GeoAI machine learning-based intelligent downscaling seasonal representation and analysis method for surface temperature

Abstract

The present invention develops a method for seasonal characterization and analysis of surface temperature based on GeoAI machine learning intelligent downscaling. This method can interpolate and downscale the missing surface temperature areas affected by cloud factors through urban landscape characteristic index products and incomplete satellite surface temperature inversion data. Through a unique stratified sampling mechanism, the training bias and overfitting problems caused by insufficient representativeness of the data training set are solved, and a surface temperature product with a complete spatiotemporal sequence of 10m resolution in seasons is generated. Verification shows that this method can accurately make up for the missing surface temperature data caused by factors such as cloud cover, ensure the spatial integrity and temporal continuity of the surface temperature data, truly reflect the surface temperature of the study area when the satellite passes, and effectively characterize the seasonal spatiotemporal evolution of the urban thermal environment. In addition, the present invention also develops a feature importance intelligent evaluation framework based on GeoAI, which can parse and screen landscape features that are strongly correlated with surface temperature, provide scientific support and reference for the accurate characterization, modeling and optimization of the urban thermal environment, and has important application value.

Description

Technical Field

[0001] The invention relates to the fields of machine learning and remote sensing technology, and provides a surface temperature intelligent downscaling seasonal characterization and analysis method based on GeoAI machine learning. Background Art

[0002] The urban thermal environment refers to the reactions and transformations that occur in the thermal field of urban built-up areas after the natural environment is transformed into an urban landscape under the influence of human activities. Land surface temperature (LST) is a comprehensive reflection of the interaction between the atmosphere and land features, as well as the exchange, migration, and transformation of energy. A complete and continuous record of LST and its associated urban spatial landscape characteristics are fundamental to characterizing, studying, and managing the urban thermal environment.

[0003] Currently, urban thermal environment research and management at the city scale primarily focuses on characterizing and analyzing the urban heat island effect. Land surface temperature, a crucial reference for characterizing and analyzing urban heat islands and thermal environments, relies primarily on the temporal and spatial scale advantages of satellite remote sensing data to invert land surface temperature. Remote sensing inversion data is unaffected by geographical location and has a complete spatial scale, but is obscured by cloud cover and cannot provide the complete temporal and spatial series necessary for a complete and accurate characterization of the urban thermal environment. Furthermore, urban spatial characteristics are limited by the scarcity of GIS information data, making it impossible to accurately characterize the landscape factors that influence and shape the urban thermal environment. Therefore, a method for accurately characterizing, analyzing, and managing the urban thermal environment is crucial. Summary of the Invention

[0004] To address the above technical issues, the present invention provides a method for intelligent downscaling of surface temperature seasonality, characterization, and analysis based on GeoAI machine learning. This method uses satellite imagery and urban landscape features to reconstruct surface temperature and analyze the seasonality of urban thermal environments, achieving a seamless spatiotemporal integration of seasonal urban thermal environment characterization and laying a foundation for urban thermal environment research. Compared to numerical simulations, the visual intelligence model constructed in this invention, based on machine learning and remote sensing visual imagery, can produce high-spatial-resolution surface temperature products and urban landscape features while maintaining accuracy.

[0005] The present invention comprises the following steps:

[0006] S1: Obtain the Landsat-8 surface temperature products within the study area on a daily basis, preprocess them, and screen effective daily surface temperature products;

[0007] S2: Obtain landscape characteristic index products related to urban land use and land cover (LULC) on an annual basis and preprocess them;

[0008] S3: For any valid daily surface temperature data in the daily surface temperature product, screen the valid surface temperature pixels in the study area on that day, extract urban land use and land cover data, surface temperature data, and Sentianl-2 satellite band data with consistent coordinates based on the geographic spatial location of the valid pixels, perform stratified random sampling, and construct a basic training sample set;

[0009] S4: Based on the basic training sample set, a multi-spatial-scale U-Net model is trained to estimate the probability index of buildings and roads, and generate the smart city building intensity index and smart road intensity index to characterize the distribution of buildings and roads in the urban space;

[0010] S5: Comprehensive landscape characteristic index product, based on the geographic spatial location of effective pixels, extracts comprehensive landscape characteristics consistent with the spatial location of stratified random sampling to generate a comprehensive training sample set;

[0011] S6: Based on the comprehensive training sample set, a random forest model is trained to reconstruct the LST data. The surface temperature data of any pixel position within the study area on that day are reconstructed to generate a daily surface temperature product with a complete spatial sequence.

[0012] S7: Perform weighted average and normalization on the LST products of spring, summer, autumn and winter to generate the LST seasonal variation index;

[0013] S8: Based on the comprehensive training sample set, a random forest model is trained for landscape feature importance analysis to generate feature contribution analysis, screen important landscape feature index products for the heat island effect, and analyze and guide potential spatial governance solutions.

[0014] Furthermore, in step S1, the preprocessing includes: surface temperature image acquisition, date screening, quality screening, study area clipping and resampling.

[0015] Furthermore, step S1 specifically includes:

[0016] S1.1: With the center of the study area as the center, set a buffer zone with a radius of r based on the spatial scale of the study urban area. r should be greater than the maximum distance between the center of the circle and the urban boundary. Obtain the Landsat-8 surface temperature product through the GEE platform. Using the buffer zone as the study area, mask all surface temperature products and crop them to match the spatial extent of the study area.

[0017] S1.2: Extract the pixel quality band QA_Pixel of the Landsat-8 product, analyze the daily surface temperature fluctuations in the study area in different seasons and the degree of cloud influence, further screen the quality of the mask-extracted Landsat-8 surface temperature product, and obtain daily surface temperature images of qualified quality.

[0018] S1.3 For qualified daily surface temperature images, a bilinear sampling method is used to resample them to a resolution of 10 m to make them consistent with the spatial resolution of the Sentinal-2 satellite.

[0019] Furthermore, in step S1.2, the quality screening criteria are that the cloud coverage of a single surface temperature image is less than 20% and the surface temperature image covers more than 50% of the study area.

[0020] The formula for calculating the cloud cover rate of a single image is as follows:

[0021]

[0022] Among them, w i represents the cloud coverage rate on the i-th day, that is, the percentage of missing values ​​in pixels on the i-th day. represents the number of valid pixels in the Landsat-8 surface temperature product of the i-th day after cropping, that is, the effective image coverage area, w ref Represents the number of pixels in the study area, that is, the area of ​​the study area.

[0023] The validity of any spatial location pixel is determined as follows: If the cloud mask (cloudsBit Mask) and shadow mask (cloudShadowBit Mask) in QA_Pixel are both equal to 0, then If the cloud mask cloudsBitMask in QA_Pixel is equal to 1 or the shadow mask cloudShadowBitMask is equal to 1, then Where r and c represent the horizontal and vertical coordinates of the pixel position respectively.

[0024] Furthermore, in step S2, the landscape characteristic index products include: Sentinel satellite bands, Remote Sensing Satellite Index (RSEI vegetation index), elevation data, night light data, urban land use and surface cover data, building distribution outlines, and road outline data.

[0025] Furthermore, step S2 specifically includes:

[0026] S2.1: Using the GEE platform, obtain 10m resolution band data (B2, B3, B4, B8A, B8, B9, B11, B12) from the Sentinel-2 satellite database as band index indicators. Use the band data obtained in S2.1 to generate the RSEI vegetation index products (NDVI, NDBI, NDWI) as the basic data set channels 1-11;

[0027] S2.2: Obtain urban land use and land cover data (LULC) with a spatial resolution of 10m from the Dynamicworld data set from the GEE platform as land use functional zoning factor indicators and as channels 12-21 of the basic dataset;

[0028] S2.3: From the GEE platform, obtain the SRTM elevation data (DEM) with a spatial resolution of 30 m from the Earthdata database as the physical form factor indicator and as the 22nd channel of the basic dataset;

[0029] S2.4: From the GEE platform, obtain the VIIRS sensor spatial resolution 400 night light data (NLT) as the socioeconomic factor indicator and the 23rd channel of the basic data set;

[0030] S2.5: From the GEE platform, obtain the built-up area building volume data (BV) and building height data (BH) with a spatial resolution of 100m from the GHSL database as the physical form factors of the urban built-up area and as the 24th and 25th channels of the basic dataset;

[0031] S2.6: Download the latest street data, building distribution data, street levels, and building functional attributes for the study area from the OSM platform. Filter the street data and building distribution data, extract street and building distribution vector data, and perform vector-to-raster conversion and morphological processing. These data will serve as the basic data for characterizing the physical form factors of road distribution and building distribution in urban built-up areas. These data will be used to generate physical form factors of built-up areas, including the smart city building index and the smart road intensity index. These features will be used to generate the smart city building index (Built Footprints) and the smart road intensity index (Road Footprints) as channels 26 and 27 of the comprehensive dataset. The obtained street vector data will be used to obtain the block outlines of the study area.

[0032] S2.7 uses the data obtained in S2.5 and S2.6 to calculate and generate the physical form factors of the built-up area; the physical form factors of the built-up area include: building density (BD) of the urban built-up area, sky width (SVF) of the urban built-up area, and floor ratio (FAR) of the urban built-up area, which serve as channels 28-30 of the comprehensive dataset.

[0033] S2.8 performs preprocessing operations on the image products obtained in S2.1-S2.7.

[0034] Furthermore, the band data obtained in step S2.1, the specific method of generating the RSEI vegetation index product is:

[0035] Using the band, the NDVI expression is generated as follows:

[0036]

[0037] Among them, NDVI stands for Normalized Difference Vegetation Index, NIR stands for Near Infrared Band, i.e. Band 8, and RED stands for Red Light Band, i.e. Band 4.

[0038] Using the band, generate NDWI and its expression is:

[0039]

[0040] Among them, NDWI stands for Normalized Difference Water Index, NIR stands for Near Infrared Band, i.e. Band 8, and GREEN stands for Green Light Band, i.e. Band 3.

[0041] Using the bands, the NDBI expression is generated as follows:

[0042]

[0043] Among them, NDBI stands for Normalized Building Index, NIR stands for near-infrared band, i.e. band 8, and SWIR stands for shortwave red light band, i.e. band 11.

[0044] Furthermore, in step S2.1, the pre-processing method is:

[0045] This includes annual median calculations for periodic band and RSEI index products to generate time series consistent index products;

[0046] Furthermore, in step S2.6

[0047] The street data and building distribution data obtained from the OSM platform were screened using street levels and building functional attributes. Street data and building distribution data that were consistent with the remote sensing image content were extracted, and the obtained street and building distribution data were converted from vectors to raster in QGIS. Morphological processing was performed on the rasterized street and building distribution data to achieve noise reduction. The building distribution data after morphological processing was converted into a binary image. The area value of the building distribution was 1, and the background area value was 0. These images served as the basic data for the physical morphological factors representing the distribution of buildings in urban built-up areas. Similarly, the road distribution data after morphological processing was also converted into a binary image. The area value of the road distribution was 1, and the background area value was 0. These images served as the basic data for the physical morphological factors representing the distribution of roads in urban built-up areas.

[0048] Furthermore, the morphological processing in step S2.6 is as follows:

[0049] The rasterized street data is first binarized, then morphological dilation is performed to merge multi-directional roads, and then morphological skeleton extraction and dilation are performed to accurately depict the distribution of roads.

[0050] The rasterized building distribution data is first binarized and then the partition morphological closing and filling operations are performed.

[0051] Furthermore, in step S2.7

[0052] For building density (BD), the calculation formula is as follows:

[0053]

[0054] Among them, A b represents the total area of ​​all buildings in the block, A S Represents the total area of ​​the block.

[0055] For the floor area ratio (FAR) of urban built-up areas, the calculation formula is as follows:

[0056]

[0057] in, and are the building area and number of floors of the jth building in the i-th block, n represents the total number of buildings, s i Represents the block area of ​​the i-th block.

[0058] For the sky width (SVF) of urban built-up areas, considering that the complex spatial environment within the urban area has an impact on the absorption and emission of long-wave radiation, the building height data (BH) and the elevation data (DEM) of the study area were made into a digital surface model and input into the SAGA GIS tool. At the same time, the maximum search radius was set to 100m to obtain the sky width (SVF) of urban built-up areas.

[0059] In a three-dimensional urban environment, the SVF calculation formula for a point is as follows:

[0060]

[0061] where β is the angle from the center point to the maximum obstacle height at a maximum distance equal to the constant search radius (R).

[0062] Furthermore, in step S2.7, the preprocessing includes downsampling all images and landscape index products with a spatial resolution greater than 10 meters to 10 meters using a bilinear sampling method to generate images and landscape index products with consistent spatial resolution. Furthermore, the RSEI index products for the study area are cropped to generate index products with consistent spatiotemporal resolution.

[0063] Furthermore, step S3 specifically includes:

[0064] S3.1 Use the daily surface temperature data of any day in the valid daily surface temperature product to select the pixel linear index coordinates of the valid surface temperature pixels in the study area on that day, and read the LULC data, valid surface temperature data and Sentianl-2 satellite band product data with consistent geographical locations into the table to form the basic data set;

[0065] S3.2 uses the surface temperature data in the basic dataset as the stratification standard and performs stratified sampling. A suitable sampling interval z is set, and sampling is performed k times. Each time, m samples are randomly selected from each stratum to construct the training sample set. Assuming the LST value interval is Z, the final number of generated samples is M = Z / z*k*m. The smaller the z value, the smaller the sampling interval and the fewer samples are drawn, and vice versa.

[0066] S3.3 Draw a probability distribution histogram of the sampled samples to determine whether the sample distribution is uniform. Continuously adjust K and M to make the sample distribution approach a balanced distribution. This will serve as the basic sample training set and as input for subsequent models.

[0067] Furthermore, step S4 specifically includes:

[0068] S4.1 Construct a multi-spatial scale U-Net model based on convolutional neural network (CNN) to generate the smart city building index and smart road intensity index.

[0069] S4.2 is based on the basic sample training set, takes the geographic spatial position of each sample as the index, extracts the exponential feature image of the image area with a neighborhood pixel size of [64,64] as the sample, and constructs the CNN sample training set.

[0070] S4.3 uses the landscape feature index product of channels 1 to 25 in the CNN sample training set as x and the road distribution data as y road Train the multi-spatial scale U-Net model. Evaluate the model training effect by calculating the model training accuracy. Use the model with the smallest MSE and R 2 Higher models generate smart road strength indices.

[0071] S4.4 uses the landscape feature index product of channels 1 to 25 in the CNN sample training set as x and the building distribution data as y built Train the multi-spatial scale U-Net model. Evaluate the model training effect by calculating the model training accuracy. Use the model with the smallest MSE and R 2 Higher models generate smart road strength indices.

[0072] Furthermore, the U-Net model described in S4.3 is set to a depth of 2, a number of model classification categories of 2, an adaptive matrix solver, a maximum number of traversals of 30, a mini-batch size of 32 for each iterative training, an initial learning rate of 1e-5, a learning decay rate of 0.1, and a number of learning decay traversals of 10. The mathematical relationship of the multi-spatial-scale U-Net model is as follows:

[0073]

[0074] where m CNN2road represents a multi-spatial-scale U-Net model for generating intelligent road strength index, represents the intelligent road strength index generated by the U-Net model, and x represents the data in the CNN sample training set.

[0075] Furthermore, the U-Net model described in S4.4 is set to a depth of 2, a number of model classification categories of 2, an adaptive matrix solver, a maximum number of traversals of 30, a mini-batch size of 32 for each iterative training, an initial learning rate of 1e-5, a learning decay rate of 0.1, and a number of learning decay traversals of 10. The mathematical relationship of the multi-spatial-scale U-Net model is as follows:

[0076]

[0077] where m CNN2built represents a multi-spatial-scale U-Net model for generating smart city building indices, represents the smart city building index generated by the U-Net model, and x represents the data in the CNN sample training set.

[0078] Furthermore, step S5 specifically includes:

[0079] S5.1 uses the pixel index coordinates of valid pixels in the study area to superimpose the smart road intensity index and smart city building index with consistent geographical locations onto the basic dataset to form a comprehensive dataset;

[0080] S5.2 extracts the corresponding building and road feature indices in the comprehensive data set through the geographic location index of the basic training set to generate a comprehensive training set to generate a daily surface temperature product with a complete spatiotemporal sequence.

[0081] Furthermore, step S6 specifically includes:

[0082] S6.1 Use the landscape characteristic index product of channels 1-30 in the comprehensive sample training set as X and LST as y LST Train a random forest model.

[0083] S6.2 Use the trained random forest model to reconstruct all pixels on any valid surface temperature date within the study area to generate a daily surface temperature product with a complete spatiotemporal sequence.

[0084] Furthermore, in the random forest model of S6.1, the number of decision trees is set to 50 and the minimum leaf size is set to 5. The mathematical relationship is as follows:

[0085]

[0086] where m RF2LST represents the random forest model for reconstructing surface temperature, represents the surface temperature data reconstructed by the random forest model, and X represents the landscape characteristic index product in the comprehensive sample training set.

[0087] Furthermore, step S7 specifically includes:

[0088] S7.1 For all reconstructed surface temperature data within the same season, perform a weighted average based on the percentage of missing pixel values ​​due to cloud cover or other reasons before interpolation to obtain the surface temperature data for that season;

[0089] S7.2 Normalize the surface temperature data for the four seasons;

[0090] S7.3 Calculate the local Moran index for the normalized land surface temperature data to obtain the LST seasonal variation index.

[0091] Furthermore, the formula for weighted average of surface temperature in the same season is as follows:

[0092]

[0093] in, represents the simulated LST value on day i, w lst i Indicates the percentage of missing values ​​on that day, w lst i A value of 1 indicates complete absence, where i is any date containing a surface temperature. Represents the surface temperature in the kth season, where k is 1, 2, 3, or 4, corresponding to spring, summer, autumn, and winter.

[0094] Furthermore, the formula for normalizing the four seasons is as follows:

[0095]

[0096] Among them, LST k Represents the heat island index for a particular season, Represents the minimum value of the water body in the study area on that day, represents the 95th percentile value of the surface temperature of this sky and earth. Represents the normalized seasonal heat island index.

[0097] Furthermore, step S8 specifically includes:

[0098] S8.1 uses the features in the comprehensive training set as x, and uses the corresponding pixel position index of x to extract the heat island index of spring, summer, autumn and winter as Construct a training set for feature seasonality estimation.

[0099] S8.2 trains random forest models for the heat island index in different seasons and analyzes the contribution of various landscape factors to the heat island in different seasons.

[0100] S8.3 Rank the feature contributions to identify important landscape features associated with the urban heat island effect. Then, based on the physical significance of the urban space mapped by the features, analyze and guide potential urban space optimization solutions.

[0101] Compared with the existing technology, the advantages and positive effects of the present invention are: the present invention can fully integrate and develop the temporal, spatial and spectral information advantages of open source satellite image data, urban land use and surface cover data and GIS database, and combine them with intelligent GeoAI algorithms to fuse them into an urban surface temperature dataset with the optimal temporal and spatial scale of daily, monthly, seasonal and annual sequences to characterize the seasonal variation characteristics of the urban thermal environment.

[0102] Through GeoAI intelligent perception, the present invention establishes a mathematical logic model between urban landscape index and surface temperature, thereby obtaining urban-scale surface temperature downscaling data with complete spatiotemporal coverage and high spatial resolution. This method effectively solves the following three problems: First, the lack of urban spatial GIS information database leads to incomplete representation of urban spatial structure, which in turn affects the accurate assessment of urban thermal environment; second, the variety of urban landscape characteristic indices in remote sensing images and the lack of standardization make it difficult to accurately depict urban spatial structure and landscape characteristics; third, surface temperature data is easily affected by clouds and satellite operation cycles, resulting in incomplete spatiotemporal scales and low spatial resolution. Through the complete simulation of daily surface temperature, the present invention provides strong support for the accurate characterization and analysis of changes in urban thermal environment.

[0103] The present invention effectively overcomes the impact of clouds on urban thermal environment observations and ensures the spatiotemporal continuity of surface temperature data. Starting from machine learning intelligent perception, urban spatial characteristics are intelligently and accurately characterized, and on this basis, high-resolution downscaling of surface temperature data is achieved. In addition, the present invention also provides a feature importance analysis function to help users deeply explore the inherent mechanisms of the urban thermal environment, reveal key urban layout and spatial structure characteristics, provide a scientific basis for urban planning departments, and support the formulation of more targeted spatial optimization and governance strategies. The present invention makes full use of open source urban spatial information and remote sensing data to provide a highly universal and easy-to-promote solution for urban thermal environment research. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 Flowchart of the GeoAI machine learning-based intelligent downscaling seasonal characterization and analysis method for surface temperature provided by the present invention.

[0105] Figure 2 Provides intelligent characterization products for downscaling urban landscapes and surface temperatures based on GeoAI.

[0106] Figure 3 It is a product for analyzing the contribution of urban landscape factors to summer heat islands and winter cold islands. DETAILED DESCRIPTION

[0107] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0108] The purpose of this invention is to develop a method for intelligent downscaling and seasonal characterization and analysis of land surface temperature based on GeoAI machine learning. This method can interpolate missing values ​​for areas where surface temperature is missing due to cloud cover using available urban landscape characteristic index products and incomplete satellite land surface temperature inversion data, and downscale the land surface temperature product. To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0109] According to the present invention, to achieve this object, the present invention comprises the following steps:

[0110] S1: Obtain the Landsat-8 surface temperature product within the study area on a daily basis, preprocess it, and screen qualified daily surface temperature pixels;

[0111] Furthermore, step S1 specifically:

[0112] This study used a buffer zone with a radius of 30 km, centered at the city center, based on the city's spatial scale. Landsat-8 surface temperature products were obtained through the GEE platform. The buffer zone served as the study area, and all surface temperature products were masked and cropped to match the study area's spatial extent. Furthermore, the original Landsat-8 surface temperature products were further screened using a quality screening criteria of less than 20% cloud cover per image and image units covering more than 50% of the study area. Daily surface temperature images with qualified dates were obtained. In this example, qualified daily surface temperature images were obtained on days 8, 56, 72, 88, 200, 232, 264, 280, 296, 312, and 328. Furthermore, these qualified daily surface temperature images were resampled to 10 m resolution using a bilinear sampling method, to match the spatial resolution of the Sentinal-2 satellite.

[0113] Preferably, in step S1.2, the quality screening criteria for a single surface temperature image is that the cloud coverage is less than 20% and the surface temperature image covers more than 50% of the study area.

[0114] Furthermore, the calculation formula for the cloud cover rate of a single image is as follows:

[0115]

[0116] Among them, w i represents the cloud coverage rate on the i-th day, that is, the percentage of missing values ​​in pixels on the i-th day. represents the number of valid pixels in the Landsat-8 surface temperature product of the i-th day after cropping, that is, the effective image coverage area, w ref Indicates the number of pixels in the study area, that is, the area of ​​the study area. In this example, i = 8, 56, 72, 88, 200, 232, 264, 280, 296, 312, 328.

[0117] The validity of any spatial location pixel is determined as follows:

[0118] If the cloud mask cloudsBitMask and shadow mask cloudShadowBitMask in QA_Pixel are both equal to 0, then If QA_Pixel cloud mask cloudsBitMask is equal to 1 or shadow mask cloudShadowBitMask is equal to 1, then Where r and c represent the horizontal and vertical coordinates of the pixel position respectively.

[0119] S2: Obtain Sentianl-2 satellite band products, open source GHSL, Dynamic World and OSM landscape characteristic index products related to urban land use and surface cover on an annual basis and preprocess them.

[0120] Furthermore, step S2 specifically includes: obtaining the 2018 Sentinel-2 satellite band data (B2, B3, B4, B8A, B8, B9, B11, B12) through the Google Earth Engine (GEE) platform and using the band data obtained by the Sentinel-2 satellite to generate RSEI vegetation index products (NDVI, NDBI, NDWI); obtaining the 2018 night light data of the VIIRS sensor with a spatial resolution of 400m from the GEE platform, which can be used to measure the status of social and economic development; obtaining the SRTM elevation data with a spatial resolution of 30m from the Earthdata database from the GEE platform to reflect the changes in the terrain of the study area; and obtaining the 2018 night light data of the VIIRS sensor with a spatial resolution of 400m from the GEE platform. Using the GEE platform, we obtained 2018 urban land use and land cover data from the Dynamicworld database at a spatial resolution of 10m to serve as indicators for land use functional zoning. We also obtained built-up area building volume (BV) and building height (BH) data from the GHSL database at a spatial resolution of 100m from the GEE platform to serve as indicators for physical form factors in built-up areas. We also obtained building and road outline vector data from the OSM platform, converted them to raster data, and processed them using morphological techniques to serve as the basis for physical form factors in built-up areas. These data were used to generate the Smart City Building Index (Built Footprints) and the Smart Road Strength Index (Road Footprints). Using these obtained physical form factors, we generated the following built-up area physical form factors: building density (BD), sky width (SVF), and floor area ratio (FAR). The resulting landscape characteristic index products were uniformly cropped to the study area and interpolated to a 10m resolution using the nearest neighbor interpolation method.

[0121] Furthermore, the specific method for generating the RSEI vegetation index product is:

[0122] Using the band, the NDVI expression is generated as follows:

[0123]

[0124] Among them, NDVI stands for Normalized Difference Vegetation Index, NIR stands for Near Infrared Band, i.e. Band 8, and RED stands for Red Light Band, i.e. Band 4.

[0125] Using the band, generate NDWI and its expression is:

[0126]

[0127] Among them, NDWI stands for Normalized Difference Water Index, NIR stands for Near Infrared Band, i.e. Band 8, and GREEN stands for Green Light Band, i.e. Band 3.

[0128] Using the bands, the NDBI expression is generated as follows:

[0129]

[0130] Among them, NDBI stands for Normalized Building Index, NIR stands for near-infrared band, i.e. band 8, and SWIR stands for shortwave red light band, i.e. band 11.

[0131] Furthermore, the specific method for morphological processing after converting the building and road outline vectors into raster is as follows:

[0132] The rasterized street data is first binarized, assigning all grayscale values ​​greater than or equal to 1 to 1. Each pixel is then morphologically dilated by 15 pixels to its surroundings, merging multiple roads. Morphological skeleton extraction and further dilation by the same number of pixels are then performed to accurately depict the distribution of roads.

[0133] The rasterized building distribution data is first binarized, and all grayscale values ​​greater than or equal to 1 are assigned to 1; then, by finding connected areas, the partition morphological closing and filling operations are performed. In this example, a total of 7523 connected areas are found.

[0134] For the building density (BD) of urban built-up areas, the calculation formula is as follows:

[0135] BD=A b / A S ,

[0136] Among them, A b represents the total area of ​​all buildings in the block, A S Represents the total area of ​​the block.

[0137] For the floor area ratio (FAR) of urban built-up areas, the calculation formula is as follows:

[0138]

[0139] Among them, A i and F i are the area and number of floors of the i-th building, n represents the total number of buildings, and S represents the area of ​​the block.

[0140] For the sky width (SVF) of urban built-up areas, considering the complex spatial environment within a certain urban area, such as dense old residential areas or high-rise buildings, which affect the absorption and emission of long-wave radiation, the building height data (BH) and the elevation data (DEM) of the study area were made into a digital surface model and input into the SAGA GIS tool. At the same time, the maximum search radius was set to 100m to obtain the sky width (SVF) of urban built-up areas.

[0141] In a three-dimensional urban environment, the SVF calculation formula for a point is as follows:

[0142]

[0143] where β is the angle from the center point to the maximum obstacle height at a maximum distance equal to the constant search radius R.

[0144] S3: For any valid daily surface temperature data of the daily surface temperature product, screen the valid surface temperature pixels in the study area on that day, extract the LULC, surface temperature and Sentianl-2 satellite band information with consistent coordinates according to the geographic spatial location of the valid pixels, perform stratified random sampling, and construct a basic training sample set;

[0145] Furthermore, step S3 specifically includes: reading the LULC data, effective surface temperature data and Sentianl-2 satellite band product data with consistent geographical locations into a table to form a basic data set through the pixel linear index coordinates of the effective pixels in the study area; performing stratified sampling based on the surface temperature data in the basic data set as the stratification standard, setting an appropriate sampling interval of 10, iterating 100 times during sampling, and extracting 20 samples from each layer each time.

[0146] S4: Based on the basic training sample set, a multi-spatial-scale U-Net model is trained to estimate the probability index of buildings and roads, and generate the smart city building intensity index and smart road intensity index to characterize the distribution of buildings and roads in the urban space;

[0147] Furthermore, in this example, step S4 specifically includes constructing a multi-spatial-scale U-Net model based on a convolutional neural network (CNN) to generate the Smart City Building Index and the Smart Road Strength Index. Based on the basic sample training set, the index feature images of the image region with a pixel size of [64,64] are extracted from the geographic spatial location of each sample, and the CNN sample training set is constructed.

[0148] Furthermore, the landscape feature index product of channel 1 to channel 25 in the CNN sample training set is used as x, and the road distribution data is used as y roadTrain a multi-spatial scale U-Net model. In this example, a model with a training accuracy of 0.85 is selected to generate the intelligent road strength index. Channels 1 to 25 in the CNN sample training set are used as x, and building distribution data is used as y. built Train a multi-spatial-scale U-Net model. In this example, a model with a training accuracy of 0.83 is selected to generate the smart city building index.

[0149] Furthermore, the multi-spatial scale U-Net model for generating the intelligent road strength index is set to a depth of 2, a model classification category of 2, an adaptive matrix solver, a maximum number of iterations of 1, a mini-batch size of 32 per iterative training, an initial learning rate of 1e-5, a learning decay rate of 0.1, and a learning decay iteration number of 1. The mathematical relationship is as follows:

[0150]

[0151] where m CNN2road represents a multi-spatial-scale U-Net model for generating intelligent road strength index, represents the intelligent road strength index generated by the U-Net model, and x represents the data in the CNN sample training set.

[0152] Furthermore, the multi-spatial scale U-Net model for generating the smart city building index is set to a depth of 2, a model classification category of 2, an adaptive matrix solver, a maximum number of iterations of 1, a mini-batch size of 32 for each iterative training, an initial learning rate of 1e-5, a learning decay rate of 0.1, and a learning decay iteration number of 1. The mathematical relationship is as follows:

[0153]

[0154] where m CNN2built represents a multi-spatial-scale U-Net model for generating smart city building indices, represents the smart city building index generated by the U-Net model, and x represents the data in the CNN sample training set.

[0155] S5: Comprehensive landscape characteristic index product, based on the geographic spatial location of effective pixels, extracts comprehensive landscape characteristics consistent with the spatial location of stratified random sampling to generate a comprehensive training sample set;

[0156] The landscape characteristic index products in step S5 include three categories: the first category is satellite band data, including eight bands of Sentinel-2, namely B2, B3, B4, B8A, B8, B9, B11, and B12; the second category is remote sensing satellite index data (RSEI index product), including normalized vegetation index (NDVI), normalized water index (NDWI), and normalized building index (NDBI); the third category is surface cover data, including nine different types of land feature cover data; the fourth category is social-physical data, including the generation of smart city building index (Built Footprints), smart road intensity index (Road Footprints), built-up area building volume data (BV), built-up area building height data (BH), urban built-up area building density (BD), urban built-up area sky width (SVF) and urban built-up area floor ratio (FAR), elevation data (DEM) describing terrain undulations, and night light data (NTL) describing socio-economic development.

[0157] Furthermore, in this example, step S5 specifically involves overlaying the geographically consistent smart road intensity index and smart city building index onto the base dataset using the pixel index coordinates of valid pixels within the study area to form a comprehensive dataset. Furthermore, the corresponding building and road characteristic indices in the comprehensive dataset are extracted using the geographic location indexes of the base training set to generate a comprehensive training set, which is then used to generate a daily land surface temperature product with a complete spatiotemporal sequence.

[0158] S6: Based on the comprehensive training sample set, a random forest model is trained to reconstruct the LST data. The surface temperature data of any pixel position within the study area on that day are reconstructed to generate a daily surface temperature product with a complete spatial sequence.

[0159] Furthermore, in this example, step S6 specifically includes: using the landscape characteristic index product of channels 1-30 in the comprehensive sample training set as X, LST as y LST A random forest model is trained. In this example, a random forest model with an RMSE of 0.003 and an r-sequare of 0.956 is used to reconstruct all pixels within the study area on any valid date for surface temperature, generating a daily surface temperature product with a complete spatiotemporal sequence.

[0160] Furthermore, in the random forest model of S6.1, the number of decision trees is set to 50 and the minimum leaf size is set to 5. The mathematical relationship is as follows:

[0161]

[0162] where m RF2LST represents the random forest model for reconstructing surface temperature, represents the surface temperature data reconstructed by the random forest model, and X represents the landscape characteristic index product in the comprehensive sample training set.

[0163] S7: Take the weighted average of the LST products of spring, summer, autumn and winter and normalize them to generate the LST seasonal variation index.

[0164] Furthermore, in this example, step S7 specifically includes: for the reconstructed surface temperature data for several days in the same season, weighting the proportion of missing pixel values ​​due to cloud cover and other reasons before interpolation, performing weighted averaging to obtain the surface temperature data for that season. In this example, Zhengzhou is used as the study area. Based on the local climate, days 60 to 150 are classified as spring, days 151 to 243 are classified as summer, days 244 to 334 are classified as autumn, and days 335 to 365 and days 1 to 60 are classified as winter.

[0165] The surface temperature data of the four seasons were normalized. The local Moran index of the normalized surface temperature data was calculated using ArcGIS software to obtain the LST seasonal variation index.

[0166] Furthermore, the formula for weighted average of surface temperature in the same season is as follows:

[0167]

[0168] in, represents the simulated LST value on day i, w i Indicates the percentage of missing values ​​on that day, w i Equal to 1 means complete absence, where i = 8, 56, 72, 88, 200, 232, 264, 280, 296, 312, 328. Represents the surface temperature in the kth season, where k is 1, 2, 3, or 4, corresponding to spring, summer, autumn, and winter.

[0169] Furthermore, the formula for normalizing the four seasons is as follows:

[0170] y lst =(lst-lst water ) / (lst 95％Percentile -lst water ),

[0171] Among them, lst represents the surface temperature data of a specific season, lst water Represents the minimum value of water in the study area on that day, lst 95％Percentile represents the 95th percentile value of the surface temperature of the sky and the earth, y lstRepresents the normalized seasonal heat island index.

[0172] S8: Based on the comprehensive training sample set, a random forest model is trained for landscape feature importance analysis to generate feature contribution analysis, screen important landscape feature index products for the heat island effect, and analyze and guide potential spatial governance solutions.

[0173] Furthermore, in this example, step S8 specifically includes using the features in the comprehensive training set as x, using the pixel position index corresponding to x, and extracting the heat island index of the four seasons of spring, summer, autumn, and winter as y i , constructing a training set for characteristic seasonal assessment. A random forest model was trained based on the seasonal heat island index, analyzing the contribution of various landscape factors to the heat island effect in each season. Feature contributions were ranked to identify key landscape features associated with the heat island effect. Based on the physical significance of the urban space mapped by these features, potential urban space optimization solutions were analyzed and guided.

[0174] Furthermore, in this example, the contribution of various landscape factors to summer heat island and winter cold island is analyzed. Figure 3 As shown, land cover and socio-physical factors were significantly higher than spectral and vegetation index factors in both summer and winter environments. In the summer heat island environment, land cover (35.95%) and socio-physical factors (33.44%) contributed more significantly to UTE than spectral data (25.21%) or vegetation index (5.40%). In the winter cold environment, land cover (36.00%) and socio-physical factors (41.82%) contributed more significantly to UTE than spectral data (13.86%) or vegetation index (8.32%).

[0175] Depend on Figure 3 Among the socio-physical factors, elevation DEM and urban openness (SVF) contribute significantly to the urban heat island and cold island effects, followed by the summer night light index (NTL) and the winter building FAR. This suggests that urban space optimization requires considering regional terrain elevation to rationally plan the number, density, and height of buildings within a city to mitigate the urban heat island effect. Furthermore, GeoAI-generated intelligent simulated building and road indices generally outperform traditional remote sensing vegetation indices in capturing the impact of landscape factors on LST.

Claims

1. A GeoAI machine learning-based intelligent downscaling seasonal characterization and analysis method for surface temperature, characterized by: The following steps are involved: S1: Obtain the Landsat-8 surface temperature products within the study area on a daily basis, preprocess them, and screen effective daily surface temperature products; S2: Obtain landscape characteristic index products related to urban land use and land cover on an annual basis and preprocess them; S3: For the daily surface temperature data of any day in the valid daily surface temperature product, screen the valid surface temperature pixels in the study area on that day, extract urban land use and surface cover data, surface temperature data and Sentianl-2 satellite band data with consistent coordinates according to the geographic spatial location of the valid pixels, perform stratified random sampling, and construct a basic training sample set; S4: Based on the basic training sample set, a multi-spatial-scale U-Net model is trained to estimate the probability index of buildings and roads, and generate the smart city building index and smart road intensity index to characterize the distribution of buildings and roads in the urban space; S5: Comprehensive landscape characteristic index product, based on the geographic spatial location of effective pixels, extracts comprehensive landscape characteristics consistent with the spatial location of stratified random sampling to generate a comprehensive training sample set; S6: Based on the comprehensive training sample set, a random forest model is trained to reconstruct the surface temperature data. The surface temperature data of any pixel position within the study area on that day is reconstructed pixel by pixel to generate a daily surface temperature product with a complete spatial sequence. S7: Perform weighted average and normalization on the surface temperature products of spring, summer, autumn and winter to generate the surface temperature seasonal variation index; S8: Based on the comprehensive training sample set, a random forest model is trained for landscape feature importance analysis to generate feature contribution analysis, screen landscape feature index products that are important for the urban heat island effect, and analyze and guide potential spatial governance solutions.

2. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 1 is characterized in that In step S1, the preprocessing includes: surface temperature image acquisition, date screening, quality screening, study area clipping and resampling; step S1 specifically includes: S1.1: With the center of the study area as the center, set a buffer zone with a radius of r based on the spatial scale of the study urban area. r should be greater than the maximum distance between the center of the circle and the urban boundary. Obtain the Landsat-8 surface temperature product through the GEE platform. Using the buffer zone as the study area, mask all surface temperature products and crop them to the same spatial extent as the study area. S1.2: Extract the pixel quality band QA_Pixel of the Landsat-8 product, analyze the daily surface temperature fluctuations in the study area in different seasons and the degree of cloud influence, further screen the quality of the mask-extracted Landsat-8 surface temperature product, and obtain daily surface temperature images of qualified quality. S1.3 For qualified daily surface temperature images, a bilinear sampling method is used to resample them to a resolution of 10 m to make them consistent with the spatial resolution of the Sentinal-2 satellite.

3. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 2 is characterized in that In step S1.2, the quality screening criteria for a single surface temperature image is that the cloud coverage is less than 20% and the surface temperature image covers more than 50% of the study area; The calculation formula for the cloud cover rate of a single image is as follows: Among them, w i represents the cloud coverage rate on the i-th day, that is, the percentage of missing values ​​in pixels on the i-th day. represents the number of valid pixels in the Landsat-8 surface temperature product of the i-th day after cropping, that is, the effective image coverage area, w ref Indicates the number of pixels in the study area, that is, the area of ​​the study area; the criterion for determining the validity of a pixel at any spatial location is: if the cloud mask and shadow mask in QA_Pixel are both equal to 0, then If the cloud mask in QA_Pixel is equal to 1 or the shadow mask is equal to 1, then Where r and c represent the horizontal and vertical coordinates of the pixel position respectively.

4. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 1 is characterized in that The landscape characteristic index products in step S2 include: Sentinel satellite bands, remote sensing satellite indices, elevation data, nighttime light data, urban land use and surface cover data, building distribution outlines, and road outline data; step S2 specifically also includes: S2.1: Using the GEE platform, obtain band data of bands B2, B3, B4, B8A, B8, B9, B11, and B12 at a resolution of 10 m from the Sentinel-2 satellite database as band index indicators. Using the obtained band data, calculate and generate remote sensing satellite index products for the Normalized Difference Vegetation Index, Normalized Difference Water Index, and Normalized Building Index. These band index indicators and remote sensing satellite index products serve as channels 1-11 of the basic dataset. S2.2: Obtain land use and land cover data with a spatial resolution of 10m from the Dynamicworld data set on the GEE platform as indicators of land use functional zoning factors and as channels 12-21 of the basic dataset; S2.3: From the GEE platform, obtain the SRTM elevation data (DEM) with a spatial resolution of 30 m from the Earthdata database as the physical form factor indicator and as the 22nd channel of the basic dataset; S2.4: From the GEE platform, obtain the VIIRS sensor spatial resolution 400 night light data (NLT) as the socioeconomic factor indicator and the 23rd channel of the basic data set; S2.5: From the GEE platform, obtain the built-up area building volume data and building height data with a spatial resolution of 100m from the GHSL database as the physical form factors of the urban built-up area and as the 24th and 25th channels of the basic dataset; S2.6: Download the latest street data, building distribution data, street hierarchy, and building function attribute data for the study area from the OSM platform. Filter the street data and building distribution data, extract street and building distribution vector data, and perform vector-to-raster conversion and morphological processing. These data serve as the basic data for characterizing the physical form factors of road and building distribution in the urban built-up area. These data are used to generate physical form factors for the built-up area, including the Smart City Building Index and the Smart Road Intensity Index. The Smart City Building Index and the Smart Road Intensity Index are used as channels 26 and 27 of the comprehensive dataset. The obtained street vector data are then used to obtain the street block outlines for the study area. S2.7: Using the data obtained in S2.5 and S2.6, calculate and generate the physical form factor of the built-up area. The physical form factor of the built-up area includes: building density, sky openness, and floor ratio of the urban built-up area, as channels 28-30 of the comprehensive data set. In the urban 3D environment, the calculation formula for the sky openness of the urban built-up area at a certain point is: Where SVF is the sky openness of the urban built-up area, β is the angle θ from the center point to the maximum obstacle height at the maximum distance equal to the constant search radius R; S2.8: Perform pre-processing operations on the image products obtained in S2.1-S2.

7.

5. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 1 is characterized in that The step S3 specifically includes: S3.1 Use the daily surface temperature data of any day in the valid daily surface temperature product to select the pixel linear index coordinates of the valid surface temperature pixels in the study area on that day, and read the urban land use and surface cover data, valid surface temperature data and Sentianl-2 satellite band product data with consistent geographical locations into the table to form the basic data set; S3.2 uses the surface temperature data in the basic dataset as the stratification standard, performs stratified sampling, sets an appropriate sampling interval z, iterates k times during sampling, and randomly selects m samples in each layer each time to construct the training sample set; S3.3 draws the probability distribution histogram of the extracted samples to determine whether the sample distribution is uniform, and continuously adjusts k and m to make the sample distribution close to a balanced distribution, which serves as the basic sample training set and is used as the input of subsequent models.

6. The method for intelligent surface temperature characterization and seasonal analysis based on GeoAI machine learning according to claim 1 is characterized in that The step S4 specifically includes: S4.1 Build a multi-spatial-scale U-Net model based on convolutional neural networks to generate smart city building index and smart road intensity index; S4.2 Based on the basic sample training set, the geographic spatial location of each sample is used as the index, and the exponential feature image of the image area with a neighborhood pixel size of [64,64] is extracted as the sample to construct the CNN sample training set; S4.3 uses the landscape feature index product of channels 1 to 25 in the CNN sample training set as x and the road distribution data as y road Train the multi-spatial scale U-Net model, evaluate the model training effect by calculating the model training accuracy, and use the MSE with the smallest R 2 Higher models generate an intelligent road strength index; S4.4 uses the landscape feature index product of channels 1 to 25 in the CNN sample training set as x and the building distribution data as y built Train the multi-spatial scale U-Net model, evaluate the model training effect by calculating the model training accuracy, and use the MSE with the smallest R 2 Higher models generate smart city building indices; The mathematical relationship of the multi-spatial scale U-Net model in S4.3 is: where m CNN2road represents a multi-spatial-scale U-Net model for generating intelligent road strength index, represents the intelligent road strength index generated by the U-Net model, and x represents the data in the CNN sample training set; The mathematical relationship of the S4.4 multi-spatial scale U-Net model is: where m CNN2built represents a multi-spatial-scale U-Net model for generating smart city building indices, represents the smart city building index generated by the U-Net model, and x represents the data in the CNN sample training set.

7. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 1 is characterized in that The step S5 specifically includes: S5.1 uses the pixel index coordinates of valid pixels in the study area to superimpose the smart road intensity index and smart city building index with consistent geographical locations onto the basic dataset to form a comprehensive dataset; S5.2 extracts the corresponding smart city building index and smart road intensity index in the comprehensive data set through the geographic location index of the basic training set to generate a comprehensive training set to generate a daily surface temperature product with a complete spatiotemporal sequence.

8. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 1 is characterized in that The step S6 specifically includes: S6.1 uses the landscape characteristic index product of channels 1-30 in the comprehensive sample training set as X and the surface temperature as y LST Train the random forest model; the mathematical relationship of the random forest model is: Where m RF2LST represents the random forest model for reconstructing surface temperature, represents the surface temperature data reconstructed by the random forest model, and X represents the landscape characteristic index product in the comprehensive sample training set; S6.2 Use the trained random forest model to reconstruct all pixels on any valid surface temperature date within the study area to generate a daily surface temperature product with a complete spatiotemporal sequence.

9. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 1 is characterized in that The step S7 specifically includes: S7.1 For all reconstructed surface temperature data within the same season, perform a weighted average using the percentage of missing pixel values ​​due to cloud cover before interpolation to obtain the surface temperature data for that season; S7.2 normalizes the surface temperature data of the four seasons in sequence; S7.3 Calculate the local Moran index for the normalized surface temperature data to obtain the surface temperature seasonal variation index; the formula for normalization for the four seasons is: Among them, LST k Represents the heat island index for a particular season, Represents the minimum value of the water body in the study area on that day, represents the 95th percentile value of the surface temperature of this sky and earth. Represents the normalized seasonal heat island index.

10. The method for intelligent downscaling seasonal characterization and analysis of land surface temperature based on GeoAI machine learning according to claim 1 is characterized in that The step S8 specifically includes: S8.1 uses the features in the comprehensive training set as x, and uses the corresponding pixel position index of x to extract the heat island index of spring, summer, autumn and winter as Construct a training set for feature seasonality evaluation; S8.2 trains random forest models for heat island indices in different seasons and analyzes the contribution of various landscape factors to heat islands in different seasons; S8.3 Sort the feature contributions and screen the important landscape features associated with the urban heat island effect. Then, based on the physical meaning of the urban space mapped by the features, analyze and guide potential urban space optimization plans.

Images