A high-resolution method for estimating 2-meter surface temperature based on remote sensing imagery

Abstract

This invention discloses a high-resolution 2-meter surface temperature estimation method based on remote sensing imagery. The method utilizes both high-resolution and low-resolution remote sensing imagery data to extract feature information at different spatial scales, and then fuses and models them within a unified neural network framework. The model employs a feature interaction structure incorporating an attention mechanism to achieve the association and optimization of multi-scale features. The output generates a gridded 2-meter surface temperature distribution result for the corresponding spatial region. During training, a multi-objective hybrid supervision strategy is designed, combining meteorological reanalysis data and surface temperature reference data to simultaneously ensure prediction accuracy and spatial consistency. This invention enables city-level high-resolution temperature estimation in areas with sparse meteorological observation stations, effectively improving the spatial accuracy and stability of 2-meter surface temperature estimation.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing meteorological estimation and intelligent analysis of geographic information, specifically involving a high-resolution method for estimating 2-meter surface temperature based on remote sensing images. Background Technology

[0002] Current estimates of 2-meter surface temperature mainly rely on meteorological station observations or reanalysis data, with a spatial resolution typically of 0.1° (approximately 10 km), which is insufficient to reflect local temperature differences within urban areas and complex terrain regions. With the development of remote sensing technology, land surface temperature (LST) retrieved from satellite imagery can provide higher spatial resolution, offering a new approach for more precise estimation of 2-meter surface temperature. However, LST reflects surface radiation temperature, which exhibits a systematic bias compared to air temperature and is influenced by various factors such as surface type, vegetation cover, topography, and climatic conditions.

[0003] Traditional methods typically employ linear regression or empirical models to establish the relationship between LST and 2-meter air temperature (T2M), but these methods struggle to accurately characterize the nonlinear mapping between the two. Some studies have introduced machine learning algorithms (such as random forests and support vector regression) to improve fitting accuracy, but these still suffer from insufficient utilization of multi-scale image features and weak spatial generalization capabilities. Particularly in urban areas, the complex underlying surface structure and local climate effects make it difficult for traditional models to obtain stable estimation results.

[0004] Therefore, there is an urgent need for a deep learning method that can integrate multi-scale remote sensing image information and make full use of spatial features and environmental factors to achieve high-precision estimation of the air temperature 2 meters above the Earth's surface. Summary of the Invention

[0005] The purpose of this invention is to provide a high-resolution method for estimating 2-meter surface temperature based on remote sensing images, in order to solve the problems of low spatial resolution of meteorological station data, systematic deviation between surface temperature and air temperature retrieved by remote sensing, and insufficient ability of existing models to fuse multi-scale features and model nonlinear relationships.

[0006] The technical solution adopted in this invention is as follows: A high-resolution method for estimating 2-meter surface temperature based on remote sensing imagery includes the following steps: (1) For the target area, acquire high-resolution remote sensing images and low-resolution remote sensing images and perform normalization processing; for each image sampling point, extract multi-source geographic environmental features and temperature data; (2) A dual-branch residual network is used to process high-resolution and low-resolution remote sensing images respectively. The flattened high-resolution features are stitched together with the low-resolution features after global average pooling. Cross-scale attention fusion is achieved through a multi-layer Transformer encoder. High-resolution features are extracted from the fused feature sequence for feature dimensionality reduction. The dimensionality-reduced features are stitched together with the geographic environment features and then temperature regression is performed to obtain the predicted temperature value of 2 meters above the ground in the target area. (3) Design and weightedly combine the 2-meter surface temperature regression loss, the surface temperature ranking learning loss, and the spatial difference regularization loss to form the total loss function; (4) Perform online data augmentation on the training set images, adopt a dynamic target switching strategy to increase the diversity of training samples, configure the AdamW optimizer and set the learning rate and training hyperparameters, and use the total loss function to train the surface 2-meter temperature estimation model. (5) Input the remote sensing image and geographic features of the area to be predicted into the trained 2-meter surface temperature estimation model to obtain the predicted 2-meter surface temperature value of the corresponding area.

[0007] Furthermore, the multi-source geographic environmental features include static geographic features, climate zoning features, temporal features, and surface temperature reference data.

[0008] Furthermore, the dual-branch residual network includes a high-resolution branch network and a low-resolution branch network; the high-resolution branch network adopts a pre-trained ResNet-18 model, removes the last fully connected layer and global pooling layer, and retains the convolutional feature extraction part; the low-resolution branch network adopts a ResNet-18 model with independently trained parameters.

[0009] Furthermore, the high-resolution branch outputs a feature map with a dimension of 7×7×512, which is flattened into a feature sequence of 49×512; the low-resolution branch outputs a feature map with a dimension of 7×7×512, which is then subjected to global average pooling to obtain a feature vector of 1×512; the features from the two branches are concatenated to form a sequence of length 50 with a feature dimension of 512; the concatenated feature sequence is input into a cross-scale attention module using a two-layer Transformer encoder architecture, wherein the high-resolution features have position embedding encoding and the low-resolution features have scale feature embedding encoding, and multi-scale feature interaction and fusion are achieved through a multi-head self-attention mechanism.

[0010] Furthermore, each layer of the Transformer encoder contains a multi-head self-attention sublayer and a feedforward fully connected sublayer. The sublayers are connected by residual connections and layer normalization, and the output dimension is 512.

[0011] Furthermore, the regression loss for 2-meter surface temperature is calculated by averaging the predicted values ​​from multiple sub-regions and then using the ERA5 labels to calculate the mean square error, thus ensuring the overall temperature estimation accuracy. The surface temperature ranking learning loss utilizes the positive correlation between surface temperature and 2-meter surface temperature, pairing predicted values ​​from different sample sub-regions with reference surface temperature values ​​to calculate the relative ranking loss. The spatial variability regularization loss introduces a variance maximization regularization term to avoid over-smoothing of sub-region predicted values.

[0012] Furthermore, the dynamic target switching strategy is as follows: at the beginning of each training round, a time index t, t∈{0,1,…,23} is randomly selected, and the supervision target for that training round is set to T2M. t The next training round will use a different time index.

[0013] Furthermore, the output of the surface 2-meter temperature estimation model is a temperature thermogram with a 7×7 grid temperature distribution.

[0014] To achieve the above objectives, the present invention also provides a high-resolution 2-meter surface temperature estimation device based on remote sensing imagery, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned high-resolution 2-meter surface temperature estimation method based on remote sensing imagery.

[0015] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for estimating high-resolution 2-meter surface temperature based on remote sensing imagery.

[0016] The beneficial effects of this invention are: 1. By introducing a multi-scale remote sensing image fusion mechanism, the surface detail information of high-resolution images and the regional background information of low-resolution images are taken into account, thereby improving the spatial accuracy and generalization ability of temperature estimation. 2. By utilizing the cross-scale attention structure of deep learning, the complex nonlinear relationships between surface temperature, topography, vegetation, and climate elements are effectively captured; 3. The proposed model can operate stably under complex underlying surface conditions such as urban areas, plateaus, and hills, providing high-resolution temperature estimation data support for regional climate monitoring, urban heat island research, and meteorological services; 4. This method features a high degree of automation and strong scalability, and is suitable for the task of accurately estimating the surface temperature at 2 meters in different climate zones and regions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort: Figure 1 This is a schematic diagram of the structure of the surface air temperature estimation model at 2 meters provided by the present invention; Figure 2 A flowchart of a high-resolution surface temperature estimation method based on remote sensing imagery at 2 meters, provided by this invention; Figure 3 This is a schematic diagram of the estimated air temperature at 2 meters above the ground provided by the present invention; Figure 4 This is a schematic diagram of the device provided by the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a high-resolution method for estimating 2-meter surface temperature based on remote sensing imagery, which includes the following steps: Step S1: Acquisition and preprocessing of multi-source data.

[0020] S11: Remote sensing image data acquisition.

[0021] For the target area, two remote sensing images with different spatial resolutions were acquired: a high-resolution remote sensing image (Level 20, denoted as L20), with a spatial resolution of approximately 9.6 meters per pixel and an image size of 224×224 pixels, covering a geographic area of ​​approximately 2150 meters × 2150 meters; and a low-resolution remote sensing image (Level 17, denoted as L17), with a spatial resolution of approximately 76.8 meters per pixel and an image size of the same 224×224 pixels, providing broader regional background information.

[0022] The center geographic coordinates of the two image scales are strictly aligned to ensure spatial consistency. This includes high-resolution satellite imagery and publicly available remote sensing image data.

[0023] S12: Geographical environment feature extraction.

[0024] For each image sampling point, the following multi-source geographic environmental features are extracted: Static geographic features: Elevation E, extracted from the SRTM (Shuttle Radar Topography Mission) digital elevation model with an accuracy of 30 meters; geographic coordinates (λ, φ), representing longitude and latitude, respectively.

[0025] The elevation was normalized.

[0026] in, This represents the elevation after normalization.

[0027] Climate zoning characteristics: Based on the Köppen climate classification system, global climate is divided into 31 categories, denoted as C∈{1,2,…,31}. For sampling points with missing climate zone data, a nearest neighbor filling strategy is adopted: the nearest valid climate zone value is searched within a radius R=10km; if no valid value is found, it is assigned to a special category.

[0028] Time characteristics. Extracted from meteorological observation timestamps: month M∈{0,1,…,11} (coded starting from 0); date D∈{0,1,…,30}; hour H∈{0,1,…,23}.

[0029] Land surface temperature reference data: The temperature values ​​of a 7×7 spatial grid (49 points in total) were extracted from the land surface temperature (LST) data obtained by inversion from the thermal infrared band of Landsat 8 / 9 satellite using the bilinear interpolation method.

[0030] Bilinear interpolation algorithm: For the target point coordinates First, convert it to a floating-point index in the LST raster coordinate system. , ),in , These are floating-point indices for the column index and row index, respectively. Calculate the integer indices for the four corner points:

[0031]

[0032] in, , , , These represent the left column integer index, right column integer index, bottom row integer index, and top row integer index adjacent to the target point, respectively.

[0033] Extract the LST values ​​of the four corner points surrounding the target point: , , , (The subscripts represent the bottom left, bottom right, top left, and top right, respectively.)

[0034] Calculate the bilinear interpolation weights:

[0035] in, These represent the normalized fractional offsets of the target point in the row and column directions, respectively.

[0036] The interpolation result is:

[0037] The weights are calculated as follows:

[0038]

[0039]

[0040]

[0041] in, Indices representing the indices of valid corner points, ∑ i This represents summing over all valid corner points. This represents the LST value of the i-th valid corner point. This indicates the interpolation weight corresponding to that corner point. , , , These represent the bilinear interpolation weights corresponding to the four corner points: bottom left, bottom right, top left, and top right.

[0042] For corner points containing invalid values ​​(nodata or NaN), only valid corner points are used for weighted averaging, and the weights are normalized.

[0043] For the 49 sub-regions of the 7×7 grid, the center coordinates of each sub-region are calculated using offset:

[0044]

[0045] Where i and j represent the indices of the 7×7 grid in the longitude and latitude directions, respectively, i,j∈{0,1,…,6}. and Let these represent the longitude and latitude of the center point of the i-th and j-th sub-regions, respectively. and This indicates the longitude and latitude of the current sample center point. and These represent the coordinate offset step size of the sub-region center point in the longitude and latitude directions, respectively, Δλ=Δφ≈0.0000864° (corresponding to a geographical distance of approximately 9.6 meters).

[0046] S13: ERA5 reanalysis temperature data extraction.

[0047] Two-meter air temperature (T2M) was extracted from ERA5 global atmospheric reanalysis data as a monitoring label. The ERA5 data has a spatial resolution of 0.1° and a temporal resolution of hourly.

[0048] For each sampling point, the hourly temperature values ​​for 24 hours (0-23:00) are extracted and recorded as follows: The same bilinear interpolation method as LST is used to extract the temperature values ​​for the corresponding coordinates from the ERA5 grid data.

[0049] S14: Data cleaning and quality control.

[0050] Perform the following data cleaning steps: Remove samples with missing LST values ​​(including NaN or nodata); remove samples with missing T2M labels; remove samples with non-existent image file paths; remove samples with missing elevation values; truncate elevation outliers: E∈[-100,2000] meters.

[0051] The effective sample retention rate after data cleaning should be no less than 85%.

[0052] Step S2: Deep neural network model construction.

[0053] like Figure 1 As shown, the deep learning model proposed in this invention includes four core modules: a dual-branch feature extraction module, a multi-scale feature fusion module, an environmental feature embedding module, and a temperature regression prediction module.

[0054] S21: Dual-branch feature extraction module.

[0055] Two independent convolutional neural networks are used to process high-resolution and low-resolution remote sensing images respectively.

[0056] High-resolution branching networks: enter: (RGB three-channel image); Backbone network: A pre-trained ResNet-18 model is used, with the last fully connected layer and global pooling layer removed, while the convolutional feature extraction part is retained.

[0057] Output feature map: ∈ ; For an input of 224×224, = =7, meaning the output is a feature map with 512 channels and a spatial resolution of 7×7. , These represent the height and width of the high-resolution branch output feature map, respectively.

[0058] The feature extraction process of ResNet-18 can be represented as follows:

[0059] in These are the network parameters for the high-resolution branch.

[0060] Low-resolution branching networks: enter: ∈ ; Backbone network: It also uses the ResNet-18 structure, but the parameters are trained independently.

[0061] Output feature map: ∈ , = =7; and These represent the height and width of the low-resolution branch output feature map, respectively.

[0062]

[0063] in These are the network parameters for the low-resolution branch.

[0064] Feature normalization: Batch normalization is performed on the extracted feature maps;

[0065] in and These are the batch mean and variance, respectively, and ε is the numerical stability constant (usually taken as...). ), γ and β are learnable scaling and translation parameters; low-resolution features are also normalized.

[0066] S22: Multi-scale feature fusion module.

[0067] Low-resolution feature aggregation: Apply adaptive average pooling to low-resolution feature maps to compress the spatial dimension to 1×1;

[0068] The specific calculation formula is as follows:

[0069] Location coding addition: To enable the model to perceive spatial location information, two-dimensional sinusoidal location coding is used.

[0070] High-resolution position coding: This corresponds to 7×7=49 spatial locations; Low-resolution positional encoding: This corresponds to one aggregated feature; Reshape the feature map into a sequence and add positional encoding:

[0071]

[0072] Sequence concatenation: concatenating high- and low-resolution feature sequences along spatial dimensions;

[0073] Transpose to Transformer input format (batch_size, seq_len, dim):

[0074] Transformer cross-scale attention fusion: A multi-layer Transformer encoder is used to achieve cross-scale feature interaction. A single-layer Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network.

[0075] Multi-head self-attention mechanism: First, calculate the query, key, and value matrix:

[0076] in This is the learnable parameter matrix.

[0077] For multi-head attention (assuming h=8 heads), Q, K, and V are divided into h subspaces:

[0078] Where i∈{1,2,…,h}, and the dimension of each head is... =512 / 8=64.

[0079] Calculate attention score:

[0080] scaling factor This is used to prevent the gradient from vanishing due to an excessively large dot product.

[0081] Multi-head output splicing:

[0082] in ∈ This is for outputting the projection matrix.

[0083] Feedforward neural networks:

[0084] in , The dimension expansion factor is 4.

[0085] Residual connectivity and layer normalization:

[0086]

[0087] This invention employs an L=2 layer stack of Transformer encoders:

[0088] High-resolution feature extraction: Extract high-resolution features corresponding to the first 49 positions from the fused sequence;

[0089] Feature dimensionality reduction: Feature dimensionality reduction is performed using fully connected layers and activation functions;

[0090] in:

[0091] The output dimension is reduced to 64. This is the residual term.

[0092] The SiLU (Sigmoid Linear Unit) activation function is defined as follows:

[0093] Output: ∈

[0094] S23: Environmental feature embedding module.

[0095] Categorical feature embedding: Climate zone features are converted into continuous vectors using an embedding layer;

[0096] Where the embedding matrix The 31 climate zone categories are mapped to a 2-dimensional vector space.

[0097] Temporal features are also represented using embedding:

[0098]

[0099]

[0100] The dimensions of the embedding matrix are respectively . and These represent the embedding vectors for month, day, and hour, respectively.

[0101] Temporal features are fused through element-wise addition:

[0102] Continuous feature normalization:

[0103] Integrated geographic feature vector: concatenates all environmental features into a unified vector;

[0104] This represents a comprehensive geographic environment feature vector, where [;] indicates a vector concatenation operation.

[0105] Spatial expansion: The geographic feature vector is copied and extended to 49 sub-regions;

[0106] That is, each sub-region shares the same environmental characteristics.

[0107] S24: Temperature Regression Prediction Module.

[0108] Feature fusion: combining the dimensionality-reduced features with geographic features;

[0109] Multilayer perceptron (MLP) regression: Temperature regression is performed using a two-layer MLP.

[0110]

[0111] in: , ;

[0112] Final output: ∈ , representing the temperature prediction values of 49 sub-regions of a 7×7 grid.

[0113] The complete forward propagation process can be expressed as:

[0114] where θ contains all learnable parameters.

[0115] Step S3: Hybrid supervised training strategy.

[0116] The present invention designs a hybrid loss function that combines multiple supervised signals to make full use of sparse meteorological observation data and rich remote sensing LST data.

[0117] S31: Regression loss of 2-meter air temperature at the surface (T2M).

[0118] It is used to ensure the absolute accuracy of the overall temperature estimation. First, the predicted values of the 49 sub-regions are averaged and aggregated:

[0119] Calculate the mean squared error (MSE) with the ERA5 label:

[0120] where N is the batch size, from ERA5 reanalysis data.

[0121] S32: LST ranking learning loss.

[0122] Utilize the positive correlation between land surface temperature (LST) and 2-meter air temperature at the surface (T2M) for weak supervised learning through the relative magnitude relationship.

[0123] For all samples within a batch, flatten the predicted values of the 49 sub-regions of each sample and the corresponding LST reference values:

[0124]

[0125] Record the sample index to which each element belongs:

[0126] Generate all possible pairs (i,j), where i < j and S(i) ≠ S(j) (only consider pairs between different samples to avoid over-constraint within the same sample).

[0127] For each pair (i,j), calculate the ranking loss:

[0128] The relative ranking loss is defined as:

[0129] The physical meaning of this loss is: if (sign=+1), then the expectation is... ,Right now When the value is >1, the loss is 0; if (sign=-1), then the expectation is... ,Right now The loss is 0 when the value is greater than 1; if the prediction violates the relative relationship of LST, the loss will be proportional to the difference.

[0130] The total ranking loss is the average of all valid pairings:

[0131] in This is a cross-sample pairing set.

[0132] S33: Spatial difference regularization loss.

[0133] To avoid the predicted values ​​of the 49 sub-regions becoming too similar (oversmoothing), a variance maximization regularization term is introduced:

[0134] The variance is calculated as follows:

[0135] The negative sign indicates maximizing variance and encouraging reasonable temperature variations in space.

[0136] S34: Total loss function.

[0137] Three types of losses in weighted combination:

[0138] in, The regression loss is the average sub-region predicted value and the measured T2M. For loss due to intra-sample prediction variance diversity, The relative ranking loss is based on land surface temperature (LST), and it is calculated as follows: for all sub-block pairs (i, j) of different samples, if... Then the expectation is Its loss is defined as:

[0139] In a preferred embodiment of the present invention, the weights are set as follows: =1.0, =0.0 (Difference loss is optional).

[0140] Step S4: Model training and optimization.

[0141] S41: Data augmentation strategy.

[0142] To improve the model's generalization ability and robustness, online data augmentation was performed on the training set images: Geometric transformation: random horizontal flip: probability p=0.5; random vertical flip: probability p=0.5.

[0143] Color jitter: Brightness adjustment: factor ∈ [0.8, 1.2]; Contrast adjustment: factor ∈ [0.8, 1.2]; Saturation adjustment: factor ∈ [0.8, 1.2]; Hue adjustment: factor ∈ [-0.1, 0.1].

[0144] Standardization: Statistical parameters of the ImageNet pre-trained model:

[0145] Among them μ=[0.485,0.456,0.406], σ=[0.229,0.224,0.225].

[0146] The validation and test sets are only standardized, without geometric or color transformations.

[0147] S42: Dynamic target switching mechanism.

[0148] Since ERA5 provides hourly temperature labels for 24 hours, this invention employs a dynamic target switching strategy to increase the effective diversity of training samples: At the beginning of each epoch (training round), a time index t∈{0,1,…,23} is randomly selected, and the supervision objective for that epoch is set to T2M. t The next epoch will use a different time index.

[0149] This strategy allows the same sample to use different T2M labels in different epochs, effectively increasing the size of the training set by 24 times.

[0150] S43: Optimizer configuration.

[0151] The AdamW optimizer is used, and its update rules are as follows:

[0152]

[0153]

[0154]

[0155] in: Let be the gradient at step t; and These are the first-order and second-order moment estimates, respectively; , (Default value); α is the learning rate, which is set to 0.0001 in this invention; The weight decay coefficient is set to 0.0001. is the numerical stability constant.

[0156] S44: Set the learning rate and training hyperparameters (Table 1).

[0157] Employing a stepped learning rate decay strategy (StepLR):

[0158] in: =0.0001 is the initial learning rate; γ=0.8 is the decay factor; =20 is the decay step size (decays once every 20 epochs).

[0159] Table 1: Training Hyperparameter Settings

[0160] S45: Save model checkpoints.

[0161] During training, every S save =Save a model checkpoint every 500 training steps.

[0162] To avoid excessive disk space usage, a circular deletion strategy is adopted, retaining a maximum of 100 recent checkpoints.

[0163] Evaluate the model performance on the validation set and save the model with the minimum validation loss as the best model.

[0164] Step S5: Model inference and result output.

[0165] S51: Inference Input Preparation.

[0166] For the target area to be predicted: prepare remote sensing image data at two scales, L20 and L17; extract geographical features (elevation, latitude and longitude, climate zone, time); standardize the images (consistent with training); do not perform data augmentation transformations.

[0167] S52: Forward reasoning.

[0168] Load the trained model weights and set them to evaluation mode.

[0169] Output , where B is the batch size and 49 is the number of sub-regions for each sample.

[0170] S53: Spatial grid coordinate calculation.

[0171] For the center coordinates of each sample Calculate the coordinates of the 49 sub-regions in a 7×7 grid, such as... Figure 3 As shown.

[0172] The formula for calculating coordinates is:

[0173] Where i,j∈{0,1,2,3,4,5,6}.

[0174] The formatted output of the results is shown in Table 2.

[0175] Table 2: Prediction Results

[0176] Temperature values ​​are processed with precision, retaining 3 decimal places, and using half-precision floating-point numbers to reduce storage space.

[0177] Output example: longitude,latitude,elevation,slice,pred,image_id,image_path 114.17234,22.28456,15.3,0,28.523,1001, / path / to / L20 / img_001.png 114.17235,22.28456,15.3,1,28.687,1001, / path / to / L20 / img_001.png 114.17236,22.28456,15.3,2,28.431,1001, / path / to / L20 / img_001.png For large-scale city-level applications (such as 300,000 samples), the total number of output points is 300,000 × 49 = 14.7 million temperature value points, which can generate a continuous temperature distribution map with a resolution of 9.6 meters.

[0178] Step S6: Model validation and accuracy assessment.

[0179] S61: Validation data preparation.

[0180] Randomly select 20% of the original dataset as the validation set to ensure that the spatial distribution of the validation set samples is similar to that of the training set. Set a random seed to ensure the reproducibility of the data partitioning.

[0181] S62: Accuracy evaluation index.

[0182] The following statistical indicators were used to evaluate the model performance: Mean Absolute Error (MAE):

[0183] Root Mean Square Error (RMSE):

[0184] Coefficient of determination (R²):

[0185] Where M is the number of validation samples. These are observations from the weather station. This is the observed mean.

[0186] Mean Deviation (MBE):

[0187] Used to assess a systematic tendency to overestimate or underestimate.

[0188] S63: Weather station verification.

[0189] Referring to Table 3, external validation was performed using measured data from independent weather stations: the model prediction values ​​closest to the weather station location were extracted (the closest point was selected from 49 sub-regions); the observation times were matched to ensure time consistency; and the above accuracy indicators were calculated.

[0190] Table 3: Examples of Verification Results

[0191] S64: Ablation test.

[0192] See Table 4. To verify the effectiveness of each module, an ablation experiment was conducted.

[0193] Table 4: Ablation Experiment Parameters

[0194] Ablation experiments show that multi-scale fusion significantly improves accuracy compared to single-scale fusion (MAE is reduced by about 30%); Transformer fusion outperforms simple splicing (MAE is reduced by 21%); LST ranking learning loss contributes about 12% to the accuracy improvement; environmental features (climate zone, time) are crucial to generalization performance.

[0195] Step S7: Cross-region generalization test.

[0196] S71: Multi-city joint training.

[0197] To enhance the model's cross-regional generalization ability, data from multiple cities were used for joint training: Training data composition: City A (subtropical monsoon climate, Cwa): 8000 samples; City B (subtropical monsoon climate, Cfa): 12000 samples; City C (subtropical monsoon climate, Cfa): 6000 samples; City D (subtropical monsoon climate, Cfa): 4000 samples.

[0198] Total: 30,000 training samples, covering different levels of urbanization and terrain conditions.

[0199] S72: Validation of migration across climate zones.

[0200] See Table 5 for examples of migration tests conducted in cities in tropical climate zones to assess the model’s climate adaptability.

[0201] Table 5: Migration Test Data

[0202] Cross-climate zone test results show that the model maintains reasonable accuracy (MAE≤1.0°C) in unseen climate zones; the Köppen climate zone embedding effectively captures differences in climate types; and ERA5 global data ensures spatiotemporal consistency.

[0203] Example 2

[0204] This embodiment applies the trained model to the refined monitoring of the urban heat island effect in city A.

[0205] Application scenario: On August 15, 2024, during the high-temperature period of summer, estimate the temperature of the entire city of A.

[0206] Data scale: Number of image samples: 306,000; Number of output temperature points: 306,000 × 49 = 14,994,000; Spatial resolution: 9.6 meters; Coverage area: approximately 1,100 square kilometers.

[0207] Computing performance. Hardware: NVIDIA A100 GPU (40GB); Inference time: approximately 4 hours; Peak memory: 38GB; Throughput: approximately 1270 samples / minute.

[0208] Figure 3 The model outputs a 7×7 spatial gridded temperature distribution and its spatial variability within urban areas, as shown in the example. Figure 3 As shown, the model of this invention provides local temperature estimation results for different surface types (such as roads, water bodies, building complexes, etc.). Each set of results, from left to right, is as follows: Original Image: Displays the input high-resolution surface image; Temperature Overlay: The predicted 2-meter air temperature is overlaid on the original image in a semi-transparent manner to show the spatial distribution trend; Temperature Heatmap: Displays the temperature distribution of the 7×7 grid output by the model.

[0209] The results show that the model can identify significant temperature differences between roads and water bodies, with water areas generally having lower temperatures. In high-density building areas, the temperature distribution exhibits a clear spatial gradient, with higher temperatures in subgrids closer to the center of the building complex. In green or shaded areas, the model can accurately reflect local cooling effects. The prediction results are smooth and continuous, and can capture microscopic thermal environment changes at a scale of 9.6 meters.

[0210] Applications in urban planning: Identify high-temperature vulnerable areas and optimize the layout of summer resorts; assess the cooling benefits of green spaces and guide ecological space planning; analyze ventilation corridors and optimize building layout; support thermal environment assessment and assist in decision-making.

[0211] Example 3

[0212] This embodiment can integrate the model into a smart city management platform to achieve real-time high temperature early warning.

[0213] The overall architecture of the extreme high temperature early warning system of the present invention can be divided into three main layers: data acquisition layer, model inference layer and application service layer. Data flow and functional collaboration between the layers are achieved through standardized interfaces.

[0214] Data Acquisition Layer. This layer is responsible for the acquisition and integration of multi-source data, including: Remote sensing image acquisition module: used for periodically downloading or retrieving high-resolution remote sensing image data, providing input of surface information; Meteorological data interface module: Obtains real-time or historical meteorological elements through meteorological data services (such as ERA5, ground observation stations); Geographic Data Management Module: Manages auxiliary geographic information such as digital elevation models (DEM), climate zones, and administrative divisions, providing geographic feature support for the models.

[0215] The model inference layer, which is the core computational part of the system, mainly includes: Deep learning model inference engine: Deploys a trained 2-meter surface temperature estimation model, supporting GPU acceleration and batch parallel processing; It is responsible for receiving multi-source data input from the acquisition layer and generating high-resolution (9.6-meter level) temperature estimation results.

[0216] The application service layer, which enables the visualization of model results and their application in business, mainly includes: Temperature distribution visualization module: Renders the estimation results as a heat map or spatial distribution map for users to view interactively; High Temperature Warning Service Module: Automatically identifies high temperature areas and triggers tiered warnings based on set thresholds; Decision support module: Integrating with urban planning and emergency management systems, it provides thermal risk analysis, response strategy suggestions, and policy-assisted decision support.

[0217] Real-time monitoring process: Data Updates: Receive the latest remote sensing imagery and meteorological data every hour; Model inference: Batch processing to generate global temperature distribution; Early warning judgment: Identify high-temperature areas based on threshold classification; Information push: Send early warning notices to relevant departments and the public; Emergency response: Measures such as opening summer resorts have been initiated.

[0218] The key technical parameters of this invention are summarized in Table 6.

[0219] Table 6: Key Technical Parameters

[0220] Through the above-mentioned technical solution, this invention achieves high-precision and high-resolution estimation of air temperature at a depth of 2 meters above the Earth's surface, and can be widely applied in fields such as urban heat island monitoring, extreme high-temperature early warning, and climate change research.

[0221] See Figure 4 The present invention provides a high-resolution 2-meter surface temperature estimation device based on remote sensing imagery, comprising one or more processors for implementing a high-resolution 2-meter surface temperature estimation method based on remote sensing imagery as described in the above embodiments.

[0222] The embodiments of the high-resolution 2-meter surface temperature estimation device based on remote sensing imagery of this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 4 The diagram shown is a hardware structure diagram of any data processing-capable device, including the high-resolution 2-meter surface temperature estimation device based on remote sensing imagery of this invention. (Except for...) Figure 4In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0223] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0224] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0225] Accordingly, embodiments of the present invention also provide a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the high-resolution surface 2-meter temperature estimation method based on remote sensing images described in the above embodiments.

[0226] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0227] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0228] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-resolution method for estimating 2-meter surface air temperature based on remote sensing imagery, characterized in that, Includes the following steps: (1) For the target area, acquire high-resolution remote sensing images and low-resolution remote sensing images and perform normalization processing; for each image sampling point, extract multi-source geographic environmental features and temperature data; (2) A dual-branch residual network is used to process high-resolution and low-resolution remote sensing images respectively. The flattened high-resolution features are stitched together with the low-resolution features after global average pooling. Cross-scale attention fusion is achieved through a multi-layer Transformer encoder. High-resolution features are extracted from the fused feature sequence for feature dimensionality reduction. The dimensionality-reduced features are stitched together with the geographic environment features and then temperature regression is performed to obtain the predicted temperature value of 2 meters above the ground in the target area. (3) Design and weightedly combine the 2-meter surface temperature regression loss, the surface temperature ranking learning loss, and the spatial difference regularization loss to form the total loss function; (4) Perform online data augmentation on the training set images, adopt a dynamic target switching strategy to increase the diversity of training samples, configure the AdamW optimizer and set the learning rate and training hyperparameters, and use the total loss function to train the surface 2-meter temperature estimation model. (5) Input the remote sensing image and geographic features of the area to be predicted into the trained 2-meter surface temperature estimation model to obtain the predicted 2-meter surface temperature value of the corresponding area.

2. The high-resolution surface air temperature estimation method based on remote sensing imagery according to claim 1, characterized in that, The multi-source geographic environmental features include static geographic features, climate zoning features, temporal features, and surface temperature reference data.

3. The high-resolution surface air temperature estimation method based on remote sensing imagery according to claim 1, characterized in that, The dual-branch residual network includes a high-resolution branch network and a low-resolution branch network; the high-resolution branch network adopts a pre-trained ResNet-18 model, removes the last fully connected layer and global pooling layer, and retains the convolutional feature extraction part; the low-resolution branch network adopts a ResNet-18 model with independently trained parameters.

4. The high-resolution surface air temperature estimation method based on remote sensing imagery according to claim 3, characterized in that, The high-resolution branch outputs a feature map with a dimension of 7×7×512, which is flattened into a feature sequence of 49×512. The low-resolution branch outputs a feature map with a dimension of 7×7×512, which is then subjected to global average pooling to obtain a feature vector of 1×512. The features from the two branches are concatenated to form a sequence of length 50 with a feature dimension of 512. The concatenated feature sequence is then input into a cross-scale attention module that employs a two-layer Transformer encoder architecture. The high-resolution features have position embedding encoding, and the low-resolution features have scale feature embedding encoding. Multi-scale feature interaction and fusion are achieved through a multi-head self-attention mechanism.

5. The high-resolution surface air temperature estimation method based on remote sensing imagery according to claim 1, characterized in that, Each layer of the Transformer encoder contains a multi-head self-attention sublayer and a feedforward fully connected sublayer. The sublayers are connected by residual connections and layer normalization, and the output dimension is 512.

6. The high-resolution surface air temperature estimation method based on remote sensing imagery according to claim 1, characterized in that, The regression loss for 2-meter surface temperature is calculated by averaging the predicted values ​​of multiple sub-regions and then calculating the mean square error with the ERA5 label to ensure the overall temperature estimation accuracy. The surface temperature ranking learning loss utilizes the positive correlation between surface temperature and 2-meter surface temperature, and pairs the predicted values ​​and surface temperature reference values ​​of different sample sub-regions to calculate the relative ranking loss. The spatial variability regularization loss introduces a variance maximization regularization term to avoid over-smoothing of sub-region predictions.

7. The high-resolution surface air temperature estimation method based on remote sensing imagery according to claim 1, characterized in that, The dynamic target switching strategy is as follows: at the beginning of each training round, a time index t is randomly selected, t∈{0,1,…,23}, and the supervision target for that training round is set to T2M. t The next training round will use a different time index.

8. The high-resolution surface air temperature estimation method based on remote sensing imagery according to claim 1, characterized in that, The output of the surface temperature estimation model at 2 meters is a temperature heat map with a 7×7 grid temperature distribution.

9. A high-resolution surface temperature estimation device based on remote sensing imagery (2 meters), comprising a memory and a processor, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the high-resolution 2-meter surface temperature estimation method based on remote sensing imagery as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the high-resolution 2-meter surface temperature estimation method based on remote sensing imagery as described in any one of claims 1-8.

Images