A high-resolution urban high-temperature monitoring method coupling urban climate models and deep learning

Abstract

The application discloses a kind of high-resolution urban high temperature monitoring methods coupled with urban climate model and deep learning, belong to high temperature monitoring and early warning field.Aiming at the problems such as low spatial resolution of prior art, insufficient use of multi-source information and not close fusion of physical mechanism and deep learning, combined with high-resolution remote sensing image, ground meteorological observation and urban surface attribute and other multi-source information, innovatively integrate urban climate physical modeling and deep learning algorithm, realize the high-precision inversion and dynamic monitoring of urban high temperature spatial pattern.The application can consider comprehensiveness, timeliness and scientificity of data, effectively improve the monitoring precision and prediction ability of urban high temperature distribution, and provide strong data support for urban high temperature disaster warning, risk assessment and scientific management.

Description

Technical Field

[0001] This invention belongs to the field of high temperature early warning and monitoring, and more specifically relates to a high-resolution urban high temperature monitoring method that couples urban climate models and deep learning. Background Technology

[0002] With the continued intensification of global warming, the frequency and intensity of extreme heat events have increased significantly worldwide. Especially against the backdrop of rapid urbanization, the high concentration of population, industry, and infrastructure has led to increasingly severe urban thermal environment problems. High temperatures and heat waves not only seriously endanger the health of urban residents, causing a series of public health risks such as heatstroke and cardiovascular diseases, but also have a profound impact on energy supply systems, transportation operations, and normal urban order. Furthermore, persistent high temperatures can lead to ecological and environmental problems such as vegetation withering, water bodies drying up, and increased air pollution, thereby affecting the sustainable development of regions and cities. The urban heat island effect has become an important scientific and social issue of concern in the field of international urban climate change. Traditionally, urban high-temperature monitoring has mainly relied on the deployment of ground meteorological observation stations and remote sensing image inversion. Ground meteorological stations can provide continuous and accurate measurements of basic meteorological elements such as temperature and humidity, providing fundamental support for various high-temperature warnings, disaster assessments, and refined meteorological services. However, the layout of meteorological stations is usually constrained by urban spatial planning and land use constraints, making it difficult to meet the needs of large-scale, high-spatial-resolution urban thermal environment monitoring in terms of both density and distribution. Especially in urban areas and on the outskirts, the heat distribution is extremely complex due to significant differences in underlying surface type, building density, and height. Relying solely on limited meteorological station data is insufficient to reflect the fine-grained pattern of the urban high-temperature field. Furthermore, the high maintenance costs and long data update cycles of observation stations present bottlenecks in both real-time performance and wide-area coverage in dynamic environments.

[0003] To address these issues, urban climate models have seen widespread development in recent years. Models such as Urban Weather Generator (UWG) and Regional Climate Rendering (WRF) can simulate urban-scale factors like temperature, humidity, wind field, and surface energy flux, driven by digital terrain, land cover, and urban structural parameters, based on energy balance and atmospheric dynamics principles. The advantages of urban climate models lie in their clear physical basis and interpretable mechanisms, enabling them to incorporate city-specific parameters such as buildings, underlying surface, and bioclimate into their calculations, thus reflecting the dynamics of the thermal environment within and around the city. However, the spatial resolution of these models is typically limited by input geographic data, computational resources, and parameterization schemes, making it difficult to achieve the high spatial resolution of remote sensing. They also rely on complex boundary conditions, initial parameters, and long-term observational data, resulting in long computation cycles and drawbacks such as delayed results and slow updates. With the rapid development of deep learning and artificial intelligence, some studies have begun to explore the use of deep neural networks to fuse multi-source data for urban climate applications such as surface temperature retrieval and high-temperature risk assessment. Deep learning models, such as convolutional neural networks, recurrent neural networks, and attention mechanisms, can uncover complex nonlinear relationships in remote sensing imagery, land use distribution, and climate time-series data, achieving good estimation accuracy on some samples. However, existing studies often rely on a single data source, such as remote sensing imagery or single-type meteorological data, and most lack modeling of the physical processes themselves. This results in "black box" characteristics, poor scientific interpretability and physical consistency, hindering the understanding of urban climate change mechanisms and providing scientific decision-making references. Furthermore, existing deep learning methods have limited generalization performance on extreme high temperatures and anomalous meteorological events (such as extreme heat waves), especially in scenarios with imbalanced training sample distribution and scarce extreme samples, where model accuracy and stability face significant challenges. Summary of the Invention

[0004] This invention aims to address the problems of existing urban high-temperature monitoring methods, such as limited spatial resolution, single data source, insufficient fusion of multi-source data, inadequate modeling of complex urban underlying surfaces, and lack of organic integration of physical mechanisms and deep learning. It proposes a technical solution that can integrate high-resolution remote sensing images, ground meteorological observations, and urban surface attribute data, combined with urban climate physics models and deep learning technology, to achieve high-resolution, intelligent, and dynamic urban high-temperature monitoring. This provides efficient and reliable data support for refined acquisition of high-temperature distribution, extreme weather early warning, and scientific urban management.

[0005] To achieve the above objectives, the present invention employs the following technical solution: the method comprises:

[0006] Data collection and processing include acquiring high-resolution remote sensing image data from the Gaofen-1 satellite, performing geometric, radiometric, and atmospheric corrections on the remote sensing images to obtain corrected remote sensing image data with a spatial resolution of 20m; collecting meteorological observation data from ground meteorological monitoring stations, including temperature, relative humidity, wind speed, air pressure, altitude, date, latitude, and longitude; and downloading urban land cover data and building information data, whereby land cover data includes construction land, forest land, grassland, water areas, cultivated land, and unused land, and building information includes roof, building height, building structure, and building functional indicators.

[0007] Urban high temperature spatial distribution information extraction: Based on the above preprocessed and fused multi-source data, urban land cover characteristics, building attributes and remote sensing data are jointly input into an enhanced bi-branch radiation neural network, wherein the network processes shortwave and longwave radiation respectively, a shared feature extraction module is used for feature extraction, and an adaptive weighted loss function is used to dynamically adjust the sample weights. Parameters such as longwave radiation output, surface emissivity, Stefan-Boltzmann constant, surface temperature, heat capacity, and surface temperature change rate are combined with the urban energy balance model to obtain the distribution of the city's highest temperature.

[0008] Accuracy verification is performed using verification data from meteorological monitoring stations as the verification dataset. The root mean square error, mean absolute error, and systematic error evaluation methods are used to quantitatively assess the accuracy and bias of the high temperature distribution in cities retrieved by the deep learning model. The applicability and high accuracy characteristics of the model are determined by comparing it with monitoring data from typical cities or typical times.

[0009] In one approach, the geometric correction uses geographic information system (GIS) tools to register remote sensing images with the actual geographic locations of cities, ensuring that the pixels in the remote sensing images correspond precisely to the actual geographic coordinates.

[0010] In one approach, the radiometric correction includes converting digital values ​​of remote sensing images into radiance, and further incorporating sensor response characteristics to eliminate the impact of differences between the device and the imaging environment on the image radiometric accuracy.

[0011] In one approach, the atmospheric correction step employs an atmospheric correction model to remove the influence of the atmosphere on the reflectivity and temperature of ground features from remote sensing images, thereby restoring the true reflectivity and temperature of the ground surface.

[0012] In one approach, the model input samples are labeled and feature-enhanced based on the various urban land cover types and building attributes obtained.

[0013] In one approach, the adaptive weighted loss function in the enhanced bibranch radiative neural network is designed for extreme high temperatures and uneven sample distribution. By adjusting the weight allocation in real time, it enhances the model's learning ability for a small number of high-temperature and special weather samples.

[0014] In one approach, the model accuracy verification uses typical monitoring data from at least three different time points or seasons. By comparing these data with the highest temperatures measured at ground stations, the root mean square error, mean absolute error, and systematic error are calculated respectively.

[0015] In one approach, the technical route achieves an organic combination of physics-driven and data-driven approaches through collaborative modeling of shortwave and longwave radiation branches and parameter coupling with urban climate energy balance models.

[0016] Beneficial effects of this invention:

[0017] (1) A dual-branch network is adopted to process shortwave and longwave radiation respectively. Features are extracted by sharing the feature extraction module, which improves the model's adaptability to diverse meteorological conditions and the simulation accuracy of extreme climates.

[0018] (2) The adaptive weighted loss function can dynamically adjust the weights according to features such as temperature, optimize for extreme weather samples, and flexibly adjust the strategy according to the actual situation to adapt to different meteorological conditions. Therefore, it has a stronger predictive ability for urban high temperatures. Other weighted loss functions based on probability distribution, such as the Gaussian weighted function, are usually based on the statistical characteristics of the predicted values, and the weight calculation is relatively fixed. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention;

[0020] Figure 2 This is a block diagram of the present invention;

[0021] Figure 3 This is a rendering of an embodiment of the present invention. Detailed Implementation

[0022] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0023] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0024] like Figure 1 and Figure 2 As shown, the specific implementation process of a high-resolution urban high-temperature monitoring method that couples urban climate models and deep learning is as follows:

[0025] Step 1, Data Collection and Processing:

[0026] (1) Satellite data was collected, primarily using high-resolution remote sensing imagery from Gaofen-1 (20m spatial resolution). Geometric, radiometric, and atmospheric corrections were performed on the acquired satellite imagery. These preprocessing steps ensured the accuracy and consistency of the data. Specifically:

[0027] Geometric correction: Image registration is performed using Geographic Information System (GIS) technology to ensure the correspondence between the image and the actual geographical location of the city.

[0028] Radiometric correction: Converting digital values ​​(DN) to radiance to ensure the accuracy of the inversion.

[0029] Atmospheric correction: The atmospheric correction model (MODTRAN model) is used to remove atmospheric effects and obtain accurate surface reflectance and temperature information.

[0030] (2) Surface feature data: Meteorological monitoring data from meteorological monitoring stations were collected, including temperature, relative humidity, wind speed, air pressure, altitude, date, latitude, and longitude. In addition, urban land cover data from the urban land use status map were collected, including land cover types such as construction land, forest land, grassland, water area, cultivated land, and unused land. Furthermore, building information data was obtained from the Multi-Attribute Building Dataset (CMAB), including indicators such as roof, building height, building structure, building function, and other functional indicators.

[0031] Step 2: Extraction of spatial distribution information of urban high temperatures

[0032] S201. Construct an enhanced bi-branch radiation prediction neural network model. The input data for the neural network model includes the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and land surface temperature calculated based on Gaofen-1 remote sensing imagery, as well as relative humidity, wind speed, air pressure, and altitude from 80% of the meteorological monitoring stations (i.e., the training dataset). Specifically, NDVI and NDWI are calculated using Gaofen-1 remote sensing image bands to enhance the vegetation / water body response. The calculation formulas are as follows:

[0033]

[0034] In the formula, NIR represents the reflectance in the near-infrared band, Red represents the reflectance in the red band, Green represents the reflectance in the green band, and SWIR represents the reflectance in the short-wave infrared band. Furthermore, the surface temperature is calculated using a single-window algorithm formula:

[0035]

[0036] In the formula, T s Where λ is the surface temperature, 10.8 μm, and ρ is the surface reflectance, which can be calculated from the near-infrared band:

[0037]

[0038] In the formula, L λ E represents the radiation value in the near-infrared band. λ The value of solar radiation in this band (can be found online), cos(θ) is the solar altitude angle, and π represents pi.

[0039] Subsequently, the constructed neural network model includes two independent branch structures, comprising a hierarchical structure, an attention mechanism, and an adaptive weighted loss function, to achieve more accurate predictions of shortwave and longwave radiation. The hierarchical structure of both branches includes convolutional layers, pooling layers, residual connections, and fully connected layers. Specifically, the prediction formula for the shortwave radiation branch is:

[0040] Q in =f CNN (X weather )+Attention(f CNN (X weather ))

[0041] H1 = ReLU(W1 * X weather +b1)

[0042] In the formula, Q in The predicted shortwave radiation value (W / m 2 ), f CNN X is the feature extraction function of a convolutional neural network. weatherThe values ​​represent the relative humidity, wind speed, air pressure, and altitude of the meteorological station. H1 is the output feature map of the convolutional layer, W1 is the weight matrix of the first convolutional layer, with a size equal to the shape of the weight matrix: 3×3 (kernel size)×3 (number of input channels)×64 (number of output channels), b1 is the bias term of the first convolutional layer, initially set to 1, and ReLU is the activation function used to increase the nonlinearity of the model. Therefore, the output feature H2 of the pooling layer and the output feature H3 after the residual connection are calculated as follows:

[0043] H2 = Max Po olin g(H1)

[0044] H3 = H2 + X weather

[0045] The structure of the fully connected layer is as follows:

[0046] Q in =W2H3+b2

[0047] In the formula, W2 is the weight matrix of the fully connected layer, and b2 is the bias term of the fully connected layer, which is initially set to 1.

[0048] The prediction formula for the long-wave radiation branch is:

[0049] Q out =f CNN (T s ,∈)+Attention(f CNN (T s ,∈))

[0050] H′1=ReLU(W′1*[T s ,ò]+b′1)

[0051] In the formula, Q out The predicted longwave radiation value (W / m 2 ), T s Let H' be the surface temperature and ω be the surface emissivity. H′1 is the output feature of the convolutional layer; W′1 is the weight matrix of the convolutional layer for the long-wavelength radiation branch, with a size equal to the shape of the weight matrix: 3×3 (kernel size)×3 (number of input channels)×64 (number of output channels); b′1 is the bias term of the convolutional layer for the long-wavelength radiation branch, initially set to 1. Therefore, the output feature map H′2 of the pooling layer and the output feature map H′3 after residual connection are calculated as follows:

[0052] H′2 = MaxPooling(H′1)

[0053] H′3=H′2+[T,ò]

[0054] The structure of the fully connected layer is as follows:

[0055] Qout =W′2H′3+b′2

[0056] In the formula, W′2 is the weight matrix of the fully connected layer, and b′2 is the bias term of the fully connected layer, which is initially set to 1.

[0057] In the attention mechanism, the feature score is calculated as follows:

[0058]

[0059] In the formula, e i Feature scores indicate the importance of a feature. The attention weight matrix is ​​set to 64×32, H i It represents the input characteristics of shortwave and longwave radiation.

[0060] The adaptive weighted loss function differentiates the importance of different samples, especially under extreme weather conditions, enabling the model to learn and optimize predictions more effectively, thereby improving the model's reliability and accuracy. Specifically, by assigning higher weights to samples under extreme weather conditions, the model can dynamically adjust the loss calculation based on the importance of the samples.

[0061]

[0062] In the formula, L is the total loss value, representing the overall error between the model's prediction and the actual value, and y i For the true value, Here, w represents the predicted value, N is the number of samples, and w i w represents the weight of sample i. i The calculation is as follows:

[0063]

[0064] Samples with a temperature T exceeding 35℃ were assigned a higher weight, while other samples had a weight of 1 (i.e., w = 1), and the β value was set to 0.50.

[0065] S202, Constructing an urban climate model coupled with deep learning.

[0066] To improve the accuracy of urban high-temperature simulation, an enhanced bibranch radiation prediction neural network model and an Urban Weather Generator (UWG) were integrated to construct a coupled deep learning urban climate model. Based on the resolution of the input satellite imagery data, the grid size was set to 20m, thus achieving a spatial resolution of 20m for the urban high-temperature data. Furthermore, monitoring data from 80% of meteorological monitoring stations (including relative humidity, wind speed, air pressure, and altitude) and surface feature data (including urban land cover and building information) were selected as input data. The model is based on the following energy balance equation, describing the input, output, and storage relationships of surface energy:

[0067] Q net =Q in -Q out -Q storag

[0068] Q out =ò·σ·T 4

[0069]

[0070] In the formula, Q net Net radiation; Q in The shortwave radiation input is provided by an enhanced dual-branch neural network; Q out The longwave radiation output is calculated from the surface emissivity ω, the Stefan-Boltzmann constant σ, and the surface temperature T; Q storage The energy stored on the Earth's surface is calculated from the heat capacity C (based on different urban land cover types) and the rate of change of surface temperature dT / dt. Finally, the city's maximum temperature T is calculated using energy balance. max :

[0071]

[0072] In the formula, Q net For net radiation, Q out For long-wavelength radiation output, Q storage Let σ be the energy stored on the Earth's surface, σ be the Stefan-Boltzmann constant, and ò be the surface emissivity.

[0073] Step 3: Accuracy Verification

[0074] S301, Accuracy Verification Indicators

[0075] Accuracy is assessed using three widely used metrics: root mean square error (RMSE), mean absolute error (MAE), and systematic error. RMSE and MAE are used to calculate precision, while systematic error quantifies estimation bias. The formulas are as follows:

[0076]

[0077] In the formula, f i It is a grid i The average maximum temperature was obtained using the remaining 20% ​​of meteorological monitoring stations (i.e., the validation dataset); f i The model calculates the mesh. i Maximum temperature; n is the total number of monitoring stations in the grid; RMSE is the root mean square error; MAE is the mean absolute error; SE is the systematic error.

[0078] S302, Accuracy Verification Results

[0079] Here, Shenzhen is used as an example for verification (Table 1 and below). Figure 3 It was found that the error of the method of the present invention was within 5%, showing high applicability.

[0080] Table 1

[0081] Monitoring time Root mean square error Mean Absolute Error Systematic error February 6, 2024 2.63 1.12 3.17 May 15, 2024 4.06 2.11 3.53 August 5, 2024 2.18 1.01 2.1 November 12, 2024 4.38 2.98 3.61

[0082] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0083] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-resolution urban high-temperature monitoring method coupling an urban climate model and deep learning, characterized in that: The method comprises: Data collection and processing, obtaining high-resolution remote sensing image data of Gaofen-1, performing geometric correction, radiation correction and atmospheric correction on the remote sensing image to obtain corrected remote sensing image data with a spatial resolution of 20 m; collecting meteorological observation data from ground meteorological monitoring stations, including air temperature, relative humidity, wind speed, air pressure, altitude, date, latitude and longitude; downloading urban land cover data and building information data, wherein the land cover data includes construction land, forest land, grassland, water area, cultivated land and unused land, and the building information includes roof, building height, building structure and building function index; Urban high-temperature spatial distribution information extraction, based on the processed and fused multi-source data, inputting urban land cover features, building attributes and remote sensing data into an enhanced double-branch radiation neural network, wherein the network respectively processes short-wave and long-wave radiation, adopts a shared feature extraction module for feature extraction, and dynamically adjusts sample weights by using an adaptive weighted loss function; inputting the long-wave radiation output by the enhanced double-branch radiation neural network into a city climate model together with the emissivity of the ground, the Stefan-Boltzmann constant, the ground temperature, the heat capacity and the ground temperature change rate parameters to calculate the urban maximum temperature distribution; Integrating the enhanced double-branch radiation prediction neural network model and the urban weather generator to build a city climate model coupled with deep learning; Based on the resolution of the input satellite image data, the grid size is set to 20 m, and the spatial resolution of the urban high-temperature data is 20 m; Selecting 80% of the monitoring data of the meteorological monitoring stations, including relative humidity, wind speed, air pressure and altitude; and selecting the ground feature data, including urban land cover and building information, as input data; Based on the following energy balance equation, the input, output and storage relationship of the ground energy is described: ; ; ; where, is the net radiation; is the shortwave radiation input provided by the enhanced dual-branch neural network; is the longwave radiation output, provided by the surface emissivity , the Stefan-Boltzmann constant σ , the surface temperature T is calculated; is the energy stored in the surface, calculated from the heat capacity C based on different urban land cover types and the rate of change of the surface temperature dT / dt ; finally, the maximum urban air temperature is calculated using the energy balance : ; wherein is the shortwave radiation input provided by the enhanced dual-branch neural network; is the longwave radiation output, is the energy stored in the surface, σ is the Stefan-Boltzmann constant, is the surface emissivity; Accuracy verification, based on the monitoring data of the meteorological monitoring stations as the verification data set, using the root mean square error, the mean absolute error and the system error evaluation method to quantitatively evaluate the accuracy and deviation of the deep learning model for the inversion of the urban high-temperature distribution, and comparing the monitoring data of a typical city or a typical time to determine the applicability and high-precision characteristics of the model.

2. The method of claim 1, wherein the method is characterized by: The geometric correction is performed by registering the remote sensing image and the real geographical position of the city through a geographic information system tool to ensure that the remote sensing image element accurately corresponds to the actual geographical coordinates.

3. The method of claim 1, wherein the method is characterized by: The radiation correction includes converting the digital value of the remote sensing image into radiation brightness, and further combining the sensor response characteristics to eliminate the influence of the difference between the equipment and the imaging environment on the image radiation accuracy.

4. The method of claim 1, wherein the method is characterized by: The atmospheric correction step uses an atmospheric correction model to remove the influence of the atmosphere on the reflectivity and temperature of the ground object from the remote sensing image to invert the real ground reflectivity and temperature.

5. The method of claim 1, wherein the method is characterized by: According to the obtained various urban land cover types and building attribute classification, the input sample of the model is labeled and enhanced.

6. The method of claim 1, wherein the method is characterized by: The adaptive weighted loss function in the enhanced double-branch radiation neural network adjusts the weight distribution in real time to strengthen the learning ability of the model for a small amount of high-temperature and special weather samples for the extreme high-temperature and unbalanced sample distribution.

7. The method of claim 1, wherein the method is characterized by: The model accuracy is verified by comparing the simulated maximum temperature with the observed maximum temperature at the ground station, and the root mean square error, mean absolute error and systematic error are calculated.

8. The method of claim 1, wherein the method is characterized by: The shortwave and longwave radiation branches are modeled collaboratively, and the parameters are coupled with the urban climate energy balance model, realizing the organic combination of physical driving and data driving.

Images