High-temporal-spatial-resolution surface temperature reconstruction system for urban thermal environment refined monitoring

Abstract

The invention provides a high-temporal-spatial-resolution surface temperature reconstruction system for urban thermal environment refined monitoring, and relates to the field of electric digital data processing. Comprising a multi-source heterogeneous data intelligent acquisition and preprocessing module, an urban thermal environment multi-dimensional feature knowledge modeling module, a physical constraint deep learning surface temperature reconstruction module and an intelligent monitoring early warning and decision support module. The multi-source heterogeneous data intelligent acquisition and preprocessing module is responsible for collecting and preprocessing various remote sensing and ground observation data, and the urban thermal environment multi-dimensional feature knowledge modeling module constructs a knowledge system of an urban underlying surface, a three-dimensional form and a thermal process. The physical constraint deep learning surface temperature reconstruction module is used for realizing accurate reconstruction of high temporal-spatial resolution surface temperature, and the intelligent monitoring early warning and decision support module converts a reconstruction result into visual display, risk early warning and regulation and control decision suggestions; according to the system, high-precision, physically consistent and interpretable urban surface temperature reconstruction can be realized.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and specifically to a high spatiotemporal resolution surface temperature reconstruction system for refined monitoring of urban thermal environment. Background Technology

[0002] The urban heat island effect refers to the phenomenon where urban areas are significantly warmer than surrounding suburban and rural areas. This phenomenon is mainly caused by human activities, land-use change, and alterations in the thermal radiation characteristics of built infrastructure. With the acceleration of global urbanization, the urban heat island effect exacerbates the impacts of climate change, leading to increased energy consumption, decreased air quality, and a rise in heat-related diseases and mortality. Accurate monitoring and prediction of urban surface temperature distribution are crucial for urban planning, public health management, and the development of heat mitigation strategies.

[0003] Surface temperature reconstruction technology is the core of urban thermal environment monitoring, primarily relying on thermal infrared remote sensing observations. However, thermal infrared sensors are limited by weak signal strength, resulting in spatial resolution far lower than optical sensors. For example, MODIS thermal infrared data has a spatial resolution of 1 kilometer, while its optical data resolution can reach 250 meters or 500 meters; ASTER thermal infrared data has a resolution of 90 meters, while its optical data resolution is only 15 meters or 30 meters. This insufficient spatial resolution is particularly prominent when monitoring heterogeneous urban and suburban landscapes, failing to capture fine-scale temperature change characteristics and limiting the accurate characterization of complex thermal environments such as urban canyons and densely built-up areas. Furthermore, thermal infrared satellite observations also suffer from low temporal resolution and data loss due to cloud cover, making it difficult to achieve continuous spatiotemporal temperature field monitoring.

[0004] To improve the spatiotemporal resolution of land surface temperature, existing technologies mainly employ the following three algorithms. The first is spatiotemporal fusion algorithms, typically represented by STARFM and ESTARFM models. These methods generate a high spatiotemporal resolution temperature field by fusing high spatial resolution but low temporal resolution images with low spatial resolution but high temporal resolution images. Their advantage is the effective utilization of the complementarity of multi-source data; however, they rely on the assumption of spectral similarity between images, are prone to significant errors in rapidly changing urban areas, and cannot consider physical process constraints, potentially leading to reconstruction results that violate fundamental physical laws such as energy conservation. The second is machine learning-based temperature downscaling algorithms, including random forests, gradient boosting, support vector machines, and deep neural networks. These methods achieve spatial downscaling by establishing statistical relationships between temperature and underlying surface features, meteorological parameters, etc. The advantage of these methods is their ability to automatically learn complex nonlinear relationships and their fast training speed; however, they lack physical interpretability, have poor generalization ability in areas with insufficient training data coverage, and cannot guarantee the physical consistency of the reconstruction results. The third is temperature simulation methods based on physical models, which use physical processes such as urban energy balance equations and radiative transfer models to simulate land surface temperature distribution. Its advantages are that it has clear physical meaning and interpretability, while its disadvantages are that it has high computational cost, requires a large number of detailed urban geometric and thermophysical parameters as input, and the simplified physical model is difficult to adapt to the high heterogeneity of the urban underlying surface, resulting in limited accuracy in complex urban environments.

[0005] Regarding patented technologies, US Patent 10306155B2 discloses a compact, high-resolution thermal infrared imaging system. This system employs dual-band infrared imaging technology, enabling surface temperature observation with a spatial resolution of 60 meters. The shortcomings of this patent lie in its focus solely on improving spatial resolution at the hardware level, lacking a systematic consideration of multi-source data fusion, time-series reconstruction, and physical constraints. It fails to address the issues of low temporal resolution in satellite thermal infrared data and data loss due to cloud cover, and it does not address the impact of urban three-dimensional morphology on the temperature field, making it ill-suited to the complex heterogeneity of urban thermal environments. Chinese Patent CN105842755A discloses a device and method for measuring the urban residential heat island effect. This device includes a meteorological data acquisition device, an infrared image acquisition device, and a receiving device, collecting the surface temperature of the residential area through a thermal infrared imager installed on the roof of a building. The shortcomings of this patent are that it uses a single thermal infrared sensor in a fixed location, which has a limited spatial coverage and cannot achieve continuous monitoring of the urban thermal environment over a large area. It also lacks spatiotemporal fusion processing of multi-temporal data, physical model constraints and uncertainty quantification, making it difficult to provide high spatiotemporal resolution temperature reconstruction results. Furthermore, it does not consider the optimized layout and adaptive adjustment of the sensor network, resulting in insufficient spatial representativeness of the observation data. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings by proposing a high spatiotemporal resolution surface temperature reconstruction system for refined monitoring of urban thermal environment.

[0007] The present invention adopts the following technical solution:

[0008] A high spatiotemporal resolution surface temperature reconstruction system for refined monitoring of urban thermal environment includes a multi-source heterogeneous data intelligent acquisition and preprocessing module, an urban thermal environment multi-dimensional feature knowledge modeling module, a physically constrained deep learning surface temperature reconstruction module, and an intelligent monitoring, early warning and decision support module.

[0009] The multi-source heterogeneous data intelligent acquisition and preprocessing module is responsible for collecting and preprocessing various remote sensing and ground observation data. The urban thermal environment multi-dimensional feature knowledge modeling module constructs a knowledge system of urban underlying surface, three-dimensional morphology and thermal processes. The physical constraint deep learning surface temperature reconstruction module is used to achieve accurate reconstruction of surface temperature with high spatiotemporal resolution. The intelligent monitoring, early warning and decision support module transforms the reconstruction results into visualization, risk warning and control decision suggestions.

[0010] The intelligent acquisition and preprocessing module for multi-source heterogeneous data includes a multi-scale remote sensing data intelligent acquisition unit, a ground observation network adaptive deployment unit, and a heterogeneous data quality control and registration unit. The multi-scale remote sensing data intelligent acquisition unit is responsible for collaboratively acquiring multispectral images and thermal infrared observation data. The ground observation network adaptive deployment unit realizes the optimal spatial layout of sensors through an observation network optimization algorithm. The heterogeneous data quality control and registration unit performs spatiotemporal registration, format standardization, and quality screening on multi-source data.

[0011] The urban thermal environment multidimensional feature knowledge modeling module includes an urban underlying surface multidimensional attribute modeling unit, an urban three-dimensional morphology and radiation transmission modeling unit, and an urban thermal environment process knowledge graph construction unit. The urban underlying surface multidimensional attribute modeling unit is used to construct a ground feature-thermal attribute knowledge base. The urban three-dimensional morphology and radiation transmission modeling unit is used to simulate the radiation transmission process and shadow dynamic evolution of building clusters. The urban thermal environment process knowledge graph construction unit constructs a three-layer knowledge graph containing entities, relationships, and rules.

[0012] The physically constrained deep learning surface temperature reconstruction module includes a physics-guided deep neural network reconstruction unit, a multi-temporal spatiotemporal fusion and dynamic correction unit, and an uncertainty quantification and multi-scenario generation unit. The physics-guided deep neural network reconstruction unit is used to achieve surface temperature reconstruction with both physical consistency and high spatiotemporal resolution. The multi-temporal spatiotemporal fusion and dynamic correction unit is used to perform real-time deviation correction in conjunction with ground observations. The uncertainty quantification and multi-scenario generation unit performs uncertainty assessment on the temperature reconstruction results.

[0013] The intelligent monitoring, early warning, and decision support module includes a multi-dimensional spatiotemporal visualization and interactive analysis unit, a thermal risk intelligent early warning and source tracing analysis unit, and an urban thermal environment regulation decision support unit. The multi-dimensional spatiotemporal visualization and interactive analysis unit is used to intuitively display the spatial distribution and temporal variation characteristics of the urban thermal environment. The thermal risk intelligent early warning and source tracing analysis unit performs source tracing analysis of the causes of thermal anomalies, classifies early warning levels, and achieves short-term prediction. The urban thermal environment regulation decision support unit simulates the cooling effect of intervention measures based on the reconstruction results and uses optimization algorithms to generate thermal mitigation strategy suggestions.

[0014] Furthermore, the adaptive deployment unit of the ground observation network includes a multi-source sensor access processor, an observation network optimization processor, and a mobile observation co-processor. The multi-source sensor access processor is used to provide meteorological driving parameters such as air temperature, relative humidity, wind speed and direction, and high-precision true values ​​of surface temperature. The observation network optimization processor is used to calculate the optimal deployment location and density of sensors. The mobile observation co-processor integrates data from the mobile observation platform to achieve collaborative observation between the mobile platform and the fixed network.

[0015] The observation network optimization processor determines the sensor deployment locations according to the following formula:

[0016] ;

[0017] ;

[0018] Among them, S opt This represents the optimal sensor deployment location, where x represents a candidate sensor deployment location. T represents the set of all possible candidate locations for deploying sensors. s For the temperature field of the entire study area, D cur H[T] represents the currently existing observation dataset. s [D] represents the temperature field T given the observed data D. s conditional entropy, This represents the temperature value observed at location x. This represents the expected value of the observed value. This is the cost weighting coefficient. This represents the cost of deploying the sensor at x.

[0019] Furthermore, the urban thermal environment process knowledge graph construction unit includes an energy budget modeling processor, a human activity heat source identification processor, and a knowledge graph reasoning engine. The energy budget modeling processor is used to establish the urban surface energy balance equation and construct the physical constraint rules of energy budget. The human activity heat source identification processor is used to identify and quantify the spatial distribution and temporal variation of human thermal contributions. The knowledge graph reasoning engine is used to construct a three-layer knowledge graph of entities, relationships, and rules to realize the logical constraints and anomaly detection of the temperature field.

[0020] The knowledge graph reasoning engine constructs a temporal causal graph of the urban thermal environment. Where V is the node set, Let B be the causal edge set, B be the node attributes, and the strength of causal relationships be learned according to the following loss function:

[0021] ;

[0022] ;

[0023] Among them, L cau Let V be the total loss function for causal relationship learning. i V j Let ) be a directed edge in the causal graph. For attention weights, represent the weights from V i To V j The causal strength, V i (t) represents the value of Vi at time t. For causal time lag, This represents the observed true causal effect. To regularize sparsity, A is the adjacency matrix. h is the acyclic constraint coefficient, and h() is the acyclic constraint function. For V i eigenvectors, W Q For querying the matrix, v j For V j eigenvectors, W K Let d be the key matrix, d be the dimension of the query matrix and the key matrix, and N(j) represent v. j The set of candidate parent nodes.

[0024] Furthermore, the physical-guided deep neural network reconstruction unit includes a physical information embedding processor, a multi-scale feature extraction processor, and a physical consistency verification processor. The physical information embedding processor transforms physical constraints into loss function terms of the deep learning model and ensures that the temperature field output by the neural network satisfies basic physical laws. The multi-scale feature extraction processor is used to extract multi-scale spatial features and temporal features, and learn the spatial heterogeneity and temporal evolution pattern of the temperature field. The physical consistency verification processor performs energy conservation checks, temperature gradient rationality checks, and extreme value anomaly identification on the temperature field output by the neural network.

[0025] The physical information embedding processor constructs a spatially varying learnable function according to the following formula:

[0026] ;

[0027] ;

[0028] ;

[0029] Among them, L total For the total loss, L data R is the data fitting loss. m [] represents the m-th physical constraint residual. Let M be the predicted temperature at location x at time t, M be the total number of physical constraints, and w be the predicted temperature at location x at time t. m (x) represents the adaptive weight of the m-th constraint at position x. For the Sigmoid activation function, MLP m It is a small multilayer sensor. It is a three-dimensional geometric feature vector. This is the feature vector of the underlying surface type. This is the relative ratio of the physical residual to the data gradient. As the global baseline weight, For about gradient, For all learnable parameters, It is a numerically stable term.

[0030] Furthermore, the uncertainty quantification and multi-scenario generation unit includes an uncertainty assessment processor, a confidence interval calculation processor, and a multi-scenario simulation processor. The uncertainty assessment processor is used to quantify the uncertainty sources and propagation paths of the temperature reconstruction results. The confidence interval calculation processor calculates the confidence interval or probability density distribution for the reconstructed temperature of each pixel based on the uncertainty assessment results. The multi-scenario simulation processor combines the uncertainty distribution with the assumption of extreme weather conditions to generate multiple temperature field schemes such as the optimal estimation scenario, the conservative estimation scenario, and the extreme high temperature scenario.

[0031] The uncertainty assessment processor calculates the uncertainty contribution of each step according to the following formula:

[0032] ;

[0033] ;

[0034] ;

[0035] in, To reconstruct the total uncertainty of the temperature field, For the reconstructed temperature field, S is the set of error sources. For the contribution of error source k, e i e j Let Cov() be the error variables from two different error sources. For temperature versus variable z k Jacobian matrix, For variable z k variance This is the error attenuation coefficient. To start from error source z k to temperature field All computational paths, The error absorption rate on path p. The normalized weights for path p.

[0036] The beneficial effects achieved by this invention are:

[0037] This system achieves dynamic optimization of the observation network by calculating the conditional entropy and information gain of temperature field reconstruction. This overcomes the problem of insufficient spatial representativeness caused by uniform or empirical sensor deployment in existing technologies. It can automatically identify key monitoring areas and incrementally deploy sensors based on the uncertainty distribution of the temperature field, significantly improving the spatiotemporal coverage efficiency and cost-effectiveness of ground observation data. By using acyclic constraints and sparse regularization to ensure the rationality of the causal graph, it achieves automatic knowledge mining and dynamic updating, providing reliable physical prior knowledge and logical reasoning support for temperature reconstruction. By constructing a location-related weight function, it dynamically adjusts the strength of physical constraints based on local three-dimensional geometric features, underlying surface type, and the relative ratio of physical model residuals to data fitting gradients, achieving an intelligent balance between physical knowledge and data-driven approaches. By constructing a directed computation graph and tracking the propagation path of errors among nodes such as data input, physical constraints, and model parameters, it calculates the contribution of each error source and its attenuation coefficient absorbed by physical constraints, overcoming the deficiency of existing uncertainty quantification methods that can only provide the overall variance but cannot identify the dominant error sources.

[0038] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall structural framework of the present invention;

[0040] Figure 2 This is a schematic diagram of the multi-source heterogeneous data intelligent acquisition and preprocessing module of the present invention;

[0041] Figure 3 This is a schematic diagram of the urban thermal environment multidimensional feature knowledge modeling module of the present invention;

[0042] Figure 4 This is a schematic diagram of the physical constraint deep learning surface temperature reconstruction module of the present invention.

[0043] Figure 5 This is a schematic diagram of the intelligent monitoring, early warning, and decision support module of the present invention;

[0044] Figure 6 This is a schematic diagram comparing the spatial resolution improvement effect of the present invention with other methods;

[0045] Figure 7 This is a schematic diagram showing the decomposition and comparison of the sources of uncertainty in this invention;

[0046] Figure 8 This is a schematic diagram of the interactive interface information of the present invention. Detailed Implementation

[0047] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0048] Example 1.

[0049] A high spatiotemporal resolution surface temperature reconstruction system for refined monitoring of urban thermal environment, combined with Figure 1 It includes a multi-source heterogeneous data intelligent acquisition and preprocessing module, an urban thermal environment multi-dimensional feature knowledge modeling module, a physical constraint deep learning surface temperature reconstruction module, and an intelligent monitoring, early warning and decision support module;

[0050] The multi-source heterogeneous data intelligent acquisition and preprocessing module is responsible for collecting and preprocessing various remote sensing and ground observation data, providing a high-quality data foundation for subsequent modeling. The urban thermal environment multi-dimensional feature knowledge modeling module constructs a knowledge system of urban underlying surface, three-dimensional morphology and thermal processes, providing physical constraints and prior knowledge for temperature reconstruction. The physical constraint deep learning surface temperature reconstruction module integrates physical models and artificial intelligence technology to achieve accurate reconstruction of surface temperature with high spatiotemporal resolution. The intelligent monitoring, early warning and decision support module transforms the reconstruction results into visualization, risk warning and control decision suggestions, supporting the intelligent management of the urban thermal environment.

[0051] Combination Figure 2 The multi-source heterogeneous data intelligent acquisition and preprocessing module includes three units: a multi-scale remote sensing data intelligent acquisition unit, a ground observation network adaptive deployment unit, and a heterogeneous data quality control and registration unit. The multi-scale remote sensing data intelligent acquisition unit is responsible for collaboratively acquiring multispectral imagery and thermal infrared observation data, and dynamically scheduling sensor resources according to urban thermal environment monitoring needs to ensure the spatiotemporal coverage integrity of the data. The ground observation network adaptive deployment unit uniformly manages meteorological stations, IoT sensors, and mobile observation platforms, and achieves optimal spatial deployment of sensors through observation network optimization algorithms to improve the representativeness of ground truth. The heterogeneous data quality control and registration unit performs spatiotemporal registration, format standardization, and quality screening on multi-source data to eliminate fusion barriers caused by data heterogeneity and provide clean and aligned data input for subsequent modeling.

[0052] Combination Figure 3 The urban thermal environment multidimensional feature knowledge modeling module includes three units: an urban underlying surface multidimensional attribute modeling unit, an urban three-dimensional morphology and radiation transfer modeling unit, and an urban thermal environment process knowledge graph construction unit. The urban underlying surface multidimensional attribute modeling unit not only classifies land cover types such as roads, buildings, vegetation, and water bodies, but also extracts thermophysical parameters such as thermal inertia, roughness, and evapotranspiration capacity of various land cover types to construct a land cover-thermal attribute knowledge base, allowing for the selection of differentiated temperature reconstruction strategies for different underlying surface types. The urban three-dimensional morphology and radiation transfer modeling unit constructs three-dimensional geometric fields such as building height, street-valley width-to-height ratio, and sky visibility factor, and simulates the radiation transfer process and shadow dynamic evolution of building clusters to improve the temperature reconstruction accuracy in urban canyon areas and high-density areas. The urban thermal environment process knowledge graph construction unit integrates energy budget mechanisms, human activity heat sources, and meteorological driving factors to construct a three-layer knowledge graph containing entities, relationships, and rules, providing physical constraints and logical reasoning support for temperature reconstruction.

[0053] Combination Figure 4The physically constrained deep learning surface temperature reconstruction module comprises three units: a physics-guided deep neural network reconstruction unit, a multi-temporal spatiotemporal fusion and dynamic correction unit, and an uncertainty quantification and multi-scenario generation unit. The physics-guided deep neural network reconstruction unit embeds the urban energy balance equation into the loss function of the deep learning model and employs physical information neural networks or knowledge distillation methods to achieve surface temperature reconstruction that combines physical consistency with high spatiotemporal resolution. The multi-temporal spatiotemporal fusion and dynamic correction unit integrates low-resolution thermal infrared, high-resolution optical images, and ground observation data, introduces a spatiotemporal attention mechanism to capture the spatiotemporal evolution of the temperature field, and performs real-time deviation correction based on ground observations to ensure the accuracy and continuity of the reconstruction results. The uncertainty quantification and multi-scenario generation unit assesses the uncertainty of the temperature reconstruction results, outputs confidence intervals, and generates multiple schemes such as optimal estimates, conservative estimates, and extreme scenarios, providing a scientific basis for urban thermal environment risk assessment and emergency decision-making.

[0054] Combination Figure 5 The intelligent monitoring, early warning, and decision support module comprises three units: a multi-dimensional spatiotemporal visualization and interactive analysis unit, a thermal risk intelligent early warning and source tracing analysis unit, and an urban thermal environment regulation and decision support unit. The multi-dimensional spatiotemporal visualization and interactive analysis unit provides various visualization formats such as two-dimensional maps, three-dimensional urban thermal fields, spatiotemporal cubes, and thermal field evolution animations. It supports user-defined queries, multi-time period comparisons, and interactive analysis, intuitively displaying the spatial distribution and temporal variation characteristics of the urban thermal environment. The thermal risk intelligent early warning and source tracing analysis unit automatically identifies the core area of ​​the heat island and high-risk areas, conducts source tracing analysis of the causes of thermal anomalies, classifies early warning levels, and achieves short-term prediction, providing early warning information for urban thermal health risk management. The urban thermal environment regulation and decision support unit simulates the cooling effects of intervention measures such as increased greening and construction of ventilation corridors based on the reconstruction results, uses optimization algorithms to generate heat mitigation strategy suggestions, and interfaces with the urban planning and emergency management system to provide decision support for the scientific regulation and sustainable development of the urban thermal environment.

[0055] The multi-scale remote sensing data intelligent acquisition unit includes a multispectral image acquisition processor, a thermal infrared data collaborative acquisition processor, and a task-driven scheduling processor. The multispectral image acquisition processor is responsible for acquiring satellite image data in the visible, near-infrared, and short-wave infrared bands, extracting key parameters such as vegetation index and surface albedo, and providing spatial texture features and underlying surface classification basis for surface temperature reconstruction. The thermal infrared data collaborative acquisition processor acquires thermal infrared observation data from multiple platforms, providing absolute thermal intensity information of surface temperature at different spatiotemporal resolutions, and constructing a core benchmark dataset for temperature reconstruction. The task-driven scheduling processor dynamically selects the optimal data source combination and adjusts the acquisition frequency based on the real-time needs of urban thermal environment monitoring and data quality assessment results, ensuring the integrity of data coverage in key areas and time periods.

[0056] The adaptive deployment unit of the ground observation network includes a multi-source sensor access processor, an observation network optimization processor, and a mobile observation co-processor. The multi-source sensor access processor uniformly manages the data access of fixed observation equipment such as meteorological stations, road surface temperature sensors, and building exterior temperature probes, providing meteorological driving parameters such as air temperature, relative humidity, wind speed, and wind direction, as well as high-precision true values ​​of surface temperature. The observation network optimization processor calculates the optimal placement and density of sensors using spatial optimization algorithms based on the heterogeneity of the urban underlying surface, the distribution characteristics of the heat island, and the existing sensor layout, maximizing the spatial representativeness and cost-effectiveness of the observation network. The mobile observation co-processor integrates data from mobile observation vehicles, drones, and other mobile observation platforms to supplement the spatial blind spots of fixed stations and achieves collaborative observation between mobile platforms and the fixed network through trajectory planning algorithms.

[0057] The observation network optimization processor determines the sensor deployment locations according to the following formula:

[0058] ;

[0059] ;

[0060] Among them, S opt This represents the optimal sensor deployment location, where x represents a candidate sensor deployment location. T represents the set of all possible candidate locations for deploying sensors. s For the temperature field of the entire study area, D cur H[T] represents the currently existing observation dataset. s [D] represents the temperature field T given the observed data D. s conditional entropy, This represents the temperature value observed at location x. This represents the expected value of the observed value. This is the cost weighting coefficient. This represents the cost of deploying the sensor at x;

[0061] The heterogeneous data quality control and registration unit includes a spatiotemporal registration processor, a data quality assessment processor, and a format standardization processor. The spatiotemporal registration processor performs spatial reprojection and temporal synchronization on remote sensing and ground observation data from different sources, coordinate systems, and time bases to eliminate spatial misalignment and temporal deviation caused by data heterogeneity and establish a unified spatiotemporal reference framework. The data quality assessment processor performs integrity checks, outlier identification, and signal-to-noise ratio evaluation on the raw data, removes low-quality data caused by cloud contamination, sensor failures, etc., and assigns quality weights to each data source for subsequent fusion. The format standardization processor converts multi-source heterogeneous data into a unified data format and storage structure, including image format conversion, attribute field mapping, and metadata standardization, providing a plug-and-play data interface for subsequent modeling.

[0062] The urban underlying surface multidimensional attribute modeling unit includes a land cover classification processor, a thermophysical parameter inversion processor, and a land cover-thermal attribute knowledge base builder. The land cover classification processor uses the spectral index, texture features, and high-resolution imagery of multispectral images, and employs object-oriented classification or deep learning semantic segmentation methods to generate high-precision land cover classification maps for types such as roads, buildings, vegetation, water bodies, and bare land. The thermophysical parameter inversion processor, for each land cover type, inverts thermophysical parameters such as thermal inertia, surface roughness, evapotranspiration capacity, and specific heat capacity based on remote sensing observation and radiative transfer models to quantify the thermal response characteristics of different underlying surfaces. The land cover-thermal attribute knowledge base builder associates and stores land cover types with corresponding thermophysical parameters, establishing a ternary knowledge base of "land cover type - thermal attribute parameter - temperature response mode" to provide differentiated physical parameter inputs for subsequent temperature reconstruction.

[0063] The urban 3D morphology and radiation transmission modeling unit includes a 3D geometry reconstruction processor, a radiation transmission simulation processor, and a shadow dynamic evolution processor. The 3D geometry reconstruction processor integrates high-resolution stereo images, LiDAR point clouds, and building vector data to construct a 3D geometric model of urban buildings and extract morphological parameters such as building height, building density, street-valley aspect ratio, and sky visibility factor. The radiation transmission simulation processor, based on the 3D urban model and the geometric relationship of solar radiation, simulates the multiple reflections of solar shortwave radiation on building surfaces and the received radiation on the ground, calculates the net radiation flux at different times and locations, and quantifies the modulation effect of building clusters on the radiation field. The shadow dynamic evolution processor, based on the temporal changes of solar altitude angle and azimuth angle, simulates the projection range and movement trajectory of building shadows on the ground surface, identifies continuous shadow areas and alternating sunshine areas, and provides spatiotemporal dynamic constraints for radiation shading for temperature reconstruction.

[0064] The urban thermal environment process knowledge graph construction unit includes an energy budget modeling processor, a human activity heat source identification processor, and a knowledge graph inference engine. The energy budget modeling processor establishes the urban surface energy balance equation, parameterizes net radiation, sensible heat flux, latent heat flux, soil heat flux, and anthropogenic heat emissions, and constructs physical constraint rules for energy budget to ensure that the temperature reconstruction results conform to the first law of thermodynamics. The human activity heat source identification processor identifies and quantifies the spatial distribution and temporal variation of anthropogenic heat contributions such as traffic heat emissions, air conditioning waste heat, and industrial heat sources based on traffic flow data, building energy consumption statistics, and high-resolution nighttime light images, supplementing urban heat source items beyond natural thermal processes. The knowledge graph inference engine integrates ground entities, thermophysical attributes, energy budget rules, and human activity elements to construct a three-layer knowledge graph of "entity-relationship-rule". Through graph inference technology, it realizes logical constraints and anomaly detection of the temperature field, providing prior knowledge for physics-guided deep learning.

[0065] The knowledge graph reasoning engine constructs a temporal causal graph of the urban thermal environment. Where V is the node set, Let B be the causal edge set, B be the node attributes, and the strength of causal relationships be learned according to the following loss function:

[0066] ;

[0067] ;

[0068] Among them, L cau Let V be the total loss function for causal relationship learning. i V j Let ) be a directed edge in the causal graph. For attention weights, represent the weights from V i To V j The causal strength, V i (t) represents the value of Vi at time t. For causal time lag, This represents the observed true causal effect. To regularize sparsity, A is the adjacency matrix. h is the acyclic constraint coefficient, and h() is the acyclic constraint function. For V i eigenvectors, W Q For querying the matrix, v j For V j eigenvectors, W K Let d be the key matrix, d be the dimension of the query matrix and the key matrix, and N(j) represent v. j The set of candidate parent nodes;

[0069] The workflow of the knowledge graph reasoning engine is as follows: input historical observation data, calculate the true causal effect, and randomly initialize the learnable parameter W. Q W K and embed v i and v j The attention weights and loss function are calculated sequentially, and W is optimized using gradient descent. Q W K Output the learned causal graph G;

[0070] The physical-guided deep neural network reconstruction unit includes a physical information embedding processor, a multi-scale feature extraction processor, and a physical consistency verification processor. The physical information embedding processor transforms physical constraints such as the urban energy balance equation and radiative transfer equation into loss function terms or soft constraint layers of the deep learning model. It adopts a physical information neural network architecture or knowledge distillation strategy to ensure that the temperature field output by the neural network meets basic physical laws. The multi-scale feature extraction processor is used to extract multi-scale spatial features and temporal features from low-resolution thermal infrared, high-resolution optical images, and auxiliary geospatial data, and learns the spatial heterogeneity and temporal evolution pattern of the temperature field. The physical consistency verification processor performs energy conservation checks, temperature gradient rationality checks, and extreme value anomaly identification on the temperature field output by the neural network. When physical inconsistencies are detected, it triggers model parameter adjustments or post-processing corrections to ensure the physical interpretability of the reconstruction results.

[0071] The physical information embedding processor constructs a spatially varying learnable function according to the following formula:

[0072] ;

[0073] ;

[0074] ;

[0075] Among them, L total For the total loss, L data R is the data fitting loss. m [] represents the m-th physical constraint residual. Let M be the predicted temperature at location x at time t, M be the total number of physical constraints, and w be the predicted temperature at location x at time t. m (x) represents the adaptive weight of the m-th constraint at position x. For the Sigmoid activation function, MLP m It is a small multilayer sensor. It is a three-dimensional geometric feature vector. This is the feature vector of the underlying surface type. This is the relative ratio of the physical residual to the data gradient. As the global baseline weight, For about gradient, For all learnable parameters, It is a numerically stable term;

[0076] The multi-temporal spatiotemporal fusion and dynamic correction unit includes a spatiotemporal attention fusion processor, a missing data filling processor, and a ground truth correction processor. The spatiotemporal attention fusion processor uses a spatiotemporal attention mechanism to capture the correlation between images of different time phases and the spatiotemporal continuity of the temperature field. It fuses the temporal information of low-resolution thermal infrared with the spatial details of high-resolution optical images to generate a high spatiotemporal resolution temperature reconstruction result. The missing data filling processor addresses the gaps in thermal infrared observation caused by cloud cover and sensor scanning intervals by using methods such as temporal interpolation, spatial neighborhood similarity matching, and generative adversarial networks to fill in the temperature values ​​in the missing areas, forming a spatiotemporally continuous and complete temperature field. The ground truth correction processor utilizes high-precision observation data from meteorological stations and ground sensors to correct the reconstructed temperature field in real time through local bias correction, regression calibration, or Bayesian update methods, eliminating systematic biases and improving absolute accuracy.

[0077] The uncertainty quantification and multi-scenario generation unit includes an uncertainty assessment processor, a confidence interval calculation processor, and a multi-scenario simulation processor. The uncertainty assessment processor, based on input data quality, model parameter sensitivity, and the spatial distribution of reconstruction errors, uses Monte Carlo simulation, ensemble learning, or Bayesian inference methods to quantify the sources and propagation paths of uncertainty in temperature reconstruction results. The confidence interval calculation processor, based on the uncertainty assessment results, calculates the confidence interval or probability density distribution for the reconstructed temperature of each pixel, providing a reliability index for temperature estimation and supporting risk-based decision-making. The multi-scenario simulation processor, combining uncertainty distribution and extreme weather condition assumptions, generates various temperature field schemes such as optimal estimation scenarios, conservative estimation scenarios, and extreme high-temperature scenarios, providing scenario support for risk assessment of urban thermal environment, emergency plan formulation, and climate adaptation planning.

[0078] The uncertainty assessment processor calculates the uncertainty contribution of each step according to the following formula:

[0079] ;

[0080] ;

[0081] ;

[0082] in, To reconstruct the total uncertainty of the temperature field, For the reconstructed temperature field, S is the set of error sources. For the contribution of error source k, e i e j Let Cov() be the error variables from two different error sources. For temperature versus variable z k Jacobian matrix, For variable z k variance This is the error attenuation coefficient. To start from error source z k to temperature field All computational paths, The error absorption rate on path p. The normalized weights for path p;

[0083] The multidimensional spatiotemporal visualization and interactive analysis unit includes a two-dimensional thermal field rendering processor, a three-dimensional spatiotemporal cube builder, and an interactive query processor. The two-dimensional thermal field rendering processor maps the reconstructed surface temperature data into a color-intuitive heat map, supports multi-time period overlay comparison, thermal anomaly highlighting, and temperature contour line drawing, and provides the spatial distribution characteristics of the urban thermal environment from a planar perspective. The three-dimensional spatiotemporal cube builder integrates the spatial and temporal dimensions of temperature data into a three-dimensional volumetric structure, and displays the spatiotemporal evolution of the temperature field through volume drawing or slicing technology, supporting users to dynamically browse and perform profile analysis along the time axis or spatial axis. The interactive query processor provides multidimensional search functions based on spatial location, time range, temperature threshold, and land cover type, and supports users to customize temperature statistics for areas of interest, extract historical change curves, and perform multi-scenario comparative analysis.

[0084] The intelligent early warning and source tracing analysis unit for thermal risks includes a thermal risk identification processor, a causal source tracing analysis processor, and an early warning level classification processor. The thermal risk identification processor, based on the spatial superposition of reconstructed temperature field with population density and vulnerable population distribution, uses threshold models, hotspot clustering algorithms, or machine learning classifiers to automatically identify the core area of ​​the heat island, high-temperature exposure risk areas, and thermally vulnerable areas, generating a spatial distribution map of thermal risks. The causal source tracing analysis processor, combining urban underlying surface attributes, three-dimensional morphological parameters, energy budget simulation results, and meteorological driving factors, uses causal inference or contribution decomposition methods to determine whether the dominant cause of high-temperature anomalies is underlying surface type, building density, radiation shielding, anthropogenic heat emissions, or unfavorable meteorological conditions, providing a diagnostic basis for precise regulation. The early warning level classification processor, based on absolute temperature value, duration, impact range, and historical extreme value statistics, combined with meteorological forecast information, performs short-term temperature prediction, automatically classifies warning levels into blue, yellow, orange, and red, and triggers corresponding early warning information release and emergency response processes.

[0085] The urban thermal environment regulation decision support unit includes an intervention measure simulation processor, a regulation strategy optimization processor, and a decision interface server. The intervention measure simulation processor, based on the reconstructed temperature field and urban thermal environment knowledge graph, simulates the cooling effect and spatial impact range of heat mitigation measures such as increasing green space, expanding water bodies, whitening roofs, and constructing ventilation corridors. It uses scenario comparison analysis to evaluate the cost-effectiveness of different intervention schemes. The regulation strategy optimization processor, based on the criteria of maximizing cooling effect, minimizing cost, or multi-objective comprehensive optimization, uses intelligent optimization methods such as genetic algorithms, particle swarm optimization, or reinforcement learning to generate the optimal strategy combination for thermal environment regulation, such as green space layout and ventilation corridor site selection, and outputs the implementation priority ranking and expected effect evaluation. The decision interface server provides a standardized API interface and data exchange protocol to push temperature reconstruction results, risk warning information, and regulation suggestions to the urban management platform, planning information system, emergency command center, and public service APP, realizing seamless connection and real-time information sharing between the system and the urban governance system.

[0086] The i, j, and k mentioned above are ordinal numbers used to represent sequence numbers and have no actual meaning.

[0087] To verify the effectiveness, an area of ​​approximately 100 square kilometers in the central urban area of ​​a certain city was selected as the study area for the experiment, covering various underlying surface types such as high-density building areas, parks and green spaces, water bodies and roads.

[0088] Data source configuration: Landsat-8 thermal infrared data, Sentinel-2 multispectral data, and MODIS daily temperature products were collected as remote sensing inputs, and 50 meteorological stations and 120 road surface temperature sensors were deployed as a ground truth verification network.

[0089] Comparison Method Setup: The patented method was compared with three existing technologies: the ESTARFM spatiotemporal fusion algorithm, the random forest downscaling method, and the standard physical information neural network (PINN) method. All methods used the same input and validation datasets, and the training time was uniformly 48 hours.

[0090] The first step involves running the observation network optimization processor to calculate the information gain based on the initial temperature field uncertainty distribution and selecting 20 optimal locations from the candidate sites to add sensors. The second step involves constructing a time-series causal graph, inputting six months of historical observation data, and learning the causal relationships between 15 variables, including underlying surface type, building density, vegetation cover, and net radiation, iterating 5000 times until the loss function converges. The third step involves initializing the physical information embedding processor, setting five physical constraints: energy balance equation, heat diffusion equation, radiative transfer constraint, boundary layer turbulence constraint, and water evapotranspiration constraint, and training an adaptive weight network to dynamically adjust the constraint strength based on local features. The fourth step involves running a multi-temporal spatiotemporal fusion processor to fuse 30 periods of Landsat-8 imagery, 180 periods of Sentinel-2 imagery, and 180 periods of MODIS data, generating temperature field sequences with 10-meter spatial resolution and daily temporal resolution. The fifth step involves starting the uncertainty assessment processor, constructing a computational graph containing data input nodes, physical constraint nodes, and model parameter nodes, tracing error propagation paths, and calculating the contribution of each error source. The sixth step is to select independent validation period data to evaluate the reconstruction accuracy, physical consistency and computational efficiency of each method.

[0091] The data was organized and obtained Figure 6 and Figure 7 .

[0092] Example 2.

[0093] A high spatiotemporal resolution surface temperature reconstruction system for refined monitoring of urban thermal environment includes a multi-source heterogeneous data intelligent acquisition and preprocessing module, an urban thermal environment multi-dimensional feature knowledge modeling module, a physically constrained deep learning surface temperature reconstruction module, and an intelligent monitoring, early warning, and decision support module. The system is deployed on a high-performance computing server cluster equipped with dual Intel Xeon Gold 6248R processors, 512GB DDR4 memory, and eight NVIDIA A100 GPUs. It is implemented using the Ubuntu 22.04 LTS operating system and the PyTorch 2.0 deep learning framework and is applied to urban thermal environment monitoring in an area of ​​approximately 667 square kilometers.

[0094] The multi-source heterogeneous data intelligent acquisition and preprocessing module is responsible for collecting and preprocessing remote sensing and ground observation data. It acquires Landsat-9 thermal infrared band TIRS-2 data to obtain a 100-meter spatial resolution surface temperature reference field with a 16-day revisit period, and simultaneously acquires Sentinel-2 data. MSI multispectral imagery extracts NDVI (Vegetation Index), NDBI (Building Index), and surface albedo at 10-meter and 20-meter resolutions using a 5-day revisit sequence. It subscribes to 16-meter resolution data from the Gaofen-6 WFV wide-field camera as a supplement to domestic satellite data. It integrates 2-kilometer resolution 10-minute interval observations from the Himawari-9 geostationary meteorological satellite's AHI sensor to provide high temporal resolution information. It also accesses 70-meter resolution intermittent transit data from the ECOSTRESS space station payload. The task-driven scheduling processor dynamically adjusts the data acquisition frequency based on the high-temperature warning level issued by the meteorological bureau. When an orange warning is issued, the Sentinel-2 download frequency is increased from every 5 days to every 2 days. When the temperature reconstruction uncertainty in a certain area exceeds 2.5°C, it automatically triggers supplementary high-resolution image acquisition. This unit achieves API integration with the China Meteorological Data Network and the Geospatial Data Cloud, supporting parallel downloading of multi-source data.

[0095] The adaptive deployment unit of the ground observation network consists of a three-dimensional observation network comprising 65 fixed meteorological stations, 180 road surface temperature sensors, 45 building exterior wall temperature probes, and 12 mobile observation vehicles. The fixed meteorological stations use Vaisala WXT536 multi-parameter meteorological sensors to measure air temperature, relative humidity, wind speed, and wind direction at a height of 2 meters, with a temperature measurement accuracy of ±0.3°C and data transmission in real-time via 4G network at a 1-minute sampling interval. The road surface temperature sensors use German Lufft IRS31 non-contact infrared thermometers installed at a height of 2.5 meters on streetlight poles, with a measurement range of -40 to 80°C, an accuracy of ±0.5°C, and a field of view of 10°. The sensor deployment density reaches 8 per square kilometer in the CBD core area and decreases to 2 per square kilometer in the urban fringe area. The building exterior wall temperature probes use FLIR A310f thermal imaging cameras to capture thermal infrared images of building facades and street canyons every 10 minutes at a resolution of 320×240 pixels. The mobile observation vehicles are equipped with Apogee... The SI-111 precision infrared radiometer conducts mobile observations at 6:00 AM, 12:00 PM, and 6:00 PM according to a preset route. Each vehicle covers an average of 80 kilometers and approximately 1200 measurement points per day. The observation network optimization processor adopts a hybrid strategy of improved greedy algorithm and simulated annealing algorithm. Based on the initial 65 meteorological stations, 12 high uncertainty areas were identified. After three rounds of optimization iterations, the optimal locations for adding 20 sensors were determined, reducing the average uncertainty of the area from 2.8°C to 1.6°C. Compared with uniform and dense deployment, this saves 35% of the number of sensors.

[0096] The observation network optimization processor employs an uncertainty quantification method based on Gaussian process regression, dividing the study area into a 100m × 100m grid of approximately 67,000 grid cells. The candidate sensor deployment location set includes 1,500 accessible locations such as all public building rooftops, municipal streetlights, and bus stops. Information gain calculation uses a Monte Carlo sampling method to extract 1,000 possible observations from historical temperature distributions. Sensor deployment costs are comprehensively considered, including equipment purchase cost of 3,500 yuan, installation and construction cost of 1,200 yuan, annual maintenance cost of 800 yuan, and power supply and communication convenience scores. The cost weighting coefficient is set to 0.15. In each iteration of the optimization algorithm, the information gain and cost ratio are selected. Sensors were deployed at locations with the highest values. After 20 iterations, the temperature uncertainty of 95% of the grid cells was controlled within 2°C. Compared with empirical deployment schemes, the uncertainty was reduced by 42% and the total number of sensors was reduced by 28. The processor also supports seasonal adjustment strategies. During the high-temperature period in summer, 12 sensors located in parks, green spaces and water bodies were temporarily relocated to densely built-up areas to enhance monitoring of the core heat island area. As a variation of the technical means, information gain calculation can be replaced by a sensor selection algorithm based on maximizing mutual information. Uncertainty quantification can use the prediction variance of Bayesian neural networks as an alternative indicator. Optimization algorithms can include swarm intelligence algorithms such as Particle Swarm Optimization (PSO) or Differential Evolution (DE).

[0097] The heterogeneous data quality control and registration unit uses the GDAL 3.6 geospatial data processing library to uniformly convert all remote sensing images to the WGS84 coordinate system UTM 50N projection and resample them to a 10-meter standard raster. The data quality assessment processor calculates 12 quality indicators for each remote sensing image, including cloud cover ratio, atmospheric transmittance, and radiometric calibration coefficient deviation. When the cloud cover exceeds 15%, the cloud detection algorithm Fmask is activated. 4.6 Marking cloud pixel locations: For ground sensor data, outlier identification rules are established, including physical range checks (temperature between -30°C and 60°C), rate of change checks (temperature change between adjacent times not exceeding 5°C), and spatiotemporal consistency checks (temperature difference with surrounding stations not exceeding 8°C). Sensor fault detection uses an LSTM-based time series prediction model. When the deviation between observed and predicted values ​​exceeds three times the standard deviation for six consecutive hours, a sensor fault alarm is automatically triggered. A format standardization processor converts Landsat's GeoTIFF format, Sentinel-2's SAFE format, and Himawari's NetCDF format to HDF5 format for storage. A block storage strategy is adopted, with each data block containing 1024×1024 pixels, combined with the LZW lossless compression algorithm. The data interface layer uses the OGC standard WMS and WCS service protocols to support access to stored temperature data products via standard HTTP requests. As a variant implementation, the projection conversion can use Lambert's conformal conic projection (LCC) instead of UTM projection. The cloud detection algorithm can use a deep learning-based CDnet or CloudSEN12 model. The data storage format can adopt Cloud Optimized... GeoTIFF format.

[0098] The urban thermal environment multidimensional feature knowledge modeling module includes an urban underlying surface multidimensional attribute modeling unit, an urban three-dimensional morphology and radiation transfer modeling unit, and an urban thermal environment process knowledge graph construction unit. The urban underlying surface multidimensional attribute modeling unit utilizes 1-meter resolution panchromatic and 4-meter resolution multispectral images from the Gaofen-2 PMS sensor acquired in December 2023 for fine land cover classification. It employs a DeepLabv3+ deep learning semantic segmentation model, based on ImageNet pre-trained weights, and uses 12,000 manually labeled samples from the Beijing area for transfer processing. The training batch size was 16, the learning rate was 0.0001, and the optimizer was AdamW. After 80 rounds of training, the model achieved an overall classification accuracy of 92.3% on the validation set. The study area was divided into 6 major categories and 15 subcategories: high-density built-up land (31.2%), low-density built-up land (23.8%), road land (16.5%), vegetated land (18.4%), water bodies (6.3%), and bare land (3.8%). For each land cover type, thermophysical parameters were obtained from literature databases and field measurements. The average thermal inertia of the high-density built-up area was set to 2400. The parameters, including J / (m²·K·s^0.5), surface albedo 0.18, roughness length 1.2 meters, and evapotranspiration coefficient 0.15, are stored in a dedicated table in the PostgreSQL database to establish a mapping relationship between land cover type codes and parameter values. This unit also integrates a building database provided by the Beijing Municipal Commission of Planning and Natural Resources, containing the construction year, building function, and exterior wall material properties of approximately 350,000 buildings. Based on the construction year, buildings are divided into three categories: brick-concrete structures before 1980, concrete structures from 1980 to 2000, and energy-saving buildings after 2000. The thermal conductivity of the walls of buildings in different years is set to 1.2, 0.8, and 0.45 W / (m·K), respectively. This refined parameter differentiation improves the temperature reconstruction accuracy of densely built areas by 18%. As an alternative, land cover classification can use a random forest algorithm combined with the spectral time-series characteristics of multi-temporal Sentinel-2 images, and thermophysical parameter inversion can use ASTER's five thermal infrared bands for temperature emissivity separation algorithm (TES) to directly extract parameters from remote sensing data.

[0099] The urban 3D morphology and radiation transmission modeling unit uses 2022 airborne LiDAR point cloud data to construct a 3D urban model within Beijing's Fifth Ring Road. The point cloud data was acquired by a Riegl VQ-1560i airborne LiDAR system at a flight altitude of 600 meters, a scanning frequency of 550 kHz, and a point density of 20 points per square meter. TerraScan software was used for point cloud classification, and the RANSAC plane fitting algorithm was used to extract building roof surfaces. The alpha-shape algorithm was then used to generate building outline vectors. The final 3D model contains precise geometric information for 352,847 buildings. Building height statistics show an average height of 23.7 meters, with the tallest building, the China Zun Tower, at 528 meters. Street valley width-to-height ratio (H / W) calculations show that the average H / W in the CBD core area reaches 2.3, classifying it as a deep street valley, while the average H / W in the suburbs is 0.6, classifying it as an open street valley. The sky visibility factor (SVF) was calculated using a fish-shaped... The hemispherical projection method from the eye's perspective emits 1000 uniformly distributed rays upwards from the center point of each 10-meter grid. The proportion of rays not blocked by buildings is statistically analyzed. The results show that the SVF (Scattered Radiation Frequency) in the CBD area is as low as 0.35, while the SVF in open park areas is as high as 0.95. A radiative transfer simulation processor uses the DART discrete anisotropic radiative transfer model for three-dimensional radiation simulation. Hourly radiation simulations are performed for the summer solstice on July 21, 2024, considering multiple reflections between building facades and tracking up to five reflection processes. The simulation results show that at noon, the net radiation in the building shadow area is approximately 400 μm lower than that in the direct sunlight area. The W / m² shadow dynamic evolution processor overlays and analyzes the 24-hour shadow distribution sequence to identify continuous shadow areas (areas with a total daily sunshine duration of less than 4 hours) accounting for 8.3% of the total area, alternating sunshine areas accounting for 43.2%, and full sunshine areas accounting for 48.5%. These morphological parameters show significant spatial correlation with the temperature field. Analysis shows that the correlation coefficient between H / W and surface temperature is r=0.62, and the correlation coefficient between SVF and temperature is r=-0.58, both of which pass the 0.01 significance test. As a deformation implementation method, 3D modeling can use oblique photogrammetry to generate a realistic 3D model, and the radiation transfer simulation can use the SOLWEIG model specifically for radiation calculation in urban canyon environments.

[0100] The urban thermal environment process knowledge graph construction unit integrates 23 thermal environment-related variables to construct a time-series causal graph, including 5 meteorological driving variables, 8 underlying surface characteristic variables, 6 energy flux variables, and 4 temperature variables. The energy budget modeling processor adopts a simplified urban energy balance model using Town Energy Balance, dividing the urban surface into three parts: roof surface, wall surface, and road surface, and establishing energy balance equations for each. Anthropogenic heat flux is estimated based on energy consumption statistics. The anthropogenic heat flux in the CBD business district reaches 150 W / m² during weekdays, mainly from air conditioning exhaust. The human activity heat source identification processor uses the DNB band of nighttime light data NPP-VIIRS as a proxy variable for anthropogenic heat intensity to establish a statistical regression model between light intensity and anthropogenic heat flux, with a coefficient of determination R²=0.76. Traffic heat emissions are estimated based on road network traffic flow data from the Beijing Municipal Commission of Transport. The heat emission rate per vehicle is calculated as 8 kW, and the traffic heat emissions on main roads such as Chang'an Avenue reach 200-300 W / m². The knowledge graph inference engine uses Neo4j. 5.3 The graph database stores entity nodes and relation edges. Relation edges define 56 types of causal relationships, such as increased building density leading to increased sensible heat, and increased vegetation cover leading to enhanced evapotranspiration. The rule layer contains 18 physical rules, such as the energy balance must be closed and temperature cannot change abruptly. Graph reasoning uses the Cypher query language to implement multi-hop path search. This unit updates the knowledge graph hourly to incorporate the latest observation data and reconstruction results. As a technological transformation method, the energy balance model can adopt more complex SUEWS or LUCY models. Human heat estimation can adopt a bottom-up approach based on building energy consumption simulation using EnergyPlus software. The knowledge graph can be extended to knowledge graph embedding models such as ComplEx or RotatE.

[0101] The training process of the knowledge graph inference engine uses historical observation data from July to December 2023 (6 months total), comprising 720 time slices across 4 time periods daily, to form approximately 15 million training samples. A sliding window method is used to extract time lag relationships and calculate partial correlation coefficients for 6 time lags from 0 to 24 hours. A significant causal relationship is considered to exist when the absolute value of the partial correlation coefficient is greater than 0.3 and the p-value is less than 0.01. The initial causal graph contains 23 nodes and 342 edges, which are further filtered through Granger causality tests to select 128 robust causal edges. The query matrix and key matrix are set to 64 dimensions. The Adam optimizer with a learning rate of 0.001 is used, and the loss function includes a causal fitting loss weight of 1.0, an L1 sparse regularization weight of 0.01, and an acyclic constraint weight of 0.1. The training batch size is 128, and the total number of iterations is 5000 rounds, requiring approximately 12 hours on a single A100 block. The study was performed on a GPU. The resulting causal graph showed that the causal strength of vegetation cover on surface temperature was -0.68, indicating a negative causal relationship, while the causal strength of building density on temperature was +0.52. In the temperature anomaly diagnosis task, the system was able to correctly identify the dominant cause in 85% of the cases. In the temperature prediction task, the addition of causal graph constraints reduced the RMSE of the 24-hour prediction from 2.8°C to 2.1°C, improving it by 25%.

[0102] The physical constraint deep learning land surface temperature reconstruction module includes a physical information embedding processor, a multi-scale feature extraction processor, and a physical consistency verification processor. The physical information embedding processor uses a residual U-Net architecture as the base network, and the encoder part uses the first four residual blocks of ResNet-50 as the pre-trained backbone network. The input consists of five channels of data: Landsat-9 thermal infrared brightness temperature, Sentinel-2 red-near-infrared-shortwave infrared bands, and a digital elevation model (DEM). The input size is 512×512 pixels, corresponding to an actual geographical range of 5.12×5.12 kilometers. The model includes five physical constraint terms: energy balance residual, thermal diffusion residual, radiative transfer residual, boundary layer stability residual, and water vapor balance residual. The data fitting loss uses Huber loss. The function uses squared loss when the deviation between predicted and observed temperatures is less than 1°C and linear loss when it is greater than 1°C. The key innovation lies in the fact that the weights of the physical constraints are not fixed values ​​but are adaptively learned through a small MLP network. The MLP input includes nine geometric and underlying surface features, such as building height, street-valley aspect ratio, and sky visibility factor. The MLP contains three fully connected layers with 32-16-5 nodes respectively. The activation functions are ReLU in the first two layers and Sigmoid in the last layer to ensure the weights are between 0 and 1. The global baseline weights are set according to cross-validation: energy balance constraint 1.0, thermal diffusion constraint 0.8, radiative transfer constraint 0.6, boundary layer constraint 0.4, and water vapor balance constraint 0.5. Training employs mixed precision acceleration technology using FP16 floating-point format, and distributed training is performed on eight A100 chips. On the GPU, a data parallel strategy is adopted, with a batch size of 4 per GPU. The learning rate is gradually reduced from 0.001 to 0.00001 using a cosine annealing strategy. Training takes about 36 hours for 80 rounds. On the validation set, the temperature prediction RMSE converges to 1.64°C, MAE is 1.21°C, and R² is 0.94. Compared with the standard PINN method without adaptive weights, the RMSE is reduced by 24.7%. In the high-density building area of ​​CBD, the RMSE of this method is 1.82°C, while that of the standard PINN is 2.85°C, which is a particularly significant improvement. As a variant implementation, the feature extraction backbone network can be replaced with EfficientNet-B4 or Swin Transformer. The physical constraints can be extended to Navier-Stokes equation constraints that include momentum equations. The adaptive weight network can use an attention mechanism to dynamically aggregate the contributions of different constraint terms.

[0103] The multi-temporal spatiotemporal fusion and dynamic correction unit adopts a spatiotemporal Transformer architecture to achieve multi-source data fusion. Inputs include 30 periods of Landsat-9 100-meter thermal infrared sequences, 180 periods of Sentinel-2 10-meter multispectral sequences, 180 periods of MODIS 1-kilometer temperature sequences, and hourly observation sequences from meteorological stations. The spatiotemporal attention mechanism includes a spatial self-attention layer and a temporal self-attention layer, employing a multi-head attention setup with eight attention heads, each with a dimension of 64. The fusion process is divided into three stages. The first stage involves fusing MODIS data... The data was upsampled from 1 km to 100 m using bilinear interpolation. In the second stage, auxiliary features such as NDVI were extracted from 10 m optical images from Sentinel-2, and the 100 m temperature was further downscaled to 10 m using a convolutional neural network. In the third stage, 100 m thermal infrared data from Landsat-9 was fused to provide an absolute temperature benchmark and correct systematic biases in the Sentinel-2 downscaling results. Temporal fusion was performed using a GRU-gated recurrent unit to capture the temporal evolution of temperature, with the hidden layer dimension set to 256. Missing data imputation employed a Generative Adversarial Network (GAN) framework. The generator network used an encoder-decoder structure to input missing temperature maps and output complete temperature maps. The imputation performance evaluation showed that when the missing proportion was 30%, the RMSE of the imputed region was 1.95°C, compared to simple spatial interpolation. Method 2.68°C improves accuracy by 27%. The ground truth correction processor uses a local bias correction method, dividing the study area into 27 sub-regions with a grid of 5×5 kilometers. For each sub-region, the systematic bias between the model's predicted temperature and the observed temperature of the meteorological station in that region is calculated. The correction process is performed every hour to ensure that the model output is consistent with the real-time observation. The corrected RMSE is further reduced from 1.64°C to 1.38°C. The final generated 10-meter resolution daily temperature product covers 365 days from July 2023 to June 2024, with a data completeness rate of 98.7%. Compared with the independent validation set, the overall accuracy RMSE is 1.38°C and R² is 0.96. As a technical variation, spatiotemporal attention can adopt the deformable attention mechanism, temporal modeling can use Temporal Convolutional Network or Informer model, and data imputation can use the diffusion model of partial differential equation (PDE).

[0104] The uncertainty quantification and multi-scenario generation unit uses a deep ensemble method to train five temperature reconstruction models with different initial weights and training data subsets. The mean of the five predicted values ​​is calculated as the final prediction standard deviation of the ensemble model, which is used as a measure of prediction uncertainty. The computational graph constructed by the uncertainty assessment processor contains five layers of nodes from the input data source to preprocessing, feature extraction, physical constraints, and temperature output. The Jacobian matrix of each node with respect to the final output is calculated using automatic differentiation technology. Error source decomposition uses variance decomposition. For each input data source node, random perturbations that conform to its true error distribution are artificially added. 1000 perturbation propagation experiments are conducted through Monte Carlo simulation. The results show that input data quality contributes 28.5% of the total uncertainty, physical constraint bias contributes 18.2%, model parameter uncertainty contributes 22.8%, spatiotemporal fusion error contributes 19.3%, and boundary condition error contributes 11.2%. Error propagation path tracing shows that the input data error is absorbed by the physical constraint node after propagation through five layers of nodes. The error absorption rate of the energy balance constraint is approximately 45%, reaching a maximum of 62%. The confidence interval calculation uses a 95% confidence level. The mean of the predicted temperature of the ensemble model is calculated as ±1.96 standard deviations as the upper and lower bounds of the confidence interval. Spatial distribution statistics show that the average confidence interval width is 2.8°C, narrowing to 1.5°C in vegetated and water areas and reaching 4.2°C in complex building areas. The multi-scenario simulation processor designed three scenarios. The optimal estimation scenario uses the mean prediction of the ensemble model, the conservative estimation scenario uses the output of the model with the lowest predicted temperature in the ensemble model, and the extreme high temperature scenario superimposes historical extreme weather conditions on the optimal estimate. The generated extreme scenario shows that the temperature in the CBD core area may reach 48.5°C, which is 6.8°C higher than the normal scenario. As a technical variation, the ensemble method can use Dropout Monte Carlo to randomly shut down neurons during inference to estimate uncertainty. Error propagation can use the linearization method of first-order Taylor expansion. Scenario generation can use variational autoencoder (VAE) to learn the latent space representation of the temperature field.

[0105] The intelligent monitoring, early warning, and decision support module includes a multi-dimensional spatiotemporal visualization and interactive analysis unit, a thermal risk intelligent early warning and source tracing analysis unit, and an urban thermal environment regulation decision support unit. The multi-dimensional spatiotemporal visualization unit developed a 3D visualization platform based on WebGL technology, using the Cesium 1.98 3D Earth engine as the underlying rendering framework. It overlays 10-meter resolution temperature raster data onto a 0.5-meter resolution oblique photogrammetric 3D model of Beijing. Temperature data is mapped using Rainbow color mapping: 15°C is displayed as dark blue, and 55°C as dark red. The spatiotemporal cube uses volume rendering technology, with the time dimension as the vertical axis, constructing a 3D voxel structure. Each voxel stores the temperature value at a specific location and time. It supports animation playback along the time axis, displaying the daily variation and seasonal evolution of the temperature field. The interactive query processor implements combined queries using spatial selection, a time range slider, and a two-way temperature threshold slider. Users can select the CBD area, choose July 1-15, 2024, and set the temperature range to 35-45°C. The system automatically calculates the area of ​​high temperatures within this spatiotemporal range, its duration, and its impact on the population. This visualization platform is deployed on Nginx. The 1.22 web server uses Node.js 18.16 for the backend to provide RESTful API services. The frontend uses the React 18 framework to develop a responsive user interface. The system can support up to 500 concurrent users, with a peak of 200 query requests per second and an average response time of less than 800 milliseconds. As a technical variation, Unity 3D or Unreal Engine can be used to develop more realistic 3D rendering effects.

[0106] The intelligent early warning and source tracing analysis unit for thermal risk uses a multi-index comprehensive evaluation method to identify thermal risk areas. Temperature thresholds are based on the high-temperature warning standards in the meteorological disaster emergency plan: a daily maximum temperature exceeding 35°C is considered general high temperature, 37°C is relatively severe high temperature, and 40°C is extremely severe high temperature. Vulnerable population density is determined using population distribution data from the National Bureau of Statistics, defining the elderly over 65 years old, children under 5 years old, and patients with chronic diseases as thermally vulnerable populations. The density of vulnerable populations in the CBD business district and residential areas reaches 8,000 people per square kilometer. The thermal risk index is defined as the degree of temperature exceeding the standard multiplied by the duration, multiplied by the vulnerable population density, and then normalized by 1000. The risk identification algorithm uses the DBSCAN density clustering method, setting a search radius of 500 meters and a minimum sample size of 50 pixels. Analysis on July 15, 2024, showed the identification of 23 heat spots of medium to high risk, including 3 reaching the extremely high risk level. The causal source tracing analysis uses Shapley value decomposition. The method quantifies the marginal contribution of each factor. For extremely high-risk hot spots in the CBD core area, the contribution rates are calculated as follows: building density 32%, underlying surface type 28%, anthropogenic heat emissions 23%, meteorological conditions 12%, and insufficient green space 5%. The warning level classification processor accesses the 72-hour weather forecast data from the Beijing Meteorological Observatory and inputs the forecast meteorological field into the temperature reconstruction model to predict future temperatures. When the predicted daily maximum temperature exceeds 37°C for more than 3 consecutive days within the next 24 hours, a yellow warning is automatically issued. Warning information is pushed to registered heat-sensitive individuals via SMS gateway. In the summer of 2024, a total of 17 high-temperature warnings were issued, with an accuracy rate of 82% and an average warning lead time of 36 hours. As a technical variation, machine learning classifiers such as XGBoost or LightGBM can be used to replace the threshold method for risk area identification. Source tracing analysis can use SHAP values ​​or LIME methods to provide local interpretability.

[0107] The urban thermal environment regulation decision support unit developed a scenario simulation tool to evaluate the cooling effects of different intervention measures. The greening increase scenario assumes that within a 5-square-kilometer area of ​​the CBD core, vegetation coverage will increase from the current 8% to 25% through rooftop greening, vertical greening, and street greening. In the temperature reconstruction model, the NDVI albedo value of this area is modified from 0.18 to 0.24, and the evapotranspiration coefficient is increased from 0.15 to 0.42. Running the temperature reconstruction model for forward prediction shows that the average surface temperature in this area decreases from the summer peak of 42.3°C to 38.7°C, a cooling rate of 3.6°C. The high-temperature area decreases by 67%, and the cooling effect can spread to a secondary cooling zone of approximately 800 meters, with a cooling rate of approximately 1.5°C. The white roof renovation scenario assumes that 80% of the building roofs in the CBD area will be coated with high-albedo paint, increasing the roof albedo from an average of 0.15 to 0.75. The model modifies the albedo parameters of building roof pixels and adjusts the roof surface temperature calculation method to simulate... The results showed that the measure could reduce the roof temperature by 8-12°C and reduce the average temperature in the street canyon by 2.1°C by reducing the long-wave radiation from the building to the surrounding area. However, the cooling effect on open areas was limited to only about 0.5°C. The ventilation corridor scenario assumed the establishment of north-south ventilation corridors within 500 meters on both sides of the main river system. In the model, the wind speed enhancement factor in this area was set to 1.8 times and the roughness length was reduced to 0.3 meters. The temperature field was simulated under the condition that the prevailing wind direction in the afternoon of summer was southeast with a wind speed of 3.2 m / s. The results showed that the average temperature in the ventilation corridor area was 2.8°C lower than the surrounding area. The cold air transported downstream along the corridor could generate a cooling belt of up to 5 kilometers. The temperature within 1 kilometer downwind of the corridor was reduced by 1.5-2.0°C. The combined scenario superimposed the three measures of increased greening, white roofs and ventilation corridors. The superimposed simulation showed that the comprehensive cooling effect in the core area of ​​the CBD could reach 5.2°C. The area with high temperatures above 35°C in the entire study area was reduced from 231 square kilometers to 87 square kilometers, a reduction of 62%.

[0108] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A high spatio-temporal resolution land surface temperature reconstruction system for urban thermal environment refinement monitoring, characterized in that, The system comprises a multi-source heterogeneous data intelligent acquisition and preprocessing module, a city thermal environment multi-dimensional feature knowledge modeling module, a physically constrained deep learning land surface temperature reconstruction module, and an intelligent monitoring and early warning and decision support module. The multi-source heterogeneous data intelligent acquisition and preprocessing module is responsible for collecting and preprocessing various types of remote sensing and ground observation data, the city thermal environment multi-dimensional feature knowledge modeling module constructs a knowledge system of city underlying surface, three-dimensional morphology and thermal process, the physically constrained deep learning land surface temperature reconstruction module is used to realize accurate reconstruction of high spatio-temporal resolution land surface temperature, and the intelligent monitoring and early warning and decision support module converts the reconstruction results into visual display, risk early warning and regulation decision suggestion. The multi-source heterogeneous data intelligent acquisition and preprocessing module comprises a multi-scale remote sensing data intelligent acquisition unit, a ground observation network adaptive deployment unit and a heterogeneous data quality control and registration unit, the multi-scale remote sensing data intelligent acquisition unit is responsible for collaborative acquisition of multi-spectral image and thermal infrared observation data, the ground observation network adaptive deployment unit realizes optimal spatial layout of sensors through observation network optimization algorithm, and the heterogeneous data quality control and registration unit performs spatio-temporal registration, format standardization and quality screening on multi-source data. The city thermal environment multi-dimensional feature knowledge modeling module comprises a city underlying surface multi-dimensional attribute modeling unit, a city three-dimensional morphology and radiation transmission modeling unit and a city thermal environment process knowledge graph construction unit, the city underlying surface multi-dimensional attribute modeling unit is used to construct a geomatics-thermal attribute knowledge base, the city three-dimensional morphology and radiation transmission modeling unit is used to simulate the radiation transmission process and shadow dynamic evolution of building groups, and the city thermal environment process knowledge graph construction unit constructs a three-layer knowledge graph containing entities, relationships and rules. The physically constrained deep learning land surface temperature reconstruction module comprises a physically guided deep neural network reconstruction unit, a multi-temporal spatio-temporal fusion and dynamic correction unit and an uncertainty quantification and multi-scenario generation unit, the physically guided deep neural network reconstruction unit is used to realize land surface temperature reconstruction with physical consistency and high spatio-temporal resolution, the multi-temporal spatio-temporal fusion and dynamic correction unit is used to combine ground observation for real-time deviation correction, and the uncertainty quantification and multi-scenario generation unit performs uncertainty evaluation on the temperature reconstruction result. The intelligent monitoring and early warning and decision support module comprises a multi-dimensional spatio-temporal visualization and interactive analysis unit, a thermal risk intelligent early warning and source analysis unit and a city thermal environment regulation decision support unit, the multi-dimensional spatio-temporal visualization and interactive analysis unit is used to intuitively display the spatial distribution and temporal variation characteristics of city thermal environment, the thermal risk intelligent early warning and source analysis unit performs thermal anomaly cause tracing analysis, divides early warning levels and realizes short-term prediction, and the city thermal environment regulation decision support unit simulates the cooling effect of intervention measures based on the reconstruction result, and generates thermal relief strategy suggestions by using optimization algorithm.

2. The high spatio-temporal resolution land surface temperature reconstruction system for urban thermal environment refined monitoring of claim 1, wherein, The ground observation network adaptive deployment unit comprises a multi-source sensor access processor, an observation network optimization processor and a mobile observation coordination processor, the multi-source sensor access processor is used for providing meteorological driving parameters such as air temperature, relative humidity, wind speed and wind direction and high-precision ground surface temperature true value, the observation network optimization processor is used for calculating the optimal layout position and density of the sensor, and the mobile observation coordination processor integrates the data of the flow observation platform to realize the cooperative observation of the mobile platform and the fixed network. The observation network optimization processor determines the sensor deployment position according to the following formula: ； ； where S opt represents the optimal sensor deployment location, x is the candidate sensor deployment location, represents the set of candidate locations for all possible sensor deployments, T s is the temperature field of the entire study area, D cur represents the current available observation dataset, H[T s |D] represents the conditional entropy of the temperature field T s given the observation data D, represents the temperature value observed at location x, represents the expected value of the observation, is the cost weight coefficient, represents the cost of deploying a sensor at x.

3. The high spatio-temporal resolution land surface temperature reconstruction system for urban thermal environment refined monitoring of claim 2, wherein, The urban thermal environment process knowledge graph construction unit comprises an energy budget modeling processor, a human activity heat source identification processor and a knowledge graph reasoning engine, the energy budget modeling processor is used for establishing an urban ground surface energy balance equation and constructing physical constraint rules of energy budget, the human activity heat source identification processor is used for identifying and quantifying the spatial distribution and time variation of human heat contribution, and the knowledge graph reasoning engine is used for constructing a three-layer knowledge graph of entity-relation-rule, realizing logical constraint and anomaly detection of the temperature field. The knowledge graph reasoning engine constructs a time series causal graph of urban thermal environment where V is a node set, is a causal edge set, B is a node attribute, and the causal relationship strength is learned according to the following loss function: ； ； wherein L cau is the total loss function of causal relationship learning, (V i , V j ) is a directed edge in the causal graph, is the attention weight, represents the causal strength from V i to V j , V i (t) represents the value of Vi at time t, is the causal time lag, represents the observed real causal effect, is the sparsity regularization sparsity, A is the adjacency matrix, is the acyclic constraint coefficient, h() is the acyclic constraint function, is the eigenvector of V i , W Q is the query matrix, v j is the eigenvector of V j , W K is the key matrix, d is the dimension of the query matrix and the key matrix, and N(j) represents the candidate parent node set of v j .

4. The high spatio-temporal resolution land surface temperature reconstruction system for urban thermal environment refined monitoring of claim 3, wherein, The physically guided deep neural network reconstruction unit comprises a physical information embedding processor, a multi-scale feature extraction processor and a physical consistency verification processor, the physical information embedding processor converts the physical constraint into a loss function item of the deep learning model and ensures that the temperature field output by the neural network meets the basic physical law, the multi-scale feature extraction processor is used for extracting multi-scale spatial features and time sequence features, learning the spatial heterogeneity and time evolution mode of the temperature field, and the physical consistency verification processor performs energy conservation test, temperature gradient rationality check and extreme value anomaly identification on the temperature field output by the neural network. The physical information embedding processor constructs a spatially varying learnable function according to the following formula: ； ； ； where L total is the total loss, L data is the data fitting loss, R m is the mth physical constraint residual, is the predicted temperature of location x at time t, M is the total number of physical constraints, w m (x) is the adaptive weight of the mth constraint at location x, is the Sigmoid activation function, MLP m is the small multi-layer perceptron, is the three-dimensional geometric feature vector, is the underlying surface type feature vector, is the relative ratio of the physical residual and the data gradient, is the global reference weight, is the gradient with respect to is the gradient with respect to is all the learnable parameters, is the numerical stability term.

5. The high spatio-temporal resolution land surface temperature reconstruction system for urban thermal environment refined monitoring of claim 4, wherein, The uncertainty quantification and multi-scenario generation unit comprises an uncertainty evaluation processor, a confidence interval calculation processor and a multi-scenario simulation processor, the uncertainty evaluation processor is used for quantifying the uncertainty source and propagation path of the temperature reconstruction result, the confidence interval calculation processor calculates the confidence interval or probability density distribution of the reconstructed temperature of each pixel according to the uncertainty evaluation result, and the multi-scenario simulation processor generates multiple temperature field schemes such as the optimal estimation scenario, the conservative estimation scenario and the extreme high temperature scenario in combination with the uncertainty distribution and the extreme weather condition assumption; The uncertainty evaluation processor calculates the uncertainty contribution of each link according to the following formula: ； ； ； where is the total uncertainty of the reconstructed temperature field, is the reconstructed temperature field, S is the set of error sources, is the contribution of error source k, e i , e j are error variables of two different error sources, Cov() is the error covariance, is the Jacobian matrix of the temperature with respect to the variable z k , is the variance of the variable z k , is the error decay coefficient, is the total number of computational paths from error source z k to the temperature field , is the error absorption rate on path p, is the normalized weight of path p.

Images