A spatiotemporal modeling land surface temperature downscaling method and system considering energy constraint

Abstract

The application discloses a kind of spatiotemporal modeling ground surface temperature downscaling method and system considering energy constraint, comprising: 1) acquisition and preprocessing of multi-source remote sensing and meteorological data;2) feature grouping and spatiotemporal feature tensor construction;3) time feature and spatial feature extraction;4) spatiotemporal feature modeling and high-resolution ground surface temperature estimation;5) loss construction and cross-scale consistency constraint;6) model iterative training and result output.The application breaks through the modeling limitation of time variation law and spatial detail description in the prior art, realizes the collaborative promotion of time continuity and spatial fine expression of ground surface temperature.On this basis, the overall temperature deviation and the problem of insufficient physical rationality commonly existing in the prior art are also effectively avoided, and the application is significantly superior to the prior art in terms of timing stability, spatial accuracy and physical reliability.

Description

Technical Field

[0001] This invention relates to a method and system for downscaling surface temperature, and more particularly to a spatiotemporal modeling method and system for downscaling surface temperature that considers energy constraints. Background Technology

[0002] Land surface temperature (LST) is a crucial parameter describing energy exchange and thermal state on the Earth's surface, playing a central role in research and applications such as climate change monitoring, drought monitoring, evapotranspiration estimation, and surface energy balance analysis. Accurate and continuous LST information is essential for water resource management, ecological monitoring, and disaster early warning. Currently, LST is primarily obtained through satellite remote sensing inversion. While medium- and low-resolution sensors like MODIS offer high temporal resolution, their spatial resolution (approximately 1 km) is limited, making it difficult to meet the demands of refined LST analysis at regional or watershed scales. Conversely, high-resolution sensors like Landsat and Sentinel, despite their high spatial resolution (30-100 m), suffer from long revisit periods and are heavily influenced by cloud cover, hindering the formation of continuous time series. Therefore, improving spatial accuracy while maintaining temporal continuity is a key challenge in LST remote sensing. Existing LST downscaling methods can be broadly categorized into two types: empirical statistical methods and machine learning methods. The former, such as multiple linear regression and geographically weighted regression (GWR), performs spatial interpolation by establishing empirical relationships between temperature and features such as vegetation index, albedo, and topography; the latter, such as random forest (RF) and convolutional neural network (CNN), utilizes nonlinear learning capabilities to extract complex relationships from multi-source data.

[0003] However, traditional methods generally suffer from the following shortcomings: Lack of temporal dynamic information; existing methods mostly use single-period imagery as input, ignoring the continuous variation of land surface temperature over time, making it difficult to characterize short-term fluctuations and cumulative effects. Insufficient spatial feature representation; traditional convolutional structures mainly rely on local neighborhood information, making it difficult to capture spatial temperature differences caused by complex terrain and underlying surface heterogeneity. Lack of scale consistency constraints; many deep learning downscaling models only focus on pixel-level accuracy improvement, neglecting the energy conservation relationship between high and low resolution temperature fields, easily leading to physical inconsistencies and result biases. In the context of climate change, the frequent occurrence of extreme heat and drought events further highlights the importance of high-precision land surface temperature products. Especially in areas with high vegetation cover or significant topographic relief, land surface temperature is affected by multiple factors such as vegetation shading, atmospheric radiation, and energy distribution, making it difficult for traditional models to accurately reconstruct spatial details and temporal continuity. Therefore, there is an urgent need for a new method that can simultaneously model the temporal dependence and spatial heterogeneity of land surface temperature and maintain cross-scale energy consistency during downscaling to achieve high spatiotemporal resolution, physical consistency, and wide applicability of land surface temperature data. Summary of the Invention

[0004] To address the shortcomings of the aforementioned technologies, this invention provides a spatiotemporal modeling method and system for downscaling surface temperature that considers energy constraints.

[0005] To solve the above technical problems, the technical solution adopted by this invention is: a spatiotemporal modeling method for surface temperature downscaling considering energy constraints, comprising the following steps: 1) Acquisition and preprocessing of multi-source remote sensing and meteorological data to construct a normalized spatiotemporally continuous dataset; 2) Feature grouping and spatiotemporal feature tensor construction: The normalized multi-source data is divided into low-resolution temporal feature groups and high-resolution spatial feature groups according to feature attributes; 3) Temporal and spatial feature extraction: For low-resolution temporal feature groups and high-resolution spatial feature groups, spatiotemporal attention networks are used to extract features respectively to obtain temporally enhanced feature vectors and spatially enhanced feature vectors; 4) Spatiotemporal feature modeling and high-resolution surface temperature estimation: Spatiotemporal feature modeling is performed at a low-resolution scale to construct a surface temperature estimation model. The model parameters learned in the learning phase are applied to a high-resolution spatial scale to achieve spatial downscaling estimation of surface temperature. 5) Loss Construction and Cross-Scale Consistency Constraint: A cross-scale consistency constraint mechanism is introduced to align the high-resolution downscaling results with the original low-resolution temperature products. The training objective is to minimize the scale difference loss. 6) Iterative training of the model and output of results.

[0006] Preferably, multi-source remote sensing and meteorological data include: Low-resolution surface temperature data and high-resolution reference temperature; Meteorological driving data, including near-surface air temperature and dew point temperature, are provided by reanalysis data such as ERA5; Radiation-driven data, including downlink shortwave radiation and downlink longwave radiation, are provided by reanalysis data such as ERA5; Underlying surface and topographic data, including land cover type, vegetation index NDVI, digital elevation model (DEM), and slope; Albedo data.

[0007] Preferably, multi-source remote sensing and meteorological data are preprocessed, including geometric correction, reprojection and temporal alignment, and temporal interpolation and spatial reconstruction algorithms are used to fill in the missing measurement areas under cloud and rain conditions, so as to construct a normalized spatiotemporally continuous dataset.

[0008] Preferably, in step 2), the process of constructing the low-resolution temporal feature set includes: Construct temporal characteristics of land surface temperature and its meteorological and radiative driving factors at low spatial resolution scales; For any low-resolution pixel (i,j) within the study area, at the current target time t, taking that time as the end of the time window, a fixed-length time window of length k is traced back. The variables are stacked in chronological order to form a time series sample of length k, and the time feature vector is... The mathematical expression is as follows: (1) in, The surface temperature of the low-resolution surface temperature product. Near-surface temperature; and These represent downlink shortwave radiation and downlink longwave radiation, respectively; the time window contains k consecutive time steps, corresponding to time t and the k-1 times preceding it, with the time step index as follows: , In this invention, k=5, therefore, The dimension is , where 4 indicates that each time step contains 4 meteorological / radiation variables; Subsequently, pixel-by-pixel and variable-by-variable time window standardization was performed on each column of data in the matrix to form a standardized time feature matrix. This vector serves as the input matrix in the subsequent time feature extraction step.

[0009] Preferably, in step 2), the process of constructing the high-resolution spatial feature set includes: The following five high-resolution spatial static or quasi-static characteristic variables characterize the spatial heterogeneity and physical properties of the land surface: vegetation index NDVI, digital elevation model (DEM), slope (S), albedo (α), and land cover type (LC). For each high-resolution pixel within the study area Extract all five features mentioned above to form the spatial feature vector of the pixel. Stacking the feature vectors of all high-resolution pixels in the spatial dimension forms a spatial feature group, as shown below: (2) in, These represent the latitude and longitude coordinates of the pixel; t represents the time. Perform pixel-by-pixel and variable-by-variable time window standardization on each column of data in the matrix to form a standardized spatial feature matrix. This vector serves as the input matrix in the subsequent spatial feature extraction step.

[0010] Preferably, step 3) includes the following steps for extracting time features: The first step is to extract time features: Standardized time feature matrix It is expressed as follows: (3) in, The standardized low-resolution temporal feature matrix obtained in the previous step has a size of [value missing]. Rows represent time steps, and columns represent physical variables; among them, The four-dimensional normalized eigenvectors at each time step are used as matrices. a line of ; Assign a time position encoding vector to each time step within the time window, and add it element-wise to the feature vector of the corresponding time step to obtain a feature vector with time position information: (4) in, Represents low-resolution pixels At time step The corresponding four-dimensional standardized time feature vector includes surface temperature and meteorological / radiation-related features; Representative assigned to time step The time location encoding vector has the same dimension as the feature vector; This represents a time index, identifying a specific time step within a time window; Stacking the time-encoded feature vectors in chronological order yields a temporal feature matrix with positional encoding, whose size remains the same. The expression is as follows: (5) The second step is to perform multi-head temporal attention calculation on the temporal feature matrix; The time feature matrix after adding time position encoding A linear transformation is performed to construct query vectors, key vectors, and value vectors to represent the correlations between features at different time steps. Taking a single attention head as an example, its computational form is as follows: (6) in, Represented by the time feature matrix The query vector matrix obtained through linear mapping; Represented by the time feature matrix The key vector matrix obtained through linear mapping; Represented by the time feature matrix The value vector matrix obtained through linear mapping; , , These represent the learnable weight matrices used to generate the query vector, key vector, and value vector, respectively. Subsequently, the association weights between different time steps are calculated using a scaled dot product attention mechanism: in, This is the time attention weight matrix; This represents the dot product operation between the query vector matrix and the transpose of the key vector matrix, and represents the similarity between features at different time steps. The feature dimension of the key vector is represented by ; softmax represents the normalization function. The third step is to generate time-enhanced features; The value vector is weighted and summed using the attention weight matrix, and the feature mapping is completed through the output transformation matrix to obtain the temporal enhancement feature matrix of the pixel within the time window: (8) in, This is the time attention weight matrix; This represents the value vector matrix obtained by linear mapping from the time feature matrix; This represents the output transformation matrix, which is a learnable weight matrix used for linear mapping of attention-weighted features; Repeat the above calculation process for all low-resolution pixels in the study area to obtain the low-resolution temporal augmentation feature tensor.

[0011] Preferably, step 3) includes the following steps for spatial feature extraction: For high-resolution pixels A standardized spatial feature vector containing NDVI, DEM, slope, albedo, and land cover type has been obtained, and this vector is denoted as: (9) Among them, spatial feature tensor Size is , and These represent the number of rows and columns of the high-resolution spatial feature tensor, respectively. Indicates the number of spatial feature channels; The first step is to calculate channel attention. First, regarding the spatial feature tensor Perform global average pooling and global max pooling in the spatial dimension respectively to obtain two channel statistical vectors: (10) in, Indicates the channel index. The c-th channel in the spatial feature tensor represents the pixel The values ​​at each location correspond to NDVI, DEM, slope, albedo, and land cover type, respectively. The two channel statistical vectors are input into a multilayer perceptron with shared weights, and the outputs are summed and normalized using the sigmoid function to obtain the channel attention weight vector for each channel. (11) in, represents the Sigmoid activation function; MLP represents a multilayer perceptron network used for channel feature mapping; By scaling the original spatial feature tensor channel by channel using the channel attention weights, we can obtain the channel-weighted spatial features, i.e., the spatial feature tensor enhanced by the attention mechanism: (12) in, Represents the channel attention weight vector The first in The nth component represents the nth component. The importance of each spatial feature channel; The first in the spatial feature tensor Each channel in a pixel The value at; The second step is spatial attention calculation.

[0012] Based on channel weighting, to further utilize neighborhood structure information, average pooling and max pooling are performed on the attention-enhanced spatial feature tensor along the channel dimension, respectively, resulting in two single-channel spatial feature maps: (13) in, Indicates the number of spatial feature channels; and All are of size Two-dimensional spatial feature map; Will and The data is concatenated along the channel dimension and used as input to a 7×7 convolution kernel. After convolution and sigmoid activation, a spatial attention weight map is obtained, reflecting the importance of different spatial locations, as shown below: (14) in, and The average pooling and max pooling spatial feature maps obtained from formula (13); This represents a two-dimensional convolution operation with a 7×7 kernel; This represents the Sigmoid activation function; Using this spatial weight map, channel weighted features are applied. Perform pixel-by-pixel scaling to generate the final spatial augmentation features: (15) in, This refers to the spatial feature tensor enhanced by the attention mechanism. Represents the spatial attention weight map;

[0013] In subsequent steps, the pixels Spatial enhancement features It is considered as a one-dimensional feature vector and used for model construction.

[0014] Preferably, step 4), spatiotemporal feature modeling and high-resolution surface temperature estimation, specifically includes the following steps: The first step is to model the spatiotemporal features at low resolution. At low resolution scales, time-enhanced feature vectors With the aggregated spatial augmented feature vector As two independent input variables, they are jointly input into the land surface temperature estimation model to establish its mapping relationship with the low-resolution land surface temperature product. The land surface temperature estimation model is expressed as follows: (16) in For low-resolution surface temperature (e.g., MODIS LST) products; The second step is to estimate the surface temperature at high resolution. By inputting high-resolution temporal and spatial features as two input variables into the land surface temperature estimation model obtained in the above steps, the high-resolution land surface temperature estimation result is obtained: (17) in, Represents high-resolution pixels, This represents high-resolution temporal features obtained by resampling low-resolution temporal enhancement features; This represents the spatial enhancement features corresponding to high-resolution pixels.

[0015] Preferably, in step 5), the loss construction and cross-scale consistency constraints are specifically processed as follows: The first step is to establish the spatial correspondence between high and low resolution; Based on the spatial reference system and resolution ratio of the study area, the low-resolution pixels of each low-resolution surface temperature product are... Coverage on a high-resolution grid is defined as a spatial set. ; The second step is scale matching of the high-resolution estimation results; High-resolution surface temperature estimation results obtained using the aforementioned steps For sets All high-resolution pixels within Spatially averaged aggregation was performed to obtain a low-resolution estimated temperature consistent with the spatial resolution of the low-resolution surface temperature product: (18) The third step is to calculate the scale consistency loss. At the resolution scale of the low-resolution surface temperature product, the low-resolution estimated temperature obtained in the second step is... Compared with the original low-resolution surface temperature product Perform pixel-by-pixel comparisons to construct a scale consistency error term: (19) Where h and w represent the number of rows and columns of the low-resolution land surface temperature product, respectively. Estimate the temperature using the low-resolution result obtained in the second step; This is a raw, low-resolution surface temperature product; The fourth step is to calculate the main loss of the model.

[0016] The high-resolution surface temperature estimation results obtained in step 4) With the corresponding high-resolution reference temperature Perform pixel-by-pixel comparisons to construct the model's main loss term: (20) in, A set of pixels representing high-resolution land surface temperature reference values; This represents a high-resolution reference value for Earth's surface temperature. Step 5: Construction of the total loss function and parameter update.

[0017] (twenty one) in, The total loss used for training, These are the weighting coefficients.

[0018] A spatiotemporal modeling system for downscaling Earth's surface temperature considering energy constraints includes the following modules: The data preprocessing module is used to acquire multi-source remote sensing and meteorological data, preprocess the data, fill in missing data, and construct a spatiotemporally continuous data input set. The feature grouping and spatiotemporal feature tensor construction module is used to divide the data input set into low-resolution temporal feature groups and high-resolution spatial feature groups according to physical properties and resolution differences. Temporal feature tensors and spatial feature tensors are constructed respectively by time series stacking and spatial feature splicing, which serve as inputs for subsequent spatiotemporal feature modeling. The time and space feature extraction modules employ a time attention mechanism to extract time features and perform feature enhancement modeling on the time feature tensor; they also employ a combination of channel attention and spatial attention to extract spatial features and perform feature enhancement modeling on the spatial feature tensor. The spatiotemporal feature modeling and high-resolution land surface temperature module is used to fuse temporal and spatial enhancement features at low resolution scales to establish a mapping relationship between land surface temperature and its spatiotemporal driving factors; and to apply the mapping relationship to high-resolution spatial scales to achieve spatial downscaling estimation of land surface temperature. The loss construction and cross-scale consistency constraint module is used to introduce scale consistency constraints between high and low resolution during model training, aggregate high-resolution surface temperature estimation results to low-resolution scale, and align them with the original observed temperature. The model iterative training and result output module iteratively optimizes the model parameters based on the joint loss function; after training, it outputs the trained land surface temperature downscaling model and the corresponding high-resolution land surface temperature product.

[0019] This invention discloses a spatiotemporal modeling method and system for land surface temperature downscaling that considers energy constraints. By introducing a multi-head temporal attention mechanism to characterize the temporal response of land surface temperature to meteorological and radiative driving factors, and combining it with a spatial attention mechanism to enhance the expression of spatial heterogeneity caused by differences in topography, vegetation, and underlying surface, this invention overcomes the limitations of existing technologies where temporal variation patterns and spatial detail are separated, achieving a synergistic improvement in the temporal continuity and spatial refinement of land surface temperature. Furthermore, this invention introduces a cross-scale consistency constraint mechanism. By constructing a scale consistency loss function between high-resolution land surface temperature estimation results and the original low-resolution land surface temperature product, physical constraints are imposed on the overall energy distribution across different spatial scales. This ensures that the downscaling results maintain energy conservation and scale consistency while enhancing spatial detail, effectively avoiding the overall temperature shift and insufficient physical rationality problems commonly found in existing downscaling methods. Therefore, it significantly outperforms existing technologies in terms of temporal stability, spatial accuracy, and physical reliability. Attached Figure Description

[0020] Figure 1This is a schematic diagram of the technical process of the present invention.

[0021] Figure 2 This is a system module connection diagram of the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0023] It is known that existing methods have the following two major technical drawbacks: First, most methods rely on single-period images or local spatial features, which makes it difficult to simultaneously characterize the temporal variation and spatial heterogeneity of surface temperature. They cannot fully reflect the combined effects of climate fluctuations, vegetation changes, and topographic differences, resulting in poor continuity and insufficient spatial detail in the surface temperature downscaling results. Second, existing deep learning-based downscaling models mostly focus only on pixel-level accuracy and do not consider the energy conservation relationship between high and low resolutions. This can easily lead to overall temperature deviations and local anomalies after downscaling, weakening the physical consistency and generalization ability of the results.

[0024] To address the limitations of existing land surface temperature downscaling methods, this invention proposes a spatiotemporal modeling land surface temperature downscaling method and system that considers energy constraints, thereby solving the following problems: First, to address the problem that existing surface temperature downscaling methods cannot simultaneously characterize the temporal variation and spatial heterogeneity of surface temperature, resulting in poor temporal continuity and insufficient spatial detail in the reconstruction results, this invention introduces multi-head attention in the temporal dimension to obtain the temporal dependence of surface temperature on meteorological and radiation elements, and introduces spatial attention in the spatial dimension to construct feature groups such as topography and underlying surface to enhance the spatial detail representation capability, thereby achieving synergistic constraints of time and space.

[0025] Second, in response to the problem that existing downscaling models generally lack physical consistency constraints between different spatial scales, which easily leads to inconsistencies in the overall energy distribution between high-resolution estimation results and the original resolution land surface temperature products, this invention constructs a cross-scale consistency constraint mechanism. By establishing a scale consistency loss function between high-resolution estimation results and low-resolution land surface temperature products, the energy conservation relationship between scales is maintained while enhancing detailed information, thereby improving the physical rationality and stability of downscaling results.

[0026] The overall strategy for the energy-constrained spatiotemporal modeling surface temperature downscaling method proposed in this invention includes: First, multi-source remote sensing and meteorological data are acquired and preprocessed, undergoing geometric correction, reprojection, and temporal alignment. Temporal difference and spatial reconstruction algorithms are then used to fill in missing data areas under cloud and rain conditions, constructing a spatiotemporally continuous dataset. Subsequently, the normalized multi-source data is divided into low-resolution temporal feature groups and high-resolution spatial feature groups based on their characteristic attributes. A spatiotemporal attention network is used to extract features separately, obtaining temporal and spatial enhanced feature vectors, thereby constructing a high-resolution land surface temperature estimation model. Further, a cross-scale consistency constraint mechanism is introduced to align the high-resolution downscaling results with the original low-resolution temperature products, maintaining energy conservation and physical consistency by minimizing scale difference loss. Finally, a joint loss function is used to complete model training, achieving high-precision downscaling estimation of land surface temperature.

[0027] Furthermore, such as Figure 1 As shown in the overall technical flowchart, this invention mainly consists of the following six steps: 1) acquisition and preprocessing of multi-source remote sensing and meteorological data; 2) feature grouping and spatiotemporal feature tensor construction; 3) temporal and spatial feature extraction; 4) spatiotemporal feature modeling and high-resolution surface temperature estimation; 5) loss construction and cross-scale consistency constraints; 6) model iterative training and result output.

[0028] Among them, 1) the acquisition and preprocessing of multi-source remote sensing and meteorological data, specifically as follows: First, remote sensing and meteorological data from different sources were collected, including: Low-resolution land surface temperature (LST) data (such as 1km resolution MODIS LST products) and high-resolution reference temperatures (such as 90m resolution Landsat LST products). Meteorological driving data, such as near-surface air temperature and dew point temperature, can be provided by reanalysis data such as ERA5; Radiation-driven data, including downlink shortwave radiation and downlink longwave radiation, can also be provided by reanalysis data such as ERA5; Underlying surface and topographic data, including land cover type, Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM), slope, etc. Albedo data can be obtained using surface albedo products (such as MODIS MCD43A3).

[0029] The above data undergoes preprocessing including geometric correction, reprojection, and time alignment. First, geometric correction is performed, and all data are projected to the same spatial coordinate system. Relevant terrain parameters, such as slope, are automatically derived from the DEM through terrain analysis (e.g., SAGA GIS, GDAL, or ArcGIS). Time alignment is performed on data with different temporal resolutions, such as daily composite. For data with different spatial resolutions, corresponding data for 30km and 1km are obtained using bilinear interpolation and average aggregation, respectively. In addition, for multiphase data, linear interpolation or regression based on NDVI time series is used to reconstruct missing temperature values. Finally, a normalized spatiotemporally continuous dataset is generated.

[0030] 2) Feature grouping and spatiotemporal feature tensor construction, the specific processing is as follows: The first step is to construct a low-resolution temporal feature set; This step constructs temporal characteristics of land surface temperature and its meteorological and radiative driving factors at a low spatial resolution scale. Low-resolution land surface temperature products, such as MODIS LST, have a spatial resolution of approximately 1 km.

[0031] To ensure the consistency of temporal characteristics across spatial scales, the meteorological and radiation driving factors introduced were selected from ERA5 reanalysis data, including variables such as near-surface air temperature, downwave radiation, and downwave radiation. Through spatial resampling and aggregation processing, these variables were unified to the same spatial resolution scale as low-resolution surface temperature products such as MODIS LST.

[0032] For any low-resolution pixel within the study area with geographic coordinates (i,j), at the current target time t, taking this time as the end of the time window, a time window of fixed length k is traced back. Preferably, the time window length k is set to 5 days. The above four types of variables are stacked in chronological order to form a time series sample of length k. For pixel (i,j) at time t, its temporal feature vector... It consists of the observations of a total of 4 variables across the aforementioned k time steps. The mathematical expression is as follows: (1) in, Surface temperature for low-resolution surface temperature products (such as MODIS LST); The near-surface air temperature is derived from REA5 reanalysis data, representing the background atmospheric temperature field in the near-surface layer. and These represent downlink shortwave radiation and downlink longwave radiation, respectively, derived from REA5 reanalysis data; the time window contains k consecutive time steps, corresponding to time t and the k-1 times preceding it, i.e., the time step index is... , In this invention, k=5, therefore, The dimension is , where 4 indicates that each time step contains 4 meteorological / radiation variables.

[0033] Subsequently, in order to eliminate the differences in units and ranges of values ​​among different spatial variables, pixel-by-pixel and variable-by-variable time window standardization was performed on the data in each column of the matrix, so that the values ​​of each variable at k time steps have a uniform scale, thus forming a standardized time feature matrix. This vector serves as the input matrix in the subsequent time feature extraction step.

[0034] The second step is to construct a high-resolution spatial feature set.

[0035] This invention aims to select the following five high-resolution spatial static or quasi-static feature variables to characterize the spatial heterogeneity and physical properties of the land surface: vegetation index (NDVI), digital elevation model (DEM), slope (S), albedo (α), and land cover type (LC), to compensate for the topographic constraints and vegetation cover differences that cannot be reflected in time series features. For each high-resolution pixel within the study area... Extract all five features mentioned above to form the spatial feature vector of the pixel. Stacking the feature vectors of all high-resolution pixels in the spatial dimension forms a spatial feature group, as shown below: (2) in, These represent the latitude and longitude coordinates of the pixel; t represents time; NDVI is the normalized vegetation index; S represents the slope; α is the albedo; DEM is the digital elevation model; and LC represents the land cover type (such as woodland, grassland, bare land, etc.). Among them, NDVI and albedo are remote sensing products aligned with time t (such as MODIS 16-day composite NDVI products and MCD43 series albedo products), while DEM, slope, and land cover type are treated as time-invariant static features and can be obtained from the processing results of step 1).

[0036] Similarly, to eliminate differences in units and ranges of values ​​among different spatial variables, pixel-by-pixel and variable-by-variable time window standardization is performed on each column of data in the matrix, so that the values ​​of the variable at k time steps have a uniform scale, forming a standardized spatial feature matrix. This vector serves as the input matrix in the subsequent spatial feature extraction step.

[0037] 3) Extraction of temporal and spatial features, specifically processed as follows: This step uncovers the dynamic dependencies within the time series and the spatial structural differences between different pixels, including temporal feature extraction steps and spatial feature extraction steps.

[0038] A) Temporal feature extraction.

[0039] The first step is to extract time features.

[0040] The temporal attention module is used to capture the temporal correlation between surface temperature sequences and meteorological features at different time steps. For any low-resolution pixel (i, j) within the study area, the standardized temporal feature matrix of meteorological / radiation variables within a time window of length k (k=5 in this invention) has been obtained in the aforementioned temporal feature construction steps. The standardized time feature matrix is ​​represented as follows: (3) in, The standardized low-resolution temporal feature matrix obtained in the previous step has a size of [value missing]. The rows represent time steps, and the columns represent physical variables. The four-dimensional normalized eigenvectors at each time step are used as matrices. a line of The columns correspond to the surface temperatures of low-resolution surface temperature products (such as MODIS LST). Near-ground temperature Downlink shortwave radiation and downwave radiation .

[0041] To display the chronological order information, this invention assigns a time position encoding vector to each time step within the time window, and adds it element-wise to the feature vector of the corresponding time step to obtain a feature vector with time position information: (4) here, Represents low-resolution pixels At time step The corresponding four-dimensional standardized time feature vector includes surface temperature and meteorological / radiation-related features; Representative assigned to time step The time location encoding vector has the same dimension as the feature vector; This represents a time index, identifying a specific time step within a time window.

[0042] Stacking the time-encoded feature vectors in chronological order yields a temporal feature matrix with positional encoding, whose size remains the same. The expression is as follows: (5)

[0043] The second step is to perform multi-head temporal attention calculation on the temporal feature matrix.

[0044] The time feature matrix after adding time position encoding A linear transformation is performed to construct query vectors, key vectors, and value vectors to represent the correlations between features at different time steps. Taking a single attention head as an example, its computational form is as follows: (6) in, Represented by the time feature matrix The query vector matrix obtained by linear mapping describes the attention required by each time step for information from other time steps when calculating time attention. Represented by the time feature matrix The key vector matrix obtained through linear mapping is used to represent the feature information contained in each time step; Represented by the time feature matrix The value vector matrix obtained through linear mapping is used to participate in information fusion during the attention weighting process; , , These represent the learnable weight matrices used to generate the query vector, key vector, and value vector, respectively. They can be randomly initialized during the initial training phase and updated by optimizing the objective function during model training.

[0045] Subsequently, the association weights between different time steps are calculated using a scaled dot product attention mechanism: in, This is the temporal attention weight matrix, representing the strength of correlation between different time steps within a time window; This represents the dot product operation between the query vector matrix and the transpose of the key vector matrix, and represents the similarity between features at different time steps. represents the feature dimension of the key vector; softmax represents the normalization function, which normalizes the attention weights in the time dimension.

[0046] The third step is to generate time-enhanced features.

[0047] The value vector is weighted and summed using the attention weight matrix, and feature mapping is performed through the output transformation matrix to obtain the temporal enhancement feature matrix of the pixel within the time window. This matrix represents the temporal information of the pixel at multiple time steps and their interrelationships. The calculation form can be expressed as: (8) in, This is the time attention weight matrix; This represents the value vector matrix obtained by linear mapping from the time feature matrix; The output transformation matrix is ​​a learnable weight matrix used to linearly map the attention-weighted features. It can be randomly initialized during the initial training phase and updated by optimizing the objective function during model training.

[0048] Repeating the above calculation process for all low-resolution pixels within the study area yields a low-resolution temporal augmentation feature tensor. This tensor will be used as a temporal input in subsequent steps to provide crucial information supporting the trend of surface temperature changes and its response to meteorological drivers.

[0049] B) Spatial feature extraction.

[0050] This step combines the spatial feature set constructed in the previous steps with channel attention and spatial attention joint modeling to characterize the differences in the contribution of different spatial variables and different pixel locations to the spatial pattern of surface temperature, thereby improving the model's sensitivity to changes in topographic relief, vegetation cover, and underlying surface type.

[0051] In the spatial feature construction step, for high-resolution pixels A standardized spatial feature vector containing NDVI, DEM, slope, albedo, and land cover type has been obtained. For ease of description, this vector is denoted as: (9) Here, the spatial feature tensor Size is , and These represent the number of rows and columns of the high-resolution spatial feature tensor, respectively, which are determined by the coverage area of ​​the study area and the spatial resolution of the data. This invention uses 5 spatial variables to represent the number of spatial feature channels, therefore, we take... =5.

[0052] The first step is to calculate channel attention.

[0053] First, regarding the spatial feature tensor Perform global average pooling and global max pooling in the spatial dimension respectively to obtain two channel statistical vectors: (10) in, Indicates the channel index. The c-th channel in the spatial feature tensor represents the pixel The values ​​at each location correspond to NDVI, DEM, slope, albedo, and land cover type, respectively.

[0054] The two channel statistical vectors are input into a multilayer perceptron (MLP) with shared weights, and the outputs are summed and normalized using the sigmoid function to obtain the channel attention weight vector for each channel. (11) in, represents the Sigmoid activation function; MLP represents a multilayer perceptron network used for channel feature mapping.

[0055] By scaling the original spatial feature tensor channel by channel using the channel attention weights, we can obtain the channel-weighted spatial features, i.e., the spatial feature tensor enhanced by the attention mechanism: (12) in, Represents the channel attention weight vector The first in The nth component represents the nth component. The importance of each spatial feature channel; The first in the spatial feature tensor Each channel in a pixel The value at that location.

[0056] The second step is spatial attention calculation.

[0057] Based on channel weighting, to further utilize neighborhood structure information, average pooling and max pooling are performed on the attention-enhanced spatial feature tensor along the channel dimension, respectively, resulting in two single-channel spatial feature maps: (13) in, This invention uses 5 spatial variables to represent the number of spatial feature channels, therefore, we take... =5; and All are of size Two-dimensional spatial feature map.

[0058] Will and The data is concatenated along the channel dimension and used as input to a 7×7 convolution kernel. After convolution and sigmoid activation, a spatial attention weight map is obtained, reflecting the importance of different spatial locations, as shown below: (14) in, and The average pooling and max pooling spatial feature maps obtained from formula (13); This represents a two-dimensional convolution operation with a 7×7 kernel; This represents the Sigmoid activation function.

[0059] Using this spatial weighting graph, channel weighted features are applied. Perform pixel-by-pixel scaling to generate the final spatial augmentation features: (15) in, This refers to the spatial feature tensor enhanced by the attention mechanism. Represents the spatial attention weight map. and Having the same dimensions, meaning the overall size remains the same. However, the numerical distribution has already taken into account the weights between variables and the differences in neighborhood spatial structure.

[0060] In subsequent steps, the pixels Spatial enhancement features It is considered as a one-dimensional feature vector and used for model construction.

[0061] 4) Spatiotemporal feature modeling and high-resolution surface temperature estimation, as detailed below: The first step is to model the spatiotemporal features at low resolution.

[0062] During the model training phase, to establish the mapping relationship between land surface temperature and its spatiotemporal driving features, low-resolution land surface temperature data (such as MODIS land surface temperature) is used as a supervision signal to construct model training samples at its spatial resolution scale. Specifically, for low-resolution pixel (i, j), the temporal augmentation feature vector of the corresponding low-resolution pixel has been obtained through the aforementioned steps. Characterizing the temporal variation of surface temperature and its response to meteorological drivers, as well as its high-resolution spatial features. Spatial augmentation feature vectors for corresponding resolution pixels are obtained by spatial aggregation within the coverage area of ​​low-resolution pixels. .

[0063] At low resolution scales, time-enhanced feature vectors With the aggregated spatial augmented feature vector As two independent input variables, they are jointly input into the land surface temperature estimation model to establish its mapping relationship with low-resolution land surface temperature (such as MODIS LST) products. The model is expressed in the following form: (16) here, Indicates low-resolution pixels; For low-resolution surface temperature (such as MODIS LST) products.

[0064] The second step is to estimate the surface temperature at high resolution.

[0065] The model parameters learned during the training phase are applied to a high-resolution spatial scale to achieve spatial downscaling estimation of land surface temperature. Specifically, high-resolution temporal and spatial features are input as two variables into the land surface temperature estimation model obtained in the above steps to obtain high-resolution land surface temperature estimation results. (17) in, Represents high-resolution pixels, This represents high-resolution temporal features obtained by resampling low-resolution temporal enhancement features; This represents the spatial enhancement features corresponding to high-resolution pixels.

[0066] Based on the temporal and spatial enhancement features generated in the aforementioned steps, the estimation of low-resolution surface temperature to high-resolution surface temperature is achieved through formula (17).

[0067] 5) Loss model construction and cross-scale consistency constraints are handled as follows: This step is used to establish scale consistency constraints between high-resolution land surface temperature estimation results and MODIS raw land surface temperature products during the model training phase. It introduces low-resolution supervision information to improve the ability to express high-resolution spatial details while ensuring that the estimation results are consistent with the original radiation observations in terms of energy and scale, which is in line with the basic remote sensing principles of land surface temperature scale transformation and energy conservation.

[0068] The first step is to establish the spatial correspondence between high and low resolution.

[0069] Based on the spatial reference system and resolution ratio of the study area, the low-resolution pixels of each low-resolution surface temperature (e.g., MODIS LST) product are... Coverage on a high-resolution grid is defined as a spatial set. The spatial set is determined once before model training begins and remains unchanged during subsequent training and inference, serving as a fixed template for high- and low-resolution spatial mapping.

[0070] The second step is scale matching of the high-resolution estimation results.

[0071] High-resolution surface temperature estimation results obtained using the aforementioned steps For sets All high-resolution pixels within Spatially averaged aggregation is performed to obtain a low-resolution estimated temperature consistent with the spatial resolution of low-resolution surface temperature products (such as MODIS LST): (18) This aggregation operation is only used for scale consistency constraints during the model training phase. Its purpose is to ensure that the high-resolution temperature field can be reverted to the MODIS observation scale after spatial aggregation, thereby avoiding systematic bias caused by scale mismatch.

[0072] The third step is to calculate the scale consistency loss.

[0073] At a low-resolution surface temperature (such as MODIS LST) product resolution scale, the low-resolution estimated temperature obtained in the second step is... Compared with raw low-resolution surface temperature (such as MODIS LST) products Perform pixel-by-pixel comparisons to construct a scale consistency error term: (19) Where h and w represent the number of rows and columns of a low-resolution land surface temperature (such as MODIS LST) product, respectively. Estimate the temperature using the low-resolution result obtained in the second step; This is a raw, low-resolution surface temperature (e.g., MODIS LST) product.

[0074] The fourth step is to calculate the main loss of the model.

[0075] The high-resolution surface temperature estimation results obtained in step 4) With the corresponding high-resolution reference temperature Perform pixel-by-pixel comparisons to construct the model's main loss term: (20) in, This represents the set of pixels with high-resolution land surface temperature reference values. In different implementation cases, this set can be all pixels of the study area or a subset thereof. This represents a high-resolution land surface temperature reference value, which may be derived from ground-based measured data, high-resolution land surface temperature products, etc.

[0076] Step 5: Construction of the total loss function and parameter update. (twenty one) in, The total loss used for training, is a weighting coefficient used to adjust the influence of the scaling consistency term on the overall loss function.

[0077] 6) Model iterative training and result output, as detailed below: After obtaining the aforementioned total loss function, the model iterative training and parameter update phase begins. This step uses an optimization algorithm to automatically update the parameters of the high-resolution land surface temperature estimation model, enabling the model to gradually reduce errors and improve the accuracy of land surface temperature retrieval over multiple iterations.

[0078] Specifically, the total loss calculated in step 5) is input into the optimizer to automatically adjust the model's internal parameters, enabling the model to gradually approach the reference temperature at high-resolution scales while maintaining consistency with the MODIS LST product at low-resolution scales. The training process is executed in a set batch loop, with each round sequentially performing the following steps: inputting spatiotemporal features, performing forward surface temperature estimation, calculating the main loss and scale consistency loss, synthesizing the total loss, and updating the model parameters.

[0079] As the number of iterations increases, the model gradually learns the temporal and spatial patterns of surface temperature variation and the inherent consistency across multiple scales. Training terminates when preset conditions are met, such as reaching a set number of training epochs, the validation error no longer decreasing, or the loss stabilizing at a low level. At this point, the model's parameters are determined as the final result and saved. The trained model can be directly used for high-resolution surface temperature estimation at any time. After inputting temporal and spatial features, it can output pixel-level high-resolution surface temperature results for the entire study area. The final output includes: a trained surface temperature estimation model, high-resolution surface temperature products, and a batch temperature downscaling workflow suitable for operational use, providing reliable foundational data for subsequent surface thermal environment analysis, thermal anomaly monitoring, and temperature time series construction.

[0080] In another embodiment of the present invention, a spatiotemporal modeling surface temperature downscaling system considering energy constraints is provided, such as... Figure 2 As shown, it includes six modules: data preprocessing module, feature grouping and spatiotemporal feature tensor construction module, temporal and spatial feature extraction module, spatiotemporal feature modeling and high-resolution land surface temperature module, loss construction and cross-scale consistency constraint module, and model iterative training and result output module.

[0081] The data preprocessing module is used to acquire multi-source remote sensing data, meteorological reanalysis data, and underlying surface auxiliary data, and to perform geometric correction, reprojection, temporal alignment, and spatial resolution matching on the data. This module also fills in missing surface temperature data under cloud and rain conditions through temporal interpolation or regression reconstruction, constructing a spatiotemporally continuous data input set.

[0082] The feature grouping and spatiotemporal feature tensor construction module is used to divide the input data set into low-resolution temporal feature groups and high-resolution spatial feature groups according to physical properties and resolution differences. This module constructs temporal feature tensors and spatial feature tensors respectively by stacking time series data and concatenating spatial features, which serve as inputs for subsequent spatiotemporal feature modeling.

[0083] The temporal and spatial feature extraction modules perform feature enhancement modeling on the temporal and spatial feature tensors, respectively. Specifically, the temporal feature extraction uses a temporal attention mechanism to characterize the temporal response of surface temperature to meteorological and radiation-driven factors; the spatial feature extraction uses a combination of channel attention and spatial attention to enhance the model's ability to express topography, vegetation cover, and underlying surface heterogeneity.

[0084] The spatiotemporal feature modeling and high-resolution land surface temperature module is used to establish a mapping relationship between land surface temperature and its spatiotemporal driving factors by fusing temporal and spatial augmentation features at low-resolution scales. After model training, this module applies the mapping relationship to a high-resolution spatial scale to achieve spatial downscaling estimation of land surface temperature.

[0085] The loss construction and cross-scale consistency constraint module introduces scale consistency constraints between high and low resolution during model training, aggregating high-resolution surface temperature estimates to a low-resolution scale and aligning them with the original observed temperatures. This module constructs a scale consistency loss term, ensuring that the downscaling results maintain spatial detail while satisfying energy conservation and physical consistency requirements.

[0086] The model iterative training and result output module iteratively optimizes the model parameters based on the joint loss function, gradually reducing high-resolution estimation errors and improving model stability. After training, this module outputs the trained land surface temperature downscaling model and the corresponding high-resolution land surface temperature product for subsequent operational applications.

[0087] Therefore, this invention proposes a spatiotemporal modeling method for land surface temperature downscaling that integrates multi-head attention and spatial attention mechanisms by introducing a land surface temperature downscaling framework that coordinates temporal and spatial dimensions. Under low-resolution land surface temperature monitoring conditions, this method can simultaneously characterize the temporal response of land surface temperature to meteorological and radiation-driven factors, as well as the spatial heterogeneity caused by differences in topography, vegetation, and underlying surface. It achieves synergistic constraints on the temporal variation of land surface temperature and the expression of spatial details, significantly improving the temporal continuity, spatial refinement, and overall stability of the downscaling results.

[0088] This invention introduces a cross-scale consistency constraint mechanism in the process of land surface temperature downscaling. It constructs a scale consistency loss function between high-resolution land surface temperature estimation results and original low-resolution land surface temperature products. By constraining the high-resolution results to regress to the low-resolution observation scale after spatial aggregation, it achieves energy conservation and physical consistency between different spatial scales. This effectively solves the problems of overall energy shift and insufficient physical rationality in high-resolution reconstruction results in existing downscaling methods. While enhancing the ability to express spatial details, it ensures the physical reliability and generalization stability of downscaling results in the energy distribution and scale conversion process.

[0089] The above embodiments are not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention are also within the protection scope of the present invention.

Claims

1. A spatiotemporal modeling method for surface temperature downscaling considering energy constraints, characterized in that: Includes the following steps: 1) Acquisition and preprocessing of multi-source remote sensing and meteorological data to construct a normalized spatiotemporally continuous dataset; 2) Feature grouping and spatiotemporal feature tensor construction: The normalized multi-source data is divided into low-resolution temporal feature groups and high-resolution spatial feature groups according to feature attributes; 3) Temporal and spatial feature extraction: For low-resolution temporal feature groups and high-resolution spatial feature groups, spatiotemporal attention networks are used to extract features respectively to obtain temporally enhanced feature vectors and spatially enhanced feature vectors; 4) Spatiotemporal feature modeling and high-resolution surface temperature estimation: Spatiotemporal feature modeling is performed at a low-resolution scale to construct a surface temperature estimation model. The model parameters learned in the learning phase are applied to a high-resolution spatial scale to achieve spatial downscaling estimation of surface temperature. 5) Loss Construction and Cross-Scale Consistency Constraint: A cross-scale consistency constraint mechanism is introduced to align the high-resolution downscaling results with the original low-resolution temperature products. The training objective is to minimize the scale difference loss. 6) Iterative training of the model and output of results.

2. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 1, characterized in that: Multi-source remote sensing and meteorological data, including: Low-resolution surface temperature data and high-resolution reference temperature; Meteorological driving data, including near-surface air temperature and dew point temperature, are provided by reanalysis data such as ERA5; Radiation-driven data, including downlink shortwave radiation and downlink longwave radiation, are provided by reanalysis data such as ERA5; Underlying surface and topographic data, including land cover type, vegetation index NDVI, digital elevation model (DEM), and slope; Albedo data.

3. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 2, characterized in that: Preprocessing of multi-source remote sensing and meteorological data, including geometric correction, reprojection and temporal alignment, is performed. Temporal interpolation and spatial reconstruction algorithms are used to fill in the missing areas under cloud and rain conditions, and a normalized spatiotemporally continuous dataset is constructed.

4. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 3, characterized in that: In step 2), the construction process of the low-resolution temporal feature set includes: Construct temporal characteristics of land surface temperature and its meteorological and radiative driving factors at low spatial resolution scales; For any low-resolution pixel (i,j) within the study area, at the current target time t, taking that time as the end of the time window, a fixed-length time window of length k is traced back. The variables are stacked in chronological order to form a time series sample of length k, and the time feature vector is... The mathematical expression is as follows: （1） in, The surface temperature of the low-resolution surface temperature product. Near-surface temperature; and These represent downlink shortwave radiation and downlink longwave radiation, respectively; the time window contains k consecutive time steps, corresponding to time t and the k-1 times preceding it, with the time step index as follows: , In this invention, k=5, therefore, The dimension is , where 4 indicates that each time step contains 4 meteorological / radiation variables; Subsequently, pixel-by-pixel and variable-by-variable time window standardization was performed on each column of data in the matrix to form a standardized time feature matrix. This vector serves as the input matrix in the subsequent time feature extraction step.

5. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 4, characterized in that: In step 2), the construction process of the high-resolution spatial feature set includes: The following five high-resolution spatial static or quasi-static characteristic variables characterize the spatial heterogeneity and physical properties of the land surface: vegetation index NDVI, digital elevation model (DEM), slope (S), albedo (α), and land cover type (LC). For each high-resolution pixel within the study area Extract all five features mentioned above to form the spatial feature vector of the pixel. Stacking the feature vectors of all high-resolution pixels in the spatial dimension forms a spatial feature group, as shown below: （2） in, These represent the latitude and longitude coordinates of the pixel; t represents the time. Perform pixel-by-pixel and variable-by-variable time window standardization on each column of data in the matrix to form a standardized spatial feature matrix. This vector serves as the input matrix in the subsequent spatial feature extraction step.

6. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 5, characterized in that: In step 3), time feature extraction Includes the following steps: The first step is to extract time features: Standardized time feature matrix It is expressed as follows: （3） in, The standardized low-resolution temporal feature matrix obtained in the previous step has a size of [value missing]. Rows represent time steps, and columns represent physical variables; among them, The four-dimensional normalized eigenvectors at each time step are used as matrices. a line of ; Assign a time position encoding vector to each time step within the time window, and add it element-wise to the feature vector of the corresponding time step to obtain a feature vector with time position information: （4） in, Represents low-resolution pixels At time step The corresponding four-dimensional standardized time feature vector includes surface temperature and meteorological / radiation-related features; Representative assigned to time step The time location encoding vector has the same dimension as the feature vector; This represents a time index, identifying a specific time step within a time window; Stacking the time-encoded feature vectors in chronological order yields a temporal feature matrix with positional encoding, whose size remains the same. The expression is as follows: （5） The second step is to perform multi-head temporal attention calculation on the temporal feature matrix; The time feature matrix after adding time position encoding A linear transformation is performed to construct query vectors, key vectors, and value vectors to represent the correlations between features at different time steps. Taking a single attention head as an example, its computational form is as follows: （6） in, Represented by the time feature matrix The query vector matrix obtained through linear mapping; Represented by the time feature matrix The key vector matrix obtained through linear mapping; Represented by the time feature matrix The value vector matrix obtained through linear mapping; , , These represent the learnable weight matrices used to generate the query vector, key vector, and value vector, respectively. Subsequently, the association weights between different time steps are calculated using a scaled dot product attention mechanism: in, This is the time attention weight matrix; This represents the dot product operation between the query vector matrix and the transpose of the key vector matrix, and represents the similarity between features at different time steps. The feature dimension of the key vector is represented by ; softmax represents the normalization function. The third step is to generate time-enhanced features; The value vector is weighted and summed using the attention weight matrix, and the feature mapping is completed through the output transformation matrix to obtain the temporal enhancement feature matrix of the pixel within the time window: （8） in, This is the time attention weight matrix; This represents the value vector matrix obtained by linear mapping from the time feature matrix; This represents the output transformation matrix, which is a learnable weight matrix used for linear mapping of attention-weighted features; Repeat the above calculation process for all low-resolution pixels in the study area to obtain the low-resolution temporal augmentation feature tensor.

7. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 6, characterized in that: Step 3) involves spatial feature extraction, which includes the following steps: For high-resolution pixels A standardized spatial feature vector containing NDVI, DEM, slope, albedo, and land cover type has been obtained, and this vector is denoted as: （9） Among them, spatial feature tensor Size is , and These represent the number of rows and columns of the high-resolution spatial feature tensor, respectively. Indicates the number of spatial feature channels; The first step is to calculate channel attention; First, regarding the spatial feature tensor Perform global average pooling and global max pooling in the spatial dimension respectively to obtain two channel statistical vectors: （10） in, Indicates the channel index. The c-th channel in the spatial feature tensor represents the pixel The values ​​at each location correspond to NDVI, DEM, slope, albedo, and land cover type, respectively. The two channel statistical vectors are input into a multilayer perceptron with shared weights, and the outputs are summed and normalized using the sigmoid function to obtain the channel attention weight vector for each channel. （11） in, represents the Sigmoid activation function; MLP represents a multilayer perceptron network used for channel feature mapping; By scaling the original spatial feature tensor channel by channel using the channel attention weights, we can obtain the channel-weighted spatial features, i.e., the spatial feature tensor enhanced by the attention mechanism: （12） in, Represents the channel attention weight vector The first in The nth component represents the nth component. The importance of each spatial feature channel; The first in the spatial feature tensor Each channel in a pixel The value at; The second step is spatial attention calculation. Based on channel weighting, to further utilize neighborhood structure information, average pooling and max pooling are performed on the attention-enhanced spatial feature tensor along the channel dimension, respectively, resulting in two single-channel spatial feature maps: （13） in, Indicates the number of spatial feature channels; and All are of size Two-dimensional spatial feature map; Will and The data is concatenated along the channel dimension and used as input to a 7×7 convolution kernel. After convolution and sigmoid activation, a spatial attention weight map is obtained, reflecting the importance of different spatial locations, as shown below: （14） in, and The average pooling and max pooling spatial feature maps obtained from formula (13); This represents a two-dimensional convolution operation with a 7×7 kernel; This represents the Sigmoid activation function; Using this spatial weight map, channel weighted features are applied. Perform pixel-by-pixel scaling to generate the final spatial augmentation features: （15） in, This refers to the spatial feature tensor enhanced by the attention mechanism. Represents the spatial attention weight map; In subsequent steps, the pixels Spatial enhancement features It is considered as a one-dimensional feature vector and used for model construction.

8. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 7, characterized in that: Step 4), spatiotemporal feature modeling and high-resolution surface temperature estimation, specifically includes the following steps: The first step is to model the spatiotemporal features at low resolution. At low resolution scales, time-enhanced feature vectors With the aggregated spatial augmented feature vector As two independent input variables, they are jointly input into the land surface temperature estimation model to establish its mapping relationship with the low-resolution land surface temperature product. The land surface temperature estimation model is expressed as follows: （16） in For low-resolution surface temperature (e.g., MODIS LST) products; The second step is to estimate the surface temperature at high resolution. By inputting high-resolution temporal and spatial features as two input variables into the land surface temperature estimation model obtained in the above steps, the high-resolution land surface temperature estimation result is obtained: （17） in, Represents high-resolution pixels, This represents high-resolution temporal features obtained by resampling low-resolution temporal enhancement features; This represents the spatial enhancement features corresponding to high-resolution pixels.

9. The spatiotemporal modeling surface temperature downscaling method considering energy constraints according to claim 8, characterized in that: In step 5), the loss construction and cross-scale consistency constraints are handled as follows: The first step is to establish the spatial correspondence between high and low resolution; Based on the spatial reference system and resolution ratio of the study area, the low-resolution pixels of each low-resolution surface temperature product are... Coverage on a high-resolution grid is defined as a spatial set. ; The second step is scale matching of the high-resolution estimation results; High-resolution surface temperature estimation results obtained using the aforementioned steps For sets All high-resolution pixels within Spatially averaged aggregation was performed to obtain a low-resolution estimated temperature consistent with the spatial resolution of the low-resolution surface temperature product: （18） The third step is to calculate the scale consistency loss. At the resolution scale of the low-resolution surface temperature product, the low-resolution estimated temperature obtained in the second step is... Compared with the original low-resolution surface temperature product Perform pixel-by-pixel comparisons to construct a scale consistency error term: （19） Where h and w represent the number of rows and columns of the low-resolution land surface temperature product, respectively. Estimate the temperature using the low-resolution result obtained in the second step; This is a raw, low-resolution surface temperature product; The fourth step is to calculate the main loss of the model. The high-resolution surface temperature estimation results obtained in step 4) With the corresponding high-resolution reference temperature Perform pixel-by-pixel comparisons to construct the model's main loss term: （20） in, A set of pixels that represents a high-resolution land surface temperature reference value; This represents a high-resolution reference value for Earth's surface temperature. Step 5: Construction of the total loss function and parameter update. （21） in, The total loss used for training, These are the weighting coefficients.

10. A spatiotemporal modeling surface temperature downscaling system considering energy constraints as described in claim 9, characterized in that: Includes the following modules: The data preprocessing module is used to acquire multi-source remote sensing and meteorological data, preprocess the data, fill in missing data, and construct a spatiotemporally continuous data input set. The feature grouping and spatiotemporal feature tensor construction module is used to divide the data input set into low-resolution temporal feature groups and high-resolution spatial feature groups according to physical properties and resolution differences. Temporal feature tensors and spatial feature tensors are constructed respectively by time series stacking and spatial feature splicing, which serve as inputs for subsequent spatiotemporal feature modeling. The time and space feature extraction modules employ a time attention mechanism to extract time features and perform feature enhancement modeling on the time feature tensor; they also employ a combination of channel attention and spatial attention to extract spatial features and perform feature enhancement modeling on the spatial feature tensor. The spatiotemporal feature modeling and high-resolution land surface temperature module is used to fuse temporal and spatial enhancement features at low resolution scales to establish a mapping relationship between land surface temperature and its spatiotemporal driving factors; and to apply the mapping relationship to high-resolution spatial scales to achieve spatial downscaling estimation of land surface temperature. The loss construction and cross-scale consistency constraint module is used to introduce scale consistency constraints between high and low resolution during model training, aggregate high-resolution surface temperature estimation results to low-resolution scale, and align them with the original observed temperature. The model iterative training and result output module iteratively optimizes the model parameters based on the joint loss function; After training is complete, the trained land surface temperature downscaling model and the corresponding high-resolution land surface temperature product are output.

Images