A super-resolution image reconstruction method and system for radiative transfer simulation

Abstract

This invention discloses a super-resolution image reconstruction method and system for radiative transfer simulation. The method includes: constructing input features and labels through strict spatiotemporal pairing and multi-source data fusion; employing a residual learning network based on Unet++ to learn the residual from the coarse-resolution baseline brightness temperature to the pixel-level resolution target brightness temperature. The network undergoes multi-task learning modification and is trained within an effective mask using a combined loss function of SmoothL1 and SSIM to simultaneously constrain amplitude and structural errors; finally, systematic biases are eliminated through validation set-driven linear piecewise calibration. This invention overcomes the problems of over-smoothing, physical inconsistencies, and low computational efficiency inherent in traditional interpolation methods for radiative transfer simulation, achieving high-precision and high-efficiency super-resolution reconstruction and meeting the needs of near-real-time quantitative inversion of massive amounts of spaceborne data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of satellite remote sensing and high-performance computing, and particularly relates to a super-resolution image reconstruction method and system for radiation transfer simulation. BACKGROUND

[0002] Forward simulation using radiation transfer models (such as RTTOV, CRTM) is the core of satellite quantitative remote sensing. These models need to input regular latitude and longitude gridded data provided by numerical weather prediction models, and the original coarse spatial resolution limits the spatial resolution of the simulation results.

[0003] To improve simulation accuracy, the prior art usually uses nearest neighbor sampling or bilinear interpolation, inverse distance weighted interpolation and other methods to reconstruct input parameters for each satellite pixel from the background field. However, the nearest neighbor sampling method has low accuracy and is prone to deviation; while various interpolation methods can improve accuracy, their inherent smoothing effect will weaken boundaries and extreme values, resulting in structural errors at sea-land edges and other places, and without physical constraints may lead to physical inconsistency. In addition, these methods have high computational complexity, significantly increasing resource consumption and processing time, making it difficult to balance accuracy and efficiency when processing high spatio-temporal resolution massive data.

[0004] Therefore, there is an urgent need in the art for a method that can effectively improve the spatial resolution and accuracy of radiation transfer simulation images without significantly increasing computational overhead. SUMMARY

[0005] In view of the defects in the prior art, the present application provides a super-resolution image reconstruction method for radiation transfer simulation, comprising the following steps:

[0006] Constructing a residual learning network configured to learn the residual Δy between the baseline prediction y corresponding to the multi-channel input X and the multi-channel label Y, wherein the multi-channel label Y is the radiation transfer simulation result at the target satellite observation pixel level resolution, the baseline prediction y is obtained by processing the radiation transfer model forward result of the original background field spatial resolution; base base

[0007] The residual learning network is a network based on Unet++ architecture and modified by multi-task learning, which has a classification output head connected in parallel at the decoding end, shares the encoder and main decoding features with the regression residual head, is used to perform an abnormal pixel detection task, and outputs spatial attention weights to adjust the loss weight of residual regression;

[0008] ​​The residual learning network is trained using a combined loss function, the combined loss function including a SmoothL1 loss for constraining the amplitude error and a structural similarity, SSIM, loss for constraining the structural error;

[0009] The input data is processed using the trained residual learning network to reconstruct a super-resolution radiative transfer simulation image.

[0010] Before constructing and training the residual learning network, the data preparation step is further included:

[0011] The radiative transfer model forward file of the target satellite observation field resolution label time and the original background field spatial resolution is strictly time-stamped matched;

[0012] Under the unified latitude and longitude network reference, the geographic information file, the original background field spatial resolution forward brightness temperature, the ground surface emissivity file and the multi-source sensor zenith angle are spatially registered and sampled to the target satellite observation pixel to assemble the multi-channel input X and the multi-channel label Y.

[0013] After the data preparation step, the data quality control step is further included:

[0014] The input validity validX and the label validity validY are calculated for each target satellite observation pixel, and a pixel-level validity mask mask is formed;

[0015] When the data set slice is divided, only the slice with an average mask value not less than 0.8 is retained for training or evaluation.

[0016] The residual learning network is constructed based on the Unet++ architecture, and the encoder of the Unet++ network adopts the ResNet34 structure.

[0017] The residual learning network is a Unet++ network modified by multi-task learning, which has a classification output head connected in parallel at the decoding end, shares the encoder and the main decoding feature with the regression residual head, and only branches at the last layer.

[0018] The form of the combined loss function is: , wherein L SmoothL1 is the SmoothL1 loss, SSIM is the structural similarity index, and α and γ are weight coefficients; and the loss function is calculated within the range defined by the pixel validity mask mask.

[0019] Before the residual learning step, the normalization step is further included: the statistical information of the input and the output is calculated with the training set, the multi-channel input X and the multi-channel label Y are respectively standardized, and the baseline prediction ybase Standardize using the same scale with output statistics.

[0020] The process includes a post-calibration step after inference using the trained network: linear piecewise fitting calibration is performed on the prediction results on the validation set, multiple intervals are defined with the satellite zenith angle as the independent variable, linear parameters are fitted in each interval, and the corresponding parameters are selected for calibration based on the pixel value segment during inference.

[0021] In the post-calibration step, parameter a is fitted within the k-th interval. k and b k The calibrated output is: During reasoning, linear interpolation is used between segments to ensure continuity.

[0022] This invention also proposes a super-resolution image reconstruction system for radiative transfer simulation, comprising:

[0023] A residual learning network module is used to construct a residual learning network, which is configured to learn a baseline prediction y corresponding to a multi-channel input X. base The residual Δy between the multi-channel label Y and the target satellite observation pixel-level resolution radiative transfer simulation result, and the baseline prediction Δy. base The results were obtained by processing the forward modeling results of the radiative transfer model with spatial resolution of the original background field;

[0024] The training module is used to train the residual learning network using a combined loss function, which includes a SmoothL1 loss for constraining magnitude error and a Structural Similarity (SSIM) loss for constraining structural error.

[0025] The reconstruction module is used to process the input data using the trained residual learning network to reconstruct a super-resolution radiative transfer simulation image.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] High precision. Through the residual learning framework, the network does not need to directly regress complex absolute brightness temperature values, but instead learns residuals that are relatively easy to model, reducing the learning difficulty. Combined with structure-aware combined loss (SmoothL1 + SSIM), it effectively suppresses interpolation artifacts and structural biases, and the reconstructed image is more accurate in both amplitude and local texture.

[0028] High efficiency. This method requires only one neural network forward propagation and simple residual reconstruction during the inference phase. Compared with traditional methods that require complex pixel-by-pixel interpolation operations, the computational overhead is significantly reduced and the processing speed is greatly improved, which can meet the requirements of near real-time applications.

[0029] Physical consistency and robustness. The scheme uses the output of the physical model (RTTOV) as a baseline, inheriting the interpretability of the physical model; through an effective masking mechanism and validation set-driven post-calibration, invalid data is systematically filtered out and systematic biases are eliminated, enhancing the model's generalization ability and the reliability of the output results under different regions and observation conditions. Attached Figure Description

[0030] The above and other objects, features, and advantages of exemplary embodiments of the present disclosure will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the present disclosure are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein:

[0031] Figure 1 This is an overall flowchart illustrating the super-resolution image reconstruction method provided according to an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the architecture of the Unet++ residual learning network used in an embodiment of the present invention; and

[0033] Figure 3 This is a block diagram illustrating the structure of the Unet++ residual learning network used in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0035] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0036] It should be understood that although the terms first, second, third, etc., may be used to describe... in the embodiments of the present invention, these... should not be limited to these terms. These terms are only used to distinguish... For example, first... may also be referred to as second... without departing from the scope of the embodiments of the present invention, and similarly, second... may also be referred to as first...

[0037] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0038] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0039] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0040] The optional embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] Example 1

[0042] This invention discloses a super-resolution image reconstruction method based on radiative transfer simulation, comprising the following steps:

[0043] Data preparation and alignment steps: The target satellite observation field resolution label time number and the original background field spatial resolution radiative transfer model forward model file are strictly matched with the same timestamp. Under the unified latitude and longitude grid reference, the geographic information file, the original background field spatial resolution RTTOV forward model brightness temperature, the surface emissivity file and the multi-source sensor zenith angle are spatially registered, sampled to the target satellite observation pixels, and assembled into multi-channel input X and multi-channel label Y.

[0044] Residual learning steps: The multi-channel input X is fed into a residual learning network based on the Unet++ architecture, and the residual learning network learns the baseline prediction y corresponding to the multi-channel input X. base The standardized residual Δy between the multi-channel label Y and the multi-channel label Y, where Δy = y - y base The baseline prediction y baseThe brightness temperature is obtained by interpolating the original background field spatial resolution RTTOV forward modeling to the target satellite observation pixels;

[0045] Model training steps: Within the range defined by the effective pixel mask, the residual learning network is trained using a combination loss function of SmoothL1 loss and structural similarity SSIM loss. The effective pixel mask is formed by performing a logical AND operation between the input validity validX and the label validity validY.

[0046] Post-calibration step: Perform linear piecewise fitting calibration on the prediction results of the residual learning network on the validation set. Define multiple intervals with the satellite zenith angle as the independent variable, fit linear parameters in each interval, and select the corresponding parameters for calibration according to the pixel value segment during inference. Output the reconstructed super-resolution radiative transfer simulation image.

[0047] The data preparation and alignment steps also include a sample screening step: calculating the scene-level effective coverage rate after sampling the original background field spatial resolution to the target satellite pixel-level resolution infrared channel interpolation; if the effective coverage rate is less than 80%, the sample at that time is removed.

[0048] The process includes a slicing step after the data preparation and alignment step and before the residual learning step: non-overlapping slices are made on the original target satellite observation pixel grid at a step size of 256×256 pixels, and the multi-channel input X, the multi-channel label Y, and the effective pixel mask are written into the HDF5 slice.

[0049] In the residual learning step, the residual learning network is a Unet++ network modified for multi-task learning. It has a classification output head connected in parallel at the decoding end, shares the encoder and main decoding features with the regression residual head, and only branches at the last layer.

[0050] The encoder of the Unet++ network adopts the ResNet34 structure.

[0051] In the model training step, the combined loss function takes the following form: , where L SmoothL1 SSIM is the SmoothL1 loss function, SSIM is the structural similarity index, and α and γ are weighting coefficients.

[0052] In the model training step, only slices with an average mask value of not less than 0.8 are retained for training or evaluation when the dataset is sliced; during the training process, if any channel of a slice has a NaN value or the effective mask of the pixel is all 0, the online filtering will not participate in the current iteration.

[0053] Prior to the residual learning step, a normalization step is included: calculating statistical information of the input and output using the training set, standardizing the multi-channel input X and the multi-channel label Y respectively, and normalizing the baseline prediction y. base Standardize using the same scale with output statistics.

[0054] In the post-calibration step, parameter a is fitted within the k-th interval. k and b k The calibrated output is: During reasoning, linear interpolation is used between segments to ensure continuity.

[0055] Example 2

[0056] One embodiment of the present invention provides a super-resolution image reconstruction method for radiative transfer simulation. For example... Figure 1 As shown, the specific implementation steps of this method are as follows:

[0057] S101: Strict Time Matching and Sample Screening. The labeled time periods of the target satellite observation field (e.g., 2km resolution) are strictly matched with the RTTOV forward modeling files of the original background field (e.g., 25km resolution). Verification is performed to ensure the simultaneous existence of corresponding satellite observation geographic information data (GEO files) and the static surface emissivity file for the current month. Subsequently, the effective pixel coverage of the scene is calculated after interpolating the coarse-resolution background field to satellite pixel-level resolution. If the coverage is below 80%, samples from that entire time period are discarded to ensure the quality and representativeness of the training data.

[0058] S102: Spatial Alignment and Feature Construction. All input data are spatially registered under a unified latitude and longitude grid reference. Specifically, GEO files are directly sampled to target satellite pixels according to latitude and longitude; coarse-resolution RTTOV forward modeled brightness temperatures are bilinearly interpolated to target satellite pixels using regular axes, and this result serves as the baseline data y_base; the surface emissivity file is similarly interpolated to target satellite pixels; and the zenith angle from multiple sensor sources is downsampled to target satellite pixels. Finally, these processed data are assembled into a multi-channel (e.g., 8-channel) input feature map X, and the ground truth brightness temperature observed by the satellite is used as the multi-channel (e.g., 3-channel) label Y.

[0059] S103: Pixel Validity Check and Mask Generation. For each satellite pixel, calculate the validity (validX and validY) of its input X and label Y, for example, checking if the data contains NaN or is outside a reasonable range. Perform a logical AND operation on validX and validY to generate the final pixel-level valid mask. When subsequently cutting the data into patches, only patches with an average mask value (mean(mask)) ≥ 0.8 are retained for training and evaluation. During training, if any channel in a patch contains a NaN value or the mask is all 0, the patch is filtered out online and does not participate in the current training iteration to ensure the stability of gradient calculation.

[0060] S104: Slicing and Piece Loading. The entire scene's input X, label Y, and mask are sliced ​​non-overlapping into 256×256 pixel slices with a stride of 256. The slice data, along with their corresponding timestamps, starting row and column numbers, are written into an HDF5 format slice file for efficient parallel data loading and management later.

[0061] S105: Normalized Statistics and Baseline Construction. Calculate the mean (μ) of the input X and label Y using all samples in the training set. x, μ y ) and standard deviation (σ) x , σ y During training and inference, the input and output data are standardized: X norm =(X - μ x ) / σ x Y norm = (Y - μ y ) / σ y Simultaneously, the baseline data y obtained in step S102 is... base Standardization is also performed using statistics for label Y: y base _ norm = (y base - μ y ) / σ y .

[0062] S106: Residual Learning Network Construction and Training. A residual learning network with Unet++ as the backbone and ResNet34 as the encoder is constructed. The network input is the standardized X. norm The goal is to learn the standardized residual: Δy norm = Y norm -y base _ norm .like Figure 2As shown, the standard Unet++ is modified by adding a classification output head in parallel at its decoding end. This output head shares encoder and most decoder features with the original regression residual head, forming a multi-task learning architecture to improve feature discrimination capability. The network training uses the AdamW optimizer, and the learning rate is scheduled using a cosine annealing strategy.

[0063] like Figure 3 As shown, a pixel-level classification branch is connected in parallel at the decoding end of the Unet++ residual learning backbone. The two branches share the main features extracted by the backbone network. The residual regression branch is used to output the three-channel brightness-temperature residuals. The classification branch results, after being mapped to single-layer spatial weights A through a small convolution, are used to output a probability map of special class pixels (such as the land-sea boundary). The classification loss function L... cls Using cross-entropy can effectively address class imbalance.

[0064]

[0065] CE is the cross-entropy calculation, N is the total number of positions, c is the shared parameter between the two branches of a certain category, and the gradient of the loss function flows back to the backbone network together, which improves the stability of the backbone for RTTOV residual prediction in special scenarios.

[0066] S107: Combined Loss Calculation. The loss function is calculated only within the pixel region where the effective mask = 1. The total loss consists of two weighted components: a SmoothL1 loss, used to robustly constrain pixel value errors; and a 1-SSIM loss, used to constrain the similarity of local image structures (brightness, contrast, and structure). The total residual regression loss is... Through multi-task learning, classification loss L will also be incorporated. cls The total loss at this point is The classification output head is connected in parallel at the Unet++ decoding end, sharing encoder and decoding features with the regression residual head to form a multi-task learning mechanism; the classification head is used for abnormal pixel detection (such as cloud detection and surface type classification), and the output is used as spatial attention weights to adjust the loss weights of residual regression; the classification head is trained through cross-entropy loss to improve the model's ability to discriminate complex scenes.

[0067] To adapt to the physical characteristics of different channels and to impose directional constraints on thermal boundaries in radiometric images, this invention introduces a joint optimization strategy of adaptive channel uncertainty weights and gradient difference loss within the effective pixel mask.

[0068] Gradient difference loss: The spatial first derivative of the predicted brightness temperature is compared pixel by pixel to maintain the boundary of the temperature front with equal gradient strength. Let channel c be (3.8, 10.8, 12µm) and pixel coordinates be (i,j).

[0069]

[0070] Where M is the effective pixel mask. It is the first derivative.

[0071] Channel uncertainty adaptive weights: Introduce learnable adaptive weights σ for each channel. c The network automatically learns the relative weights of each channel through backpropagation to uniformly scale the amplitude loss and structural loss of that channel, and adds a regularization term to prevent degradation.

[0072] The total loss function is now:

[0073]

[0074] Where L reg The internal α, γ, and λ are set to initial reference ratios (0.4, 0.2, 0.2), and σ is σ. 3.8 σ 10.8 σ 12 , λ is the regularization term, and λ is the weight of the classification branch, set to 0.2.

[0075] S108: Validation Set-Driven Post-Calibration. After training, the network output is calibrated on the validation set. Using the satellite zenith angle as the independent variable, the set is divided into several intervals. Within each interval, the network's predicted value Y is linearly regressed against the true value to obtain the calibration parameter a for that interval. k and b k During inference, the interval to which each pixel belongs is determined based on its zenith angle value, and the corresponding parameters are applied for calibration. If the pixel value falls between two intervals, linear interpolation can be used to determine the parameters to ensure the continuity of the output.

[0076] The satellite zenith angle determines the observation path length and absorption / scattering intensity, and is the main driving factor for brightness temperature system bias. Within physically similar intervals, the relationship between predicted and true values ​​can be sufficiently approximated as linear. After the main model training is completed, this invention performs adaptive segmentation on the validation set, using the residual variance change points as boundaries. Within each segment, linear regression fitting is applied, applying only affine correction to the amplitude, without changing the spatial structure and relative gradient, and without introducing new geometric or texture artifacts. Within the linearly segmented intervals, a k This can be viewed as the model output's local sensitivity (equivalent transmittance) to the true brightness temperature, b kThe system bias acting on the brightness temperature domain, which is approximated by the combined effects of upward atmospheric radiation, calibration error, and residual emissivity error (not fully fitted by the model), is estimated piecewise with respect to the satellite zenith angle and linearly interpolated between segments to eliminate the cold / warm bias. The segmentation is based on the satellite zenith angle, as it directly affects atmospheric path length and absorption intensity. An adaptive segmentation strategy is proposed, using the residual variance change point as the boundary to improve calibration accuracy.

[0077] S109: Result Output and Evaluation. The inference results are de-standardized to obtain the final reconstructed brightness temperature image. The reconstructed image is output, and metrics such as mean absolute error (MAE), root mean square error (RMSE), bias, and correlation coefficient are calculated to quantitatively evaluate the reconstruction effect.

[0078] Through the above steps, the present invention achieves high-precision and high-efficiency super-resolution reconstruction of radiative transfer simulation brightness temperature images.

[0079] Example 3

[0080] This invention also proposes a super-resolution image reconstruction system for radiative transfer simulation, comprising:

[0081] A residual learning network module is used to construct a residual learning network, which is configured to learn a baseline prediction y corresponding to a multi-channel input X. base The residual Δy between the multi-channel label Y and the target satellite observation pixel-level resolution radiative transfer simulation result, and the baseline prediction Δy. base The results were obtained by processing the forward modeling results of the radiative transfer model with spatial resolution of the original background field;

[0082] The training module is used to train the residual learning network using a combined loss function, which includes a SmoothL1 loss for constraining magnitude error and a Structural Similarity (SSIM) loss for constraining structural error.

[0083] The reconstruction module is used to process the input data using the trained residual learning network to reconstruct a super-resolution radiative transfer simulation image.

[0084] Example 4

[0085] This disclosure provides a non-volatile computer storage medium storing computer-executable instructions that can perform the steps described in the above embodiments.

[0086] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0087] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0088] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (AN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0089] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0090] The units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0091] The preferred embodiments of the present invention have been described above to make the spirit of the present invention clearer and easier to understand, and are not intended to limit the present invention. All modifications, substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope summarized by the appended claims.

Claims

1. A super-resolution image reconstruction method for radiative transfer simulation, characterized in that, Includes the following steps: A residual learning network is constructed, which is configured to learn the baseline prediction y corresponding to the multi-channel input X. base The residual Δy between the multi-channel label Y and the target satellite observation pixel-level resolution radiative transfer simulation result, and the baseline prediction Δy. base The results were obtained by processing the forward modeling results of the radiative transfer model with spatial resolution of the original background field; The residual learning network is a network based on the Unet++ architecture and modified by multi-task learning. It has a classification output head connected in parallel at the decoding end, which shares the encoder and main decoding features with the regression residual head. It is used to perform the abnormal pixel detection task and outputs spatial attention weights to adjust the loss weights of residual regression. The residual learning network is trained using a combined loss function, which includes SmoothL1 loss for constraining magnitude error, SSIM loss for constraining structural error, and gradient difference loss for constraining spatial gradient consistency. The trained residual learning network is used to process the input data in order to reconstruct a super-resolution radiative transfer simulation image. After using the trained network for inference, a post-calibration step is also included: linear piecewise fitting calibration is performed on the prediction results on the validation set, multiple intervals are defined with the satellite zenith angle as the independent variable, linear parameters are fitted in each interval, and the corresponding parameters are selected for calibration according to the pixel value segment during inference.

2. The method as described in claim 1, characterized in that, Before constructing and training the residual learning network, a data preparation step is also included: Strictly match the timestamps of the target satellite observation field resolution label with the forward modeling file of the radiative transfer model of the original background field spatial resolution; Under a unified latitude and longitude grid benchmark, the geographic information file, the original background field spatial resolution forward modeling brightness temperature, the surface emissivity file, and the zenith angle of the multi-source sensor are spatially registered and sampled to the target satellite observation pixels to assemble the multi-channel input X and the multi-channel label Y.

3. The method as described in claim 2, characterized in that, Following the data preparation steps, a data quality control step is also included: For each target satellite observation pixel, calculate the input validity validX and label validity validY, and form a pixel-level valid mask. When segmenting the dataset, only slices with an average mask value of not less than 0.8 are retained for training or evaluation.

4. The method as described in claim 1, characterized in that, The residual learning network is built on the Unet++ architecture, and the encoder of the Unet++ network adopts the ResNet34 structure.

5. The method as described in claim 4, characterized in that, The residual learning network is a Unet++ network modified for multi-task learning. It has a classification output head connected in parallel at the decoding end, which shares the encoder and main decoding features with the regression residual head, and only branches at the last layer.

6. The method as described in claim 1, characterized in that, The combined loss function is in the form of: , where L SmoothL1 The loss function is SmoothL1, where SSIM is the structural similarity index, and α and γ are weighting coefficients. Furthermore, the loss function is calculated within the range defined by the effective pixel mask.

7. The method as described in claim 1, characterized in that, Before the residual learning step, a normalization step is also included: calculating the statistical information of the input and output using the training set, standardizing the multi-channel input X and the multi-channel label Y respectively, and normalizing the baseline prediction y. base Standardize using the same scale with output statistics.

8. The method as described in claim 1, characterized in that, In the post-calibration step, parameter a is fitted within the k-th interval. k and b k The calibrated output is: During reasoning, linear interpolation is used between segments to ensure continuity.

9. A super-resolution image reconstruction system for radiative transfer simulation, comprising: A residual learning network module is used to construct a residual learning network, which is configured to learn a baseline prediction y corresponding to a multi-channel input X. base The residual Δy between the multi-channel label Y and the target satellite observation pixel-level resolution radiative transfer simulation result, and the baseline prediction Δy. base The results were obtained by processing the forward modeling results of the radiative transfer model with spatial resolution of the original background field; The training module is used to train the residual learning network using a combined loss function, which includes a SmoothL1 loss to constrain magnitude error and a SSIM loss to constrain structural error. After inference using the trained network, a post-calibration step is also included: linear piecewise fitting calibration is performed on the prediction results on the validation set, multiple intervals are defined with the satellite zenith angle as the independent variable, linear parameters are fitted in each interval, and the corresponding parameters are selected for calibration according to the pixel value segment during inference. The reconstruction module is used to process the input data using the trained residual learning network to reconstruct a super-resolution radiative transfer simulation image.

Images