A high-resolution monthly land surface temperature reconstruction method, device and equipment

Abstract

The present disclosure relates to the technical field of remote sensing data processing and earth observation, and provides a high-resolution monthly land surface temperature reconstruction method, device and equipment. The method can include the following steps: performing spatial grid processing and quality control on multi-source input data, constructing a monthly reference field, extracting and aggregating fine-scale residuals, generating an initial monthly land surface temperature field, constructing a time window containing a target month and several adjacent months before and after the target month, constructing a multi-channel input representation for each time step in the time window, performing spatial feature extraction and interpretable time series aggregation on the multi-channel input representation, and generating refined monthly land surface temperature data by using fused spatio-temporal features. The technical scheme provided by one or more embodiments of the present disclosure can stably provide land surface temperature information with a monthly time scale and a hundred-meter spatial scale, and can meet the actual application requirements in the fields of urban thermal environment monitoring, agricultural drought diagnosis and ecological phenology tracking.

Description

Technical Field

[0001] This disclosure relates to the fields of remote sensing data processing and Earth observation technology, specifically to a high-resolution method, apparatus, and equipment for reconstructing monthly Earth surface temperature. Background Technology

[0002] Land surface temperature (LST) is an important physical quantity in the land and atmosphere system, and it is of great significance for drought monitoring, urban heat island assessment, and surface process analysis. Currently, satellite thermal infrared remote sensing is the main means of obtaining large-scale, long-term LST information, but existing technologies generally suffer from the problem of balancing temporal continuity and spatial resolution.

[0003] Coarse-resolution continuous land surface temperature products, such as MODIS (Moderate Resolution Imaging Spectroradiometer) and VIIRS (Visible Infrared Imaging Radiometer Suite), provide regional or global coverage information with high temporal continuity. However, their spatial resolution is typically around 1 kilometer, making it difficult to characterize fine-scale thermal differences in urban blocks and agricultural plots. Therefore, researchers have further developed seamless monthly land surface temperature datasets, such as GSHTD (Global Seamless High-resolution Thermal Dataset). By interpolating and quality-controlling coarse-resolution observations, GSHTD provides temporally complete monthly land surface temperature benchmark information, which can serve as a low-frequency reference for time-series reconstruction. However, the spatial resolution of GSHTD data is still comparable to the aforementioned coarse-resolution products, making it difficult to meet the needs of fine-scale applications at the 100-meter scale.

[0004] High-resolution thermal infrared observation data, represented by Landsat 5, Landsat 7, and Landsat 8, can provide thermal infrared observation information on a scale of about 60 to 120 meters. However, due to the 16-day revisit cycle and cloud contamination, the effective observation time series is usually sparse. Summary of the Invention

[0005] In view of this, this disclosure provides a method, apparatus and equipment for high-resolution monthly land surface temperature reconstruction in one or more embodiments, which can stably provide land surface temperature information with a time scale of monthly and a spatial scale of hundreds of meters, and can meet the practical application needs in fields such as urban thermal environment monitoring, agricultural drought diagnosis and ecological phenology tracking.

[0006] In a first aspect, this disclosure provides a high-resolution monthly land surface temperature reconstruction method. The method includes: uniformly processing and quality-controlling multiple initial land surface temperature data to obtain candidate land surface temperature data, wherein the initial land surface temperature data includes monthly average continuous land surface temperature data and high-resolution satellite land surface temperature data; using annual periodic data from the candidate land surface temperature data, performing harmonic fitting to generate a continuous daily resolution reference field and a monthly reference field; using high-resolution satellite observation data from the candidate land surface temperature data to perform fine-scale residual extraction and aggregation to determine a fine-scale residual estimate; and combining the monthly reference field with the fine-scale residual estimate. The data are overlaid to generate an initial monthly land surface temperature field. A time window is constructed, centered on the target month and including the target month and several neighboring months before and after it. For each time step within the time window, a multi-channel input representation with the same channel structure is constructed using the initial land surface temperature data, the monthly reference field, and the fine-scale residual estimation. Spatial feature extraction and interpretable temporal aggregation are performed on the multi-channel input representation for each time step within the time window to generate fused spatiotemporal features for the target month. The fused spatiotemporal features are then used to perform fusion correction on the initial monthly land surface temperature field to generate refined monthly land surface temperature data. The step of using high-resolution satellite observation data from the candidate surface temperature data to extract and aggregate fine-scale residuals and determine fine-scale residual estimates includes: calculating the difference between the daily satellite observation temperature and the daily harmonic reference value pixel by pixel based on the high-resolution satellite observation data and the continuous daily resolution reference field to generate daily fine-scale thermal spatial deviations; and statistically aggregating the daily fine-scale thermal spatial deviations within the annual cycle to determine the fine-scale residual estimates. The step of performing spatial feature extraction and interpretable temporal series aggregation on the multi-channel input representation for each time step within the time window to generate fused spatiotemporal features for the target month includes: extracting features from the multi-channel input representation using a parameter-shared encoder-decoder network to determine the spatial feature map for each time step; and using an interpretable temporal series aggregation module based on the Kolmogorov-Arnold theorem to perform weighted fusion of the spatial feature maps for each time step within the time window to obtain the fused spatiotemporal features.

[0007] Secondly, this disclosure provides a high-resolution monthly land surface temperature reconstruction device, comprising: a multi-source data acquisition unit, used to uniformly process and quality control multiple initial land surface temperature data to obtain candidate land surface temperature data, wherein the initial land surface temperature data includes monthly average continuous land surface temperature data and high-resolution satellite land surface temperature data; a monthly reference field creation unit, used to perform harmonic fitting using annual periodic data in the candidate land surface temperature data to generate a continuous daily resolution reference field and a monthly reference field; a fine-scale residual estimation unit, used to extract and aggregate fine-scale residuals using high-resolution satellite observation data in the candidate land surface temperature data to determine fine-scale residual estimates; and an initial temperature field generation unit, used to combine the monthly reference field with the fine-scale residual estimates. The system is configured to: overlay data to generate an initial monthly land surface temperature field; a time window setting unit to construct a time window centered on the target month, including the target month and several neighboring months before and after it; a multi-channel input setting unit to construct a multi-channel input representation with the same channel structure for each time step within the time window, using the initial land surface temperature data, the monthly reference field, and the fine-scale residual estimation; a feature extraction and aggregation unit to perform spatial feature extraction and interpretable temporal aggregation on the multi-channel input representation for each time step within the time window, generating fused spatiotemporal features for the target month; and a refinement result output unit to perform fusion correction on the initial monthly land surface temperature field using the fused spatiotemporal features, generating refined monthly land surface temperature data. The step of using high-resolution satellite observation data from the candidate surface temperature data to extract and aggregate fine-scale residuals and determine fine-scale residual estimates includes: calculating the difference between the daily satellite observation temperature and the daily harmonic reference value pixel by pixel based on the high-resolution satellite observation data and the continuous daily resolution reference field to generate daily fine-scale thermal spatial deviations; and statistically aggregating the daily fine-scale thermal spatial deviations within the annual cycle to determine the fine-scale residual estimates. The step of performing spatial feature extraction and interpretable temporal series aggregation on the multi-channel input representation for each time step within the time window to generate fused spatiotemporal features for the target month includes: extracting features from the multi-channel input representation using a parameter-shared encoder-decoder network to determine the spatial feature map for each time step; and using an interpretable temporal series aggregation module based on the Kolmogorov-Arnold theorem to perform weighted fusion of the spatial feature maps for each time step within the time window to obtain the fused spatiotemporal features.

[0008] Thirdly, this disclosure provides an electronic device including a memory and a processor. The memory is used to store a computer program, which, when executed by the processor, implements the above-described high-resolution monthly land surface temperature reconstruction method.

[0009] Fourthly, this disclosure provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the above-described high-resolution monthly land surface temperature reconstruction method.

[0010] The technical solutions provided by one or more embodiments of this disclosure replace the conventional method of directly fusing in the original temperature space with harmonic residual decomposition. By utilizing the temporal smoothness of the harmonic reference field, the annual cycle temperature change is expressed as a continuous smooth function. This allows the high-resolution observation residual to mainly characterize the local spatial deviation of surface thermals, rather than superimposing seasonal amplitude changes. Therefore, it is possible to robustly reconstruct fine-scale thermal spatial models without relying on high-resolution images of cloud-free areas in the same period of the target month.

[0011] This disclosure provides a technical solution through one or more embodiments. By employing a temporal aggregation mechanism with time offset as an explicit independent variable, the contribution of multi-month time-series information is not arbitrarily and implicitly determined by the model, but rather adaptively allocated under the constraints of temporal relationships. This preserves the data-driven modeling capability while avoiding disordered interference from multi-temporal information on the target month. Therefore, it can improve the stability and coherence of monthly reconstruction even when observations are sparse and irregularly distributed.

[0012] This disclosure provides a technical solution through one or more embodiments. By employing a KAN temporal aggregation module parameterized by cubic B-spline basis functions (i.e., replacing fixed weights with learnable mapping functions on network connection edges), the temporal weight function is constrained within a function space that possesses smoothness, local support, and explicit reviewability. This allows it to learn the strength of the influence of different time offsets on the target month based on sample data, without exhibiting uncontrolled and drastic fluctuations under sparse sample conditions, unlike pure implicit depth mapping. Therefore, it can balance data-driven capabilities with constraint stability, reducing the risk of overfitting sampling artifacts.

[0013] The technical solutions provided by one or more embodiments of this disclosure, by using a robust aggregation method of effective residuals within the year to uniformly process high-resolution satellite observations, can reduce the impact of observation differences between different sensors on the consistency of reconstruction results to a certain extent, thereby improving the stability of multi-sensor fusion reconstruction under long-term series conditions. Attached Figure Description

[0014] The features and advantages of the embodiments of this disclosure will be more clearly understood by referring to the accompanying drawings, which are illustrative and should not be construed as limiting the present disclosure in any way. In the drawings: Figure 1 A schematic diagram illustrating the steps of a high-resolution monthly land surface temperature reconstruction method according to one embodiment of the present disclosure is shown. Figure 2 A schematic diagram of the functional modules of a high-resolution monthly land surface temperature reconstruction device in one embodiment of this disclosure is shown. Figure 3 A schematic diagram of the structure of an electronic device according to one embodiment of the present disclosure is shown. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0016] In related technologies, seamless monthly land surface temperature datasets, such as GSHTD, and high-resolution thermal infrared observation data, such as Landsat 5, Landsat 7, and Landsat 8, can theoretically complement each other in terms of both temporal continuity and spatial resolution. However, existing spatiotemporal fusion methods, such as STARFM (Spatial and Temporal Adaptive Reflectance Fusion Model) and ESTARFM (Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model), typically require a high-resolution reference image without cloud cover near the target time as an anchor point, which is often difficult to meet in monthly reconstruction scenarios. Existing statistical and machine learning methods usually model each time step independently, making it difficult to effectively utilize the temporal context information between adjacent months. Existing deep learning temporal modules, such as MLP (Multilayer Perceptron) or LSTM (Long Short-Term Memory), mostly adopt implicit weight mapping, which lacks interpretability of temporal relationships and is prone to overfitting under incomplete temporal conditions caused by cloud pollution.

[0017] In related technologies, from the perspective of the technical approach for the specific task of reconstructing monthly land surface temperature sequences, the core ideas of existing methods can be summarized into the following two categories.

[0018] The first category comprises time smoothing and annual cycle fitting methods, typically represented by ATC (Annual Temperature Cycle) methods and their derivatives. These methods are centered around a pre-defined sinusoidal annual cycle function or a fixed form of time constraint. By applying smoothing constraints to the surface temperature series along the time axis, they maintain the continuity of the annual temperature curve and suppress temporal noise. Their main problem lies in the fact that the temporal constraint is fixed and independent of data distribution, lacking the ability to utilize local fine-scale thermal differences contained in high-resolution observations. They easily erase real local thermal anomalies and high-frequency spatial differences during the smoothing process, resulting in insufficient detail recovery and inadequate response to anomalous local thermal signals.

[0019] The second category comprises data-driven methods based on machine learning and deep learning, typically including regression or spatiotemporal fusion networks based on time-series modules such as MLP and LSTM. These methods learn complex mapping relationships from multi-source inputs and possess strong potential for detail recovery. Their main problems are: the time-series relationships are implicitly learned from the network parameters, the time-series weights cannot be directly examined, and there is a lack of explicit smoothing constraints and structural limitations; under conditions of sparse, irregular, and significantly cloud-affected high-resolution thermal infrared observations, problems such as weight instability, overfitting to random sampling patterns, and insufficient cross-regional generalization can easily occur.

[0020] In summary, among the related technologies, the first type of method has good intra-year smoothing constraint capability but insufficient fine-scale utilization, while the second type of method has strong data-driven fine-scale recovery capability but insufficient stability.

[0021] In view of this, the high-resolution monthly land surface temperature reconstruction method provided by one or more embodiments of this disclosure can simultaneously take into account the smoothing constraint of the annual temperature curve and the data-driven utilization of effective observation information in the monthly reconstruction task.

[0022] Please see Figure 1 The present disclosure provides a high-resolution monthly land surface temperature reconstruction method according to one embodiment, which may include the following steps S1-S8.

[0023] S1: Multiple initial surface temperature data are uniformly processed by spatial grid and quality control to obtain candidate surface temperature data. The initial surface temperature data includes monthly average continuous surface temperature data and high-resolution satellite surface temperature data.

[0024] In this embodiment, coarse-resolution monthly continuous land surface temperature data (e.g., GSHTD, MODIS MOD11A2, VIIRS LST), high-resolution satellite observation data (e.g., Landsat 5, Landsat 7, Landsat 8), and other digital elevation data can all be part of the initial land surface temperature data. After spatial grid processing (e.g., projecting to a 100-meter-level UTM grid) and quality control (e.g., removing cloud and cloud shadow contamination pixels) of the initial land surface temperature data, spatially consistent multi-source input data, i.e., candidate land surface temperature data, can be obtained.

[0025] S2: Using the annual cycle data in the candidate surface temperature data, perform harmonic fitting to generate a continuous daily resolution reference field and a monthly reference field.

[0026] In this embodiment, the twelve monthly average surface temperature values ​​from the candidate surface temperature data are used as input, and a preset harmonic regression model is fitted to obtain a continuous daily resolution reference field. The preset harmonic regression model may include an annual average temperature term, a fundamental harmonic term, and a second harmonic term. By integrating the continuous daily resolution reference field, a monthly reference field can be obtained.

[0027] In a practical application example, for each spatial grid, with the monthly average surface temperature value of 12 months at a coarse resolution as input, the following harmonic regression model can be fitted.

[0028] ; in, For years, The average annual temperature and The fundamental harmonic coefficient, and These are the second harmonic coefficients, and the specific values ​​of these parameters are estimated using the least squares method based on the 12-month average. From this, a continuous daily resolution reference field can be obtained. The monthly reference field formed by its integral This monthly reference field can capture large-scale seasonal signals, providing a low-frequency reference for subsequent residual decomposition.

[0029] S3: Using the high-resolution satellite observation data in the candidate surface temperature data, perform fine-scale residual extraction and aggregation to determine the fine-scale residual estimate.

[0030] In this embodiment, based on high-resolution satellite observation data and a continuous daily resolution reference field, the difference between the daily satellite observation temperature and the daily harmonic reference value can be calculated pixel by pixel, generating a daily fine-scale thermodynamic spatial deviation. Statistical aggregation of the daily fine-scale thermodynamic spatial deviations within the annual cycle allows for the determination of fine-scale residual estimates.

[0031] In a practical application example, for all valid high-resolution observations within each spatial grid, the residuals can be calculated pixel by pixel using the following formula.

[0032] ; in, Indicates different observation days, Representing different spatial pixels, For high-resolution satellite temperature observation, This is the harmonic reference value for that day. This represents fine-scale thermo-space deviation.

[0033] Preferably, within each year, when pixels When there are at least two valid observations, the median of the annual residuals can be used as the final robust fine-scale residual estimate. The median aggregation formula is as follows.

[0034] .

[0035] Median aggregation can simultaneously suppress atmospheric and registration residual errors and naturally fuse multi-sensor (e.g., Landsat 5 / 7 / 8) observations without explicit cross-calibration.

[0036] Alternatively, in addition to aggregating the median, the effective residuals within the year can be aggregated using the mean, or representative residual values ​​can be calculated separately for each quarter. Quarterly estimation is suitable for land surface types with significant seasonal variations.

[0037] It is important to note that temporal smoothing methods, such as ATC, which rely on fixed-form annual periodic functions or time constraints, while maintaining the continuity of the temperature curve within the year, fail to adequately utilize the local fine-scale differences contained in high-resolution observations. They are prone to processing real local thermal anomalies along with large-scale seasonal signals during the smoothing process. This proposed approach, however, robustly aggregates the effective residuals within the year, explicitly extracting recurring fine-scale spatial biases from multiple high-resolution observations as pixel-by-pixel residual estimates, which are then superimposed onto the reference field. In this way, while preserving the smoothness of the annual cycle, it explicitly retains repeatable and stable fine-scale thermo-spatial patterns, avoiding the smoothing out of real local thermal signals.

[0038] S4: The monthly reference field is superimposed with the fine-scale residual estimate to generate the initial monthly surface temperature field.

[0039] In this embodiment, by superimposing the harmonic reference monthly mean field with the fine-scale residual estimate, an initial monthly surface temperature field constrained by annual cycle smoothing can be constructed. This initial monthly surface temperature field can provide a monthly temperature field with spatial details (e.g., at the 100-meter level) without relying on the high-resolution imagery of the target month, serving as input for subsequent deep fusion and refinement.

[0040] Optionally, while generating the initial monthly surface temperature field, a monthly residual diagnostic layer, observation counts, and overlay mask can be generated to characterize the observation support.

[0041] In a practical application example, the initial monthly surface temperature field can be generated using the following formula.

[0042] ; in, As the monthly baseline field, For fine-scale residual estimation, For the initial monthly surface temperature field, parameters Indicates the target month, parameter Represents pixels in different spatial locations.

[0043] It is important to note that spatiotemporal fusion methods, such as STARFM and ESTARFM, typically require a high-resolution reference image with no cloud cover near the target time as an anchor point. If such a reference image is lacking for the target month, their reconstruction capabilities are significantly limited, a condition often difficult to meet in monthly reconstruction scenarios. This proposed solution addresses this by using a harmonic reference field to handle the low-frequency annual cycle signal and fine-scale residuals to handle the high-frequency spatial bias signal, modeling continuous temperature variations and local spatial thermal differences within the year separately. Even when no concurrent high-resolution observations are available for the target month, reconstruction can still be completed using the annual cycle reference and the accumulated fine-scale residuals, fundamentally eliminating the reliance on concurrent reference images for the target month.

[0044] S5: Construct a time window centered on the target month, including the target month and several neighboring months before and after it.

[0045] In this implementation, the time series window for the target month can be two months before and after, for a total of five time steps. Depending on the actual application scenario, the time series window can also be set to one month before and after, for a total of three time steps, or three months before and after, for a total of seven time steps. By adjusting the length of the time window, a trade-off can be struck between computational complexity and the scope of time series information utilization.

[0046] S6: For each time step within the time window, construct a multi-channel input representation with the same channel structure using the initial surface temperature data, the monthly reference field, and the fine-scale residual estimation.

[0047] In this embodiment, a multi-channel input representation with the same channel structure can be constructed for each time step within the time window. The multi-channel input representation shares the same channel definition and arrangement across all time steps, facilitating subsequent shared encoding processing.

[0048] In some implementations, the multi-channel input representation may include the monthly reference field corresponding to the target time step, the initial monthly surface temperature field corresponding to the target time step, the residual spatial layer corresponding to the high-resolution satellite observation data within the target time step, the observation availability mask for the target time step, and the digital elevation data corresponding to the target time step. The observation availability mask indicates whether valid high-resolution satellite observation data exists for each pixel within the target time step, and the residual spatial layer can be set to zero if there are no corresponding valid observations at the target time step.

[0049] S7: For the multi-channel input representation of each time step within the time window, perform spatial feature extraction and interpretable temporal aggregation to generate fused spatiotemporal features for the target month.

[0050] In this embodiment, for the multi-channel input representation of each time step within the time window, a parameter-sharing encoder-decoder network (e.g., a U-Net network) can be used to extract features and determine the spatial feature map of each time step.

[0051] Specifically, an encoder-decoder network (e.g., the U-Net network) can consist of three parts: an encoder, a bottleneck layer, and a decoder. The encoder can contain four downsampling stages, each consisting of two 3×3 convolutional-batch normalization-ReLU layers. The number of output channels for each downsampling stage can be 32, 64, 128, and 256, respectively. At the end of each downsampling stage, a max-pooling operation with a stride of 2 can be applied for double spatial downsampling. The bottleneck layer can consist of two 3×3 convolutional-batch normalization-ReLU layers with 512 channels. The decoder can contain four upsampling stages. Each upsampling stage first performs bilinear interpolation on the feature map, then concatenates it with the feature map of the corresponding stage in the encoder along the channel dimension via skip connections, and finally fuses it through two 3×3 convolutional-batch normalization-ReLU layers. The number of output channels for each upsampling stage can be 256, 128, 64, and 32, respectively. The decoder ends with a 1×1 convolutional layer mapping the number of channels to... Dimensions (e.g., ), corresponding to the first Spatial feature map at each time step .in, , These represent the height and width of the feature map, respectively, and are consistent with the spatial dimensions of the input at the corresponding time step. All time steps share the same set of U-Net parameters, and the spatial thermal structures of different months are encoded into the same feature space, facilitating subsequent temporal aggregation processing.

[0052] Alternatively, depending on the specific application scenario, spatial feature extraction can employ not only the U-Net encoding / decoding structure, but also U-Net variants with attention mechanisms, Transformer structures, or other network structures capable of extracting multi-scale spatial features, in order to achieve subsequent temporal aggregation and refined reconstruction.

[0053] In this embodiment, an interpretable temporal aggregation module based on the Kolmogorov-Arnold (KAN) theorem is used to perform weighted fusion of the spatial feature maps of each time step within the time window to obtain fused spatiotemporal features.

[0054] Specifically, the core idea of ​​the KAN theorem is to replace the fixed weights on the network edges with learnable unary functions, each of which can be parameterized by a linear combination of spline basis functions. Unlike MLP, which applies fixed activation functions to nodes, the mapping function on each edge of the KAN network is an explicit unary function. After training, it can be directly plotted as input-output curves for manual review, thus possessing complete interpretability.

[0055] In some implementations, an interpretable temporal aggregation module based on the Kolmogorov-Arnold theorem is used to perform weighted fusion of the spatial feature maps of each time step within the time window to obtain the fused spatiotemporal features. This includes: calculating the time offset of each reference time step relative to the target month within the time window; using the time offset as the sole input, calculating the scalar temporal weights corresponding to each reference time step through the KAN temporal weight mapping function parameterized by cubic B-spline basis functions; normalizing the scalar temporal weights along the time dimension to obtain normalized weights; and performing element-wise weighted summation of the normalized weights and the spatial feature maps of the corresponding time steps to obtain the fused spatiotemporal features.

[0056] Optionally, depending on the actual application scenario, the time series aggregation module can parameterize the time series weight function using cubic B-spline basis functions, or it can construct the time series weight function using Fourier basis functions, radial basis functions, or other function bases with smooth constraints, as long as interpretable time series aggregation based on time offset relationships can still be achieved.

[0057] In a practical application example, the first time within the time window Each time step allows for the calculation of its time offset relative to the target month. .in, and These are the month sequence numbers (in integer months) for the current time step and the target month, respectively. With the input as the sole input, the scalar time series weight corresponding to this time step can be calculated using the following KAN time series weight mapping function.

[0058] ; in, To define in the interval The first uniform node sequence on the upper A cubic B-spline basis function, The total number of basis functions ( The number of node intervals is given in this embodiment. ), These are the learnable coefficients updated via backpropagation during training. The cubic B-spline basis functions are cubic polynomials within each node interval, and the overall expression has... Continuity. Each basis function is non-zero only within a finite local interval (local support), thus forming a time-series weight function. It combines global smoothness with piecewise local adjustability, and is not prone to drastic fluctuations under sparse observation conditions. The above parameterization is equivalent to a single-layer KAN network with an input dimension of 1 and an output dimension of 1, whose only learnable parameters are coefficient vectors. The number of parameters is extremely small.

[0059] Obtain the scalar weights at each time step. Then, normalization (softmax) along the time dimension yields the normalized weights in the following form. .

[0060] .

[0061] Normalized weights Then compare with the spatial feature map of the corresponding time step By performing element-wise weighted summation, we can obtain the fused spatiotemporal features for the target month in the following form. .

[0062] .

[0063] Due to the time-series weighting function With time offset as the sole independent variable, independent of the input content, this function can be directly plotted as a function of time offset after training. The model is a one-dimensional curve. Therefore, the contribution distribution of different time offsets to the target month can be manually reviewed and qualitatively interpreted. For example, it can be determined whether the time series weights monotonically decrease as the offset increases, or whether they are concentrated within ±1 month, thereby verifying the rationality of the model's time series behavior.

[0064] It is important to note that methods based on deep time series modules such as MLP and LSTM implicitly learn their time series weights from the network parameters. The weight relationships are hidden within the network and cannot be directly examined. This can easily lead to drastic weight fluctuations and overfitting when observations are sparse or irregularly distributed. This approach, however, uses the time offset of each time step relative to the target month as an explicit independent variable and employs cubic B-spline basis functions to parameterize the time series weight mapping. This restricts the time series weights to a function space that simultaneously possesses smoothness and local support. Therefore, the time series weights can adaptively adjust according to the data while being explicitly constrained by the function space structure. After training, they can be directly plotted as curves and manually reviewed, making the distribution of the contribution of different time offsets to the target month interpretable.

[0065] S8: Using the fused spatiotemporal features, the initial monthly surface temperature field is fused and corrected to generate refined monthly surface temperature data.

[0066] In this embodiment, the spatiotemporal features are input into a preset decoding module, which can output a single-channel residual correction map with the same spatial size as the initial monthly land surface temperature field. Using this single-channel residual correction map, residual correction can be performed on the initial monthly land surface temperature field. By overlaying the residual correction result pixel-by-pixel with the initial monthly land surface temperature field, refined monthly land surface temperature data can be obtained.

[0067] In this embodiment, the model parameters of the preset decoding module can be determined through supervised training. During the training process, the mean square error between the prediction result and the reference high-resolution observation can be used as the loss function. Random missing simulation can be applied to the input of non-target months (randomly setting the residual layer of some time steps to zero and marking its availability mask as unavailable) to improve robustness under sparse observation conditions.

[0068] In a practical application example, spatiotemporal features will be fused. (Number of channels is) The input decoding module can output the same data as the initial reconstructed field. Single-channel residual correction plot with the same spatial dimensions The decoding module can consist of two convolutional blocks and a 1×1 convolutional layer connected sequentially. Each convolutional block can contain a 3×3 convolutional layer (outputting 64 channels), a batch normalization layer, and a ReLU activation layer. The 1×1 convolutional layer can compress the number of channels to 1. Referring to the following formula, the residual correction result can be superimposed pixel-by-pixel with the initial reconstructed field to obtain refined monthly land surface temperature data.

[0069] .

[0070] The technical solutions provided by one or more embodiments of this disclosure replace the conventional method of directly fusing in the original temperature space with harmonic residual decomposition. By utilizing the temporal smoothness of the harmonic reference field, the annual cycle temperature change is expressed as a continuous smooth function. This allows the high-resolution observation residual to mainly characterize the local spatial deviation of surface thermals, rather than superimposing seasonal amplitude changes. Therefore, it is possible to robustly reconstruct fine-scale thermal spatial models without relying on high-resolution images of cloud-free areas in the same period of the target month.

[0071] This disclosure provides a technical solution through one or more embodiments. By employing a temporal aggregation mechanism with time offset as an explicit independent variable, the contribution of multi-month time-series information is not arbitrarily and implicitly determined by the model, but rather adaptively allocated under the constraints of temporal relationships. This preserves the data-driven modeling capability while avoiding disordered interference from multi-temporal information on the target month. Therefore, it can improve the stability and coherence of monthly reconstruction even when observations are sparse and irregularly distributed.

[0072] This disclosure provides a technical solution through one or more embodiments. By employing a KAN temporal aggregation module parameterized by cubic B-spline basis functions (i.e., replacing fixed weights with learnable mapping functions on network connection edges), the temporal weight function is constrained within a function space that possesses smoothness, local support, and explicit reviewability. This allows it to learn the strength of the influence of different time offsets on the target month based on sample data, without exhibiting uncontrolled and drastic fluctuations under sparse sample conditions, unlike pure implicit depth mapping. Therefore, it can balance data-driven capabilities with constraint stability, reducing the risk of overfitting sampling artifacts.

[0073] The technical solutions provided by one or more embodiments of this disclosure, by using a robust aggregation method of effective residuals within the year to uniformly process high-resolution satellite observations, can reduce the impact of observation differences between different sensors on the consistency of reconstruction results to a certain extent, thereby improving the stability of multi-sensor fusion reconstruction under long-term series conditions.

[0074] Please see Figure 2 This disclosure also provides a high-resolution monthly land surface temperature reconstruction apparatus, the apparatus comprising: The multi-source data acquisition unit 100 is used to uniformly perform spatial grid processing and quality control on multiple initial surface temperature data to obtain candidate surface temperature data. The initial surface temperature data includes monthly average continuous surface temperature data and high-resolution satellite surface temperature data. The monthly reference field creation unit 200 is used to perform harmonic fitting using the annual cycle data in the candidate surface temperature data to generate a continuous daily resolution reference field and a monthly reference field. The fine-scale residual estimation unit 300 is used to extract and aggregate fine-scale residuals using high-resolution satellite observation data from the candidate surface temperature data, and to determine the fine-scale residual estimate. The initial temperature field generation unit 400 is used to superimpose the monthly reference field with the fine-scale residual estimate to generate an initial monthly surface temperature field. The time window setting unit 500 is used to construct a time window centered on the target month, which includes the target month and several adjacent months before and after it. The multi-channel input setting unit 600 is used to construct a multi-channel input representation with the same channel structure for each time step within the time window, using the initial surface temperature data, the monthly reference field, and the fine-scale residual estimation. The feature extraction and aggregation unit 700 is used to perform spatial feature extraction and interpretable temporal aggregation on the multi-channel input representation for each time step within the time window, and generate fused spatiotemporal features for the target month. The refinement result output unit 800 is used to perform fusion correction on the initial monthly land surface temperature field using the fusion spatiotemporal features to generate refined monthly land surface temperature data.

[0075] In one embodiment, the monthly reference field creation unit 200 is specifically used to: take the twelve monthly average surface temperature values ​​from the candidate surface temperature data as input, fit a preset harmonic regression model to obtain the continuous daily resolution reference field, wherein the preset harmonic regression model includes an annual average temperature term, a fundamental harmonic term, and a second harmonic term; and perform integral calculation on the continuous daily resolution reference field to obtain the monthly reference field.

[0076] In one embodiment, the fine-scale residual estimation unit 300 is specifically used to: calculate the difference between the daily satellite observation temperature and the daily harmonic reference value pixel by pixel based on the high-resolution satellite observation data and the continuous daily resolution reference field, and generate a daily fine-scale thermodynamic spatial deviation; and statistically aggregate the various daily fine-scale thermodynamic spatial deviations within the annual cycle to determine the fine-scale residual estimate.

[0077] In one implementation, the multi-channel input represents the monthly reference field corresponding to the target time step, the initial monthly surface temperature field corresponding to the target time step, the residual spatial layer corresponding to the high-resolution satellite observation data within the target time step, the observation availability mask for the target time step, and the digital elevation data corresponding to the target time step; wherein, the observation availability mask is used to indicate whether each pixel has valid high-resolution satellite observation data within the target time step, and the residual spatial layer is set to zero when there is no corresponding valid observation value at the target time step.

[0078] In one embodiment, the feature extraction and aggregation unit 700 is specifically used to: extract features from the multi-channel input representation using a parameter-sharing U-Net encoder-decoder network to determine the spatial feature map at each time step; and use an interpretable temporal aggregation module based on the Kolmogorov-Arnold theorem to perform weighted fusion of the spatial feature maps at each time step within the time window to obtain the fused spatiotemporal features.

[0079] In one embodiment, the feature extraction and aggregation unit 700 includes a temporal aggregation subunit, which is specifically used for: calculating the time offset of each reference time step relative to the target month within the time window; using the time offset as the only input, calculating the scalar temporal weight corresponding to each reference time step through the KAN temporal weight mapping function parameterized by cubic B-spline basis functions; normalizing the scalar temporal weight along the time dimension to obtain normalized weights; and performing element-wise weighted summation of the normalized weights and the spatial feature map of the corresponding time step to obtain the fused spatiotemporal features.

[0080] In one embodiment, the refinement result output unit 800 is specifically used to: input the fused spatiotemporal features into a preset decoding module and output a single-channel residual correction map with the same spatial size as the initial monthly land surface temperature field; use the single-channel residual correction map to perform residual correction on the initial monthly land surface temperature field; and superimpose the residual correction result with the initial monthly land surface temperature field pixel by pixel to obtain the refined monthly land surface temperature data.

[0081] The various units described in the above embodiments can be implemented by a computer chip or by a product with a certain function. For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0082] Please see Figure 3This disclosure also provides an electronic device, which includes a memory and a processor. The memory is used to store a computer program, which, when executed by the processor, implements the above-described high-resolution monthly land surface temperature reconstruction method.

[0083] This disclosure also provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the above-described high-resolution monthly land surface temperature reconstruction method.

[0084] The processor can be a central processing unit (CPU). It can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations thereof.

[0085] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the methods in the above-described embodiments.

[0086] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0087] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0088] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, embodiments of apparatus, devices, and storage media are basically similar to method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0089] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0090] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A high-resolution method for reconstructing monthly land surface temperature, characterized in that, The method includes: Multiple initial surface temperature data are uniformly processed and quality controlled by spatial grid to obtain candidate surface temperature data. The initial surface temperature data includes monthly average continuous surface temperature data and high-resolution satellite surface temperature data. Using the annual cycle data in the candidate surface temperature data, harmonic fitting is performed to generate a continuous daily resolution reference field and a monthly reference field. Using high-resolution satellite observation data from the candidate surface temperature data, fine-scale residual extraction and aggregation are performed to determine the fine-scale residual estimate; The monthly baseline field is superimposed with the fine-scale residual estimate to generate an initial monthly land surface temperature field; Centered on the target month, construct a time window that includes the target month and several adjacent months before and after it; For each time step within the time window, a multi-channel input representation with the same channel structure is constructed using the initial surface temperature data, the monthly reference field, and the fine-scale residual estimation. For the multi-channel input representation at each time step within the time window, spatial feature extraction and interpretable temporal aggregation are performed to generate fused spatiotemporal features for the target month. Using the fused spatiotemporal features, the initial monthly land surface temperature field is fused and corrected to generate refined monthly land surface temperature data; The step of using high-resolution satellite observation data from the candidate surface temperature data to extract and aggregate fine-scale residuals and determine fine-scale residual estimates includes: calculating the difference between the daily satellite observation temperature and the daily harmonic reference value pixel by pixel based on the high-resolution satellite observation data and the continuous daily resolution reference field to generate daily fine-scale thermal spatial deviations; and statistically aggregating the daily fine-scale thermal spatial deviations within the annual cycle to determine the fine-scale residual estimates. The step of performing spatial feature extraction and interpretable temporal series aggregation on the multi-channel input representation for each time step within the time window to generate fused spatiotemporal features for the target month includes: extracting features from the multi-channel input representation using a parameter-shared encoder-decoder network to determine the spatial feature map for each time step; and using an interpretable temporal series aggregation module based on the Kolmogorov-Arnold theorem to perform weighted fusion of the spatial feature maps for each time step within the time window to obtain the fused spatiotemporal features. The method employs an interpretable temporal aggregation module based on the Kolmogorov-Arnold theorem to perform weighted fusion of the spatial feature maps of each time step within the time window to obtain the fused spatiotemporal features. This includes: calculating the time offset of each reference time step relative to the target month within the time window; using the time offset as the sole input, calculating the scalar temporal weights corresponding to each reference time step through the KAN temporal weight mapping function parameterized by cubic B-spline basis functions; normalizing the scalar temporal weights along the time dimension to obtain normalized weights; and performing element-wise weighted summation of the normalized weights and the spatial feature maps of the corresponding time steps to obtain the fused spatiotemporal features.

2. The method according to claim 1, characterized in that, The step of using annual cycle data from the candidate surface temperature data to perform harmonic fitting and generate continuous daily resolution reference fields and monthly reference fields includes: Using the twelve monthly average surface temperature values ​​from the candidate surface temperature data as input, a preset harmonic regression model is fitted to obtain the continuous daily resolution reference field. The preset harmonic regression model includes an annual average temperature term, a fundamental harmonic term, and a second harmonic term. The monthly reference field is obtained by integrating the continuous daily resolution reference field.

3. The method according to claim 1, characterized in that, The multi-channel input includes the monthly reference field corresponding to the target time step, the initial monthly surface temperature field corresponding to the target time step, the residual spatial layer corresponding to the high-resolution satellite observation data within the target time step, the observation availability mask of the target time step, and the digital elevation data corresponding to the target time step. The observation availability mask is used to indicate whether each pixel has valid high-resolution satellite observation data within the target time step, and the residual spatial layer is set to zero when there is no corresponding valid observation value at the target time step.

4. The method according to claim 1, characterized in that, The process of using the fused spatiotemporal features to perform fusion correction on the initial monthly land surface temperature field to generate refined monthly land surface temperature data includes: The fused spatiotemporal features are input into a preset decoding module, which outputs a single-channel residual correction map with the same spatial size as the initial monthly surface temperature field. The initial monthly surface temperature field is corrected using the single-channel residual correction map. The residual correction result is superimposed pixel by pixel with the initial monthly land surface temperature field to obtain the refined monthly land surface temperature data.

5. A high-resolution monthly surface temperature reconstruction device, characterized in that, The device includes: The multi-source data acquisition unit is used to uniformly process and control the spatial grid of various initial surface temperature data to obtain candidate surface temperature data. The initial surface temperature data includes monthly average continuous surface temperature data and high-resolution satellite surface temperature data. The monthly reference field creation unit is used to perform harmonic fitting using the annual cycle data in the candidate surface temperature data to generate a continuous daily resolution reference field and a monthly reference field. The fine-scale residual estimation unit is used to extract and aggregate fine-scale residuals using high-resolution satellite observation data from the candidate surface temperature data, and to determine the fine-scale residual estimate. An initial temperature field generation unit is used to superimpose the monthly reference field with the fine-scale residual estimate to generate an initial monthly surface temperature field. The time window setting unit is used to construct a time window centered on the target month, which includes the target month and several adjacent months before and after it. A multi-channel input setting unit is used to construct a multi-channel input representation with the same channel structure for each time step within the time window, using the initial surface temperature data, the monthly reference field, and the fine-scale residual estimation. The feature extraction and aggregation unit is used to perform spatial feature extraction and interpretable temporal aggregation on the multi-channel input representation for each time step within the time window, and generate fused spatiotemporal features for the target month. The refinement result output unit is used to perform fusion correction on the initial monthly land surface temperature field using the fusion spatiotemporal features to generate refined monthly land surface temperature data. The step of using high-resolution satellite observation data from the candidate surface temperature data to extract and aggregate fine-scale residuals and determine fine-scale residual estimates includes: calculating the difference between the daily satellite observation temperature and the daily harmonic reference value pixel by pixel based on the high-resolution satellite observation data and the continuous daily resolution reference field to generate daily fine-scale thermal spatial deviations; and statistically aggregating the daily fine-scale thermal spatial deviations within the annual cycle to determine the fine-scale residual estimates. The step of performing spatial feature extraction and interpretable temporal series aggregation on the multi-channel input representation for each time step within the time window to generate fused spatiotemporal features for the target month includes: extracting features from the multi-channel input representation using a parameter-shared encoder-decoder network to determine the spatial feature map for each time step; and using an interpretable temporal series aggregation module based on the Kolmogorov-Arnold theorem to perform weighted fusion of the spatial feature maps for each time step within the time window to obtain the fused spatiotemporal features. The method employs an interpretable temporal aggregation module based on the Kolmogorov-Arnold theorem to perform weighted fusion of the spatial feature maps of each time step within the time window to obtain the fused spatiotemporal features. This includes: calculating the time offset of each reference time step relative to the target month within the time window; using the time offset as the sole input, calculating the scalar temporal weights corresponding to each reference time step through the KAN temporal weight mapping function parameterized by cubic B-spline basis functions; normalizing the scalar temporal weights along the time dimension to obtain normalized weights; and performing element-wise weighted summation of the normalized weights and the spatial feature maps of the corresponding time steps to obtain the fused spatiotemporal features.

6. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory being used to store a computer program that, when executed by the processor, implements the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 4.

Images