Adaptive design method based on urban street disaster scenario generation

Abstract

The application provides an adaptive design method based on urban street disaster scene generation, and belongs to the technical field of urban disaster simulation, comprising: multi-element data acquisition and processing, extracting disaster occurrence probability characteristics; training the diffusion model based on the LoRA method and the ControlNet model; generating street disaster simulation images based on the trained model based on the disaster occurrence probability characteristics and the street view images; inputting the generated images and related data into a large language model to generate transformation suggestions and transformed street simulation images; performing image semantic segmentation using an open vocabulary image segmentation model, performing closed-loop optimization and evaluation; and multi-scene adaptive strategy induction and classification. The application introduces disaster-space conflict detection and multi-agent automatic feedback mechanism, so that the generated results have visual authenticity, quantifiable evaluation and automatic optimization capability, and provides technical support for urban resilience planning and climate adaptive design.

Description

Technical Field

[0001] This invention belongs to the field of urban disaster simulation technology, and in particular relates to an adaptive design method based on the generation of urban street disaster scenarios. Background Technology

[0002] In recent years, influenced by global climate change, the frequency and intensity of extreme weather events have been increasing, posing a significant threat to urban infrastructure and residents' lives. According to statistics from the World Meteorological Organization (WMO), the number of climate-related disasters worldwide has increased fivefold in the past 50 years. Extreme precipitation, floods, heat waves, droughts, typhoons, and cold waves have become major factors affecting urban safety and resilience. Meanwhile, rapid urbanization has led to a high concentration of population and critical infrastructure, further exacerbating disaster risks.

[0003] Currently, urban disaster scenario simulation mainly relies on physical modeling, GIS system analysis, and traditional computer graphics methods. These methods suffer from long modeling cycles, high data requirements, and difficulty in intuitively representing complex dynamic disaster processes, thus limiting their effectiveness in disaster early warning, urban planning, and emergency management.

[0004] With the development of artificial intelligence technology, generative AI technologies such as Generative Adversarial Networks (GANs) and diffusion models have been initially applied to visual simulation of single-type disasters (such as floods). However, existing research mostly focuses on specific disaster types, lacking universal simulation capabilities for multi-hazard scenarios, and the linkage mechanism between GANs and urban policy-making and spatial transformation strategies is still imperfect. In addition, traditional AI-generated image models mostly aim for image realism, lacking deep integration with the needs of improving urban resilience.

[0005] Therefore, there is an urgent need for a method that can combine multi-source urban data to quickly and realistically simulate multi-hazard urban street scene disaster scenarios, and propose spatial transformation suggestions based on the simulation results, so as to improve the climate adaptability and disaster resistance of cities. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned issues of lacking universal simulation capabilities for multi-hazard scenarios, an imperfect linkage mechanism between urban policy formulation and spatial transformation strategies, and the lack of deep integration between AI-generated image models and the needs for enhancing urban resilience. This invention provides a method that can rapidly and realistically simulate multi-hazard urban street scene disaster scenarios by combining multi-source urban data, and propose spatial transformation suggestions based on the simulation results. This adaptive design method, based on urban street disaster scenario generation, aims to improve urban climate adaptability and disaster resilience. The method includes the following steps: Step 1: Collect and standardize natural disaster event data and urban street view image data, and extract the disaster occurrence probability characteristics of different disaster types in each city.

[0007] Step 2: The LoRA method is used to train the diffusion model, and combined with the ControlNet model, the depth map or edge contour information of the street corner image is introduced to improve the disaster representation ability and structural stability of the generated image.

[0008] Step 3: Based on the disaster occurrence probability characteristics and street view images, use the trained model to generate street view disaster simulation images of the city under different disaster scenarios.

[0009] Step 4: Input the disaster simulation image and related data into the large language model to generate street view space transformation suggestions, and generate the transformed street view simulation image based on the transformation suggestions.

[0010] Step 5: Use an open-vocabulary image semantic segmentation model to identify elements and perform proportion statistics on street scene images before and after the disaster and after the reconstruction. Then, use a visual big language model to evaluate the quality of the generated images. When the evaluation score is lower than the preset threshold, the parameter adjustment strategy is automatically triggered. The image is regenerated based on the adjusted parameters until the maximum number of iterations is reached or the evaluation score reaches the threshold.

[0011] Step 6: Perform semantic vector processing and cluster analysis on the modification suggestions to form standardized adaptive strategies for different disaster types.

[0012] Furthermore, step 1 specifically includes the following steps: Step 1.1: Obtain disaster event data through authoritative national channels; Step 1.2: Based on disaster event data, the frequency threshold method is used to screen floods, heavy rainfall, strong winds and hail, drought, typhoons, earthquakes, snow disasters and low temperatures as disaster types through statistical analysis of the frequency of various disasters and the affected population; Step 1.3: Select multiple cities as city samples based on the degree of disaster impact; Step 1.4: Collect street view images of the selected urban centers and disaster-prone areas through a public platform, and perform angle standardization and image clarity filtering. Image clarity filtering includes removing images containing watermarks, excessive distortion, and atypical scenes. Step 1.5: Crop the collected images to a uniform size, use bilinear interpolation algorithm to scale the images and save them in a fixed format, and establish a structured image database by labeling information such as disaster type, city of shooting and shooting angle, so as to provide standard input for model training in Step 2; Step 1.6: Based on historical disaster data, calculate the annual average probability of occurrence for each type of disaster in each city, use the maximum likelihood estimation method to fit the probability distribution of disaster occurrence, and encode the probability values ​​into a continuous disaster control vector for subsequent adjustment of the intensity of disaster feature injection during the generation of the diffusion model.

[0013] Furthermore, step 2 specifically includes the following steps: Step 2.1: Collect real disaster scene photos of each filtered disaster type from the Internet, and based on the image database's storage requirements, retain multiple standard sample images for each disaster type, and finally construct multiple subset datasets; Step 2.2: Based on the table Diffusion basic diffusion model, the LoRA method is used to fine-tune and train each subset of data to generate LoRA fine-tuned sub-models corresponding to each disaster type; Step 2.3: Introduce the ControlNet model to improve the structural stability of the generated images and the controllability of disaster features; Step 2.4: After training is completed, image generation tests are performed on the generative models for each disaster type, including using the Laplacian operator to evaluate image sharpness, manually evaluating the integrity of disaster features, and using SSIM (Structural Similarity Index) to evaluate the fidelity of street scene structure.

[0014] Furthermore, during the processing in steps 2.2 and 2.3, the learning rate, regularization parameters, and disaster feature weights are dynamically adjusted to ensure that the model highly reproduces the disaster morphology while maintaining a reasonable layout of street scene elements.

[0015] Furthermore, step 3, generating the disaster simulation image specifically includes the following steps: Step 3.1: Input a street view image of the city to be simulated, and determine the types of disasters to be generated based on the probability of occurrence of different types of disasters in the city; Step 3.2: Call the LoRA fine-tuning sub-model corresponding to the disaster type, and generate images under the control of the ControlNet control model; Step 3.3: Input the continuous disaster probability vector into the diffusion model, adjust the diffusion noise injection intensity, and realize the continuous evolution of disaster severity with probability changes; Step 3.4: Output a street view simulation image with disaster characteristics; Step 3.5: Using the CLIP-based open vocabulary image segmentation model, semantic segmentation is performed on the original street view image and the generated disaster scene image through a dual-channel processing method. The local vector comparison method is used to detect conflicts in the segmentation results of the two channels. The detection results are structured and represented as a disaster impact-urban function conflict matrix to characterize the degree of conflict faced by various urban functional elements under different disaster conditions.

[0016] Furthermore, in step 4, the generation of the reconstruction strategy includes the following steps: Step 4.1: Input the generated disaster scenario map, disaster occurrence probability data, and disaster impact-urban function conflict matrix into the large language model; Step 4.2: Utilize a large language model to assess the degree of disaster impact in the disaster scenario, identify affected street scene elements, and generate targeted spatial transformation suggestions; Step 4.3: Convert the generated modification suggestion text into a prompt word format that can be recognized by the Stable Diffusion basic diffusion model. The Stable Diffusion basic diffusion model uses the original street view image as a base, combines the prompt words, generates a simulated image of the street view after modification, and displays it. Step 4.4: Collect all renovation suggestions, summarize and extract common strategies through text embedding and cluster analysis, and classify them according to disaster type and response mechanism to provide a data foundation for subsequent evaluation of renovation effects.

[0017] Furthermore, step 5 specifically includes the following steps: Step 5.1: Select an open-vocabulary semantic segmentation model to perform pixel-level segmentation of the input image; Step 5.2: Define custom category terms for each disaster type; Step 5.3: Input the simulated street scene before, the simulated disaster scene after, and the street scene after the transformation into the open vocabulary semantic segmentation model, and perform pixel-level segmentation on each image, outputting the pixel ratio of each type of element; Step 5.4: Use the GPT-4o model to evaluate the quality of the generated image. Score it from three dimensions: visual realism, disaster feature integrity, and street view structure fidelity. Compare the scores with a preset threshold. If the score is lower than the preset threshold, the parameter adjustment mechanism is automatically triggered. The image is regenerated by adjusting the ControlNet guidance strength, sampling steps, and prompt word weights until the image evaluation score reaches the threshold or the maximum number of iterations is reached. This completes the iterative optimization loop and improves the efficiency and accuracy of image generation. Step 5.5: Compare the changes in the proportion of each main category of elements before, after, and after the disaster, and analyze the effect characteristics of each renovation measure under different disaster scenarios in combination with the disaster type, so as to provide guidance for the subsequent construction of urban spatial resilience.

[0018] Furthermore, step 6 specifically includes the following steps: Step 6.1: Organize the multimodal large language model generation and transformation strategy text in a unified manner, and use an embedding algorithm to convert the text into a high-order semantic vector; Step 6.2: Calculate the similarity between the texts of each strategy using cosine similarity, merge highly similar strategies, and remove redundant descriptions; Step 6.3: Based on different disaster triggers, the strategies are divided into meteorological response type, temperature response type, and geological response type; Step 6.4: Statistically analyze the frequency of each strategy in different cities and disaster types, select the strategies that frequently appear in most mature and disaster types, define them as high-priority strategies, and form a list of key recommended strategies; Step 6.5: Match the priority strategy with the street scene transformation simulation images under each disaster scenario to form a corresponding demonstration case library, providing a reference for urban decision-makers.

[0019] Furthermore, in step 6.3, the meteorological response type specifically refers to: for floods, heavy rainfall, typhoons, and storm events, proposing strategies such as drainage system optimization, permeable paving, green infrastructure construction, canopy management, and windbreak establishment; Temperature-responsive strategies specifically include: in response to high temperature, drought, severe cold, and snowfall events, proposing strategies such as increasing green areas, rain gardens, bioretention facilities, drought-resistant landscape design, and winter maintenance plans; The geological response type specifically refers to strategies for earthquake disasters, including foundation reinforcement, seismic retrofitting of structures, and construction of elevated walkways.

[0020] Compared with the prior art, the beneficial effects of the present invention are mainly reflected in: 1. By incorporating the probability of urban disaster occurrence into the continuous control process of the generative model, an intelligent agent collaborative closed-loop system is constructed, which integrates disaster scenario simulation, spatial transformation generation, and effect evaluation. This enables continuous disaster evolution simulation and quantitative transformation optimization of urban streetscapes under different disaster risk levels.

[0021] 2. This invention breaks through the traditional generation method based on discrete disaster type selection, and introduces disaster-spatial conflict detection and multi-agent automatic feedback mechanism, so that the generated results not only have visual realism, but also have quantifiable evaluation and automatic optimization capabilities, providing technical support for urban resilience planning and climate adaptability design.

[0022] 3. The images generated by the method of the present invention can accurately reflect the characteristics of various disaster types and are applicable to different urban environments in multiple locations.

[0023] 4. The method of this invention, through street scene simulation and transformation suggestions, makes urban disaster adaptation strategies specific to the pedestrian scale, thereby improving the operability and effectiveness of implementation.

[0024] 5. The visualized disaster simulation and transformation effect diagrams in this invention facilitate public understanding and participation in urban disaster resilience construction, and help improve the transparency and scientific nature of policy making.

[0025] 6. The standardized reference system summarized in this invention facilitates different cities to formulate differentiated and refined transformation plans based on disaster characteristics. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the adaptive design method based on urban street disaster scenarios according to the present invention.

[0027] Figure 2 This is a detailed diagram of the dual-channel semantic segmentation and conflict detection technology in the method of this invention.

[0028] Figure 3 This is a diagram showing the mapping and effect evolution of continuous control parameters driven by disaster probability in the method of this invention.

[0029] Figure 4 This is a flowchart of the closed-loop process for automatic evaluation and optimization of multiple agents in the method of this invention. Detailed Implementation

[0030] The following will describe in more detail an adaptive design method based on urban street disaster scenario generation according to the present invention with reference to the schematic diagram, which illustrates the preferred embodiment of the present invention. It should be understood that those skilled in the art can modify the present invention described herein while still achieving the advantageous effects of the present invention. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the present invention.

[0031] like Figure 1 As shown, an adaptive design method based on urban street disaster scenarios includes the following steps: Step 1: Multi-source data acquisition and processing.

[0032] This invention first involves the acquisition and preprocessing of multi-source data, specifically including: (1) Disaster event data collection.

[0033] Data on natural disasters from 2014 to 2020 were collected from authoritative sources such as the National Disaster Reduction Center's data platform and the China Emergency Information Network, totaling 3,650 event records covering 27 types of disasters, including floods, typhoons, droughts, and earthquakes.

[0034] (2) Disaster type screening.

[0035] Through statistical analysis of the frequency of various disasters and the affected population, the frequency threshold method was used for screening. The screening criteria were set as ≥5 times per year and ≥10,000 people affected per year. Finally, eight major disasters were selected as research objects: floods, heavy rainfall, strong winds and hail, drought, typhoons, earthquakes, snow disasters, and low temperatures.

[0036] (3) Screening of city samples.

[0037] Using the degree of disaster impact (number of affected people) as the main indicator and combining the representativeness of geographical regions, eight cities, namely Shiyan, Wenzhou, Baiyin, Taizhou, Qujing, Huangshi, Yuxi, and Chaoyang, were selected as experimental areas.

[0038] (4) Street view image acquisition.

[0039] In eight selected cities, street view images were collected at 50-meter intervals in the central urban areas and disaster-prone areas of each city using the Baidu Street View API. The collection parameters were set as follows: pitch angle 0° (horizontal view), horizontal view range 0°-360° (one image collected every 90°), and image resolution 1024×512 pixels. After collection, angle standardization was performed (images with pitch angles between -10° and +10° were selected), image sharpness was filtered (the Laplacian operator was used to calculate the image sharpness score, and images with a score ≥100 were retained), and images containing watermarks, excessive distortion, or atypical scenes (such as indoor scenes or highway scenes) were removed.

[0040] (5) Data standardization processing.

[0041] All acquired images were uniformly cropped to 512×512 pixels, and then scaled using a bilinear interpolation algorithm before being saved as PNG format. A structured image database was established by labeling information such as disaster type, city of origin, shooting angle, and acquisition time, providing standard input for subsequent model training.

[0042] (6) Extraction of disaster probability features.

[0043] Based on historical disaster data, the average annual probability of each disaster type in each city is calculated. The maximum likelihood estimation method is used to fit the probability distribution of disaster occurrence. The probability values ​​are encoded into a continuous disaster control vector with a dimension of 8, which serves as an important reference feature for subsequent disaster scenario generation and transformation strategy inference.

[0044] Step 2: Generative AI model training (combination of LoRA and ControlNet).

[0045] To generate realistic urban street scene images with disaster characteristics, this invention employs a model training strategy that combines the LoRA (Low-Rank Adaptation) fine-tuning method with the ControlNet control network: (1) Construction of disaster image dataset.

[0046] We collected real disaster scene photos from 2010 to 2024 from publicly available internet resources (Baidu Images, disaster archive databases) and government-published disaster records, covering eight types of disasters including floods, droughts, typhoons, and earthquakes, with a total of 1,620 images collected in the original collection.

[0047] After filtering out images with abnormal angles (absolute pitch angle > 30°), unbalanced subject proportions (disaster subject proportion < 20% or > 80%), or watermarks, 100 standard sample images were retained for each disaster type. All images were uniformly cropped to 512×512 pixels and stored in PNG format, ultimately constructing 8 subsets, each containing 100 labeled images.

[0048] (2) LoRA model fine-tuning training.

[0049] On the Stable Diffusion v1.5 base diffusion model, Low-Rank Adaptation (LoRA) was used to fine-tune the training for each disaster type individually. Specifically, the weights of the U-Net backbone network (approximately 850 million parameters) of the base diffusion model were frozen, and low-rank factorization matrices were introduced only in the cross-attention layers for parameter updates. The LoRA rank was set to 64, the alpha parameter to 32, the dropout rate to 0.1, and a cosine annealing learning rate was adopted with an initial learning rate of 1. e-4 The minimum learning rate is set to 1. e-6 The training batch size was 4, and the training steps were 1,500. Through mini-batch iterative training, the specific expression of disaster features was optimized, and after training, eight LoRA fine-tuned sub-models corresponding to disaster types were generated.

[0050] (3) Integration of ControlNet control model.

[0051] To further improve the structural stability of the generated images and the controllability of disaster features, this invention introduces the ControlNet auxiliary control mechanism.

[0052] Based on the original street view image, the MiDaS depth estimation model is used to generate the corresponding depth map with a resolution of 512×512 pixels and a depth value range of 0-255; the Canny edge detection algorithm (low threshold 100, high threshold 200) is used to generate the edge contour map.

[0053] During the generation process, the ControlNet module takes depth or contour information as input and guides the diffusion process, so that the generated result retains the original street scene spatial structure while superimposing disaster features. The control weights of ControlNet are set to an adjustable range of 0.8-1.2, which are dynamically adjusted according to the disaster type: for structural disasters (such as earthquakes), depth map control is used with a weight of 1.0; for surface disasters (such as floods), edge contour map control is used with a weight of 0.8.

[0054] (4) Model parameter optimization.

[0055] During the LoRA fine-tuning and ControlNet integration process, the learning rate, regularization parameters, and disaster feature weights are dynamically adjusted. An early stopping strategy is adopted, which terminates training when the validation set loss does not decrease for 50 consecutive steps, ensuring that the model can both highly reproduce the disaster morphology and maintain a reasonable layout of street scene elements.

[0056] (5) Generate capability verification.

[0057] After training, the model's generation performance on street view images of various cities was tested through image generation tests under different disaster scenarios. The main evaluation metrics included image sharpness (scored using the Laplacian operator), disaster feature completeness (human evaluation of 1-10 points), and street view structure fidelity (SSIM structural similarity index).

[0058] Step 3: Disaster scenario simulation and reconstruction strategy generation.

[0059] After completing the generative model training, this invention enters the disaster scenario simulation and transformation strategy generation stage. The core innovation lies in introducing a dual-channel semantic segmentation and conflict detection mechanism for disaster street scenes to improve the spatial targeting of transformation strategies. The specific steps are as follows: (1) Disaster scenario simulation.

[0060] 1) Input a street view image of the city to be simulated.

[0061] 2) Determine the types of disasters to be generated (such as floods, droughts, earthquakes, etc.) based on the probability of occurrence of different types of disasters in the city.

[0062] 3) Call the LoRA fine-tuning sub-model corresponding to the disaster type and generate images under the control of ControlNet. Set the sampling step to 50 steps and the classifier free guided scale (CFG Scale) to 7.5.

[0063] 4) Reference Figure 3 In the disaster scenario simulation process, this invention inputs the urban disaster probability vector into the diffusion model control module as a continuous control parameter to participate in the generation process. This control vector is used to synchronously adjust the diffusion noise injection intensity (by adjusting the noise scheduling parameters of the DDPM sampler), the disaster LoRA weight (range 0.6-1.0), and the ControlNet guidance intensity (range 0.8-1.2), thereby realizing the continuous evolution of disaster severity with probability changes.

[0064] 5) Output street scene simulation images with obvious disaster characteristics (such as water accumulation, fractures, collapses, etc.).

[0065] (2) Disaster impact - detection of urban functional space conflict.

[0066] refer to Figure 2 This invention innovatively establishes a "dual-channel semantic segmentation + conflict detection" step for near-disaster images, which is a key technology that distinguishes it from traditional one-way semantic segmentation methods.

[0067] Specifically, semantic segmentation (dual-channel processing) is first performed on the original street view image and the generated disaster scene image. The Open Vocabulary Image Segmentation (OV-SEG) model based on CLIP is adopted, and classification labels are set including eight main categories and 32 subcategories, namely road infrastructure, road surface, vegetation, building structure, street furniture, traffic participants, vehicles, and background.

[0068] For the street view image channel, the spatial distribution characteristics of various urban functional elements are extracted, including indicators such as green coverage rate, sidewalk ratio, and traffic facility density; for the disaster image channel, the spatial distribution of disaster-affected areas is identified, such as flood-inundated areas, earthquake crack locations, and trees felled by typhoons.

[0069] Then, the "local vector comparison method" is used to check for conflicts in the segmentation results of the two channels: the semantic segmentation results of the street view image and the disaster image are represented as pixel-level category matrices S and D, respectively, with a dimension of 512×512 and each element value of 0-7 (corresponding to 8 main categories). The two images are overlaid, and a category histogram vector is calculated in each local window of 16×16 pixels. The vector for the street view window is Vs, and the vector for the disaster window is Vd. The cosine similarity between the two vectors is calculated as cos(θ)=(Vs·Vd) / (||Vs||×||Vd||). When the similarity is lower than the threshold of 0.7, it is determined that there is a disaster-spatial conflict in the local area. A conflict heatmap is generated by traversing the entire image through a sliding window (with a stride of 8 pixels), marking the coordinates and conflict type of all conflict locations.

[0070] The above conflict detection results are structured as a disaster impact-urban function conflict matrix. The matrix has a dimension of 8×8 (8 main categories), and the element values ​​represent the conflict intensity between the corresponding categories (range of 0-1), which is used to characterize the degree of conflict faced by various urban functional elements under different disaster conditions.

[0071] (3) Reconstruction strategy generation.

[0072] 1) Input the generated disaster scene images along with the disaster occurrence probability data into a multimodal large language model (such as ChatGPT-4o).

[0073] 2) Utilize large language models to conduct comprehensive analysis of disaster scenarios, including: disaster impact assessment (calculating the overall conflict index based on the conflict matrix), identification of affected street scene elements (extracting categories with conflict intensity > 0.5 from the conflict matrix), and preliminary suggestions for targeted spatial transformation (such as adding drainage systems, expanding green spaces, building elevated walkways, and renovating earthquake-resistant buildings).

[0074] (4) Visualize the renovation suggestions.

[0075] 1) Convert the generated modification suggestion text into a prompt format that can be recognized by Stable Diffusion, and generate structured prompts using keyword extraction and semantic expansion methods.

[0076] 2) Based on the original street view image, combined with prompt words, generate a simulated image of the street view after modification in the Stable Diffusion model.

[0077] 3) The generated images visually demonstrate the effects of improving streetscape resilience through modification measures after a disaster occurs.

[0078] (5) Strategy summary.

[0079] 1) Collect renovation suggestions generated for all cities and under all disaster types.

[0080] 2) By using text embedding and cluster analysis, common strategies are summarized and extracted to avoid duplication.

[0081] 3) Classify and organize data according to disaster type and response mechanism to provide a data foundation for subsequent strategy evaluation and application.

[0082] Step 4: Image semantic segmentation and performance evaluation.

[0083] To perform quantitative analysis on street scene disaster simulation images and modified images, this invention employs an open-vocabulary image semantic segmentation method and innovatively introduces an automated evaluation and optimization mechanism for the generated results, realizing the transformation from manual selection to automatic iterative optimization. The specific steps are as follows: (1) Model selection.

[0084] 1) An Open Vocabulary Semantic Segmentation Model (OV-SEG) based on the CLIP (Contrastive Language-Image Pre-training) model is adopted. This model can perform pixel-level segmentation of input images based on a custom category vocabulary, without being restricted to fixed categories in advance, and has good scalability and adaptability.

[0085] (2) Setting category labels.

[0086] To meet the needs of analyzing street scene elements in urban disaster scenarios, this invention divides the elements in street scene images into eight main categories and their subcategories, as shown in Table 1.

[0087] Table 1

[0088] (3) Image processing flow.

[0089] 1) Input the three sets of images—before simulation (original street view), after simulation (disaster scenario), and after modification (street view with improved resilience)—into the OV-SEG model respectively.

[0090] 2) Perform pixel-level semantic segmentation on each image and output the pixel ratio of each category of elements.

[0091] (4) Construct an iterative optimization closed loop refer to Figure 4 The core innovation of this invention lies in the introduction of an "automatic evaluation and optimization module for generated results," which constructs a closed-loop optimization mechanism based on AI-based automatic scoring.

[0092] Specifically, the first step is to define a multi-dimensional evaluation indicator system: The first is the semantic segmentation quantification index, including green coverage rate (the proportion of vegetation pixels), sidewalk ratio (the proportion of sidewalk pixels to road surface pixels), road infrastructure completeness (the proportion of infrastructure pixels), and protective facility density (the proportion of guardrail and fence pixels). Each index is calculated by dividing the number of pixels in the corresponding category by the total number of pixels. Second, a visual large language model scoring system is used. The GPT-4o model is employed to evaluate the quality of the generated images, scoring them across three dimensions: visual realism (1-10 points), disaster feature integrity (1-10 points), and street view structure fidelity (1-10 points). The overall score is the average of these three scores. Then, an automatic optimization process is established: when the overall score of the generated image falls below a preset threshold (set to 7.0 points) or any key indicator fails to reach the target range, the system automatically triggers a parameter adjustment mechanism. Parameter adjustment strategies include: adjusting the ControlNet guidance strength (±0.1 steps), adjusting the sampling steps (±5 steps), and modifying the cue word weights (adding weight labels to key descriptive words).

[0093] The adjusted parameters are re-inputted into the generative model to generate a new version of the image, which is then evaluated again, forming an iterative optimization closed loop.

[0094] The iteration process is set to a maximum of 5 iterations, terminating when the score reaches a threshold or the maximum number of iterations is reached. This automated evaluation and optimization mechanism achieves ideal generation results with almost no manual intervention, significantly improving generation efficiency and accuracy while reducing manual screening workload.

[0095] (4) Effectiveness evaluation indicators.

[0096] 1) Statistical analysis of changes in key indicators, such as: changes in vegetation coverage; changes in sidewalk proportion; changes in the proportion of road infrastructure; and changes in the addition or subtraction of safety facilities such as guardrails; 2) Analyze the specific effects of the renovation measures on improving the disaster resistance and livability of the streetscape.

[0097] (5) Quantitative comparative analysis.

[0098] 1) Compare the changes in the proportions of each main category of elements before, after, and after the disaster; 2) In conjunction with the type of disaster, analyze the effect characteristics of each transformation measure under different disaster scenarios to guide the subsequent construction of urban spatial resilience.

[0099] Step 5: Summarize and classify multi-scenario adaptive strategies.

[0100] Based on the streetscape renovation suggestions output by the generative model, this invention further summarizes and classifies the renovation strategies so that different cities can select the most suitable resilience enhancement measures according to their disaster characteristics. The specific steps are as follows: (1) Strategy text organization and vectorization processing.

[0101] 1) Organize and standardize the various street view modification strategy texts generated by ChatGPT-4o.

[0102] 2) An embedding algorithm is used to convert the text into a high-dimensional semantic vector, which facilitates subsequent clustering analysis.

[0103] (2) Merging and deduplication of similar strategies 1) Calculate the similarity between strategy texts using cosine similarity.

[0104] 2) Strategies with high similarity (similarity threshold ≥ 0.85) were merged, redundant descriptions were removed, and 82 standardized modification strategies were formed.

[0105] (3) Classification of scene response types.

[0106] Based on different disaster triggers, the strategies are divided into three main response types: 1) Weather-oriented.

[0107] In response to events such as floods, heavy rainfall, typhoons, and storms, strategies such as drainage system optimization, permeable paving, green infrastructure construction, canopy management, and the establishment of windbreaks are proposed.

[0108] 2) Temperature-oriented.

[0109] In response to disasters such as high temperatures, drought, severe cold, and snowfall, strategies are proposed such as increasing green areas, rain gardens, bioretention facilities, drought-resistant landscape design (Xeriscaping), and winter maintenance plans.

[0110] 3) Geological-oriented.

[0111] In response to geological disasters such as earthquakes, strategies such as foundation reinforcement, seismic retrofitting of structures, and construction of elevated walkways are proposed.

[0112] (4) Strategy frequency statistics and key strategy extraction.

[0113] 1) Statistical analysis of the frequency of occurrence of each strategy in different cities and different disaster types.

[0114] 2) Select high-priority strategies that frequently appear in most cities and disaster types to form a list of 22 key recommended strategies.

[0115] (5) Strategy application demonstration.

[0116] By mapping key strategies to simulated streetscape transformation images under various disaster scenarios, an intuitive "strategy-scenario-effect" correspondence demonstration case library is formed, providing a reference for urban decision-makers.

[0117] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. An adaptive design method based on urban street disaster scenarios, characterized in that, Includes the following steps: S1: Collect and standardize natural disaster event data and urban street view image data, and extract the probability characteristics of different disaster types in each city; S2: The LoRA method is used to train the diffusion model, and combined with the ControlNet model, the depth map or edge contour information of the street corner image is introduced to improve the disaster representation ability and structural stability of the generated image. S3: Based on the disaster occurrence probability characteristics and street view images, use the trained model to generate street view disaster simulation images of the city under different disaster scenarios; S4: Input the disaster simulation image and related data into the large language model to generate street view space transformation suggestions, and generate the transformed street view simulation image based on the transformation suggestions; S5: Use an open-vocabulary image semantic segmentation model to identify elements and perform proportion statistics on street scene images before and after disasters and after reconstruction. Use a visual big language model to evaluate the quality of the generated images. When the evaluation score is lower than the preset threshold, the parameter adjustment strategy is automatically triggered. The image is regenerated based on the adjusted parameters until the maximum number of iterations is reached or the evaluation score reaches the threshold. S6: Perform semantic vector processing and cluster analysis on the modification suggestions to form standardized adaptive strategies for different disaster types.

2. The adaptive design method based on urban street disaster scenario generation according to claim 1, characterized in that, S1 specifically includes the following steps: S11: Obtain disaster event data through authoritative national channels; S12: Based on disaster event data, the frequency threshold method is used to screen floods, heavy rainfall, strong winds and hail, drought, typhoons, earthquakes, snow disasters and low temperatures as disaster types through statistical analysis of the frequency of various disasters and the affected population. S13: Select multiple cities as city samples based on the degree of disaster impact; S14: For the selected urban centers and disaster-prone areas, street view images are collected through public platforms, and angle standardization and image clarity screening are performed. Images containing watermarks, excessive distortion, and atypical scenes are removed. S15: The collected images are cropped to a uniform size, and the images are scaled using a bilinear interpolation algorithm and saved in a fixed format. By labeling information such as disaster type, city of shooting, and shooting angle, a structured image database is established to provide standard input for model training in S2. S16: Based on historical disaster data, calculate the annual average probability of occurrence for each type of disaster in each city, use the maximum likelihood estimation method to fit the probability distribution of disaster occurrence, and encode the probability values ​​into a continuous disaster control vector for subsequent adjustment of the intensity of disaster feature injection during the generation of the diffusion model.

3. The adaptive design method based on urban street disaster scenario generation according to claim 2, characterized in that, S2 specifically includes the following steps: S21: Collect real disaster scene photos of each filtered disaster type, and based on the image database's storage requirements, retain multiple standard sample images for each disaster type, and finally construct multiple subset datasets; S22: Based on the Stable Diffusion v1.5 basic diffusion model, the LoRA method is used to fine-tune and train each subset of data to generate LoRA fine-tuned sub-models corresponding to each disaster type; S23: Introduce the ControlNet control model to improve the structural stability of generated images and the controllability of disaster features; S24: After training is completed, image generation tests are performed on the generative models for each disaster type, including using the Laplacian operator to evaluate image sharpness, manually evaluating the integrity of disaster features, and using the structural similarity index to evaluate the fidelity of street scene structures.

4. The adaptive design method based on urban street disaster scenario generation according to claim 3, characterized in that, During the processing of S22 and S23, the learning rate, regularization parameters and disaster feature weights are dynamically adjusted, and an early stopping strategy is adopted to ensure that the model highly restores the disaster morphology while maintaining a reasonable layout of street scene elements.

5. The adaptive design method based on urban street disaster scenario generation according to claim 1, characterized in that, In step S3, generating the disaster simulation image specifically includes the following steps: S31: Input a street view image of the city to be simulated, and determine the types of disasters to be generated based on the probability of occurrence of different types of disasters in the city; S32: Call the LoRA fine-tuning sub-model corresponding to the disaster type, and generate images under the control of the ControlNet control model; S33: Input the continuous disaster probability vector into the diffusion model, adjust the intensity of diffusion noise injection, and realize the continuous evolution of disaster severity with probability changes; S34: Output a street scene simulation image with disaster features; S35: An open vocabulary image segmentation model based on CLIP is adopted. The original street scene image and the generated disaster scene image are semantically segmented by a dual-channel processing method. The local vector comparison method is used to detect the conflict between the two channel segmentation results. The detection results are structured and represented as a disaster impact-urban function conflict matrix to characterize the degree of conflict faced by various urban functional elements under different disaster conditions.

6. The adaptive design method based on urban street disaster scenario generation according to claim 1, characterized in that, In step S4, the generation of the reconstruction strategy includes the following steps: S41: Input the generated disaster scenario map, disaster occurrence probability data, and disaster impact-urban function conflict matrix into the large language model; S42: Utilize large language models to assess the degree of disaster impact in disaster scenarios, identify affected street scene elements, and generate targeted spatial transformation suggestions; S43: Convert the generated modification suggestion text content into a prompt word format that can be recognized by the Stable Diffusion basic diffusion model. The Stable Diffusion basic diffusion model uses the original street view image as a base, combines the prompt words, generates a simulated image of the street view after modification, and displays it. S44: Collect all renovation suggestions, summarize and extract common strategies through text embedding and cluster analysis, and classify them according to disaster type and response mechanism to provide a data foundation for subsequent evaluation of renovation effects.

7. The adaptive design method based on urban street disaster scenario generation according to claim 1, characterized in that, S5 specifically includes the following steps: S51: Select an open-vocabulary semantic segmentation model to perform pixel-level segmentation of the input image; S52: Customize category terms for each disaster type; S53: Input the simulated street scene before, the simulated disaster scene after, and the street scene after the transformation into the open vocabulary semantic segmentation model, and perform pixel-level segmentation on each image, outputting the pixel ratio of each type of element; S54: The GPT-4o model is used to evaluate the quality of the generated images. The images are scored from three dimensions: visual realism, disaster feature integrity, and street view structure fidelity. The scores are compared with preset thresholds. If the score is lower than the preset threshold, the parameter adjustment mechanism is automatically triggered. The image is regenerated by adjusting the ControlNet guidance strength, sampling steps, and prompt word weights until the image evaluation score reaches the threshold or the maximum number of iterations is reached. This completes the iterative optimization loop and improves the efficiency and accuracy of image generation. S55: Compare the changes in the proportions of each main category of elements before, after, and after the disaster, and analyze the effectiveness characteristics of each renovation measure under different disaster scenarios in conjunction with the disaster type, so as to provide guidance for the subsequent construction of urban spatial resilience.

8. The adaptive design method based on urban street disaster scenario generation according to claim 1, characterized in that, S6 specifically includes the following steps: S61: Unify and organize the text generated by the multimodal large language model transformation strategy, and use an embedding algorithm to convert the text into a high-order semantic vector; S62: Calculate the similarity between the texts of each strategy using cosine similarity, and merge highly similar strategies and remove redundant descriptions; S63: Based on different disaster triggers, strategies are divided into meteorological response type, temperature response type, and geological response type; S64: Statistically analyze the frequency of each strategy in different cities and disaster types, filter out the strategies that frequently appear in most mature and disaster types, define them as high-priority strategies, and form a list of key recommended strategies. S55: Match this priority strategy with streetscape transformation simulation images under various disaster scenarios to form a corresponding demonstration case library, providing a reference for urban decision-makers.

9. The adaptive design method based on urban street disaster scenario generation according to claim 1, characterized in that, In S63, the meteorological response type specifically refers to: proposing strategies for drainage system optimization, permeable paving, green infrastructure construction, canopy management, and windbreak establishment in response to floods, heavy rainfall, typhoons, and storm events; The temperature-responsive type specifically refers to: strategies for increasing green areas, rain gardens, bioretention facilities, drought-resistant landscape design, and winter maintenance plans in response to high temperature, drought, severe cold, and snowfall events; The geological response type specifically refers to strategies for foundation reinforcement, seismic retrofitting of structures, and construction of elevated walkways in response to earthquake disasters.

Images