Urban blue-green space microclimate prediction system and method based on AI fusion

Abstract

The invention discloses an urban blue-green space microclimate prediction system and method based on AI fusion. The method comprises the following steps of: 1, acquiring field microclimate measured data through linkage of a portable weather station and a GPS (Global Positioning System); 2, utilizing a machine learning downscaling algorithm and a deep learning semantic segmentation technology to extract high-resolution land surface temperature and urban land surface coverage classification data from the remote sensing image; 3, constructing a time-space aligned multi-modal GIS data set, and constructing a high-precision microclimate prediction model by using an ensemble learning algorithm; and 4, introducing an SHAP interpretability framework, carrying out deep analysis on the model, and quantifying the contribution degree of various environmental elements to a microclimate prediction result and the complex nonlinear influence of the contribution degree. According to the method, unprecedented high-precision and explainable scientific decision support can be provided for formulating urban planning, landscape design and thermal environment mitigation strategies, and development of healthy, sustainable and climate-flexible cities is powerfully promoted.

Description

Technical Field

[0001] The present invention belongs to the field of urban environmental science and smart city technology, and in particular relates to an urban blue-green space microclimate prediction system and method based on AI fusion. Background Art

[0002] With the acceleration of global climate change and urbanization, the urban heat island (UHI) effect is becoming increasingly significant, posing a serious threat to urban ecological environments, energy consumption, and resident health. Urban blue-green spaces, as crucial components of urban ecosystems, play an indispensable role in regulating local microclimates and mitigating thermal stress. Therefore, scientifically and accurately predicting and assessing the spatiotemporal distribution characteristics of urban blue-green space microclimates has become a core scientific issue in urban sustainable development planning and climate-adaptive design. Currently, methods for studying urban microclimates primarily include in-situ fixed-site observations, mobile measurements, remote sensing inversion, and numerical simulations. While in-situ observations can provide highly accurate time series data, their spatial representation is poor, making it difficult to capture the high spatial heterogeneity of microclimates within complex urban morphology. Remote sensing technology, particularly thermal infrared remote sensing, can provide macroscopic land surface temperature (LST) distributions, but its poor penetration prevents direct measurement of air temperature above the surface and is often limited by satellite revisit periods and spatial resolution. Although traditional numerical simulation methods (such as CFD and ENVI-met) can simulate three-dimensional flow fields and thermal environments, their model construction is complex, the computational cost is extremely high, and they are extremely sensitive to the setting of boundary conditions and parameterization schemes, making them difficult to apply to large-scale, multi-scenario rapid assessments. In recent years, machine learning and artificial intelligence technologies have provided a new paradigm for solving the above problems. However, existing research still faces significant challenges: First, the depth and breadth of data fusion are insufficient. Most studies rely on single or limited data sources and fail to effectively integrate field measurement data that can reflect the "real" state of the microclimate with remote sensing data that can provide spatial background information. Second, the accuracy and intelligence level of environmental factor extraction need to be improved, especially in the generation of high-resolution LST products and the refined classification of urban surface cover. Traditional methods are inefficient and have limited accuracy. Third, the model lacks interpretability. The machine learning models used in most studies are like "black boxes". Although they can achieve high prediction accuracy, they cannot reveal the complex, nonlinear internal connections and driving mechanisms between input environmental factors (such as green space form and water body size) and output microclimate parameters (such as temperature). This greatly limits the transformation and application of research results to urban planning and design practices. Summary of the Invention

[0003] The present invention aims to overcome the aforementioned shortcomings of the existing technology by providing a technologically advanced, comprehensive, and reliable AI-integrated urban blue-green space microclimate prediction system and method. By deeply integrating multi-source heterogeneous data and combining cutting-edge AI prediction and interpretation techniques, the present invention aims to achieve high-precision, high-resolution predictions of urban microclimates, scientifically quantify the impact of key environmental design elements, and provide intuitive and quantitative decision-making support for urban planning.

[0004] The present invention provides the following technical solutions: The urban blue-green space microclimate prediction system based on AI fusion includes: Data acquisition and synchronization module, used to synchronously collect on-site microclimate measurement data and GPS spatiotemporal coordinate data; Remote sensing data intelligent processing module, used to perform land surface temperature (LST) downscaling inversion and deep learning-based pixel-level semantic segmentation on multi-source remote sensing images to extract urban surface environmental element data; The multidimensional data spatiotemporal integration and feature engineering module is used to align, fuse, and extract multi-scale spatial features from on-site microclimate measurement data, downscaled LST data, and urban land cover environmental element data from remote sensing images under a unified spatiotemporal benchmark to construct a comprehensive geographic information system (GIS) training dataset. The microclimate intelligent prediction and explainability analysis module builds and trains a machine learning prediction model based on the GIS training dataset, uses the explainable artificial intelligence (XAI) method to analyze the trained model, and quantifies the contribution and impact mechanism of various environmental factors on the microclimate.

[0005] The urban blue-green space microclimate prediction method based on AI fusion includes the following steps: Step S1, on-site microclimate data collection: obtain microclimate parameters and synchronously record GPS time and space coordinates; Step S2, remote sensing environmental element extraction: performing land surface temperature (LST) inversion on low- and medium-resolution thermal infrared satellite images and downscaling processing in combination with high-resolution optical images, and performing deep learning semantic segmentation on high-resolution remote sensing images to identify and quantify urban surface cover types and extract remote sensing environmental element data; Step S3, multidimensional data integration and feature construction: the microclimate parameter data obtained in step S1 and the remote sensing environmental element data extracted in step S2 are matched and aligned with the multimodal data based on unified timestamps and spatial coordinates to construct a GIS dataset containing multi-scale spatial environmental features; Step S4: Microclimate prediction model construction and validation: Based on the GIS dataset integrated in step S3, a microclimate prediction model is constructed using an integrated learning algorithm, and the model is trained, tuned, and performance evaluated; Step S5: Model interpretability analysis and contribution quantification: Use an interpretable machine learning framework to deeply interpret the trained prediction model, analyze and quantify the contribution of each environmental factor to the microclimate prediction results and their nonlinear relationship.

[0006] Furthermore, in step S1, a portable micro-weather station is constructed that integrates a semiconductor temperature sensor, a capacitive humidity sensor, a three-dimensional ultrasonic wind speed and direction sensor, a photoelectric total radiation sensor, and a differential GPS or real-time dynamic positioning module to perform dynamic or fixed-point microclimate measurements, and the data is transmitted in real time to a cloud central database through a wireless communication module for storage, cleaning, and quality control.

[0007] Furthermore, in step S2, the land surface temperature (LST) inversion adopts a thermal radiation transfer equation algorithm or a single window algorithm, and adopts a downscaling algorithm or a statistical downscaling model based on machine learning to fuse the multispectral information of the optical image to generate LST data; at the same time, a deep learning semantic segmentation network based on a convolutional neural network is used to perform pixel-level semantic segmentation on the sub-meter remote sensing image to identify and quantify the urban surface cover type, and calculate the corresponding quantitative indicators.

[0008] Furthermore, in step S3, the discrete on-site microclimate measurement point data are associated with the downscaled LST data in raster format and the urban land cover type data extracted by remote sensing using a unified timestamp and high-precision GPS coordinates, and the multimodal data are aligned; and then imported into the geographic information system (GIS) software. By establishing a multi-level buffer for each measurement point and performing buffer analysis, neighbor analysis, and spatial statistics, the environmental characteristic variables at different spatial scales are extracted, thereby constructing a multi-attribute, multi-scale, spatialized comprehensive GIS dataset.

[0009] Furthermore, in step S4, an extreme gradient boosting XGBoost algorithm or a random forest RF algorithm is used to build a prediction model; the XGBoost algorithm corrects the residual of the previous tree by iteratively adding decision trees, and its optimization objective function is: ; in, For the The total objective function of the iterations, is the loss function, is the true value, For the front The cumulative prediction value of the tree, For the The predicted output of the tree, is a regularization term used to control model complexity and prevent overfitting. It is defined as: ; in is the penalty coefficient, T is the total number of leaf nodes, is the coefficient of the L2 regularization term, is the weight or score of the jth leaf node, Represents the square of the score, and j is the index variable for traversing all leaf nodes.

[0010] Furthermore, in step S5, the trained prediction model is interpreted using the Shapley addition interpretation SHAP framework based on game theory, which assigns a marginal contribution value to each input feature in a single prediction, and the calculation formula is: ; in, Representation characteristics For the SHAP value of this prediction, is the complete set of all input features, yes Does not contain features A feature subset of Is the model using only a subset of features The predicted output value under the condition of ; by drawing SHAP dependency diagram, force diagram and summary diagram visualization means, the global and local, linear and nonlinear relationships between each environmental feature and the prediction target, as well as the interaction between features are intuitively revealed.

[0011] An electronic device comprises: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method according to any one of claims 2 to 7.

[0012] A computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the method according to any one of claims 2 to 7 can be implemented.

[0013] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows: 1) Improving the depth and breadth of data fusion: This invention innovatively integrates high-precision, timely, point-based ground-based data with macroscopic, surface-level remotely sensed land temperature data and mesoscopic, refined remotely sensed land cover data, constructing a comprehensive, cross-scale, multimodal spatiotemporal dataset. This fusion effectively overcomes the limitations of a single data source and significantly enhances the ability to comprehensively characterize complex urban environments. 2) Intelligent and refined extraction of environmental elements: Advanced machine learning LST downscaling technology and deep learning semantic segmentation techniques replace traditional, inefficient manual or semi-automatic methods. This enables rapid, high-precision, and automated extraction of urban thermal environment information and land cover types, providing high-quality key input data for in-depth analysis of the intrinsic relationship between urban form and microclimate. 3) High Precision and Strong Generalization Capability of the Prediction Model: Utilizing ensemble learning algorithms such as XGBoost and Random Forest, the model effectively captures the complex, highly nonlinear mapping relationships between multi-source, heterogeneous environmental factors and microclimate parameters. The model boasts high prediction accuracy and, as demonstrated by cross-validation, strong generalization capabilities, effectively avoiding overfitting and ensuring the reliability of prediction results across different regions and scenarios. 4) A revolutionary breakthrough in model interpretability: The SHAP framework was systematically introduced for the first time into the field of urban blue-green space microclimate prediction, making the previously opaque "black box" machine learning model completely transparent and interpretable. It can also scientifically quantify and visualize the specific contributions and effects of various environmental factors on microclimate prediction. 5) Enabling scientific decision-making in urban planning and design: The high-precision prediction results and in-depth interpretable analysis provided by this invention can provide urban planners, landscape designers, and policymakers with intuitive, reliable, and quantitative scientific decision-making basis. For example, it can accurately evaluate the cooling effect of different green space layouts and water body configuration schemes, thereby accurately optimizing the layout and form of urban blue-green spaces during the planning and design stage to maximize their ecological benefits, effectively alleviate the urban heat island effect, and improve the quality of the human living environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a schematic diagram of the process and results of extracting land cover types from high-resolution remote sensing images using a deep learning semantic segmentation network in the present invention; Figure 3 The ROC characteristic curve of solar radiation in the XGBoost model prediction model performance diagram of the present invention; Figure 4 The ROC characteristic curve of wind speed in the XGBoost model prediction model performance diagram of the present invention; Figure 5 The ROC characteristic curve of wind direction in the XGBoost model prediction model performance diagram of the present invention; Figure 6 The ROC characteristic curve of temperature in the XGBoost model prediction model performance diagram of the present invention; Figure 7The ROC characteristic curve of relative humidity in the XGBoost model prediction model performance diagram of the present invention; Figure 8 A summary graph of the solar radiation prediction SHAP importance in the SHAP interpretability analysis results output by the XGBoost model of the present invention; Figure 9 A summary graph of the wind speed prediction SHAP importance in the SHAP interpretability analysis results output by the XGBoost model of the present invention; Figure 10 A summary graph of the wind direction prediction SHAP importance in the SHAP interpretability analysis results output by the XGBoost model of the present invention; Figure 11 A summary graph of the temperature prediction SHAP importance in the SHAP interpretability analysis results output by the XGBoost model of the present invention; Figure 12 A SHAP importance summary graph of relative humidity prediction in the SHAP interpretability analysis results output by the XGBoost model of the present invention; Figure 13 The SHAP dependency graph of solar radiation prediction in the SHAP interpretability analysis results output by the XGBoost model of the present invention; Figure 14 This is the temperature prediction SHAP dependency graph in the SHAP interpretability analysis results output by the XGBoost model of the present invention. DETAILED DESCRIPTION

[0015] 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.

[0016] 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.

[0017] An urban blue-green space microclimate prediction system based on AI fusion, which includes: Data acquisition and synchronization module, used to synchronously collect on-site microclimate measurement data and GPS spatiotemporal coordinate data; Remote sensing data intelligent processing module, used to perform land surface temperature (LST) downscaling inversion and deep learning-based pixel-level semantic segmentation on multi-source remote sensing images to extract urban surface environmental element data; The multidimensional data spatiotemporal integration and feature engineering module is used to align, fuse, and extract multi-scale spatial features from on-site microclimate measurement data, downscaled LST data, and urban land cover environmental element data from remote sensing images under a unified spatiotemporal benchmark to construct a comprehensive geographic information system (GIS) training dataset. The microclimate intelligent prediction and explainability analysis module is used to build and train a machine learning prediction model based on the GIS training data set, analyze the trained model using the explainable artificial intelligence (XAI) method, and quantify the contribution and impact mechanism of various environmental factors on the microclimate. The urban blue-green space microclimate prediction method based on AI fusion has a core process including high-precision data collection, multi-source data intelligent processing and fusion, high-precision model construction, and explainability analysis.

[0018] Step S1: Acquisition of high-precision and high-density on-site microclimate data.

[0019] This step builds and deploys a portable micro-weather station system integrated with high-precision sensors. This system conducts mobile measurements along representative routes within urban green and blue spaces (such as parks, waterfronts, street canyons, and plazas) or at fixed locations at key nodes. It simultaneously collects microclimate parameters such as air temperature, relative humidity, wind speed, wind direction, and global radiation at a high temporal frequency (e.g., 1-10 seconds per time interval). Using a built-in differential GPS (DGPS) or real-time kinematic (RTK) module, it obtains centimeter-precise geographic coordinates and timestamps. This collected data constitutes the "Gold Standard" (Ground Truth) labels for model training and validation.

[0020] Step S2: Intelligent processing of remote sensing images and extraction of environmental elements.

[0021] This step processes two types of remote sensing data in parallel. First, to analyze the surface thermal environment, thermal infrared imagery from medium-resolution satellites such as Landsat 8 / 9, synchronized with in-situ measurements, is acquired. The raw land surface temperature (LST) is inverted using the radiative transfer equation method or a single-window algorithm. Subsequently, the LST is downscaled to a high spatial resolution of 10 meters using machine learning downscaling models such as random forests and fused with auxiliary variables such as NDVI and surface albedo provided by high-resolution optical imagery such as Sentinel-2. Downscaling is a key data fusion technique, the core goal of which is to leverage the rich surface details in high-resolution optical imagery (such as Landsat and Sentinel-2) to sharpen and refine land surface temperature (LST) data acquired by medium- and low-resolution thermal infrared satellites (such as MODIS). This is not a simple image magnification process; it is based on a core physical principle: there is a close physical correlation between surface temperature and land cover characteristics (such as vegetation density, building density, and surface albedo). The processing pipeline begins by calculating a series of predictors reflecting these surface characteristics, such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Building Index (NDBI), from high-resolution optical imagery. These high-resolution predictor maps are then spatially aggregated to match the resolution of the original coarse-resolution land surface temperature (LST) imagery. At this coarse scale, a quantitative relationship model between LST and these predictors is established using machine learning methods. Once trained, the model is applied to the original, unaggregated, high-resolution predictor imagery, generating a preliminary, high-resolution LST prediction map with fine spatial texture. However, to ensure macroscopic energy conservation and temperature accuracy of the downscaling results, a crucial final step—residual correction—is required. This step calculates the coarse-scale average error (i.e., residual) between the original coarse LST and the predicted high-resolution LST. This residual is then appropriately distributed back to the high-resolution prediction map, ultimately generating a high-resolution land surface temperature product that retains the original radiometric accuracy of the thermal infrared data while retaining the fine spatial detail of the optical imagery. Secondly, focusing on surface cover, the pre-trained KNN algorithm, using the SinoLC-1 public dataset, performs pixel-level classification on the images, accurately extracting various surface cover types, including vegetation, water bodies, buildings, roads, and bare soil. Based on the segmentation results, a series of quantitative environmental indicators are further calculated, such as the fractional vegetation cover (FVC), building density, water area ratio, sky openness (SVF), and normalized difference vegetation index (NDVI) within each buffer zone.

[0022] Step S3: Multi-dimensional and multi-scale data integration and alignment.

[0023] In a geographic information system (GIS) platform, the discrete point microclimate data from step S1 are spatiotemporally aligned with the continuous raster LST data, land cover classification data, and various quantitative environmental indicator layers generated in step S2, using the high-precision spatiotemporal coordinates of the field measurement points as a reference. Multi-level concentric buffers (e.g., with radii of 25, 50, 100, and 200 meters) are created for each measurement point. Spatial analysis tools are then used to extract features such as the average LST value, the area percentage of various land cover types, and the average NDVI within each buffer. This approach aims to capture the comprehensive impact of the surrounding environment on the microclimate of the measurement point at different spatial scales, ultimately constructing a comprehensive, multi-attribute, multi-scale GIS feature dataset.

[0024] Step S4: Construct a highly robust microclimate prediction model.

[0025] Based on the comprehensive GIS dataset generated in step S3, the extracted multiscale environmental features and downscaled LST are used as the model's input features (X), and field-measured microclimate parameters (such as temperature) are used as the target predictor variables (Y). Ensemble learning algorithms such as extreme gradient boosting (XGBoost) or random forest (RF) are used to construct the prediction model. These algorithms excel at processing high-dimensional data, effectively capturing nonlinear relationships and complex interactions between features, and exhibit strong resistance to overfitting. Model hyperparameters are optimized using methods such as k-fold cross-validation and grid search, and the model's predictive performance is rigorously evaluated using metrics such as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²).

[0026] Step S5: Deep interpretability analysis based on the SHAP framework.

[0027] To open the "black box" of the machine learning model, this step uses the SHAP (Shapley Additive Explanations) interpretability framework to perform a posteriori interpretation of the trained XGBoost or RF model. SHAP can calculate an accurate contribution value (SHAP value) for each input feature in each prediction. This value represents the contribution of the feature to the prediction result from the baseline value (the average value of all sample predictions) to the current value. By generating SHAP summary graphs, dependency graphs, force graphs and other visual results (such as Figure 8-14 As shown), the following functions can be achieved: (1) Global ranking of the importance of each environmental factor in affecting microclimate; (2) Intuitively reveal the nonlinear relationship between a single factor (e.g., vegetation coverage) and a microclimate parameter (e.g., temperature), such as identifying the threshold or saturation point of cooling benefits; (3) Explore the synergistic or antagonistic effects (interaction effects) among different environmental factors.

[0028] Example: The AI-integrated urban blue-green space microclimate prediction system described in this invention can be physically implemented using cloud servers and edge computing devices. The core system includes a data acquisition and synchronization module, a remote sensing data intelligent processing module, a multidimensional data spatiotemporal integration and feature engineering module, and a microclimate intelligent prediction and interpretability analysis module.

[0029] Reference Figure 1 The specific operation steps of the prediction method of the present invention are as follows: Step S1, Data Acquisition: In City A's central park and its surrounding areas, the research team carried a portable micro-meteorological station and conducted mobile measurements along a pre-defined path covering various underlying surfaces, including grasslands, undergrowth, waterfronts, and plaza pavements. The weather station simultaneously collected data such as air temperature (accuracy ±0.1°C), relative humidity (accuracy ±1.5%RH), 3D wind speed, and GPS coordinates (horizontal accuracy <2cm in RTK mode) every 2 seconds (see Table 1). Table 1. Summary of microclimate measurement data .

[0030] Data is uploaded to the central database deployed on the cloud server in real time via the 5G network.

[0031] Step S2, remote sensing data processing: obtain a Landsat 9OLI / TIRS-2 image and a Sentinel-2 L2A image with clear weather and closest time to the measurement day. Use the radiation transfer equation method to invert the 30-meter resolution LST. Then, build a random forest downscaling model, using the 10-meter resolution NDVI, MNDWI, surface albedo, etc. provided by Sentinel-2 as prediction variables to downscale the 30-meter LST to 10 meters. At the same time, obtain a 0.5-meter resolution aerial remote sensing image of the area, and use the KNN algorithm pre-trained on the SinoLC-1 public dataset and fine-tuned on local samples for semantic segmentation to accurately extract the surface cover types of water bodies, vegetation, and built-up areas ( Figure 2 Based on this classification map, we calculated the NDVI, FVC, building density and other indicator layers.

[0032] Step S3, Data Integration: In ArcGIS Pro, the approximately 5,000 mobile measurement point data obtained in Step S1 were imported as a point layer. Using the "Extract Multi-Values ​​to Points" tool, the values ​​of the 10-meter-resolution LST layer, FVC layer, building density layer, and other raster data generated in Step S2 were precisely matched to each measurement point. Subsequently, the "Buffer" tool was used to create four buffer levels for each point, with radii of 25, 50, 100, and 200 meters. The percentage of each type of land cover (vegetation, water, buildings, etc.) within each buffer was calculated. Ultimately, each measurement point was associated with its own LST value and more than 10 environmental features at four different scales, forming a comprehensive dataset in a wide table format containing approximately 20 feature columns.

[0033] Step S4, XGBoost model construction and training: The comprehensive dataset generated in step S3 was randomly divided into a training set and a test set in a ratio of 7:3. An XGBoost regression model was constructed with 20 environmental features as input (X) and the measured temperature on site as output (Y). The model's hyperparameters (such as n_estimators=500, learning_rate=0.05, max_depth=7, etc.) were tuned using 5-fold cross-validation and Bayesian optimization algorithms. The final model was evaluated on the test set, and the coefficient of determination (R²) between its prediction results and the measured temperature reached 0.92, and the root mean square error (RMSE) was 0.25°C, showing extremely high prediction accuracy (see Table 2, Figure 3-7 ); Table 2 XGBoost model performance data .

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Urban blue-green space microclimate prediction system based on AI fusion, characterized by the system include: Data acquisition and synchronization module, used to synchronously collect on-site microclimate measurement data and GPS spatiotemporal coordinate data; Remote sensing data intelligent processing module, used to perform land surface temperature (LST) downscaling inversion and deep learning-based pixel-level semantic segmentation on multi-source remote sensing images to extract urban surface environmental element data; The multidimensional data spatiotemporal integration and feature engineering module is used to align, fuse, and extract multi-scale spatial features from on-site microclimate measurement data, downscaled LST data, and urban land cover environmental element data from remote sensing images under a unified spatiotemporal benchmark to construct a comprehensive geographic information system (GIS) training dataset. The microclimate intelligent prediction and explainability analysis module builds and trains a machine learning prediction model based on the GIS training dataset, uses the explainable artificial intelligence (XAI) method to analyze the trained model, and quantifies the contribution and impact mechanism of various environmental factors on the microclimate.

2. The urban blue-green space microclimate prediction method based on AI fusion is characterized by: The steps include: Step S1, on-site microclimate data collection: obtain microclimate parameters and synchronously record GPS time and space coordinates; Step S2, remote sensing environmental element extraction: performing land surface temperature (LST) inversion on low- and medium-resolution thermal infrared satellite images and downscaling processing in combination with high-resolution optical images, and performing deep learning semantic segmentation on high-resolution remote sensing images to identify and quantify urban surface cover types and extract remote sensing environmental element data; Step S3, multidimensional data integration and feature construction: the microclimate parameter data obtained in step S1 and the remote sensing environmental element data extracted in step S2 are matched and aligned with the multimodal data based on unified timestamps and spatial coordinates to construct a GIS dataset containing multi-scale spatial environmental features; Step S4: Microclimate prediction model construction and validation: Based on the GIS dataset integrated in step S3, a microclimate prediction model is constructed using an integrated learning algorithm, and the model is trained, tuned, and performance evaluated; Step S5: Model interpretability analysis and contribution quantification: Use an interpretable machine learning framework to deeply interpret the trained prediction model, analyze and quantify the contribution of each environmental factor to the microclimate prediction results and their nonlinear relationship.

3. The urban blue-green space microclimate prediction method based on AI fusion according to claim 2 is characterized in that: In step S1, a portable micro-weather station is constructed that integrates a semiconductor temperature sensor, a capacitive humidity sensor, a three-dimensional ultrasonic wind speed and direction sensor, a photoelectric global radiation sensor, and a differential GPS or real-time dynamic positioning module to perform dynamic or fixed-point microclimate measurements, and the data is transmitted in real time to a cloud-based central database through a wireless communication module for storage, cleaning, and quality control.

4. The urban blue-green space microclimate prediction method based on AI fusion according to claim 2 is characterized in that: In step S2, the land surface temperature (LST) inversion adopts a thermal radiation transfer equation algorithm or a single window algorithm, and adopts a downscaling algorithm or a statistical downscaling model based on machine learning to fuse the multispectral information of the optical image to generate LST data; at the same time, a deep learning semantic segmentation network based on a convolutional neural network is used to perform pixel-level semantic segmentation on the sub-meter remote sensing image, identify and quantify the urban surface cover type, and calculate the corresponding quantitative indicators.

5. The urban blue-green space microclimate prediction method based on AI fusion according to claim 2 is characterized in that: In step S3, using a unified timestamp and high-precision GPS coordinates, the discrete on-site microclimate measurement point data are associated with the downscaled LST data in raster format and the urban land cover type data extracted by remote sensing, and the multimodal data is aligned; The data are then imported into the Geographic Information System (GIS) software. By establishing a multi-level buffer zone for each measurement point and performing buffer zone analysis, neighbor analysis, and spatial statistics, environmental characteristic variables at different spatial scales are extracted, thereby constructing a multi-attribute, multi-scale, spatialized comprehensive GIS dataset.

6. The urban blue-green space microclimate prediction method based on AI fusion according to claim 2, characterized in that: In step S4, the extreme gradient boosting XGBoost algorithm or the random forest RF algorithm is used to build a prediction model; the XGBoost algorithm corrects the residual of the previous tree by iteratively adding decision trees, and its optimization objective function is: ； in, For the The total objective function of the iterations, is the loss function, is the true value, For the front The cumulative prediction value of the tree, For the The predicted output of the tree, is a regularization term used to control model complexity and prevent overfitting. It is defined as: ； in is the penalty coefficient, T is the total number of leaf nodes, is the coefficient of the L2 regularization term, is the weight or score of the jth leaf node, Represents the square of the score, and j is the index variable for traversing all leaf nodes.

7. The urban blue-green space microclimate prediction method based on AI fusion according to claim 2, characterized in that: In step S5, the trained prediction model is interpreted using the Shapley addition interpretation SHAP framework based on game theory, which assigns a marginal contribution value to each input feature in a single prediction. The calculation formula is: ； in, Representation characteristics For the SHAP value of this prediction, is the complete set of all input features, yes Does not contain features A feature subset of Is the model using only a subset of features The predicted output value under the condition of ; by drawing SHAP dependency diagram, force diagram and summary diagram visualization means, the global and local, linear and nonlinear relationships between each environmental feature and the prediction target, as well as the interaction between features are intuitively revealed.

8. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method according to any one of claims 2 to 7.

9. A computer-readable storage medium, characterized in that The readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the method according to any one of claims 2 to 7 can be implemented.

Images