embeddings of both the real and generated images follow a Gaussian distribution.

[0031] Furthermore, in step (7), the evaluation of the emotional enhancement effect of the landscape optimization scheme includes:

[0032] (71) Obtain detection images by taking pictures in the target area according to the preset sampling method;

[0033] (72) Input the acquired detection image into the model fine-tuned in step (5) to generate an optimized landscape visualization effect under the guidance of positive sentiment score;

[0034] (73) Use a multimodal large-scale language model to evaluate the sentiment scores of landscape images before and after optimization;

[0035] (74) Descriptive statistics were performed on the sentiment scores.

[0036] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. It constructs a dual-modal fusion analysis model of "text sentiment + image semantics" to achieve large-scale collection of tens of millions of public sentiment data samples; 2. It realizes the automated batch generation of urban park landscape space scene update schemes, further improving the matching degree between the generated schemes and the public's emotional needs, and effectively compressing the design cycle. Attached Figure Description

[0037] Figure 1 This is a flowchart of the landscape space fine-tuning generation and verification process based on generative AI of the present invention.

[0038] Figure 2 This invention presents a graph illustrating the coupling relationship between public sentiment and landscape environment based on social media text-image data.

[0039] Figure 3 These are visual effect diagrams before and after optimization of the present invention. Detailed Implementation

[0040] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0041] like Figure 1 As shown, the intelligent generation method for urban park landscape renewal driven by public sentiment, as described in this invention, includes the following steps:

[0042] Step 1: Obtain user comments related to the city park from social media platforms, including text, images, and time / location metadata, as raw data for obtaining the "text-image" correlation of public sentiment;

[0043] First, the selection of locations for city parks is based on the following criteria:

[0044] It is already built, open to the public, and has a certain number of visitors; it is well-developed and has complete infrastructure; there are a large number of public user reviews on the online platform, covering a wide range of topics and of high quality, reflecting the diversity of modern urban parks;

[0045] Secondly, Python was used to retrieve user comments from social media platforms, including text and images, with each dataset stored in a separate folder;

[0046] The folder contains a comment text and its corresponding image to ensure a one-to-one correspondence between text and image in subsequent processing;

[0047] Any user comments containing only a single data type (e.g., only text or only images) are not included in the retrieval scope.

[0048] Step 2: Clean the user comment data;

[0049] First, the text data cleaning standards include: removing comments published too early, such as those from 10 years ago, to ensure the timeliness of the data and accurately reflect the current state of the park; removing text containing errors, duplicate comments, Baidu Encyclopedia citations, dialects, irrelevant content, or text with fewer than 30 words; and excluding irrelevant content such as promotional advertisements from travel agencies and hotels, and duplicate descriptions of scenic spots.

[0050] Secondly, the image data cleaning standards include: the images should focus on the city park, excluding portraits, close-ups of objects, animals, plants, large crowds, or any content unrelated to the city park; the images should reflect the human eye's eye level, excluding images taken from above, below, by drone, or with excessively large or small focal lengths; and images with poor lighting conditions, such as indoor scenes, overexposed or underexposed scenes, winter and cloudy scenes, and night or dusk scenes, should be excluded.

[0051] Step 3: Quantify the sentiment of user comment text;

[0052] First, the VADER tool was used to conduct a sentiment analysis of user reviews related to urban integrated parks; the sentiment score was divided into seven discrete levels, ranging from -3 to +3; where -3 represents the most negative and +3 represents the most positive.

[0053] This study focuses on analyzing user reviews with positive sentiment scores, i.e., sentiment scores ≥ 0, and further subdivides them into four categories with sentiment values ​​of 0, 1, 2, and 3. Table 1 shows the case studies of sentiment value classification of user reviews from 15 representative urban parks in Nanjing from 2016 to 2023.

[0054] To ensure consistency in size, color, perspective, and content purity of the training dataset, the cleaned user review image data needs to be standardized, including the following aspects: Size: The image resolution is uniformly set to 768×512 to ensure the consistency of the input data; Color: Images with severe color or grayscale distribution deviations are removed to prevent them from negatively impacting the training process; Perspective: Images are taken at a level viewpoint of 1.5-2 meters to ensure no image flipping and to maintain a specific angle to guarantee the accuracy and stability of landscape elements; Content purity: Images containing homogeneous content, watermarks, or other irrelevant elements are removed to ensure image cleanliness and usability.

[0055] Table 1. Classification of Affective Score Levels

[0056] Step 4: Establish the "text-image" association of user comments and create a supervised fine-tuning training set mainly composed of user comments reflecting positive emotions;

[0057] To further identify landscape optimization targets in the images, the WD 1.4 tagger was used for label recognition. A corresponding txt document was generated for each image, containing all identified tags and their delimiters, ensuring that key information about the image's visual features and constituent elements was effectively labeled.

[0058] The labels predicted by this tagger may include irrelevant content outside of landscape elements. Therefore, it is necessary to set a reasonable label extraction threshold or conduct further manual verification to ensure the quality of landscape element labels.

[0059] To prevent overfitting due to semantically redundant labels during model training, it is necessary to review and appropriately delete semantically similar labels.

[0060] Based on the previous sentiment analysis results, the training dataset was grouped according to sentiment scores of 0, 1, 2, and 3, and the Booru dataset label manager was used to manage the labels.

[0061] To achieve targeted supervised learning, this study assigns a specific trigger word to each training set, with positive sentiment scores of 0, 1, 2, and 3. During the training phase, the trigger word serves as the sentiment-oriented target, and the sentiment preference score is incorporated into the model as an additional higher-order feature. Subsequently, when generating images using the fine-tuned model, the trigger word with the specific sentiment score prompts the model to generate landscape images that correspond to the learned sentiment-feature correspondence, thereby generating images that align with the corresponding sentiment tendency.

[0062] Step 5: Input the training set into the Stable Diffusion model with low-rank adaptation LoRA fine-tuning, add a low-rank adaptation matrix to the attention layer of the U-Net architecture, freeze the original weights, train only the newly added low-rank parameters, and perform landscape fine-tuning and generation.

[0063] In the PyTorch 2.0.1 environment, LoRA v1.8.3 is used as the training framework. A total of four fine-tuning tasks need to be performed, each task corresponding to a sentiment score. The sentiment score levels include 0, 1, 2 and 3.

[0064] We chose to fine-tune the official open-source base model Stable-Diffusion-v1-5.

[0065] Step 6: Quantitatively and qualitatively evaluate the fine-tuned landscape optimization results;

[0066] Quantitative assessments include: Inception Score (IS) and Fraser Inception Distance (FID).

[0067] The Inception Score (IS) measures the entropy of the classification probability distribution of the generated image at the output of the Inception model, as well as the KL divergence (Kullback-Leibler divergence) between the distribution of each image and the overall distribution. By averaging these values ​​using weighted scores, the concentration of the class distribution and its global diversity can be captured simultaneously, thereby evaluating the quality and diversity of the generated images.

[0068] The Fraser Inception Distance (FID) directly measures the difference between a generated image and the real dataset. It calculates the Fraser distance between two multivariate Gaussian distributions of features extracted from the intermediate layers of the Inception network. This method assumes that the embeddings of both the real and generated images follow a Gaussian distribution. The distance quantifies how close the feature distribution of the generated image is to the feature distribution of the real image.

[0069] Qualitative evaluation includes: Realism and accuracy: the model's ability to reproduce realistic visual effects; Stability and diversity: the model's ability to generate images with consistent or diverse features under similar or different input conditions; Controllability and generalization: the model's ability to generate images that meet specific requirements (including style, elements, and scene conditions) based on text prompts.

[0070] Taking Nanjing Baima Park as a case study, a sampling method with a road network interval of 50 meters was used, and a total of 180 baseline images were collected from 45 sampling points. The baseline images were then optimized using the LoRA fine-tuning model of the original SD model. Compared with the original SD model, the LoRA fine-tuned model performed better in generating park landscape images. The inception score (IS value) increased from 8.57 to 9.46, indicating that the generated output has greater diversity. The Fraser inception distance (FID value) decreased from 5.79 to 3.32, indicating that the distribution of the generated images is closer to the distribution of the real images.

[0071] Step 7: Examine the emotional enhancement effect of the landscape optimization plan;

[0072] Images were obtained through on-site sampling and photography; sampling points were spaced more than 50 meters apart, and four-way street view photography was used at a height of 1.6 meters; good weather and lighting conditions were ensured during photography to minimize exposure problems.

[0073] The acquired detection images are processed in batches and intelligently using the generative AI model established above, and optimized landscape visualization effects are generated under the guidance of tourists' positive emotion ratings.

[0074] To verify the effectiveness of the fine-tuning model in enhancing positive emotions, a multimodal large language model (LLM) was used to evaluate the emotional scores of landscape images before and after optimization. To ensure the unbiasedness of the evaluation results, the optimization stage of the evaluated images was concealed from the LLM in the cue word part. Using 180 baseline images and 180 optimized images of Baima Park in Nanjing as a case study, LLM was used to evaluate the emotional information of the images during the optimization stage. The average landscape emotional score of LLM before and after optimization increased from 1.41 (standard deviation = 0.75) to 2.27 (standard deviation = 0.67), an increase of 60.99%.

[0075] After descriptive statistics of the emotion scores in the two stages, the paired T-test or Wilcoxon signed-rank test was used to analyze the significance of the difference in emotion before and after optimization. The p-value calculated by the Wilcoxon signed-rank test was less than 0.001, indicating that there was a statistically significant difference in public sentiment between the two stages.

[0076] Figure 3 To optimize the visual effects before and after.

Claims

1. A public sentiment-driven intelligent generation method for urban park landscape renewal, characterized in that, Includes the following steps: (1) Obtain user comments related to urban parks from social media platforms, including text, images, and time / location metadata, as raw data for obtaining the "text-image" association of public sentiment; (2) Clean the publicly available user comment data and exclude data that does not meet the training criteria; (3) Quantify the sentiment of user comment text based on natural language processing to provide a basis for sentiment linking landscape images in the comments; (4) Establish the "text-image" association of user comments and create a supervised fine-tuning training set consisting of user comments with positive sentiments; (5) The established training set is used for the LoRA-tuned Stable Diffusion model. A low-rank adaptation matrix is ​​added to the attention layer of the U-Net architecture, the original weights are frozen, and only the newly added low-rank parameters are trained to perform landscape fine-tuning and generation. (6) Evaluate the results of the fine-tuned landscape optimization; (7) Examine the emotional enhancement effect of the landscape optimization scheme.

2. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 1, characterized in that, In step (1), publicly available comment data of users is obtained from social media platforms, including text and images. Each data is stored in a separate folder, which contains a comment text and its corresponding image. User comments containing only a single data type are not included in the acquisition scope.

3. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 1, characterized in that, Step (2) involves cleaning the publicly available user comment data, including: Text data cleaning: Remove comments posted too early; remove text containing errors, duplicate comments, Baidu Encyclopedia citations, dialects, irrelevant content, or text shorter than the specified word count; exclude advertisements and duplicate descriptions of scenic spots; Image data cleaning: Images should focus on the city park, excluding portraits, close-ups of objects, animals, plants, large crowds, or any content unrelated to the city park; images should reflect the human eye's eye level, excluding images taken from above, below, by drone, or with excessively large or small focal lengths; images with poor lighting conditions should also be excluded.

4. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 1, characterized in that, In step (3), sentiment analysis tools are used to perform sentiment quantification analysis on user comments related to the urban integrated park.

5. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 1, characterized in that, In step (4), the steps for establishing the "text-image" association of user comments and creating the landscape optimization training set are as follows: (41) Use a tagger to identify the tags of each landscape image associated with the user's comments. Generate a corresponding document for each image, which contains all the identified visual tags and preset delimiters. The visual features and key information of the constituent elements of the image are effectively marked. (42) Set a label confidence threshold or perform further manual verification to filter out low-confidence non-landscape-related labels, and merge or delete semantically repetitive or highly similar labels; (43) Based on the sentiment analysis results, the training dataset is grouped according to the sentiment score, and the Booru dataset label manager is used to complete the label management; (44) Assign a specific trigger word to each training set. During the training phase, the trigger word is used as an emotion-oriented target, and the emotion preference score is incorporated into the model as an additional high-order feature. When generating images using the fine-tuned model, the trigger word with a specific emotion score prompts the model to generate landscape images that correspond to the corresponding emotion tendency based on the learned emotion-feature correspondence.

6. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 1, characterized in that, In step (5), inputting the training set into the Stable Diffusion model fine-tuned by low-rank adaptive LoRA includes: In a PyTorch 2.0.1 environment, LoRA v1.8.3 was used as the training framework to perform four fine-tuning tasks, each corresponding to a sentiment score. Fine-tuning based on the open-source model Stable-Diffusion-v1-5.

7. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 1, characterized in that, In step (6), the evaluation of the fine-tuned landscape optimization results includes: Quantitative evaluation includes Inception Score (IS) and Fraser Inception Distance (FID). Inception Score (IS) measures the entropy of the classification probability distribution of the generated image at the output of the Inception model and the KL divergence between the distribution of each image and the overall distribution. Fraser Inception Distance (FID) measures the difference between the generated image and the real dataset. Qualitative evaluation: The model's ability to generate images with consistent or diverse features under similar or different input conditions; the model's ability to generate images that meet specific requirements based on text prompts.

8. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 7, characterized in that, The Inception Score (IS) is a weighted score that averages the entropy of the classification probability distribution at the model output and the KL divergence between the distribution of each image and the overall distribution. It captures the concentration of the class distribution and its global diversity, and evaluates the quality and diversity of the generated images.

9. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 7, characterized in that, The Fraser distance (FID) is calculated between two multivariate Gaussian distributions of features extracted from the intermediate layers of the Inception network, assuming that the embeddings of both the real and generated images follow a Gaussian distribution.

10. The intelligent generation method for urban park landscape renewal driven by public sentiment as described in claim 1, characterized in that, In step (7), the evaluation of the emotional enhancement effect of the landscape optimization scheme includes: (71) Obtain detection images by taking pictures in the target area according to the preset sampling method; (72) Input the acquired detection image into the model fine-tuned in step (5) to generate an optimized landscape visualization effect under the guidance of positive sentiment score; (73) Use a multimodal large-scale language model to evaluate the sentiment scores of landscape images before and after optimization; (74) Descriptive statistics were performed on the sentiment scores.

Images