Outdoor thermal comfort evaluation system and method based on AI multi-source fusion and visual compensation

Abstract

The invention discloses an outdoor thermal comfort evaluation system and method based on AI multi-source fusion and visual compensation. The method provided by the invention comprises the following steps: 1, synchronously acquiring on-site microclimate physical parameters, GPS space-time coordinates and a three-dimensional streetscape image of a first person view angle; 2, obtaining pedestrian subjective perception data; thirdly, pixel-level analysis is carried out on the streetscape image, and view indexes of elements are quantized to serve as key features of a visual compensation mechanism; 4, constructing a time-space aligned multi-modal comprehensive data set, and constructing a high-precision thermal comfort prediction model; and 5, introducing an SHAP framework to deeply explain the model, and quantifying and revealing the independent and interactive influence of microclimate physical parameters and visual environment elements on the subjective thermal comfort. According to the method, a refined and explainable scientific decision-making tool is provided for people-oriented urban planning and climate adaptability design by quantifying the visual cooling effect.

Description

Technical Field

[0001] This invention lies at the intersection of urban planning, environmental science, human settlements, and artificial intelligence. Specifically, it relates to an innovative system and method for accurately assessing and predicting subjective thermal comfort in complex urban outdoor spaces. This technical solution integrates on-site microclimate physical measurements, subjective perception surveys, and visual environment analysis based on streetscape imagery to focus on studying and quantifying the role of visual compensation mechanisms in regulating thermal perception. Background Art

[0002] Under the dual pressures of global climate change and rapid urbanization, the urban heat island effect is increasingly exacerbating, severely impacting the environmental quality of urban outdoor spaces and the health and well-being of residents. Therefore, scientifically and accurately assessing the thermal comfort of pedestrians in urban open spaces has become a critical component in achieving healthy cities and sustainable development goals. Current thermal comfort research methods, such as physical parameter-based indices (e.g., UTCI and PET), while theoretically integrating multiple meteorological factors, exhibit significant limitations in practical applications. First, these models are inherently "physical deterministic," assuming that human thermal perception is entirely determined by four physical parameters: temperature, humidity, wind speed, and radiation. This completely ignores the complex psychological and physiological differences of humans as perceptual subjects. Research and practice have repeatedly demonstrated significant discrepancies between comfort predicted by physical indices alone and actual subjective perceptions (TCVs), with weak correlations. This suggests the existence of a significant "perception gap," a nonlinear and non-unique mapping between the physical environment and psychological perceptions. Second, existing research fails to effectively incorporate and quantify visual factors, which play a crucial role in modulating thermal perception. Urban blue-green spaces (vegetation and water bodies) not only reduce temperatures through physical processes (shading and evapotranspiration), but their beautiful landscapes themselves can also create a "cooling sensation." This positive visual perception can effectively "compensate" or alleviate some of the physiological discomfort caused by high temperatures, known as the "visual compensation effect." However, traditional methods lack effective means to accurately and quantitatively describe these visual elements from a pedestrian's first-person perspective, leading to the long-overlooked nature of this key mechanism in models. Furthermore, with the development of artificial intelligence technology, while some studies have begun to attempt to use machine learning methods to predict thermal comfort, most still suffer from the following problems: Data sources remain limited, failing to effectively integrate objective physical data, subjective perception data, and refined visual environment data; and failing to provide designers with clear, quantitative optimization guidance. Therefore, a new thermal comfort evaluation method and system that can integrate multi-source data and quantify the visual compensation effect is urgently needed. Summary of the Invention

[0003] The core purpose of this invention is to overcome the aforementioned shortcomings of the existing technology and provide a technologically advanced, data-driven, and highly interpretable outdoor thermal comfort evaluation system and method based on artificial intelligence (AI) multi-source data fusion and visual compensation mechanisms. By deeply integrating three types of heterogeneous data—the objective physical environment, subjective psychological perception, and refined visual landscape—and combining them with cutting-edge AI prediction and interpretation technologies, this system aims not only to achieve high-precision predictions of outdoor thermal comfort but also to scientifically quantify and reveal the role of "visual compensation," a key psychological mechanism, providing intuitive and quantitative decision-making support for creating healthier and more pleasant living environments.

[0004] The present invention provides the following technical solutions: Outdoor thermal comfort evaluation system based on AI multi-source fusion and visual compensation, including: Multi-source data synchronous acquisition module, used to collect on-site microclimate data, GPS spatiotemporal coordinate data, first-person perspective street view image data, and pedestrian subjective thermal comfort voting TCV data; The intelligent extraction module of visual environmental elements uses a deep learning semantic segmentation model to perform pixel-level analysis on the first-person perspective street view image data, automatically identify and quantify each visual element, and calculate the corresponding visual field index; The multimodal data integration and feature engineering module cleans, aligns, fuses, and extracts multi-scale spatial features from the on-site microclimate data, visual element data, and pedestrian subjective thermal comfort voting data based on a unified spatiotemporal benchmark to construct a comprehensive geographic information system (GIS) training dataset. The thermal comfort prediction and interpretability analysis module builds and trains a machine learning prediction model based on the GIS training dataset to predict subjective thermal comfort, and uses the explainable artificial intelligence (XAI) method to analyze the model and quantify the contribution and influence mechanism of each factor.

[0005] The outdoor thermal comfort evaluation method based on AI multi-source fusion and visual compensation includes the following steps: Step S1: On-site synchronous collection of multi-source data: obtaining microclimate physical parameters, GPS spatiotemporal coordinates, first-person perspective street view images, and pedestrian subjective thermal comfort voting (TCV) data; Step S2: Intelligent extraction of visual environment elements: Based on the first-person perspective street view image, deep learning semantic segmentation technology is used to identify and quantify environmental visual elements from the pedestrian's perspective, and their visual field index is calculated; Step S3: Multimodal data integration and feature construction: The multi-source data obtained in step S1 and the visual feature data extracted in step S2 are matched and integrated based on unified spatiotemporal coordinates to construct a comprehensive GIS dataset containing multi-scale environmental features. Step S4: Integrate and align the multimodal data to form a unified GIS dataset; and use machine learning to construct an outdoor thermal comfort prediction model to reflect the nonlinear relationship between microclimate, visual elements and subjective thermal perception, and evaluate the visual compensation effect.

[0006] Furthermore, in step S1, dynamic microclimate measurement is performed through a portable micro-weather station that integrates a high-precision temperature sensor, a capacitive humidity sensor, a three-dimensional ultrasonic wind speed and direction sensor, and a black globe thermometer; a differential GPS or real-time dynamic RTK positioning module is used to obtain sub-meter precision spatiotemporal coordinates; and a wide-angle motion camera is used to synchronously record street view videos.

[0007] Furthermore, in step S2, a pre-trained deep learning semantic segmentation model is used to perform pixel-level classification on single-frame images extracted from street view videos to identify corresponding environmental elements; and the field of view index is calculated based on pixel statistics of the segmentation results.

[0008] Furthermore, in step S3, the discrete on-site measurement point data are associated with the view index extracted from the corresponding street view image using a unified timestamp and GPS coordinates; and in the geographic information system (GIS) software, a multi-level buffer zone is established for each measurement point to extract environmental characteristic variables at different spatial scales, thereby constructing a multi-attribute, multi-scale comprehensive GIS dataset.

[0009] Furthermore, in step S4, an extreme gradient boosting (XGBoost) algorithm or a random forest (RF) algorithm is used to construct a prediction model. The input features of the prediction model include physical environment parameters, visual field index, and personal parameters. The output target is the thermal comfort voting (TCV) level of the pedestrian. The optimization objective function of the XGBoost algorithm is: ; in, For the The output of a decision tree, is a regularization term containing L1 and L2 norms, which is used to control the complexity of the model to prevent overfitting; i represents the i-th sample, n is the total number of samples, is the true value of the i-th sample, and It represents the cumulative prediction results of the first t-1 decision trees for the sample Is the loss function used to measure the and The predicted value and the true value together The error between is a regularization term that aims to control the model complexity of the t-th tree to prevent overfitting. The L1 norm promotes sparsity by penalizing the absolute value and sum of parameters, while the L2 norm shrinks parameters by penalizing the sum of squares of parameters, thereby smoothing the model and improving generalization ability.

[0010] An electronic device comprises: a memory and a graphics processing unit (GPU); the memory is used to store computer-executable instructions, and the processor and GPU are used to execute the computer-executable instructions to implement the method according to any one of claims 2 to 6.

[0011] A computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, can implement the method according to any one of claims 2 to 6.

[0012] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following significant and groundbreaking beneficial effects: (1) The visual compensation mechanism was scientifically quantified and confirmed for the first time: Through innovative data fusion and model construction, this paper successfully quantified and verified the concept of "visual compensation" that has long existed in psychology and landscape aesthetics in urban thermal environment research; it revealed that blue-green landscapes with high GVF and WVF can significantly improve people's subjective thermal comfort in hot environments, providing a new gain path for urban planning that goes beyond pure physical cooling.

[0013] (2) Construction of a comprehensive "physical-psychological-visual" data set: This invention breaks through the data barriers of traditional research and constructs a comprehensive spatiotemporal data set that includes three dimensions of information: objective physical measurement, subjective psychological voting, and first-person perspective quantitative vision through high-precision synchronous acquisition. Based on the depth and breadth of this data set, it greatly improves the comprehensive characterization capability of complex urban living environments and provides a data foundation for thermal comfort research.

[0014] (3) Realizes intelligent and refined extraction of environmental elements from the pedestrian's perspective: The present invention adopts deep learning semantic segmentation technologies such as PSPNet to realize the automatic, high-precision, pixel-level recognition and quantification of environmental elements in street scene images; the series of visual indicators such as GVF and WVF calculated in this way are based on the real perception perspective of pedestrians and finely depict the micro-environment. Its accuracy, efficiency and human-scale attributes far exceed those of traditional remote sensing or manual mapping methods.

[0015] (4) The accuracy and generalization ability of the prediction model are significantly improved: Based on ensemble learning algorithms such as XGBoost and the rich features of multi-source fusion, the prediction model constructed by the present invention can capture the complex nonlinear relationship between physical, visual and personal factors and subjective thermal comfort; compared with traditional models that only rely on physical parameters (such as UTCI), this model has significantly improved the prediction accuracy of subjective TCV, and has stronger generalization ability, and the results are closer to real-world human perception.

[0016] (5) The model is interpretable: This paper introduces the SHAP framework, which can clearly reveal whether each environmental factor (such as temperature, wind speed, green view rate, and water view) contributes positively or negatively to thermal comfort, as well as the magnitude of the contribution; quantitative attribution analysis can provide urban designers with clear optimization directions.

[0017] (6) Driving the transformation of urban planning to “perception-friendly”: Based on the system and method of the present invention, designers can simulate and evaluate the thermal comfort of different design schemes. During the planning stage, they can optimize the layout, form, and proportion of blue and green spaces, build a “visually cool” urban microenvironment, and promote the transformation of urban planning from a simple functional layout to a “perception-friendly” city that focuses on multi-sensory experience, ultimately significantly improving the quality of outdoor life of residents. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the overall method of the outdoor thermal comfort evaluation method based on artificial intelligence multi-source data fusion and visual compensation mechanism according to the present invention; Figure 2 This is a schematic diagram of the process and results of using the deep learning semantic segmentation network (PSPNet) to extract visual environment elements from street scene images in this invention, showing the conversion from original images to pixel-level classification maps; (a) is a partial result of PSPNet street scene semantic segmentation, and (b) is a PSPNet street scene semantic segmentation example; Figure 3 is the relationship diagram between UTCI and thermal comfort; Figure 4 is a comparison diagram of UTCI and thermal comfort distribution over time; Figure 5 Comparison of the temporal distribution of MUTCI and thermal comfort; Figure 6 This is the SHAP summary diagram of the thermal comfort prediction model, which is used to show the global importance ranking and influence direction of each feature; Figure 7 The histogram of the thermal comfort prediction model is used to reveal the nonlinear relationship and threshold effect between it and thermal comfort; Figure 8Schematic diagram of ROC curve of thermal comfort prediction model; Figure 9 Schematic diagram of the ROC curve for the validation data obtained using Sanbao as the validation site. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0020] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.

[0021] The outdoor thermal comfort evaluation system with multi-source data fusion and visual compensation mechanism includes the following modules: A multi-source data synchronization acquisition module is configured to simultaneously collect on-site microclimate data, high-precision GPS spatiotemporal coordinate data, first-person perspective street view image data, and pedestrian subjective thermal comfort voting (TCV) data through a portable mobile weather station, street view camera equipment, and an electronic questionnaire terminal; a visual environment element intelligent extraction module configured to utilize a deep learning semantic segmentation model to perform pixel-level analysis on the street view image data, automatically identify and quantify visual elements such as the sky, vegetation, water, and buildings, and calculate corresponding visual field indices (SVF, GVF, WVF, BVF); a multimodal data integration and feature engineering module configured to clean, align, fuse, and extract multi-scale spatial features from the microclimate data, visual element data, and subjective voting data based on a unified spatiotemporal benchmark to construct a comprehensive geographic information system (GIS) training dataset; The thermal comfort prediction and interpretable analysis module is configured to build and train a machine learning prediction model based on the GIS training dataset to predict subjective thermal comfort, and use an explainable artificial intelligence (XAI) method to deeply analyze the model to quantify the contribution and impact mechanism of various physical and visual environmental elements.

[0022] The outdoor thermal comfort evaluation method based on multi-source data fusion and visual compensation mechanism includes the following steps: Step S1: On-site synchronous collection of multi-source heterogeneous data In this step, a portable mobile measurement system integrated with high-precision sensors is constructed and deployed. This system conducts mobile measurements along representative paths in typical urban outdoor spaces (e.g., waterfront trails, parks, pedestrian streets, residential plazas, etc.). Four types of data are collected simultaneously at a high temporal frequency (e.g., 1 Hz): (1) Microclimate physical data: including air temperature, relative humidity, wind speed and direction, and globe temperature (used to calculate mean radiant temperature (MRT)); (2) High-precision spatiotemporal coordinate data: Use RTK or DGPS modules to obtain centimeter-level geographic coordinates and precise timestamps; (3) First-person perspective street view image data: Use a wide-angle action camera (such as GoPro) at the same height as the human eye to continuously record high-definition video; (4) Pedestrian subjective perception data: The accompanying researchers conducted a synchronous structured questionnaire survey on pedestrians on site to obtain their thermal comfort votes (TCV), thermal sensation votes (TSV) and demographic information.

[0023] Step S2: Intelligent extraction of visual environment elements based on deep learning This step involves deep processing of street view image data to quantify the visual environment. First, frames are extracted from the video recorded in S1 to obtain high-quality still images corresponding to TCV survey points. Next, a state-of-the-art deep learning semantic segmentation network (such as PSPNet or DeepLabV3+), pre-trained on large urban scene datasets (e.g., Cityscapes, ADE20K) and fine-tuned using local data, is used to automatically classify the images at the pixel level. This network accurately identifies the object class (e.g., sky, tree, shrub, grass, water, building, road, etc.) that each pixel in the image belongs to. Based on the segmentation results, a series of key visual metrics are derived from the pedestrian's perspective by calculating the proportion of each pixel in each category. These include the sky view factor (SVF), the vegetation view factor (GVF, which can be further divided into tree, shrub, and grass view factors), the water view factor (WVF), and the building view factor (BVF).

[0024] Step S3: Multimodal data integration and feature engineering Within a geographic information system (GIS) platform, using high-precision spatiotemporal coordinates as the "key," the microclimate data and TCV questionnaire data collected in step S1 are precisely matched and spatiotemporally aligned with the visual field index data extracted in step S2. Each TCV survey sample point is associated with a comprehensive set of attribute data: physical environment (temperature, MRT, etc.), visual environment (SVF, GVF, WVF, etc.), and subjective perception (TCV value). To capture the scale effects of the surrounding environment, multi-level buffer zones (e.g., 10m, 30m, 50m) are established for each point. GIS spatial analysis tools are then used to extract other environmental characteristics within the buffer zones (e.g., green space ratio, building density, etc.). Ultimately, a comprehensive, multi-dimensional, multi-scale, and multi-modal GIS feature dataset is constructed.

[0025] Step S4: Build a highly robust thermal comfort prediction model Based on the comprehensive dataset generated in step S3, physical and visual environmental parameters, as well as personal information, are used as input features (X) for the model, and pedestrians' subjective thermal comfort votes (TCV) are used as the target predictor variable (Y). Ensemble learning algorithms with strong fit for high-dimensional, nonlinear data, such as Extreme Gradient Boosting (XGBoost) or Random Forest (RF), are used to construct the prediction model. Model hyperparameters are fine-tuned using methods such as k-fold cross-validation and grid search, and the model's predictive performance is rigorously evaluated using classification evaluation metrics such as accuracy, precision, recall, and F1 score. The interaction between physical and visual elements is explored, such as the degree to which a high GVF alleviates thermal discomfort under the same high temperature conditions, thereby scientifically quantifying the "visual compensation" mechanism.

[0026] Example: In order to verify the effectiveness of the Modified Universal Thermal Climate Index (MUTCI) proposed in the present invention after adjusting for outdoor thermal comfort, we conducted a correlation analysis on its results. Figure 3-Figure 5 As shown in FIG, the correlation between the adjusted MUTCI and [another variable, such as “actual thermal sensation vote” or “raw UTCI”] is shown. The consistency of the curve fluctuations is shown in the figure, which proves the effectiveness of the method of the present invention.

[0027] Reference Figure 1 The specific operation steps of the method of the present invention are as follows: Step S1, data acquisition: On a typical sunny afternoon in summer (2:00 PM-4:00 PM), a portable mobile weather station equipped with temperature, humidity, wind speed, a black globe thermometer, and an RTK-GPS module was carried out in a waterfront park along a preset route (covering tree-lined avenues, waterfront platforms, sunny lawns, etc.). The weather station recorded data at a frequency of 1 Hz. At the same time, a GoPro Hero 10 camera fixed at eye level recorded street view video at 4K resolution. Two investigators simultaneously conducted TCV questionnaire surveys on citizens who were resting or participating in activities along the route. A total of 215 valid questionnaires were collected, and the exact interview start time was recorded for each questionnaire. The questionnaire data details are shown in Table 1; sample data examples and statistics are shown in Tables 2 and 3; Table 1 Questionnaire voting ; Table 2 Data sample details ; Table 3 Data sample details .

[0028] Step S2, Visual Feature Extraction: Import the video recorded by GoPro into the processing program. According to the interview timestamp of each questionnaire, extract the corresponding static image frames from the video stream. Input these images into a PSPNet model pre-trained on the Cityscapes dataset. Figure 2 As shown, the PSPNet model used was trained on the Cityscapes dataset. The Cityscapes dataset contains a wealth of urban streetscape images, meticulously annotated with up to 19 semantic categories. The 12 categories used in this example (sky, trees, shrubs, grass, water, buildings, roads, sidewalks, walls, fences, poles, and vehicles) are based on the capabilities of this pre-trained model and are filtered or aggregated based on the requirements of outdoor thermal comfort assessment. The model outputs pixel-level category predictions. Even if not all categories appear in a single image, its internal processing logic still identifies these pre-set categories. The program automatically counts the total number of tree pixels, shrub pixels, grass pixels, and water pixels in each image and divides them by the total number of pixels in the image to obtain the tree visual field index (TVF), shrub visual field index (SVF), grass visual field index (GVF_grass), and water visual field index (WVF) for that point, respectively. The total vegetation visual field index (GVF) = TVF + SVF + GVF_grass.

[0029] Step S3, Data Integration: In ArcGIS Pro, we first generated a timestamped measurement track point layer based on the RTK-GPS data. We then matched the 215 questionnaires to the track points based on the interview times. Next, we added the visual field indices (GVF, WVF, etc.) calculated in Step S2 as attributes to the corresponding questionnaire sample points based on the timestamp. At this point, each sample point possessed three types of data: microclimate, subjective voting, and visual indicators. Finally, we created a 20-meter-radius buffer zone for each point, overlaid with the city GIS basemap, and calculated the building density within the buffer zone as an additional environmental feature.

[0030] Step S4: Model Construction and Training: The 215 sample datasets generated in step S3 were randomly divided into training and test sets with a ratio of 8:2. An XGBoost classification model was constructed with a multi-class logistic regression (multi:softmax) optimization objective function to predict the TCV voting results of respondents on a five-level scale, ranging from -2 (cold) to +2 (hot).

[0031] Input features include air temperature, relative humidity, wind speed, mean radiant temperature (MRT), GVF, WVF, SVF, building density, and the thickness of the respondent's clothing (clo value). The model's hyperparameters (key parameters such as max_depth, learning_rate, n_estimators, subsample, and colsample_bytree) were optimized using 5-fold cross-validation and grid search. The final model, evaluated on the test set, achieved an overall prediction accuracy of 88.6%, a significant improvement over the UTCI model (65.1%) that uses only physical parameters.

[0032] In order to further understand the model decision-making mechanism and reveal the contribution and influence direction of each feature on thermal comfort prediction, this paper introduces the SHAP (Shapley Additive exPlanations) framework for interpretability analysis after the model training is completed. The specific analysis results are shown in the following example: Figure 6-Figure 9 As shown: Figure 6 This is the SHAP summary diagram of the thermal comfort prediction model, which is used to show the global importance ranking and influence direction of each feature; Figure 7 The characteristic dependence diagram of the thermal comfort prediction model is used to reveal the nonlinear relationship and threshold effect between it and thermal comfort; Figure 8 The ROC curve diagram of the thermal comfort prediction model is used to evaluate the performance of the model in various classification tasks; Figure 9 The figure is a schematic diagram of the ROC curve of the validation data, which further verifies the generalization ability of the model.

[0033] In summary, this invention not only achieves high-precision prediction of outdoor thermal comfort by systematically integrating multi-source data and applying advanced artificial intelligence technology, but more importantly, it scientifically quantifies and explains the key role of the visual compensation mechanism for the first time, providing an unprecedented and powerful decision-making support tool for truly people-oriented and perception-friendly urban environmental planning and design.

Claims

1. Outdoor thermal comfort evaluation system based on AI multi-source fusion and visual compensation, characterized by: include: Multi-source data synchronous acquisition module, used to collect on-site microclimate data, GPS spatiotemporal coordinate data, first-person perspective street view image data, and pedestrian subjective thermal comfort voting TCV data; The intelligent extraction module of visual environmental elements uses a deep learning semantic segmentation model to perform pixel-level analysis on the first-person perspective street view image data, automatically identify and quantify each visual element, and calculate the corresponding visual field index; The multimodal data integration and feature engineering module cleans, aligns, fuses, and extracts multi-scale spatial features from the on-site microclimate data, visual element data, and pedestrian subjective thermal comfort voting data based on a unified spatiotemporal benchmark to construct a comprehensive geographic information system (GIS) training dataset. The thermal comfort prediction and interpretability analysis module builds and trains a machine learning prediction model based on the GIS training dataset to predict subjective thermal comfort, and uses the explainable artificial intelligence (XAI) method to analyze the model and quantify the contribution and influence mechanism of each factor.

2. The outdoor thermal comfort evaluation method based on AI multi-source fusion and visual compensation according to claim 1 is characterized in that: The following steps are involved: Step S1: On-site synchronous collection of multi-source data: obtaining microclimate physical parameters, GPS spatiotemporal coordinates, first-person perspective street view images, and pedestrian subjective thermal comfort voting (TCV) data; Step S2: Intelligent extraction of visual environment elements: Based on the first-person perspective street view image, deep learning semantic segmentation technology is used to identify and quantify environmental visual elements from the pedestrian's perspective, and their visual field index is calculated; Step S3: Multimodal data integration and feature construction: The multi-source data obtained in step S1 and the visual feature data extracted in step S2 are matched and integrated based on unified spatiotemporal coordinates to construct a comprehensive GIS dataset containing multi-scale environmental features. Step S4: Integrate and align the multimodal data to form a unified GIS dataset; and use machine learning to construct an outdoor thermal comfort prediction model to reflect the nonlinear relationship between microclimate, visual elements and subjective thermal perception, and evaluate the visual compensation effect.

3. The outdoor thermal comfort evaluation method based on AI multi-source fusion and visual compensation according to claim 2 is characterized in that: In step S1, dynamic microclimate measurement is performed using a portable micro-weather station that integrates a high-precision temperature sensor, a capacitive humidity sensor, a three-dimensional ultrasonic wind speed and direction sensor, and a black globe thermometer; sub-meter precision spatiotemporal coordinates are acquired using a differential GPS or real-time kinematic RTK positioning module; and a wide-angle motion camera is used to synchronously record street view videos.

4. The outdoor thermal comfort evaluation method based on AI multi-source fusion and visual compensation according to claim 2 is characterized in that: In step S2, a pre-trained deep learning semantic segmentation model is used to perform pixel-level classification on single-frame images extracted from street view videos to identify corresponding environmental elements; and a visual field index is calculated based on pixel statistics of the segmentation results.

5. The outdoor thermal comfort evaluation method based on AI multi-source fusion and visual compensation according to claim 2 is characterized in that: In step S3, the discrete on-site measurement point data are associated with the view index extracted from the corresponding street view image using a unified timestamp and GPS coordinates; In the geographic information system (GIS) software, a multi-level buffer zone is established for each measurement point to extract environmental characteristic variables at different spatial scales in order to construct a multi-attribute, multi-scale comprehensive GIS dataset.

6. The outdoor thermal comfort evaluation method based on AI multi-source fusion and visual compensation according to claim 2 is characterized in that: In step S4, an extreme gradient boosting (XGBoost) algorithm or a random forest (RF) algorithm is used to construct a prediction model. The input features of the prediction model include physical environment parameters, visual field index, and personal parameters. The output target is the thermal comfort voting (TCV) level of the pedestrian. The optimization objective function of the XGBoost algorithm is: ； in, For the The output of a decision tree, is a regularization term containing L1 and L2 norms, which is used to control the complexity of the model to prevent overfitting; i represents the i-th sample, n is the total number of samples, is the true value of the i-th sample, and It represents the cumulative prediction results of the first t-1 decision trees for the sample Is the loss function used to measure the and The predicted value and the true value together The error between is a regularization term that aims to control the model complexity of the t-th tree to prevent overfitting. The L1 norm promotes sparsity by penalizing the absolute value and sum of parameters, while the L2 norm shrinks parameters by penalizing the sum of squares of parameters, thereby smoothing the model and improving generalization ability.

7. An electronic device, characterized in that: include: A memory and a graphics processing unit (GPU); the memory is used to store computer-executable instructions, and the processor and GPU are used to execute the computer-executable instructions to implement the method according to any one of claims 2 to 6.

8. A computer-readable storage medium, characterized in that Computer-executable instructions are stored thereon, and when the computer-executable instructions are executed by a processor, the method according to any one of claims 2 to 6 can be implemented.

Images