Land surface temperature spatial downscaling method and device fusing topographic physical priori knowledge

Abstract

The application relates to a land surface temperature spatial downscaling method and equipment fusing terrain physical priori knowledge, and belongs to the field of remote sensing technology and meteorological data processing. First, low-resolution temperature data, high-resolution temperature data, terrain and elevation data are acquired and raster matching is performed; then, multi-scale terrain features are extracted through a double-branch network, terrain priori and temperature background field are fused through a coding-decoding network with a built-in sliding window attention mechanism, and residual error offset diffusion network modeling is performed on hidden variable distribution difference values to predict temperature residual errors. The model is trained in an end-to-end mode and accuracy is ensured through double convergence determination. In application, low-resolution temperature, terrain and elevation data of a target area are input into the model, and high-resolution land surface temperature products can be output. The application effectively enhances the model's ability to capture terrain constraints, significantly improves the physical rationality and spatial details of the reconstruction results, and can be widely applied in the fields of meteorological prediction, ecological monitoring and resource management.

Description

Technical Field

[0001] This application relates to the fields of remote sensing technology and meteorological data technology, and in particular to a method and device for spatial downscaling of surface temperature by integrating prior knowledge of topographic physics. Background Technology

[0002] Land surface temperature is a core parameter for measuring the energy balance and water cycle of the Earth's surface, and it is widely used in many fields such as weather forecasting, ecological environment monitoring, and resource management. With the deepening of research on land surface processes, the industry has placed increasingly higher demands on the spatiotemporal resolution and data integrity of land surface temperature products. However, due to the contradiction between observation methods and physical simulation capabilities, datasets from a single source often struggle to meet the dual requirements of high resolution and full coverage: traditional reanalysis data, while exhibiting strong spatiotemporal continuity, have relatively coarse spatial resolution; while satellite remote sensing inversion products, although possessing high resolution, are easily affected by weather conditions and observation perspectives, resulting in significant data gaps and fragmentation.

[0003] To overcome this bottleneck, spatial downscaling technology has become a key approach to connecting multi-source data and improving product usability. Although existing downscaling methods have gradually evolved from traditional geometric interpolation to machine learning-based statistical modeling, attempting to uncover the spatial distribution patterns of temperature fields by integrating multi-source environmental factors such as topography and vegetation, they still face significant challenges in practical applications. Especially in complex areas with dramatic topographic relief, existing processing mechanisms often struggle to effectively capture the deep constraints of the underlying physical structure on temperature distribution, resulting in significant deficiencies in physical consistency and spatial detail in the reconstruction results, failing to meet the application standards for high-precision land surface simulation. Summary of the Invention

[0004] Based on this, the purpose of this application is to provide a spatial downscaling method for land surface temperature that integrates prior knowledge of terrain physics, thereby solving the technical problems of existing technologies that cannot effectively utilize terrain features as physical prior constraints, resulting in distorted high-resolution land surface temperature reconstruction results, limited model generalization ability, and difficulty in meeting the requirements of high-precision physical consistency.

[0005] The surface temperature spatial downscaling method that integrates prior knowledge of topographic physics described in this application includes the following steps:

[0006] Acquire first surface temperature data, second surface temperature data, and topographic and elevation data covering a multi-topographic sample area; the first resolution of the first surface temperature data is smaller than the second resolution of the second surface temperature data; The first surface temperature data and the terrain and elevation data are raster-matched with the second surface temperature data to achieve a one-to-one correspondence between each raster cell. The first surface temperature data, the second surface temperature data, and the terrain and elevation data are then input into the surface temperature downscaling model to be trained to obtain the trained surface temperature downscaling model. The surface temperature downscaling model is used to learn terrain physical prior constraints from the terrain and elevation data and predict temperature residuals based on these constraints. The predicted temperature residuals are then used to correct low-resolution surface temperatures, achieving high-resolution surface temperature reconstruction. The first surface temperature data and the terrain and elevation data are used together as training features; the second surface temperature data is used as the training label. Acquire the target area, the third surface temperature data with the same resolution as the first surface temperature data, and the corresponding terrain and elevation data. Input the surface temperature downscaling model and output the fourth surface temperature data.

[0007] This application also provides a computer device, including a processor, a memory, and a computer-readable program stored in the memory, wherein the computer-readable program, when executed by the processor, implements the steps of the method described in any one of the embodiments of this application.

[0008] This application's embodiments achieve significant results in improving the accuracy and physical plausibility of land surface temperature spatial downscaling by constructing a residual learning framework based on topographic physical prior constraints. During the model training phase, by using low-resolution first land surface temperature data and high-resolution topographic and elevation data as joint input features, and specifying high-resolution second land surface temperature data as a supervision label, the model is guided to break free from the limitations of traditional methods that rely solely on numerical statistics. During training, the model actively interprets geographic structural features from the topographic and elevation data and internalizes them as physical prior constraints on the temperature field distribution. This mechanism enables the model to keenly capture high-frequency details lost in low-resolution data during high-resolution temperature reconstruction and to correct the low-resolution background field by accurately predicting temperature residuals. Therefore, the output fourth land surface temperature data not only perfectly inherits the large-scale spatial trends and seamless temporal continuity of the first land surface temperature data (reanalysis data), but also effectively integrates the fine texture of the second land surface temperature data (remote sensing data). In complex areas such as mountains and hills, it completely eliminates the over-smoothing or thermal environment distortion caused by traditional interpolation methods ignoring topographic heterogeneity, and truly restores the local micro-scale temperature change characteristics. At the same time, thanks to the in-depth exploration of the underlying physical mechanisms by the prior constraints of topographic physics, the trained model has extremely strong generalization ability and stability. When facing target areas with different landform types and climate backgrounds, only the corresponding third land surface temperature data and topographic data need to be input to output a downscaling product with high-precision physical consistency without retraining. This provides highly reliable, high-resolution, and spatiotemporally continuous basic land surface temperature data support for land surface process simulation, disaster monitoring, and climate and environmental research.

[0009] To better understand and implement this application, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the spatial downscaling method for surface temperature that integrates prior knowledge of topographic physics, as described in this application. Figure 2 This is a schematic diagram of the structure of a computer device according to an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. Wherein, when the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements.

[0012] It should be understood that the embodiments described below do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0013] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, in the description of this application, unless otherwise stated, “a plurality” means two or more. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items, for example, A and / or B, which can represent: A alone, A and B together, and B alone; the character “ / ” generally indicates that the preceding and following objects are in an “or” relationship.

[0014] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, this information should not be limited to these terms, and these terms are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. Depending on the context, the word "if" as used in this application can be interpreted as "when," "when," or "in response to determination."

[0015] Please refer to Figure 1 The surface temperature spatial downscaling method based on incorporating prior knowledge of topographic physics described in this application includes the following steps: S101: Acquire first surface temperature data, second surface temperature data, and topographic and elevation data covering a multi-topographic sample area; the first resolution of the first surface temperature data is smaller than the second resolution of the second surface temperature data; S102: The first surface temperature data and the terrain and elevation data are raster-matched with the second surface temperature data to achieve a one-to-one correspondence between each raster cell; the first surface temperature data, the second surface temperature data, and the terrain and elevation data are input into the surface temperature downscaling model to be trained for training to obtain the trained surface temperature downscaling model; wherein, the surface temperature downscaling model is used to learn terrain physical prior constraints from the terrain and elevation data and predict temperature residuals based on the terrain physical prior constraints, and use the predicted temperature residuals to correct low-resolution surface temperature to achieve high-resolution surface temperature reconstruction; the first surface temperature data and the terrain and elevation data are used together as training features; the second surface temperature data is used as training labels. S103: Obtain the target area, the third surface temperature data with the same resolution as the first surface temperature data, and the corresponding terrain and elevation data; input the surface temperature downscaling model; and output the fourth surface temperature data.

[0016] This application's embodiments achieve significant results in improving the accuracy and physical plausibility of land surface temperature spatial downscaling by constructing a residual learning framework based on topographic physical prior constraints. During the model training phase, by using low-resolution first land surface temperature data and high-resolution topographic and elevation data as joint input features, and specifying high-resolution second land surface temperature data as a supervision label, the model is guided to break free from the limitations of traditional methods that rely solely on numerical statistics. During training, the model actively interprets geographic structural features from the topographic and elevation data and internalizes them as physical prior constraints on the temperature field distribution. This mechanism enables the model to keenly capture high-frequency details lost in low-resolution data during high-resolution temperature reconstruction and to correct the low-resolution background field by accurately predicting temperature residuals. Therefore, the output fourth land surface temperature data not only perfectly inherits the large-scale spatial trends and seamless temporal continuity of the first land surface temperature data (reanalysis data), but also effectively integrates the fine texture of the second land surface temperature data (remote sensing data). In complex areas such as mountains and hills, it completely eliminates the over-smoothing or thermal environment distortion caused by traditional interpolation methods ignoring topographic heterogeneity, and truly restores the local micro-scale temperature change characteristics. At the same time, thanks to the in-depth exploration of the underlying physical mechanisms by the prior constraints of topographic physics, the trained model has extremely strong generalization ability and stability. When facing target areas with different landform types and climate backgrounds, only the corresponding third land surface temperature data and topographic data need to be input to output a downscaling product with high-precision physical consistency without retraining. This provides highly reliable, high-resolution, and spatiotemporally continuous basic land surface temperature data support for land surface process simulation, disaster monitoring, and climate and environmental research.

[0017] The surface temperature spatial downscaling method that integrates prior knowledge of topographic physics described in this application uses a computer as the execution subject. The following provides a detailed description of each step.

[0018] For step S101, first surface temperature data, second surface temperature data, and terrain and elevation data covering a multi-terrain sample area are acquired; the first resolution of the first surface temperature data is less than the second resolution of the second surface temperature data.

[0019] Multi-terrain sample areas refer to geographical spatial ranges that include various landform types such as plains, hills, mountains, and plateaus. These areas have sufficient topographic relief and surface cover diversity, aiming to provide training samples for the model to learn temperature response mechanisms under different underlying surface conditions, and to ensure the model's generalization ability under complex terrain.

[0020] First-level surface temperature data refers to temperature products that cover the study area and have low spatial resolution, high temporal continuity, and seamless global characteristics. They are usually derived from numerical weather prediction model outputs or reanalysis datasets and are used to provide a large-scale temperature background field.

[0021] The second type of surface temperature data refers to temperature products with high spatial resolution obtained from the thermal infrared sensors of satellite platforms. Although it is rich in spatial details, it usually suffers from spatial gaps and fragmentation due to cloud contamination. In this scheme, it serves as a high-resolution benchmark reference.

[0022] Topographic and elevation data refers to digital elevation models (DEMs) and their derived topographic factor data, including but not limited to elevation, slope, aspect, topographic curvature, and topographic humidity index, used to characterize the physical structure of the underlying surface.

[0023] In step S101, data acquisition and preprocessing are performed. First and second surface temperature data, along with topographic and elevation data, are acquired covering a multi-terrain sample area including plains, hills, mountains, and plateaus. In this stage, the first surface temperature data uses CLDAS reanalysis surface temperature data with a spatial resolution of 6.25 km; while the second surface temperature data uses MODIS land surface temperature products with a spatial resolution of 1 km. Simultaneously, SRTM DEM data for the area is acquired and aggregated to a 1 km resolution, and elevation, gradient, and topographic relief factors are extracted. By selecting samples from multiple time periods and cloud conditions, it is ensured that the training dataset covers rich meteorological conditions and surface cover changes, providing sufficient information support for the model to learn complex nonlinear mapping relationships.

[0024] For step S102, the first surface temperature data and the terrain and elevation data are raster-matched with the second surface temperature data to achieve a one-to-one correspondence between each raster cell; the first surface temperature data, the second surface temperature data, and the terrain and elevation data are input into the surface temperature downscaling model to be trained for training to obtain the trained surface temperature downscaling model; wherein, the surface temperature downscaling model is used to learn terrain physical prior constraints from the terrain and elevation data and predict temperature residuals based on the terrain physical prior constraints, and use the predicted temperature residuals to correct low-resolution surface temperature to achieve high-resolution surface temperature reconstruction; the first surface temperature data and the terrain and elevation data are used together as training features; the second surface temperature data is used as training labels.

[0025] Raster matching refers to the process of transforming raster datasets from different sources, with different spatial resolutions, or different projection coordinate systems into the same spatial reference system and grid size through resampling and geometric correction, ensuring that corresponding pixels in space correspond one-to-one.

[0026] The surface temperature downscaling model refers to a nonlinear mapping model based on a deep neural network. Instead of directly fitting the absolute values ​​of high and low resolution temperatures, this model is configured as an encoder-decoder structure, specifically designed to extract physical constraint information from terrain features and output the temperature residual between high and low resolution data. The residual correction mechanism is then used to achieve super-resolution reconstruction from low resolution to high resolution.

[0027] Temperature residuals refer to the difference between the true high-resolution surface temperature value and the low-resolution surface temperature background field after interpolation or resampling, at the same geographical location and observation time. This value reflects the microscale temperature variation details that cannot be captured by low-resolution data due to topographic fine structure, local circulation and underlying surface heterogeneity.

[0028] Topographic physical prior constraints refer to the use of physical laws contained in topographic and elevation data, such as the decrease in temperature with increasing altitude and the influence of slope aspect on solar radiation absorption, as mandatory constraints for model learning, guiding the model to predict temperature distribution results that conform to natural physical mechanisms.

[0029] In step S102, the multi-source heterogeneous data undergoes standardization processing and model training. Specifically, the first and second surface temperature data are projected onto the same geographic coordinate system. The low-resolution first surface temperature data is resampled to the same resolution as the second surface temperature data and terrain data using bicubic convolution interpolation, completing raster matching and ensuring that each spatial grid cell contains the corresponding low-resolution temperature value, high-resolution temperature value, and terrain feature value. Subsequently, a surface temperature downscaling model to be trained is constructed. This model preferably employs a deep convolutional neural network with an encoder-decoder structure (such as the U-Net model) or an ensemble learning model such as random forest or XGBoost. During training, the first surface temperature data is stacked with terrain and elevation data (which may include three bands: altitude, slope, and aspect) to form a multidimensional input feature tensor, which is then input into the model. The second surface temperature data is used as a label for supervised learning. The forward propagation process of the model is designed specifically to extract physical structure information from terrain features. By learning prior physical constraints of the terrain, it predicts the temperature residual map between high- and low-resolution data. The loss function is optimized by calculating the difference between the temperature field after the predicted residual is corrected and the second surface temperature label, such as by using root mean square error (RMSE) or mean absolute error (MAE). This forces the model not only to fit numerical statistical relationships but also to follow the terrain's regulation of temperature, thereby achieving convergence of model parameters and obtaining the trained surface temperature downscaling model.

[0030] In one embodiment, the surface temperature downscaling model includes a bi-branch feature extraction network, an encoding / decoding network, and a residual migration diffusion network.

[0031] In this embodiment, in order to further enhance the model's ability to perceive terrain physical mechanisms and improve the accuracy of residual prediction, the surface temperature downscaling model is constructed as a deep neural network architecture with multi-network collaboration.

[0032] A dual-branch feature extraction network is a parallel deep learning network structure, typically consisting of two independent sub-networks, which are used to process terrain and elevation data of different dimensions respectively. Through convolutional kernel groups with different receptive fields, this network can extract multi-scale terrain geometric features from the original terrain data, covering both macroscopic terrain orientation and microscopic landform undulations, providing a detailed physical structure prior for subsequent temperature reconstruction.

[0033] The codec network refers to a fully convolutional network architecture consisting of an encoder and a decoder. The encoder is responsible for compressing the input features and extracting high-level semantic information through downsampling operations, while the decoder restores the feature map to its original resolution through upsampling operations. This structure is mainly used to achieve deep fusion and spatial reconstruction of cross-modal features (terrain features and temperature features).

[0034] The residual migration diffusion network refers to a network module based on the idea of ​​a probabilistic generation model or diffusion model. It is configured to learn the distribution law of residuals between high- and low-resolution temperature data. By encoding and decoding the latent variable space, the network can simulate the random diffusion process and migration characteristics of temperature residuals under complex terrain conditions, thereby generating high-frequency detail information with physical rationality.

[0035] Step S102 involves inputting the first surface temperature data, the second surface temperature data, and the topography and elevation data into the surface temperature downscaling model to be trained, to obtain the trained surface temperature downscaling model. This step includes: Step S1021: The terrain and elevation data are feature extracted using the dual-branch feature extraction network to obtain multi-scale terrain features; the multi-scale terrain features and the first surface temperature data are fused using the encoding and decoding network to output the temperature fusion features of the fused terrain prior; the first surface temperature data and the second surface temperature data are latently encoded using the residual offset diffusion network, and the predicted temperature residuals corresponding to each grid cell are obtained by combining the temperature fusion features; based on the loss error between the predicted temperature residuals and the actual temperature residuals, the dual-branch feature extraction network, the encoding and decoding network, and the residual offset diffusion network are synchronously back-iterally trained until the loss converges to obtain the trained surface temperature downscaling model.

[0036] Latent variable encoding refers to the process of compressing and representing input data in a latent space. In this embodiment, it specifically refers to mapping the first surface temperature data (low resolution) and the second surface temperature data (high resolution) to a high-dimensional manifold space to capture the nonlinear mapping relationship and uncertainty distribution between the two.

[0037] In this embodiment, the training process is initially taken over by a dual-branch feature extraction network. For the input terrain and elevation data (such as a Digital Elevation Model (DEM) and its derived slope and aspect layers), the two branches of the network use convolutional kernels of different sizes: the shallow branch uses small convolutional kernels to capture local terrain abrupt changes and gully textures, while the deep branch uses large convolutional kernels to perceive large-scale mountain range orientations and elevation gradients. After multi-level convolution and activation function processing, the network outputs multi-scale terrain feature maps containing rich geometric information. These feature maps not only include the physical morphology of the terrain but also implicitly contain prior physical knowledge of solar radiation reception angles and ventilation conditions. Subsequently, the aforementioned multi-scale terrain features, along with the first surface temperature data (low-resolution temperature background field) after raster matching, are input into the encoder-decoder network. In the encoder-decoder network, the encoder concatenates the terrain and temperature features along the channel dimension and performs multi-scale feature fusion through a series of convolutional and pooling layers. This process essentially injects the physical constraint information of the terrain into the low-resolution temperature field, generating temperature fusion features that incorporate the terrain priors. This fusion feature preserves the macroscopic temperature trends of low-resolution data while carrying spatial detail guidance from high-resolution terrain.

[0038] Simultaneously, the residual migration diffusion network begins operation. This network receives first and second surface temperature data as input, encodes them using latent variables—that is, maps the low-resolution and high-resolution temperature fields into the same latent space—to solve for the distribution of the difference between them. During the decoding and generation stage, the residual migration diffusion network uses the temperature fusion features output by the aforementioned encoding and decoding network as conditions. Through a back-diffusion process, it gradually removes noise and corrects the distribution of the residuals, ultimately predicting the temperature residuals corresponding to each grid cell. This residual solution method based on a diffusion mechanism is more adaptable to the nonlinear thermal distribution characteristics under complex terrain compared to traditional linear regression.

[0039] Finally, the loss error between the predicted temperature residual output by the model and the true temperature residual (i.e., the second surface temperature data minus the interpolated first surface temperature data) is calculated. This loss error is then used to simultaneously update the weight parameters of the bi-branch feature extraction network, the encoding / decoding network, and the residual offset diffusion network through a backpropagation algorithm. The training process continues iteratively until the loss function stabilizes and converges on the validation set. At this point, the model parameters are saved, signifying that the trained surface temperature downscaling model has the ability to reconstruct a high-resolution temperature field from low-resolution temperature and topographic data.

[0040] In summary, this embodiment, by introducing a generative architecture combining a dual-branch feature extraction network and a residual offset diffusion network, significantly overcomes the problems of insufficient physical mechanism and detail generation capability in existing technologies when dealing with drastically undulating terrain. The dual-branch feature extraction network, through a multi-scale perception mechanism, can accurately interpret the physical constraints of terrain elevation and slope aspect on energy balance. This allows the model to move beyond simple numerical fitting during training, learning the deep causal relationship between terrain structure and temperature distribution, thus effectively suppressing temperature inversion errors in shaded mountain areas and sunny slopes. Simultaneously, the encoding / decoding network, by deeply fusing low-resolution temperature background fields and prior terrain features, ensures the continuity and trend consistency of the reconstruction results in large-scale space, avoiding the block artifacts caused by traditional interpolation methods. More importantly, the residual offset diffusion network utilizes latent variable encoding technology to capture the random diffusion patterns of temperature residuals, generating high-frequency details that conform to natural physical processes, greatly improving image clarity and texture realism. Through the closed-loop optimization mechanism of multi-network collaborative training, the fourth land surface temperature data finally output in this embodiment not only achieved a leap in spatial resolution, but also reached the business standard of high-precision land surface simulation in terms of physical consistency and detail reproduction, truly realizing the generation of high-quality land surface temperature products covering all weather and all terrain.

[0041] In one embodiment, the multi-scale terrain features include several levels of terrain feature maps with progressively smaller resolutions; the dual-branch feature extraction network includes a numerical flow branch and a texture flow branch; and the terrain and elevation data include elevation data, elevation gradient data, and terrain relief data.

[0042] The numerical flow branch refers to a sub-network path in the dual-branch feature extraction network that is specifically designed to process continuous numerical elevation information. This branch focuses on extracting low-frequency semantic features that reflect macroscopic terrain trends and absolute altitude through conventional convolution and pooling operations, aiming to provide large-scale terrain background constraints for temperature reconstruction.

[0043] The texture flow branch refers to a sub-network path in the dual-branch feature extraction network that is specifically used to process the rate of change of terrain and surface roughness. This branch uses special convolution operators such as dilated convolution to focus on capturing high-frequency geometric texture features such as the edges, corners and local undulations of the terrain, aiming to enhance the model's ability to perceive micro-terrain structures.

[0044] Elevation gradient data refers to derived data obtained by spatial differentiation of a digital elevation model (DEM), which typically includes aspect and slope information. This data quantifies the rate of vertical change of terrain per unit horizontal distance and is directly related to the intensity and distribution of solar shortwave radiation received by the Earth's surface.

[0045] Topographic relief data refers to the elevation difference between the highest and lowest points in a region within a certain analysis window. This data reflects the relative degree of dissection and erosion of the land surface and is an important indicator for measuring topographic complexity and local circulation development conditions.

[0046] Step S1021 involves extracting features from the terrain and elevation data using the dual-branch feature extraction network to obtain multi-scale terrain features, including: Step S10211: Perform convolution and average pooling operations on the elevation data through the numerical stream branch to extract continuous elevation data and obtain the first terrain feature.

[0047] The numerical stream branch primarily receives raw elevation data (DEM). Considering that elevation data is essentially a continuous numerical matrix, this branch first uses standard convolutional kernels to filter the data, removing noise and initially extracting trends in elevation changes. Subsequently, cascaded average pooling layers compress the feature maps. This operation effectively smooths local anomalies and preserves the primary topographic features reflecting the main framework of the mountain range and the contours of the basin. This processing ensures that the model can identify large-scale elevation effects such as overall low temperatures in the plateau and high temperatures in the basin, laying a macroscopic foundation for subsequent temperature correction.

[0048] Step S10212: Perform convolution and dilated convolution operations on the elevation gradient data and the terrain undulation data through the texture flow branch to extract the terrain texture data and obtain the second terrain feature.

[0049] The texture flow branch processes elevation gradient data and terrain relief data in parallel. Unlike the numerical flow branch, the texture flow branch introduces dilated convolution operations. By setting different dilation rates, it significantly expands the receptive field of the convolution kernel without increasing the number of parameters or sacrificing resolution. This allows the network to keenly capture terrain texture details such as steep cliffs, narrow valleys, and gentle hill transition zones, generating secondary terrain features. For example, in areas with abrupt slope changes, this branch can extract strong edge features, thereby guiding the model to assign greater weights to temperature residuals in these areas to reconstruct temperature abrupt changes caused by terrain shading or enhanced radiation effects.

[0050] Step S10213: After fusing the first terrain feature and the second terrain feature, perform multiple downsampling operations to generate several levels of terrain feature maps with progressively smaller resolutions, thus obtaining multi-scale terrain features.

[0051] The first terrain feature output from the numerical stream branch and the second terrain feature output from the texture stream branch are fused along the channel dimension. This fusion process typically employs feature concatenation or element-wise addition, ensuring that the fused feature map possesses both macroscopic elevation and microscopic texture information. Subsequently, this fused feature is fed into a series of downsampling modules consisting of convolutional and max-pooling layers, performing multiple progressive downsampling operations. Each downsampling generates a terrain feature map at a specific level; as the level deepens, the resolution of the feature map decreases progressively, but the semantic abstraction level increases progressively. Ultimately, this series of feature maps at different scales constitutes a multi-scale terrain feature pyramid, providing subsequent encoding and decoding networks with comprehensive physical prior constraints from global topography to local landforms.

[0052] This embodiment successfully achieves hierarchical decoupling and accurate extraction of terrain physical information by constructing a dual-branch feature extraction mechanism that combines numerical and texture flows. The numerical flow branch effectively preserves the macroscopic control effect of absolute altitude on the vertical temperature lapse rate through average pooling operations, avoiding temperature anomalies that violate thermodynamic laws in complex mountainous areas. The texture flow branch utilizes the high-resolution perception capability of dilated convolution to enhance the model's accuracy in recognizing micro-topography such as slope, aspect, and terrain undulations, enabling the reconstruction results to delicately depict the thermal differences between sunny and shady slopes and the local temperature changes caused by valley winds. Simultaneously, by fusing multi-scale features and constructing a feature pyramid, the model possesses strong cross-scale modeling capabilities, capable of grasping large-scale temperature field distribution trends while reproducing small-scale details of ground feature thermal radiation. This embodiment significantly improves the physical consistency and spatial detail representation of the downscaling model in areas with drastically dissected terrain.

[0053] In one embodiment, the encoding / decoding network is a sliding window transform residual prediction encoding / decoding network; the encoding / decoding network includes several layers of encoding layers and several layers of decoding layers connected in sequence; each of the encoding layers has a built-in sliding window transform multi-head attention module; the sliding window transform multi-head attention module is used to capture local spatial details and global long-distance spatial dependencies between surface temperature and terrain features.

[0054] The sliding window transformation multi-head attention module refers to an improved Transformer attention mechanism computation unit. This module divides the input feature map into several non-overlapping local windows and performs self-attention computation within the windows. At the same time, it introduces window shifting operations between windows to achieve cross-window information interaction. This design aims to significantly reduce the computational cost of traditional global attention while effectively capturing local spatial details and long-distance spatial dependencies across the entire domain.

[0055] The sliding window transform residual prediction codec network refers to a U-shaped network architecture that integrates the aforementioned sliding window attention module. This network overcomes the limitation of traditional convolutional networks being restricted to local receptive fields and uses a self-attention mechanism to dynamically establish the correlation between arbitrary long-distance grid units. It is specifically used to predict the residual distribution of high-resolution temperature relative to the low-resolution background field, thereby achieving high-fidelity spatial detail restoration.

[0056] In one embodiment, step S1021, which involves fusing the multi-scale terrain features with the first surface temperature data via the encoding / decoding network to output fused temperature features based on the fused terrain priors, includes: Step S10212: Input the first surface temperature data into the encoding / decoding network; the encoding / decoding network includes several sequentially connected encoding layers and several decoding layers; the several encoding layers and the several terrain features Figure 1 One-to-one correspondence.

[0057] The first surface temperature data (low-resolution temperature background field) after raster matching and interpolation is input into the encoding / decoding network. This network is constructed as a symmetrical U-shaped structure, consisting of several stacked encoder layers and decoder layers. Notably, the encoder layers in this embodiment are not traditional convolutional blocks, but rather incorporate a sliding window transform multi-head attention module. This module first divides the input feature map into fixed-size windows (e.g., 8×8 pixels), and independently calculates the attention weights between the query, key, and value within each window, thereby efficiently capturing subtle correlations between terrain and temperature in the local neighborhood. Subsequently, a cyclic shift operation changes the window division method, creating overlapping areas between adjacent windows, thus enabling cross-window information flow and establishing long-distance spatial dependencies across the entire map (such as the blocking effect of mountain ranges on monsoons). This design allows the model to focus on both micro-scale terrain texture and large-scale climate background.

[0058] Step S10213: The input data is feature extracted through the encoding layers at each level and fused with the corresponding level of the terrain feature map in the multi-scale terrain features to obtain fused data, which is then transmitted to the next level of encoding layer until it is output to the decoding layer; the input data is upsampled one by one through the decoding layers at each level to restore spatial details, and finally the temperature fusion feature adapted to the target resolution is output; the target resolution is the resolution of the second surface temperature data.

[0059] In step S10213, the network begins a hierarchical feature fusion and upsampling process. During the encoding phase, each encoding layer extracts temperature features while simultaneously performing precise alignment fusion with the multi-scale terrain features output by the aforementioned dual-branch network. Specifically, the temperature feature map output by the Nth encoding layer is channel-stitched or feature-calibrated with the Nth-layer terrain feature map of matching resolution. This skip connection-style fusion strategy ensures that the higher-level semantic information (temperature trend) is always guided by the lower-level physical information (terrain constraints). As data is passed down through the encoder, the size of the feature map gradually decreases, but the level of feature abstraction continuously increases, ultimately forming a deep fused feature rich in physical priors. During the decoding phase, several decoding layers sequentially upsample the features from the deepest layer of the encoder. Each decoding layer also utilizes a sliding window transform multi-head attention module to refine the upsampled features, gradually restoring the spatial details lost due to downsampling. Through deconvolution or interpolation operations, the resolution of the feature map is gradually restored to the same target resolution as the second land surface temperature data (i.e., the high-resolution remote sensing product). Finally, the network outputs a temperature fusion feature adapted to this target resolution. This feature not only includes the accurate temperature baseline provided by the low-resolution data, but also incorporates high-frequency residual information guided by terrain prior constraints, laying a solid foundation for the final accurate reconstruction.

[0060] This embodiment successfully overcomes the bottleneck of traditional convolutional neural networks in balancing receptive field and computational efficiency by introducing a sliding window transformation multi-head attention mechanism. This mechanism enables the model to dynamically establish long-distance dependencies between terrain structure and temperature distribution. For example, when predicting the temperature at the edge of a plateau, the model can refer to the blocking effect of mountains hundreds of kilometers away, thus significantly improving the physical plausibility of the temperature field on a large scale. Simultaneously, the jump fusion strategy combining multi-scale terrain features effectively prevents gradient vanishing and information loss in deep networks, ensuring that temperature abrupt changes caused by micro-topography can be accurately captured. Ultimately, this feature fusion method based on the attention mechanism achieves an optimal balance between the accuracy of macroscopic trends and the richness of microscopic details in the output temperature fusion features, greatly reducing artifacts and biases in temperature inversion in complex mountainous areas, and providing core algorithmic support for generating high-precision, high-reliability surface temperature products.

[0061] In one embodiment, step S1021 involves encoding latent variables in the first and second surface temperature data using the residual offset diffusion network, and solving for the predicted temperature residuals corresponding to each grid cell using the temperature fusion features; based on the loss error between the predicted and actual temperature residuals, synchronously iteratively training the dual-branch feature extraction network, the encoding / decoding network, and the residual offset diffusion network until the loss converges, to obtain the trained surface temperature downscaling model. This step includes: Step S10214: The first surface temperature data and the second surface temperature data are respectively encoded by variational latent variables through the residual offset diffusion network to generate corresponding conditional latent variables and target latent variables.

[0062] Variational latent variable encoding refers to a feature compression and distribution fitting method based on variational autoencoder (VAE). This process maps input data to a predefined probability distribution space (usually Gaussian distribution) and outputs a latent variable vector representing the mean and variance of the distribution, aiming to capture the uncertainty and complex statistical characteristics in the data generation process.

[0063] Conditional latent variables refer to the latent spatial representation generated by variational latent variable encoding of the first surface temperature data (low-resolution background field). This variable serves as a priori input for the diffusion process, carrying large-scale temperature trends and macroscopic energy balance information of the area to be reconstructed, and is used to constrain the direction and range of residual generation.

[0064] The target latent variable refers to the latent space representation generated by variational latent variable encoding of the second surface temperature data (high-resolution ground truth). This variable represents the true distribution of the high-resolution temperature field under ideal conditions and serves as a supervisory signal to guide the model in learning an accurate mapping from low resolution to high resolution.

[0065] In this step, the residual migration diffusion network initiates two encoding channels in parallel. The first channel receives the first land surface temperature data (low resolution), analyzes it using a variational latent variable encoder, and generates the corresponding conditional latent variables. This process essentially compresses the macroscopic temperature background field into a latent vector containing statistical distribution parameters, preserving the large-scale warming and cooling trends. The second channel receives the second land surface temperature data (high-resolution ground truth), performs the same variational encoding operation, and generates the target latent variable. This step not only achieves dimensionality reduction and feature extraction but, more importantly, unifies the high- and low-resolution data into the same latent space coordinate system, laying the foundation for subsequent calculations of their distribution migrations.

[0066] Step S10215: Using the conditional latent variable, the target latent variable, and the temperature fusion feature as input parameters for the latent variable mapping function of the pre-constructed adaptive terrain physical prior constraints in the residual offset diffusion network, perform latent variable mapping operation to obtain the latent variable distribution difference.

[0067] The latent variable mapping function refers to the core computational unit in the residual offset diffusion network, typically a multilayer perceptron (MLP) or a convolutional neural network. This function is configured to receive conditional latent variables, target latent variables, and fused features as input, solve for the distribution difference between them in the latent space through nonlinear transformation, and decode to generate the corresponding temperature residual.

[0068] This step constructs a latent variable mapping function. The input to this function contains three key elements: the aforementioned conditional latent variables (low-resolution features), the target latent variables (high-resolution features), and temperature fusion features from the encoding / decoding network (topographic physical prior). During the mapping operation, the network first calculates the distribution difference between the conditional and target latent variables, which is geometrically represented as a "distance" in the latent space. Subsequently, the temperature fusion features intervene as a physical constraint, modulating the offset direction of the latent variables to ensure that the generated residuals not only conform to the data distribution patterns but also strictly follow topographic physical mechanisms (e.g., forcibly generating a cooling offset in high-altitude areas). Through this mapping operation adapted to the topographic physical prior constraints, the network outputs a latent variable distribution difference, which encodes all the high-frequency information needed to supplement the transition from low to high resolution.

[0069] Step S10216: Based on the latent variable distribution difference, solve for the predicted temperature residual between the corresponding raster cells of the first land surface temperature data and the second land surface temperature data; use the difference between the corresponding raster cells of the second land surface temperature data and the first land surface temperature data as the true temperature residual, calculate the loss error between the predicted temperature residual and the true temperature residual, and synchronously update the network parameters of the dual-branch feature extraction network, the encoding / decoding network, and the residual offset diffusion network based on backpropagation of the loss error; iterate training until the loss converges to obtain the trained land surface temperature downscaling model.

[0070] Decoding is performed based on the latent variable distribution difference to solve for the predicted temperature residual corresponding to each grid cell. Then, the loss error between this predicted temperature residual and the true temperature residual (i.e., the second surface temperature data minus the interpolated first surface temperature data) is calculated. To comprehensively optimize model performance, the loss function is typically designed as a composite form, including both L1 or L2 loss (measuring numerical accuracy) and SSIM loss (measuring structural similarity). During backpropagation, based on the calculated loss error, all network parameters of the bi-branch feature extraction network, the encoding / decoding network, and the residual offset diffusion network are updated synchronously. This end-to-end joint training mechanism allows the three sub-modules—terrain feature extraction, temperature feature fusion, and residual probability modeling—to mutually feed back and collaboratively optimize. Iterative training continues until the loss function converges on the validation set, signifying that the model has successfully learned the downscaling mapping law of surface temperature under complex terrain.

[0071] This embodiment transforms the traditional deterministic regression problem into a probabilistic generation problem by introducing variational latent variable encoding and latent variable mapping mechanisms, significantly enhancing the model's ability to handle complex nonlinear relationships. Utilizing the distributional difference between conditional and target latent variables as the driving signal for residual generation, the model can accurately capture the random fluctuations and shifts in the temperature field caused by micro-topography, effectively solving the systematic biases that traditional methods are prone to in areas with dramatic undulations. Simultaneously, embedding temperature fusion features as physical priors into the latent variable mapping function ensures that the generated residual distribution has clear physical meaning, avoiding artifacts and illusions common in unconstrained generation models. Through synchronous backward iterative training of all network parameters, seamless coupling between topographic physical constraints and data-driven features is achieved, resulting in the final output of fourth-order land surface temperature data achieving leading levels in numerical accuracy, spatial detail, and physical consistency, fully meeting the stringent requirements of high-precision land surface process simulation and climate effect assessment.

[0072] In one embodiment, step S10216, the step of iteratively training until the loss converges to obtain the trained surface temperature downscaling model, includes: Step S102161: When the loss error of several consecutive iterations meets the preset first convergence condition, obtain the pixel error between the surface temperature data reconstructed by the surface temperature downscaling model and the second surface temperature data.

[0073] The number of consecutive rounds can be set according to the actual situation; for example, it can be 10 consecutive rounds.

[0074] The preset first convergence condition refers to a numerical threshold condition used to initially determine whether the model training has entered a stable period. It is usually defined as the total loss error decreasing by less than a certain small threshold (e.g., 0.001) within a certain number of consecutive training epochs, or the loss value fluctuating smoothly within a specific range without a significant downward trend. This condition is used to filter out the oscillating phase in the early stage of training and avoid premature termination of training.

[0075] Pixel error refers to the statistical deviation of the reconstructed land surface temperature data output by the model from the second land surface temperature data (high-resolution true value) in the corresponding raster cells. Commonly used evaluation indicators include mean absolute error (MAE), root mean square error (RMSE), or mean bias, which are used to quantify the absolute accuracy of the reconstruction results in a physical sense.

[0076] This step monitors the changes in loss error during training in real time. When the loss error of several consecutive iterations (e.g., 10 iterations) meets the preset first convergence condition, it indicates that the main weight parameters of the model have been adjusted to near the optimal range, the loss surface tends to flatten, and training has entered the fine-tuning stage. At this point, instead of relying solely on the numerical changes of the loss function, the accuracy verification process is initiated. The current model is forward-inferred on the validation set to obtain the reconstructed surface temperature data, and the pixel error between this data and the second surface temperature data used as the training label is calculated (e.g., the mean absolute error (MAE) of the two is calculated).

[0077] Step S102162: When the pixel error does not meet the preset second convergence condition, continue iterative training; when the pixel error meets the preset second convergence condition, the trained surface temperature downscaling model is obtained.

[0078] The preset second convergence condition refers to the final accuracy threshold set based on pixel error. This threshold is set according to specific business application standards (e.g., requiring MAE to be less than 0.5K). Only when the reconstruction error of the model meets this stringent standard in terms of physical accuracy is the model training considered complete, thereby ensuring the reliability of the output product.

[0079] This step compares the calculated pixel error with the preset second convergence condition. If the pixel error is still greater than the preset accuracy threshold (e.g., MAE > 0.5K), it indicates that although the model has converged in terms of loss, it has not yet reached the expected standard in terms of actual physical accuracy, which may indicate overfitting or underfitting. Iterative training will continue, and strategies such as learning rate decay or regularization adjustment may be triggered. Conversely, if the pixel error meets the preset second convergence condition (e.g., MAE ≤ 0.5K), it proves that the model has converged not only in terms of statistical loss but also in terms of the accuracy of the physical temperature values. At this point, the system determines that training is complete, locks and saves the current model parameters, and generates the final usable downscaling model of land surface temperature.

[0080] This embodiment effectively solves the "pseudo-convergence" or "overfitting" problems caused by relying solely on loss function values ​​to determine convergence in traditional training by constructing a dual convergence determination mechanism of "initial convergence judgment based on loss function + secondary convergence judgment based on pixel accuracy". By introducing physically meaningful pixel error as a hard indicator in the later stages of training, the model is forced to seek the optimal balance between statistical optimization and physical accuracy, ensuring that the final output of the fourth land surface temperature data has high numerical accuracy and reliability. This strict training termination strategy significantly improves the model's generalization adaptability in different seasons and climate zones, avoids reconstruction bias caused by insufficient or overtraining, and enables the generated operational products to stably meet the stringent industry standards of high-precision land surface simulation and quantitative remote sensing inversion.

[0081] In one embodiment, the residual offset diffusion network is built on a variational autoencoder and incorporates a latent variable mapping function adapted to the prior physical constraints of the terrain. The variational autoencoder includes a low-resolution encoder, a ground truth encoder, and a decoder. The low-resolution encoder is used to encode the input first surface temperature data to generate conditional latent variables. The truth encoder is used to encode the input second surface temperature data to generate target latent variables; The latent variable mapping function is used to calculate the latent variable distribution difference based on conditional latent variables, target latent variables, and temperature fusion features. The decoder is used to output the predicted temperature residual of the raster cells corresponding to the first land surface temperature data and the second land surface temperature data based on the latent variable distribution difference.

[0082] Variational autoencoders are a type of generative deep learning architecture consisting of an encoder and a decoder. The core idea is to map the input data into a latent variable space that follows a specific probability distribution (usually a Gaussian distribution), and then resample from this latent space to generate new data samples. In this embodiment, the architecture is used to model the difference in probability distribution between high and low resolution temperature data, rather than a simple deterministic mapping.

[0083] The low-resolution encoder refers to a coding subnetwork in the variational autoencoder, which is specifically responsible for receiving the first surface temperature data (low-resolution reanalysis data). This encoder compresses the low-resolution temperature field and converts it into statistical distribution parameters (mean and variance) that characterize the macro temperature trend through a combination of convolutional layers and fully connected layers, i.e., conditional latent variables.

[0084] The truth encoder refers to another coding subnetwork in the variational autoencoder, which is specifically responsible for receiving the second land surface temperature data (high-resolution remote sensing truth). This encoder is structurally symmetrical with the low-resolution encoder and is used to encode the high-resolution true temperature field into the corresponding target latent variable, which serves as the target distribution for supervised learning.

[0085] The decoder refers to the generator component in the variational autoencoder, located at the end of the network. This component receives the latent variable distribution difference after mapping operation as input, decodes it back to the original physical space through deconvolution or upsampling operations, and finally outputs the specific predicted temperature residual value.

[0086] The latent variable mapping function refers to the nonlinear transformation function embedded in the latent space of the variational autoencoder. This function is configured to receive conditional latent variables, target latent variables, and externally introduced temperature fusion features (topographic physical priors), calculate the interaction relationship between the three through a multilayer perceptron or feature modulation layer, and output the latent variable distribution difference used to correct the temperature field.

[0087] In this embodiment, the residual migration diffusion network is constructed based on the theoretical framework of variational autoencoders, but its core innovation lies in breaking the traditional VAE paradigm of relying solely on a single input for reconstruction. Instead, it constructs a special structure of dual encoders and a single decoder. Specifically, the network first processes the input first land surface temperature data through a low-resolution encoder. This encoder extracts the large-scale atmospheric background and energy balance information contained in the low-resolution data and transforms it into a conditional latent variable. This variable serves as a priori constraint for the generation process, determining the overall trend and warm / cold distribution pattern of the reconstructed temperature field. Simultaneously, the ground truth encoder processes the second land surface temperature data, extracting the true detailed distribution from the high-resolution data and generating a target latent variable. During the training phase, these two latent variables are treated as two probabilistic expressions of the same physical quantity (land surface temperature) at different observation scales. The latent variable mapping function then intervenes, which not only calculates the simple difference between the conditional latent variable and the target latent variable but also introduces the temperature fusion features from the encoder-decoder network (including physical prior information such as terrain elevation, slope, and aspect) as a modulation factor into the mapping process. In this way, the latent variable mapping function can calculate a latent variable distribution difference that adapts to the prior constraints of terrain physics. This difference includes the statistical corrections required for the transition from low to high resolution, while being strictly constrained within the boundaries of terrain physics (e.g., ensuring that unreasonable warming residuals are not generated in high-altitude areas). Finally, the decoder receives this latent variable distribution difference and decodes it into specific values ​​in physical space, outputting the predicted temperature residuals for each raster cell. These predicted residuals are then used to correct the low-resolution temperature background field, achieving high-resolution surface temperature reconstruction. Through this network structure design based on a combination of variational inference and physical constraints, this embodiment achieves deep coupling between data-driven approaches and physical mechanisms.

[0088] In summary, this embodiment fundamentally improves the modeling capability of downscaling models for complex nonlinear transformations of temperature fields by constructing a dual-channel encoding and latent variable mapping architecture based on variational autoencoders. By utilizing low-resolution encoders and ground truth encoders to extract statistical features at different scales, the model can understand the intrinsic relationship between high- and low-resolution data at the probability distribution level, effectively avoiding the information loss problem common in traditional regression methods. In particular, the introduction of a latent variable mapping function adapted to terrain physical prior constraints successfully embeds terrain features as physical regularization terms into the transformation process of the latent space, forcing the generated residual distribution to conform to natural geographical laws and significantly eliminating temperature artifacts and physical inconsistencies in complex mountainous areas. Finally, the decoder accurately outputs the predicted temperature residuals, achieving high-fidelity correction of the low-resolution background field, resulting in an optimal balance between statistical accuracy and physical realism in the output fourth-level surface temperature data.

[0089] In one embodiment, before synchronously updating the network parameters of the dual-branch feature extraction network, the encoding / decoding network, and the residual offset diffusion network, the optimal hyperparameter configuration is determined by traversing a preset parameter range using a grid search method. The optimal hyperparameters are used for network initialization and parameter iterative training. The preset parameter range includes: learning rate 1e-5~6e-5, batch size 32~96, basic number of network channels 128~256, Swing embedding dimension 128~256, window size 4~16, and number of training iterations 50000~200000. The iterative training process uses the AdamW optimizer, and the learning rate scheduling strategy uses Warmup preheating combined with cosine annealing decay.

[0090] Among them, grid search refers to an exhaustive search method in hyperparameter tuning. This method systematically traverses all possible parameter combinations within a given discrete parameter range, evaluates the model performance of each parameter group through cross-validation, and selects the configuration that performs best on the validation set. It aims to eliminate the subjectivity and randomness of manually setting parameters based on experience.

[0091] The AdamW optimizer is an improved stochastic gradient descent optimization algorithm. Based on the traditional Adam optimizer, this algorithm decouples weight decay from gradient updates and applies it directly to the weight parameters, effectively solving the problem of L2 regularization failure in the Adam optimizer. It has stronger generalization ability and faster convergence speed.

[0092] Warmup refers to a learning rate scheduling strategy. In the early stages of training, the model weights are randomly initialized, and an excessively large learning rate can easily lead to training instability or even divergence. The warmup strategy achieves a smooth start to the training process by linearly or non-linearly increasing the learning rate from a minimum value to a preset base learning rate over several steps at the beginning of training.

[0093] Cosine annealing decay refers to a dynamic learning rate adjustment strategy. During training, the learning rate is no longer kept constant, but gradually decays from its initial value to near zero according to the curve of a cosine function. This strategy allows for a larger update pace in the early stages of training to quickly approach the optimal solution region, while in the later stages of training, a finer search is performed with smaller step sizes, which helps the model converge to a deeper local optimum.

[0094] This embodiment introduces a grid search-based full-parameter space optimization mechanism, completely eliminating the blindness and randomness of traditional manual parameter tuning. This ensures the model enters training with an optimal initial state under complex terrain samples, significantly shortening the iteration cycle required for convergence. Combining the AdamW optimizer with a weight decay strategy effectively improves the model's generalization ability, preventing overfitting common in complex mountainous data and maintaining high accuracy even when facing unfamiliar regions. More importantly, the learning rate scheduling strategy combining Warmup and cosine annealing successfully avoids gradient explosion and oscillation problems in deep neural network training, achieving extremely smooth training and prompting the model to converge to a better local extremum. In summary, this embodiment significantly improves the reconstruction accuracy and robustness of the fourth land surface temperature data, providing a stable and reliable high-performance algorithm model for operational use.

[0095] For step S103, acquire the target area, the third surface temperature data with the same resolution as the first surface temperature data, and the corresponding terrain and elevation data, input the surface temperature downscaling model, and output the fourth surface temperature data.

[0096] The third type of surface temperature data refers to the real-time or historical surface temperature dataset with lower resolution to be processed. This data usually comes from numerical weather prediction model outputs or reanalysis data, and has full spatial coverage and good physical consistency, but lacks the ability to characterize fine local features. It is the basic input for upscaling reconstruction in this invention.

[0097] The fourth type of surface temperature data refers to the final product output during the model inference stage, which has a spatial resolution of sub-kilometer or even meter level. This data not only inherits the macroscopic temperature trend and continuity of the low-resolution reanalysis data, but also integrates the thermal differences caused by high-resolution terrain details, and is a complete temperature field that fills the gaps in the original high-resolution data.

[0098] In step S103, the trained model is deployed in the target area for practical application. Third surface temperature data (with the same resolution as the first surface temperature data, e.g., 0.1°C) and corresponding terrain and elevation data for the target area at the current time or any time series are acquired. This set of low-resolution temperature data and high-resolution terrain data is input into the model trained in step S102. The model utilizes the learned terrain physical prior constraints to directly output fourth surface temperature data. This fourth data has a spatial resolution comparable to the second surface temperature data (e.g., reaching 1 km or higher), and its spatial distribution seamlessly integrates the details of microclimate changes caused by terrain, thus achieving high-fidelity reconstruction from coarse to fine resolution.

[0099] In one embodiment, step S103, which involves acquiring a target area, third surface temperature data with a resolution consistent with the first surface temperature data, and corresponding terrain and elevation data, inputting the surface temperature downscaling model, and outputting fourth surface temperature data, includes: Step S1031: Based on the distribution of each grid cell of the second surface temperature data, the third surface temperature data and the corresponding terrain and elevation data are divided into one-to-one grid cells.

[0100] First, the current third-level surface temperature data and its corresponding topographic and elevation data for the target area are acquired. Since the coverage area, projection method, or grid size of the target area may not be entirely consistent with the training samples in practical applications, the grid cell distribution of the second-level surface temperature data (i.e., the high-resolution ground truth during training, such as the MODIS 1km product) is used as a standard reference. A grid alignment operation is performed, resampling the third-level surface temperature data (low resolution) to the same resolution and grid system as the second-level surface temperature data using bilinear interpolation. Simultaneously, the topographic and elevation data of the corresponding area are also cropped and resampled to this standard grid. This process ensures that the data input to the model achieves a pixel-level one-to-one correspondence in spatial location, eliminating spatial misalignment problems caused by projection deviations or resolution mismatches.

[0101] Step S1032: Input the third surface temperature data and the data of each grid cell in the corresponding topographic and elevation data into the surface temperature downscaling model to obtain the predicted temperature of each grid cell; and obtain the fourth surface temperature data based on the predicted temperature of each grid cell.

[0102] This step loads the trained land surface temperature downscaling model. This model is configured to receive low-resolution temperature data and topographic features as input and output high-resolution temperature residuals. Preprocessed third-order land surface temperature data, along with corresponding topographic and elevation data, are input into the network according to the tensor format required by the model. During inference, a two-branch feature extraction network first extracts topographic features, then an encoding / decoding network fuses topographic priors and the temperature background field, and finally, a residual offset diffusion network predicts the temperature residuals of each raster cell based on the learned latent variable mapping relationship. The predicted temperature residuals output by the model are directly superimposed onto the upsampled third-order land surface temperature data to obtain the final high-resolution predicted temperature values ​​for each raster cell. The discrete predicted temperature values ​​are then reassembled into a standard-format raster image, generating fourth-order land surface temperature data. This data not only fills the gaps in the original observations but also improves the spatial resolution to a level comparable to satellite remote sensing products.

[0103] The inference process provided in this embodiment effectively solves the spatial benchmark inconsistency problem commonly encountered in operational applications of multi-source heterogeneous data through a strict data alignment strategy based on standard grid cell distribution. This ensures the geometric accuracy of the model input data, thereby guaranteeing the reliability of the inference results. Utilizing the trained deep learning model for forward propagation computation achieves efficient end-to-end conversion from low-resolution input to high-resolution output, demonstrating strong engineering feasibility. The generated fourth land surface temperature data successfully overcomes the physical limitations of a single data source, retaining the full spatial coverage advantage and macroscopic physical consistency of low-resolution data while incorporating microscopic thermal details under high-resolution terrain constraints, achieving all-weather, complete, and high-precision land surface temperature product generation. This embodiment significantly improves the spatiotemporal precision of land surface temperature monitoring in complex terrain areas, providing solid data support for precision agriculture, mountain disaster early warning, and climate change research.

[0104] Please refer to Figure 2 This application also provides a computer device 301, including: a processor 302, a memory 303, and a computer program 304 stored in the memory 303 and executable on the processor 302. When the processor 302 executes the computer program 304, it implements the steps of the method as described in any one of the embodiments of this application.

[0105] The processor 302 may include one or more processing cores. The processor 302 connects to various parts within the computer device 301 via various interfaces and lines. It executes various functions of the computer device 301 and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 303, and by accessing data stored in the memory 303. Optionally, the processor 302 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 302 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required to be displayed on the touch screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 302 and may be implemented as a separate chip.

[0106] The memory 303 may include random access memory (RAM) or read-only memory. Optionally, the memory 303 may include a non-transitory computer-readable storage medium. The memory 303 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 303 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch instructions), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory 303 may also be at least one storage device located remotely from the aforementioned processor 302.

[0107] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the method described in any one of the embodiments of this application. That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The computer program may include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The aforementioned storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in computer-readable media may be appropriately added to or subtracted from the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, computer-readable media may not include electrical carrier signals and telecommunication signals, in accordance with legislation and patent practice.

[0108] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and this application also intends to include these modifications and variations.

Claims

1. A spatial downscaling method for surface temperature that integrates prior knowledge of topographic physics, characterized in that, Includes the following steps: Acquire first surface temperature data, second surface temperature data, and topographic and elevation data covering a multi-topographic sample area; The first resolution of the first surface temperature data is smaller than the second resolution of the second surface temperature data; The first surface temperature data and the terrain and elevation data are raster-matched with the second surface temperature data to achieve a one-to-one correspondence between each raster cell. The first surface temperature data, the second surface temperature data, and the terrain and elevation data are then input into the surface temperature downscaling model to be trained to obtain the trained surface temperature downscaling model. The surface temperature downscaling model is used to learn terrain physical prior constraints from the terrain and elevation data and predict temperature residuals based on these constraints. The predicted temperature residuals are then used to correct low-resolution surface temperatures, achieving high-resolution surface temperature reconstruction. The first surface temperature data and the terrain and elevation data are used together as training features; the second surface temperature data is used as the training label. Acquire the target area, the third surface temperature data with the same resolution as the first surface temperature data, and the corresponding terrain and elevation data. Input the surface temperature downscaling model and output the fourth surface temperature data.

2. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 1, characterized in that, The surface temperature downscaling model includes a bi-branch feature extraction network, an encoding / decoding network, and a residual migration diffusion network. The steps of inputting the first surface temperature data, the second surface temperature data, and the topography and elevation data into the surface temperature downscaling model to be trained to obtain the trained surface temperature downscaling model include: The dual-branch feature extraction network is used to extract features from the terrain and elevation data to obtain multi-scale terrain features; The multi-scale terrain features and the first surface temperature data are fused using the encoding and decoding network to output the temperature fusion features based on the fused terrain prior. The first surface temperature data and the second surface temperature data are latently encoded using the residual offset diffusion network, and the predicted temperature residuals for each grid cell are obtained by combining the temperature fusion features. Based on the loss error between the predicted temperature residuals and the actual temperature residuals, the dual-branch feature extraction network, the encoding and decoding network, and the residual offset diffusion network are synchronously trained in reverse iterative manner until the loss converges, resulting in the trained surface temperature downscaling model.

3. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 2, characterized in that, The steps include: encoding latent variables in the first and second surface temperature data using the residual offset diffusion network, and solving for the predicted temperature residuals for each grid cell using the temperature fusion features; based on the loss error between the predicted and actual temperature residuals, synchronously iteratively training the dual-branch feature extraction network, the encoding / decoding network, and the residual offset diffusion network until the loss converges, to obtain the trained surface temperature downscaling model; and including: The residual offset diffusion network is used to perform variational latent variable encoding on the first surface temperature data and the second surface temperature data respectively to generate corresponding conditional latent variables and target latent variables. The conditional latent variable, the target latent variable, and the temperature fusion feature are used together as input parameters for the latent variable mapping function of the pre-constructed adaptive terrain physical prior constraints in the residual migration diffusion network, and the latent variable mapping operation is performed to obtain the latent variable distribution difference. Based on the latent variable distribution difference, the predicted temperature residuals of the corresponding raster cells of the first and second land surface temperature data are calculated; the difference between the corresponding raster cells of the second and first land surface temperature data is used as the true temperature residual, and the loss error between the predicted temperature residual and the true temperature residual is calculated. Based on the backpropagation of the loss error, the network parameters of the dual-branch feature extraction network, the encoding / decoding network, and the residual offset diffusion network are updated synchronously; iterative training is performed until the loss converges, and the trained land surface temperature downscaling model is obtained.

4. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 3, characterized in that, The steps for iterative training until the loss converges to obtain the trained land surface temperature downscaling model include: When the loss error of several consecutive iterations meets the preset first convergence condition, the pixel error between the surface temperature data reconstructed by the surface temperature downscaling model and the second surface temperature data is obtained. If the pixel error does not meet the preset second convergence condition, the training continues iteratively; when the pixel error meets the preset second convergence condition, the trained surface temperature downscaling model is obtained.

5. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 2, characterized in that, The multi-scale terrain features include several levels of terrain feature maps with progressively smaller resolutions; the dual-branch feature extraction network includes a numerical flow branch and a texture flow branch; the terrain and elevation data include elevation data, elevation gradient data, and terrain relief data. The steps of extracting features from the terrain and elevation data using the dual-branch feature extraction network to obtain multi-scale terrain features include: The elevation data is convolutionally and averagely pooled by the numerical stream branch to extract continuous elevation data and obtain the first terrain feature. The elevation gradient data and the terrain relief data are convolved and dilated convolutionally performed by the texture flow branch to extract the terrain texture data and obtain the second terrain feature. After fusing the first terrain feature and the second terrain feature, multiple downsampling operations are performed to generate several levels of terrain feature maps with progressively smaller resolutions, resulting in multi-scale terrain features.

6. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 2, characterized in that, The step of fusing the multi-scale terrain features and the first surface temperature data through the encoding and decoding network to output the fused temperature fusion features based on the fused terrain prior includes: The first surface temperature data is input into the encoding and decoding network; the encoding and decoding network includes several layers of encoding layers and several layers of decoding layers connected in sequence; the several layers of encoding layers correspond one-to-one with the several layers of terrain feature maps; The input data is feature extracted through each level of the encoding layer and fused with the corresponding level of the terrain feature map in the multi-scale terrain features to obtain fused data, which is then transmitted to the next level of the encoding layer until it is output to the decoding layer. The input data is upsampled one by one through each level of the decoding layer to restore spatial details, and finally outputs temperature fusion features adapted to the target resolution. The target resolution is the resolution of the second surface temperature data.

7. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 6, characterized in that, The encoding / decoding network is a sliding window transform residual prediction encoding / decoding network; each of the encoding layers has a built-in sliding window transform multi-head attention module; the sliding window transform multi-head attention module is used to capture local spatial details and global long-distance spatial dependencies between surface temperature and terrain features.

8. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 7, characterized in that, The residual migration diffusion network is built on a variational autoencoder and incorporates a latent variable mapping function adapted to the prior physical constraints of the terrain. The variational autoencoder includes a low-resolution encoder, a ground truth encoder, and a decoder. The low-resolution encoder is used to encode the input first surface temperature data to generate conditional latent variables. The truth encoder is used to encode the input second surface temperature data to generate target latent variables; The latent variable mapping function is used to calculate the latent variable distribution difference based on conditional latent variables, target latent variables, and temperature fusion features. The decoder is used to output the predicted temperature residual of the raster cells corresponding to the first surface temperature data and the second surface temperature data based on the latent variable distribution difference.

9. The method for spatial downscaling of surface temperature based on incorporating prior knowledge of topographic physics as described in claim 1, characterized in that, The steps of acquiring third surface temperature data with a target area and resolution consistent with the first surface temperature data, along with corresponding topographic and elevation data, inputting them into the surface temperature downscaling model, and outputting fourth surface temperature data include: Based on the distribution of each grid cell of the second surface temperature data, the third surface temperature data and the corresponding terrain and elevation data are respectively divided into one-to-one grid cells. The third surface temperature data and the data of each grid cell in the corresponding topographic and elevation data are input into the surface temperature downscaling model to obtain the predicted temperature of each grid cell; based on the predicted temperature of each grid cell, the fourth surface temperature data is obtained.

10. A computer device, characterized in that, It includes a processor, a memory, and a computer-readable program stored in the memory, which, when executed by the processor, implements the steps of the method as described in any one of claims 1-9.

Images