A multi-image super-resolution imaging method for Antarctic passive microwave brightness temperature data

Abstract

The application relates to a multi-image super-resolution reconstruction method for passive microwave brightness temperature of the South Pole, comprising the following steps: collecting microwave low-resolution brightness temperature images of continuous frames in the same space South Pole coverage area, and constructing a low-resolution brightness temperature image sequence; performing brightness temperature feature supplementing on the multi-time low-resolution brightness temperature image sequence based on a local time window, and obtaining local time enhanced features; wherein the local time window is centered on the brightness temperature image of the current frame, and the brightness temperature images of a certain number of adjacent frames before and after are combined; introducing a bidirectional time propagation mechanism, globally modeling the local time enhanced features, and obtaining feature representation after global time modeling; and performing super-resolution reconstruction on the brightness temperature image based on a residual learning mechanism. Compared with the prior art, the application can improve the spatial resolution while maintaining the brightness temperature time consistency and physical rationality, so that the melting process of the South Pole ice sheet and ice shelf surface can be finely monitored.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a multi-image super-resolution imaging method for Antarctic passive microwave brightness temperature data. Background Technology

[0002] Against the backdrop of global warming, the combined effects of rising near-surface temperatures and large-scale atmospheric circulation anomalies have significantly accelerated the melting process of the Antarctic ice sheet and ice shelves. Ice shelf melting not only directly affects the mass balance of the ice sheet-ice shelf system but may also weaken the structural stability of ice shelves through mechanisms such as meltwater infiltration and hydraulic fracturing, inducing ice shelf breakup and accelerating upstream glacial flow, thus having a profound impact on global sea-level changes. Therefore, continuous, detailed, and reliable monitoring of the melting process and its spatiotemporal variations of the Antarctic ice sheet and ice shelves is a crucial technical foundation for polar climate change research and ice sheet dynamics analysis.

[0003] Due to the extreme climate and harsh observation conditions in Antarctica, ground-based measurement methods are severely limited in terms of spatial coverage and temporal continuity. Remote sensing has become the primary technology for monitoring the melting of the Antarctic ice sheet and ice shelves. Existing technologies mainly rely on data from various sensors, including optical / infrared remote sensing, synthetic aperture radar (SAR), and passive microwave remote sensing. Optical and infrared remote sensing can relatively intuitively identify wet areas and meltwater distribution on the ice surface under clear sky conditions, but their observations are significantly affected by polar nights and cloud cover, making it difficult to achieve continuous monitoring throughout the year and in all weather conditions. SAR remote sensing has high spatial resolution and all-weather observation capabilities, but its backscattered signals are easily affected by factors such as changes in the incident angle, surface roughness, and snow layer structure, resulting in some uncertainty in the interpretation of melting information.

[0004] In contrast, passive microwave remote sensing is unaffected by illumination and cloud cover, offering advantages such as wide coverage, high revisit frequency, and high sensitivity to liquid water content, and has been widely used for monitoring melting of polar ice and snow surfaces. However, limited by microwave radiometer antenna size and observation geometry, the spatial resolution of passive microwave remote sensing data is typically low; for example, the spatial resolution of commonly used frequency bands is usually between 12.5 and 25 kilometers. This low spatial resolution makes it difficult to characterize the fine-scale spatial heterogeneity of ice shelf surface melting, easily leading to significant mixed pixel effects, thus limiting its accuracy in regional-scale ice sheet and ice shelf process studies and multi-source remote sensing collaborative analysis.

[0005] To improve the spatial resolution of passive microwave remote sensing data, existing technologies mainly fall into two categories: analytical model-based methods and data-driven deep learning methods.

[0006] One category is spatial resolution enhancement methods based on analytical models. These methods typically use the antenna directivity gain function or observation footprint model of a microwave radiometer as prior information, and utilize the spatial overlap relationship between different observation footprints to perform weighted inversion or reconstruction of low-resolution brightness temperature observations. Representative methods include the Backus–Gilbert (BG) inversion method and the Scatterometer / Radiometer Image Reconstruction (SIR) imaging method. Theoretically, these methods can achieve a certain degree of spatial resolution improvement while maintaining radiometric consistency. However, their performance is highly dependent on the accurate characterization of antenna modes and observation geometry. In practical applications, they often have to be approximated using empirical models, which easily introduces radiometric bias and edge blurring. At the same time, these methods mainly rely on the spatial overlap information between single-scan observations, making them difficult to directly apply to time-averaged gridded brightness temperature products, and they also have certain limitations in terms of resolution improvement and model generalization ability.

[0007] Another category is data-driven super-resolution reconstruction methods, especially deep learning-based super-resolution techniques. These methods construct end-to-end neural network models to learn the nonlinear mapping relationship between low-resolution and high-resolution brightness-temperature images, thereby achieving spatial resolution enhancement without explicitly relying on antenna directivity functions and observation geometry parameters. Based on the amount of input observation information, existing methods can be further divided into single-image super-resolution methods and multi-image super-resolution methods. Single-image super-resolution methods utilize only single-temporal low-resolution observations for reconstruction, simplifying the model structure to some extent. However, they struggle to effectively mitigate the mixed-pixel effect under complex surface conditions and have limited robustness to temporal variations and anomalous observations. Multi-image super-resolution methods introduce multi-temporal observation information, utilizing redundant features in the temporal dimension to improve reconstruction quality. However, in practical applications, they still face challenges such as difficulty in feature alignment, significant interference from anomalous frames, and insufficient physical consistency constraints.

[0008] Especially in the Antarctic ice sheet and ice shelf region, due to the complex surface types, the predominance of low-frequency components in brightness temperature temporal changes, and significant noise interference, existing super-resolution methods still have shortcomings in terms of maintaining temporal consistency, suppressing local anomalies, and constraining physical rationality.

[0009] Therefore, there is an urgent need for a passive microwave brightness temperature super-resolution method that can collaboratively model in both time and space dimensions, and take into account both high-resolution reconstruction effect and physical consistency, in order to meet the practical needs of fine monitoring of the melting of Antarctic ice sheet and ice shelf surfaces. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a multi-image super-resolution imaging method for Antarctic passive microwave brightness temperature data.

[0011] The objective of this invention can be achieved through the following technical solutions: A multi-image super-resolution reconstruction method for Antarctica based on passive microwave brightness temperature includes: S1. Collect consecutive frames of microwave low-resolution brightness temperature images within the same space Antarctic coverage area and construct a low-resolution brightness temperature image sequence. S2. The multi-temporal low-resolution brightness temperature image sequence is supplemented with brightness temperature features based on a local temporal window to obtain local temporal enhancement features; wherein, the local temporal window is obtained by combining the brightness temperature images of a number of adjacent frames before and after, with the brightness temperature image of the current frame as the center. S3. Introduce a bidirectional temporal propagation mechanism to perform global modeling on the local temporal enhancement features, and obtain the feature representation after global temporal modeling; S4. Based on the residual learning mechanism, perform super-resolution reconstruction of brightness temperature images.

[0012] Preferably, in step S1, acquiring consecutive frames of microwave low-resolution brightness temperature images within the same Antarctic coverage area to construct a low-resolution brightness temperature image sequence specifically includes: Obtain continuous coverage within the same spatial area Passive microwave low-resolution brightness temperature images of each time phase were used to construct a multi-time series of low-resolution brightness temperature images. ,in, This indicates the time frame number of the brightness temperature sequence. Indicates the first Frame brightness temperature image, It is a single-channel grayscale image, and its pixel value represents the brightness temperature intensity of the corresponding ground surface.

[0013] Preferably, in S2, the multi-temporal low-resolution brightness temperature image sequence Brightness-temperature feature supplementation based on local temporal windows is performed, specifically including: S21. Feature supplementation is performed on the brightness temperature image of a single frame using a local temporal window, for any time index. , with the first Frame brightness temperature image Construct a local time window centered on the time window. = ; S22, Set the local timing window A convolutional feature extraction network with shared parameters is used to extract brightness temperature features from each frame of brightness temperature image in the image. in, Indicated in scale Feature extraction operators; S23. Introduce a cascaded deformable convolution mechanism to perform layer-by-layer alignment of non-reference frames and obtain spatial alignment features. ; S24. Introduce a radiation similarity guidance mechanism based on normalized cross-correlation to adaptively measure and adjust the contribution weight of each time frame to the central reference frame, thus obtaining the temporal weight. ; S25. Based on spatial alignment features With the corresponding time-domain radiation weight Preliminary local fusion features are extracted using a weighted aggregation operator. , is represented as: ; S26. Introduce a temporal-spatial joint attention feature enhancement mechanism to improve the initial local fusion features. Enhancement is performed to obtain local temporal enhancement features. .

[0014] Preferably, in step S23, a cascaded deformable convolution mechanism is introduced to perform layer-by-layer alignment of non-reference frames to obtain spatial alignment features. Specifically, it includes: At every scale The offset field is estimated using the fusion features of the reference frame and adjacent frames, as follows: ,in, Representing scale The offset prediction network is constructed from the corresponding convolutional layers; Using the cascaded deformable convolution operator to perform a spatial transformation in the feature space, aligning the feature vectors of neighboring frames to the reference frame, it can be represented as: ,in, This represents the features of non-center frames within a local time window. This represents the features of the center frame used as a reference. Indicates the center position of the convolution. Indicates the offset of standard convolution sampling points. A learnable spatial offset. For convolution weights, This represents the number of sampling points; Obtain spatial alignment features : .

[0015] Preferably, step S24 introduces a radiation similarity guidance mechanism based on normalized cross-correlation to adaptively measure and adjust the contribution weight of each time frame to the central reference frame, specifically including: 1) Perform average pooling on the input window to obtain a low-frequency brightness temperature image. ; 2) The normalized cross-correlation operator is used to measure the similarity between adjacent frames and the reference frame in the low-frequency brightness temperature distribution. For two given low-frequency brightness temperature images... and Calculate low-frequency normalized cross-correlation similarity : , in, Indicates spatial location index, and Low-frequency brightness temperature images and Mean pixel intensity in the spatial domain For constant terms; 3) Introduce the Softmax operator to normalize the cross-correlation similarity of low frequencies. Mapping to the probability space yields the temporal weights. : Preferably, the time-series weights , is represented as: , in, This is a temperature coefficient used to adjust the discriminative power or smoothness of the weight distribution.

[0016] Preferably, a temporal-spatial joint attention feature enhancement mechanism is introduced in S26 to enhance the initial local fusion features. Enhancement is performed to obtain local temporal enhancement features, specifically including: 1) Temporal correlation attention modeling: Preliminary local fusion features corresponding to the central reference frame Using temporal anchors, features are mapped to a low-dimensional embedding space through a convolutional embedding operator with shared parameters, and inter-frame correlations are calculated, represented as: , ,in, and It is a convolution mapping operator; The embedded features are element-wise multiplied and summed along the channel dimension to obtain the temporal correlation response between adjacent frames and the center frame. , is represented as: ,in, This represents the c-th channel; Using the Sigmoid activation function to correlate responses Mapped to normalized temporal attention weights ; The obtained normalized temporal attention weights The corresponding frame features are applied pixel-by-pixel. After temporal correlation attention weighting, all temporal frame features are concatenated along the channel dimension and a channel compression operator is used to obtain the fused features. , is represented as: ,in, for Convolution operator, Represents element-wise product; 2) Multi-scale spatial attention modeling: Fusion features After pooling and convolution operations to generate a coarse-scale spatial response, it is upsampled stepwise and fused with shallow spatial attention features to obtain a spatial attention map. , is represented as: ,in, This represents a spatial attention generation operator consisting of multiple convolutional, pooling, and upsampling operations. 3) Joint attention feature enhancement output: utilizing the obtained spatial attention Mapping pairs of fused features Element-wise modulation is performed, and a residual enhancement term is introduced to obtain local temporal enhancement features. , is represented as: ,in, This is the convolution mapping operator.

[0017] Preferably, a bidirectional temporal propagation mechanism is introduced in S3 to enhance the local temporal features. Perform global modeling to obtain the feature representation after global temporal modeling. Specifically, it includes: 1) Backward time propagation: Starting from the end of the sequence, the time-reverse cluster update propagation features are as follows: ,in, This represents the backpropagation feature update operator, which internally consists of... Implemented by combining convolutional layers and gating units; 2) Forward temporal propagation: Based on the fusion of backward propagation features, the propagation features are updated from the temporal start frame to the last frame as follows: ,in, This represents the forward propagation feature update operator, which internally consists of... Implemented by combining convolutional layers and gating units; 3) By fusing forward ship features and backpropagation features, a feature representation after global temporal modeling is obtained: , in, for Convolutional layers are used to compress channels and fuse bidirectional spatiotemporal information.

[0018] Preferably, in step S4, super-resolution reconstruction of the brightness temperature image based on a residual learning mechanism specifically includes: For low-resolution brightness temperature images Bicubic interpolation was performed to obtain a low-frequency brightness temperature substrate. ; Using convolutional neural networks, feature representations after global temporal modeling High-frequency residual information is obtained from the prediction. ; Superimposed high-frequency residual information With low-frequency brightness temperature substrate High-resolution brightness temperature reconstruction results were obtained. .

[0019] Preferably, a joint loss function incorporating physical constraints is constructed for end-to-end optimization, wherein the joint loss function specifically includes: , in, This is the pixel reconstruction loss term, used to constrain the consistency of the super-resolution results with the true high-resolution brightness temperature at the pixel level. These are high-frequency constraint terms based on real observations; Indicates the weighting coefficient; The pixel reconstruction loss term , is represented as: , in, The super-resolution predicted image and the actual super-resolution image are respectively in the 1st... The pixel value at each pixel; This indicates the total number of pixels in each frame of the image; The time frame number of the brightness temperature sequence; Represents the stability constant; The high-frequency constraint terms based on real observations , is represented as: , in: For network prediction results The corresponding high-frequency components, For true high-resolution images The corresponding high-frequency components are represented by the difference between the high-frequency components and the low-frequency components extracted after two-dimensional Gaussian low-pass filtering of the brightness temperature image and alignment. Compared with the prior art, the present invention has the following advantages: (1) The present invention proposes a multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature. By combining local temporal modeling with global temporal propagation, it achieves synergistic enhancement of brightness temperature sequence in both time and space dimensions. It is applicable to multi-channel and multi-scale passive microwave brightness temperature data and can provide high-quality input data support for Antarctic surface melting monitoring, ice sheet-ice shelf process research and climate model driving.

[0020] (2) This invention does not rely on antenna directivity function or explicit observation geometry model, which overcomes the defect of traditional analytical model super-resolution method that is highly dependent on prior accuracy, and has stronger applicability and generalization ability.

[0021] (3) This invention effectively suppresses abnormal frames and noise interference through a low-frequency normalized cross-correlation weighting mechanism, thereby improving the stability and robustness of brightness temperature reconstruction under complex Antarctic ice sheet-ice shelf surface conditions.

[0022] (4) The present invention introduces a joint loss optimization strategy with physical constraints, which improves spatial resolution while limiting non-physical high-frequency artifacts, thus ensuring the physical rationality of the reconstructed brightness temperature results. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] Example like Figure 1 As shown, this embodiment provides a multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature, the method comprising: S1. Acquire consecutive frames of microwave low-resolution brightness temperature images within the same Antarctic coverage area to construct a low-resolution brightness temperature image sequence, specifically including: Obtain continuous coverage within the same spatial area Passive microwave low-resolution brightness temperature images of each time phase were used to construct a multi-time series of low-resolution brightness temperature images. ,in, This indicates the time frame number of the brightness temperature sequence. Indicates the first Frame brightness temperature image, It is a single-channel grayscale image, and its pixel value represents the brightness temperature intensity of the corresponding ground surface.

[0026] S2. Brightness temperature feature supplementation based on local temporal windows is performed on multi-temporal low-resolution brightness temperature image sequences to obtain local temporal enhancement features, specifically including: S21. Feature supplementation is performed on the brightness temperature image of a single frame using a local temporal window, for any time index. , with the first Frame brightness temperature image Construct a local timing window of length 3 centered on [the element]. = This window is used to characterize the changes in brightness temperature over a short timescale.

[0027] S22, Set the local timing window A convolutional feature extraction network with shared parameters is used to extract brightness temperature features from each frame of brightness temperature image in the image. ,in, Indicated in scale Feature extraction operators are used to extract features. Specifically, this includes the original resolution, resolution and The network employs three spatial scales for resolution. Through layer-by-layer downsampling, it constructs a feature pyramid with three scales to capture structural information and displacement features of the brightness temperature field at different spatial scales.

[0028] S23. To compensate for inter-frame displacement, a cascaded deformable convolution mechanism is introduced to perform layer-by-layer alignment of non-reference frames, obtaining spatial alignment features. Specifically, it includes: At every scale The offset field is estimated using the fusion features of the reference frame and adjacent frames, as follows: ,in, Representing scale The offset prediction network is constructed from the corresponding convolutional layers; By utilizing the cascaded deformable convolution operator to perform spatial transformation in the feature space, the features of neighboring frames are aligned to the reference frame, achieving precise alignment of features from neighboring frames to the reference frame. Specifically, this is expressed as follows: ,in, This represents the features of non-center frames within a local time window. This represents the features of the center frame used as a reference. Indicates the center position of the convolution. Indicates the offset of standard convolution sampling points. This is a learnable spatial offset (initial value set to 0, maximum offset range is ±2 pixels). These are the convolution weights (obtained through network training and can be initialized as uniform weights). The number of sampling points (usually taken as...) nuclear).

[0029] By using this bottom-up cascading approach to progressively transfer coarse-scale alignment information to fine-scale alignment, we can effectively obtain the feature representations of adjacent frames aligned in the reference frame coordinate system. Ultimately, we obtain the spatial alignment features. : .

[0030] S24. Introduce a radiation similarity guidance mechanism based on normalized cross-correlation to adaptively measure and adjust the contribution weight of each time frame to the central reference frame, thus obtaining the temporal weight. .

[0031] Passive microwave brightness temperature (MBBT) images of the Antarctic region are highly susceptible to interference from atmospheric water vapor variations, instantaneous sensor gain fluctuations, or extreme weather conditions, leading to non-stationary noise or anomalous observation frames in time-series observations. Directly fusing weakly correlated neighboring frames would disrupt the physical consistency of the reconstruction process and introduce radiation artifacts. Therefore, this embodiment further introduces a radiation similarity guidance mechanism based on normalized cross-correlation to adaptively measure and adjust the contribution weight of each time frame to the central reference frame.

[0032] Specifically, it includes: 1) Considering that the Antarctic passive microwave brightness temperature field is more stable in low-frequency components, the inter-frame similarity was not directly calculated at the original spatial resolution. Instead, average pooling was first performed on the input window to obtain the low-frequency brightness temperature image: ,in, The average pooling operator is used to suppress the interference of high-frequency random noise on similarity calculation. Subsequently, a normalized cross-correlation operator is used to measure the similarity between adjacent frames and the reference frame in the low-frequency brightness temperature distribution, serving as a physical indicator of temporal reliability.

[0033] 2) Subsequently, a normalized cross-correlation operator is used to measure the similarity between adjacent frames and the reference frame in the low-frequency brightness temperature distribution. For two given low-frequency brightness temperature images... and Calculate low-frequency normalized cross-correlation similarity , , in, Indicates spatial location index, and Low-frequency brightness temperature images and Mean pixel intensity in the spatial domain For constant terms; The closer the value is to 1, the more consistent the radiation characteristics of the neighboring frame are with the reference frame.

[0034] 3) Finally, in order to quantitatively evaluate the contribution of features in each frame and ensure the stability of the local feature fusion process, this embodiment introduces the Softmax operator to normalize the low-frequency cross-correlation similarity. Mapping to the probability space yields comparable time-series weights. , is represented as: , in, This is a temperature coefficient used to adjust the discriminancy or smoothness of the weight distribution. Its physical meaning lies in adjusting the discriminancy or smoothness of the weight distribution. In local time series modeling, the temperature coefficient... The value was set to 0.1 through empirical experiments to balance the utilization rate of time series information with the effect of suppressing observation noise.

[0035] S25. Based on spatial alignment features With the corresponding time-domain radiation weight Preliminary local fusion features are extracted using a weighted aggregation operator. , is represented as: This allows for the spatial alignment and weighted fusion of local temporal features.

[0036] S26. To further explore the temporal consistency and spatial structure features hidden in Antarctic passive microwave brightness temperature images, this embodiment introduces a temporal-spatial joint attention feature enhancement mechanism to enhance the preliminary local fusion features. Refinement yields more robust local temporal enhancement features. Specifically, it includes: 1) Temporal correlation attention modeling: Preliminary local fusion features corresponding to the central reference frame Using temporal anchors, features are mapped to a low-dimensional embedding space through a convolutional embedding operator with shared parameters, and inter-frame correlations are calculated, represented as: , ,in, and It is a 3×3 convolution mapping operator with one-quarter of the original feature channels, used to reduce computational complexity and enhance the ability to identify inter-frame correlation.

[0037] The embedded features are element-wise multiplied and summed along the channel dimension to obtain the temporal correlation response between adjacent frames and the center frame. , is represented as: ,in, This represents the c-th channel.

[0038] Using the Sigmoid activation function to correlate responses Mapped to normalized temporal attention weights .

[0039] The obtained normalized temporal attention weights The corresponding frame features are applied pixel-by-pixel. After temporal correlation attention weighting, all temporal frame features are concatenated along the channel dimension and a channel compression operator is used to obtain the fused features. This allows for adaptive emphasis on time-reliable frames and suppression of anomalous frames.

[0040] The fusion feature is represented as: ,in, for Convolution operator, This represents element-wise product.

[0041] 2) Multi-scale spatial attention modeling: To further enhance the expressive power of spatial structure, a multi-scale spatial attention mechanism is introduced. Spatial statistical features at different scales are extracted through parallel max pooling and average pooling operators, and a pyramid-shaped attention representation is constructed.

[0042] Specifically, regarding fusion features After generating a coarse-scale spatial response through pooling and convolution operations, it is progressively upsampled and fused with shallow spatial attention features to obtain a spatial attention map that takes into account both global background and local structure. , is represented as: ,in, This represents a spatial attention generation operator consisting of multiple convolutions, pooling, and upsampling operations.

[0043] 3) Joint attention feature enhancement output: utilizing the obtained spatial attention Mapping pairs of fused features Element-wise modulation is performed, and a residual enhancement term is introduced to obtain local temporal enhancement features. , is represented as: ,in, This is a convolutional mapping operator used to supplement fine-grained spatial information. This enhanced feature effectively highlights stable thermal structure regions while maintaining the overall continuity of the brightness-temperature field, providing high-quality feature input for subsequent global temporal propagation and super-resolution reconstruction.

[0044] S3. To maintain the consistency of brightness temperature information over a longer time scale, this invention introduces a bidirectional temporal propagation mechanism based on local feature supplementation to perform global modeling of local temporal enhancement features, obtaining a feature representation after global temporal modeling, specifically including: 1) Backward time propagation: Starting from the end of the sequence, the time-reverse cluster update propagation features are as follows: ,in, This represents the backpropagation feature update operator, which internally consists of... Implemented by combining convolutional layers and gating units; 2) Forward temporal propagation: Based on the fusion of backward propagation features, the propagation features are updated from the temporal start frame to the last frame as follows: ,in, This represents the forward propagation feature update operator, which internally consists of... Implemented by combining convolutional layers and gating units; 3) By fusing forward ship features and backpropagation features, a feature representation after global temporal modeling is obtained: , in, for Convolutional layers are used to compress channels and fuse bidirectional spatiotemporal information.

[0045] S4. Based on the residual learning mechanism, super-resolution reconstruction of brightness temperature images is performed, specifically including: For low-resolution brightness temperature images Bicubic interpolation was performed to obtain a low-frequency brightness temperature substrate. , is represented as: ,in, This indicates a bicubic interpolation operation.

[0046] Using convolutional neural networks, feature representations after global temporal modeling High-frequency residual information is obtained from the prediction. , is represented as: ,in, This indicates the prediction network from the residuals.

[0047] Superimposed high-frequency residual information With low-frequency brightness temperature substrate High-resolution brightness temperature reconstruction results were obtained. , is represented as: .

[0048] In order to suppress non-physical high-frequency artifacts while ensuring the accuracy of brightness temperature numerical reconstruction, this embodiment constructs a joint loss function with physical constraints for end-to-end optimization.

[0049] Joint loss function Based on data consistency constraints, a frequency domain prior for the physical characteristics of passive microwave brightness temperature is introduced to improve the physical rationality and stability of the super-resolution results. Specifically, this includes: , in, This is the pixel reconstruction loss term, used to constrain the consistency of the super-resolution results with the true high-resolution brightness temperature at the pixel level. These are high-frequency constraint terms based on real observations; Indicates the weighting coefficient; Pixel reconstruction loss term The robust Charbonnier loss is adopted, which is expressed as: , in, The super-resolution predicted image and the actual super-resolution image are respectively in the 1st... The pixel value at each pixel; This indicates the total number of pixels in each frame of the image; The time frame number of the brightness temperature sequence; This represents the stability constant; this loss can maintain the continuity of brightness temperature values ​​while exhibiting good robustness to abnormal residuals, and is suitable for passive microwave brightness temperature data that are significantly affected by noise and mixed pixels.

[0050] Passive microwave brightness temperature images primarily exhibit large-scale continuous thermal fields, but their high-frequency components have limited energy and are susceptible to noise, pixel mixing, and inversion errors. Optimization based solely on pixel-level loss may lead to the network generating non-physical high-frequency components in local regions. Therefore, a high-frequency constraint term based on real observations is introduced. Physical consistency constraints are imposed on the super-resolution results from the perspective of energy amplitude. This constraint only takes effect when the predicted high-frequency amplitude exceeds the actual observed high-frequency amplitude, thereby effectively limiting the introduction of non-physical high-frequency energy and avoiding excessive smoothing of real edge and structural information.

[0051] Specifically, high-frequency constraint terms Represented as: , in: For network prediction results The corresponding high-frequency components, For true high-resolution images The corresponding high-frequency components are represented by the difference between the high-frequency components and the low-frequency components extracted after two-dimensional Gaussian low-pass filtering of the brightness temperature image and alignment.

[0052] Specifically, a two-dimensional Gaussian low-pass filter is first used. The brightness temperature image is smoothed to extract low-frequency components: ,in, Indicates time Brightness temperature image, For the corresponding low-frequency components, This represents the convolution operation. In this embodiment, the bandwidth of the control filter is... The value ranges from 1 to 3 pixel kernels and can be adjusted according to sensor resolution and noise level. Correspondingly, the high-frequency component is defined as: Based on this, the network prediction results are analyzed. Compared to real high-resolution images Calculate its high-frequency components respectively and .

[0053] Next, to verify the effectiveness and applicability of the proposed multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature sequences under different channel and seasonal conditions, the experiment used the AMSR2 passive microwave brightness temperature standard product released by the Japan Aerospace Exploration Agency (JAXA) as the experimental data source.

[0054] This invention primarily utilizes AMSR2 Level-3 daily average brightness temperature data, covering the period from July 3, 2012 to December 31, 2024, encompassing the Antarctic ice sheet, ice shelf region, and surrounding sea ice area. The invention selects two frequency bands, 18.7 GHz and 36.5 GHz, which are highly sensitive to Antarctic surface melting processes, and includes both horizontal (H) and vertical (V) polarization modes. All brightness temperature data are converted to a 16-bit single-channel grayscale image with a spatial size of 632 × 664 pixels after scaling factor conversion, while removing dates with missing or abnormal observations.

[0055] In the data preprocessing stage, a sliding window method is used to crop the original image into 256 × 256 pixel image blocks (with a step size of 128 pixels) to reduce edge effects. In the absence of higher-resolution passive microwave reference data, a low-resolution image (64 × 64 pixels) with a 4x downsampling is constructed using bicubic interpolation as the model input, and further time-series samples of length 5 are constructed with a time step of 1. To balance computational efficiency and sample representativeness, this invention employs a random sampling strategy to construct the experimental dataset from the original data, dividing it into training, validation, and test sets in a 7:1:2 ratio.

[0056] (1) Parameter settings.

[0057] The model training uses a joint loss function composed of pixel reconstruction loss and high-frequency constraint loss as the optimization objective. This ensures both the accuracy of brightness temperature numerical reconstruction and the suppression of non-physical high-frequency artifacts, thereby enhancing the physical plausibility and structural fidelity of the reconstruction results. The optimizer used is Adam, with its parameters set as follows: = 0.9、 = 0.99, initial learning rate 2×10⁻⁴, combined with cosine annealing strategy and 2000-step warmup to stabilize the initial gradient. The total number of training steps is 400,000, and exponential moving average (EMA, decay coefficient 0.999) is used to smooth the model parameters. During training, performance is evaluated on the validation set every 5000 steps and the model is saved; training is terminated early when the validation PSNR does not improve after several consecutive evaluations to prevent overfitting.

[0058] (2) Evaluation indicators.

[0059] To comprehensively and quantitatively evaluate the performance of the proposed passive microwave super-resolution algorithm on the Antarctic AMSR2 brightness temperature image, this paper selects the common and widely used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) as evaluation metrics to measure the fidelity between the reconstructed image and the target image. PSNR reflects the pixel-level error between the reconstructed image and the target image; a higher value indicates that the reconstruction effect is closer to the reference image. The formula for calculating PSNR is: , in, This represents the maximum brightness temperature value of the image. The mean squared error is calculated using the following formula: , in, and In pixels of the predicted image and the target image, respectively The value at that location, and These are the number of rows and columns of the image, respectively. PSNR is suitable for quantitatively assessing pixel-level differences in images, but it is not sensitive to the structural information and detail recovery capabilities of images. SSIM, on the other hand, measures image similarity from three aspects: brightness, contrast, and structure, and better reflects the human eye's perception of image quality. The formula for calculating SSIM is: , in, and These are the mean values ​​of the predicted image and the target image, respectively. and These are the variances of the predicted image and the target image, respectively. Let be the covariance of the two. (Constant) and The definition is as follows: , , in, and They are usually set to 0.01 and 0.03 respectively. PSNR represents the dynamic range of the image. SSIM reflects the structural fidelity of the image and provides a quality assessment that is more in line with human visual perception. These two metrics comprehensively evaluate the model's ability to recover pixel and structural information during reconstruction. A higher PSNR value or an SSIM value closer to 1 indicates a greater similarity between the reconstructed image and the target image.

[0060] (3) Comparison of multi-channel super-resolution results and methods.

[0061] The super-resolution reconstruction method for Antarctic passive microwave brightness temperature data proposed in this invention was applied to the horizontal and vertical polarization channels in the 36.5 GHz and 18.7 GHz bands of AMSR2 data, respectively, and the model was trained and tested on the data of the above channels.

[0062] To verify the effectiveness and performance advantages of the method of the present invention, the experiment further compared and analyzed the method of the present invention with a variety of existing methods. The comparison methods included traditional interpolation methods such as bicubic upsampling, as well as a variety of classic and deep learning super-resolution models, including Omni, PFT, EDVR, BasicVSR, IconVSR and BasicVSR++.

[0063] As shown in Table 1, the quantitative results reveal significant differences in reconstruction performance among different methods on AMSR2 multi-channel brightness temperature data. Overall, deep learning-based methods significantly outperform traditional interpolation strategies in both PSNR and SSIM, indicating that data-driven models can more effectively characterize the implicit spatial structural features in brightness temperature images. Under various polarization and frequency band conditions, single-image super-resolution methods improve reconstruction quality to some extent compared to traditional interpolation, but their performance is limited by single-frame information, making it difficult to fully utilize the temporal correlation inherent in the brightness temperature sequence. In contrast, temporal modeling methods incorporating multi-frame information achieve higher PSNR and SSIM in most channels, demonstrating the importance of rationally utilizing observations at adjacent time points for restoring spatial details in the brightness temperature field. Further comparison of different multi-frame methods reveals that while some methods perform well in natural video scenes, their performance improvement is unstable in passive microwave brightness temperature images. This is mainly because brightness temperature images are characterized by large-scale, low-frequency changes and relatively gentle temporal evolution; directly transferring models designed for natural images or complex motion scenes often fails to adequately match their physical characteristics.

[0064] To address the aforementioned issues, the method proposed in this invention effectively improves the spatiotemporal consistency and physical rationality of brightness temperature super-resolution reconstruction by introducing joint loss constraints of local temporal modeling, bidirectional global information propagation, and physical guidance. As shown in Table 1, the algorithm of this invention achieves optimal quantitative results across all frequency bands. Specifically, in the 36.5H, 36.5V, 18.7H, and 18.7V channels, the average PSNR of the algorithm proposed in this invention is improved by 1.2597 dB, 1.8284 dB, 2.0411 dB, and 2.2105 dB, respectively, compared to all comparative methods. Further analysis reveals that the performance advantage of the algorithm proposed in this invention becomes more significant as the frequency decreases. This phenomenon is closely related to the radiation characteristics of the passive microwave brightness temperature field, which is dominated by large scale and low frequency. By introducing an NCC similarity guidance mechanism based on low-frequency features in the local temporal modeling stage, the network can adaptively suppress inconsistent or noise-interfered neighboring frame information, thereby improving the stability of temporal alignment and enhancing the consistency recovery capability of large-scale structures. Furthermore, the bidirectional global propagation strategy integrates long-term temporal redundancy information, preserving spatial details while effectively suppressing the accumulation of non-physical high-frequency artifacts by combining a physically guided joint loss function. These factors collectively contribute to the robust performance improvement of the algorithm proposed in this invention in multi-channel brightness-temperature super-resolution tasks.

[0065] Table 1. Comparison of super-resolution reconstruction performance of different methods on AMSR 2 multichannel brightness temperature data. (5) Ablation test To verify the effectiveness of the core components and loss function terms in the algorithm of this invention, a detailed ablation study was conducted on an 18.7H channel dataset, using a 4x super-resolution task as a benchmark. Considering computational overhead and experimental efficiency, this paper randomly selected 20% of the samples from the original 18.7H channel training, validation, and test sets to construct a representative experimental subset, and uniformly trained it for 200,000 iterations to ensure the fairness of performance evaluation.

[0066] Analysis of Local Temporal Modeling Module and Propagation Strategy: To verify the effectiveness of the local temporal modeling module and information propagation strategy, detailed ablation experiments were conducted in this embodiment, and the results are shown in Table 2. To ensure the fairness of comparison under different structural configurations, the weights of each loss term were set based on experience and remained consistent in all ablation experiments, with the weight of the high-frequency constraint term uniformly set to 0.1. The baseline model had the lowest performance when no alignment and guidance mechanism was introduced and only unidirectional propagation was used. When only deformable convolution alignment was added without introducing low-frequency consistency weight guidance (Experiment 2), the performance decreased slightly, indicating that the alignment module was difficult to function stably in the absence of similarity constraints. In contrast, introducing bidirectional propagation (Experiment 3) without using low-frequency consistency weight guidance resulted in a significant performance improvement, indicating that bidirectional propagation can make fuller use of temporal redundancy information. After further introducing low-frequency consistency guidance (Experiments 4 and 5), both optical flow alignment and deformable convolution alignment showed a stable improvement in reconstruction performance, with deformable convolution performing slightly better. Ultimately, the complete model, which combines deformable convolution alignment, low-frequency consistency guidance, and bidirectional propagation strategies, achieved optimal results, verifying the rationality of the design of each module and its synergistic gains.

[0067] Table 2 Ablation experiments of local temporal modeling module and propagation strategy Analysis of loss function constraint terms: To verify the impact of high-frequency constraint terms in the joint loss function on the super-resolution reconstruction effect, a systematic ablation experiment was conducted on the high-frequency constraint weights. The relevant experimental settings and quantitative results are shown in Table 3. Considering the strong regularization effect of high-frequency constraint terms on the reconstruction results, their weight settings followed the principle of "from small to large, gradually expanding": denser values ​​were used in small weight intervals to finely characterize the impact of high-frequency constraints on reconstruction performance; relatively looser value intervals were used in larger weight intervals to verify the changing trend of model performance under strong constraints. When only pixel reconstruction loss was used, the network was able to achieve high pixel accuracy (PSNR of 47.1162 dB), but slight non-physical high-frequency artifacts may still be introduced in local areas. As the weight of the high-frequency constraint term gradually increased, the overall model reconstruction performance showed a trend of first improving and then stabilizing. When the weight increased to 0.1, the PSNR reached a peak of 47.141 dB, indicating that moderate high-frequency constraints can effectively suppress abnormal high-frequency components while maintaining the spatial structure information of the brightness temperature field.

[0068] As the weight of the high-frequency constraint term continues to increase, the PSNR index shown in Table 3 decreases to some extent, indicating that excessively strong high-frequency constraints can cause over-smoothing of local edges and details in the brightness temperature field, thereby weakening structural fidelity. This phenomenon further verifies the conclusion that the weight of the high-frequency constraint term should not be set too large. The experimental results in Table 3 show that the design of the loss function combining pixel reconstruction loss and high-frequency constraint term achieves a good balance between brightness temperature numerical consistency and structural preservation. This study ultimately selects a small weight interval near the performance peak as the default parameter configuration for model training, providing a reasonable and reproducible parameter basis for the stable training and reconstruction performance of the algorithm of this invention.

[0069] Table 3. Impact of High-Frequency Constraint Loss Term Weighting on Brightness-Temperature Super-Resolution Reconstruction Results The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature, characterized in that, include: S1. Collect consecutive frames of microwave low-resolution brightness temperature images within the same space Antarctic coverage area and construct a low-resolution brightness temperature image sequence. S2. The multi-temporal low-resolution brightness temperature image sequence is supplemented with brightness temperature features based on a local temporal window to obtain local temporal enhancement features; wherein, the local temporal window is obtained by combining the brightness temperature images of a number of adjacent frames before and after, with the brightness temperature image of the current frame as the center. S3. Introduce a bidirectional temporal propagation mechanism to perform global modeling on the local temporal enhancement features, and obtain the feature representation after global temporal modeling; S4. Based on the residual learning mechanism, perform super-resolution reconstruction of brightness temperature images.

2. The multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature as described in claim 1, characterized in that, S1 involves acquiring consecutive frames of microwave low-resolution brightness temperature images within the same Antarctic coverage area to construct a low-resolution brightness temperature image sequence, specifically including: Obtain continuous coverage within the same spatial area Passive microwave low-resolution brightness temperature images of each time phase were used to construct a multi-time series of low-resolution brightness temperature images. ,in, This indicates the time frame number of the brightness temperature sequence. Indicates the first Frame brightness temperature image, It is a single-channel grayscale image, and its pixel value represents the brightness temperature intensity of the corresponding ground surface.

3. The multi-image super-resolution reconstruction method for Antarctica based on passive microwave brightness temperature as described in claim 2, characterized in that, S2 refers to the multi-temporal low-resolution brightness temperature image sequence Brightness-temperature feature supplementation based on local temporal windows is performed, specifically including: S21. Feature supplementation is performed on the brightness temperature image of a single frame using a local temporal window, for any time index. , with the first Frame brightness temperature image Construct a local time window centered on the time window. = ; S22, Set the local timing window A convolutional feature extraction network with shared parameters is used to extract brightness temperature features from each frame of brightness temperature image in the image. in, Indicated in scale Feature extraction operators; S23. Introduce a cascaded deformable convolution mechanism to perform layer-by-layer alignment of non-reference frames and obtain spatial alignment features. ; S24. Introduce a radiation similarity guidance mechanism based on normalized cross-correlation to adaptively measure and adjust the contribution weight of each time frame to the central reference frame, thus obtaining the temporal weight. ; S25. Based on spatial alignment features With the corresponding time-domain radiation weight Preliminary local fusion features are extracted using a weighted aggregation operator. , is represented as: ; S26. Introduce a temporal-spatial joint attention feature enhancement mechanism to improve the initial local fusion features. Enhancement is performed to obtain local temporal enhancement features. .

4. The multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature as described in claim 3, characterized in that, In step S23, a cascaded deformable convolution mechanism is introduced to perform layer-by-layer alignment of non-reference frames to obtain spatial alignment features. Specifically, it includes: At every scale The offset field is estimated using the fusion features of the reference frame and adjacent frames, as follows: ,in, Representing scale The offset prediction network is constructed from the corresponding convolutional layers; Using the cascaded deformable convolution operator to perform a spatial transformation in the feature space, aligning the feature vectors of neighboring frames to the reference frame, it can be represented as: ,in, This represents the features of non-center frames within a local time window. This represents the features of the center frame used as a reference. Indicates the center position of the convolution. Indicates the standard convolution sampling point offset. A learnable spatial offset. For convolution weights, This represents the number of sampling points; Obtain spatial alignment features : .

5. The multi-image super-resolution reconstruction method for Antarctica based on passive microwave brightness temperature as described in claim 3, characterized in that, S24 introduces a radiation similarity guidance mechanism based on normalized cross-correlation to adaptively measure and adjust the contribution weight of each time frame to the central reference frame, specifically including: 1) Perform average pooling on the input window to obtain a low-frequency brightness temperature image. ; 2) The normalized cross-correlation operator is used to measure the similarity between adjacent frames and the reference frame in the low-frequency brightness temperature distribution. For two given low-frequency brightness temperature images... and Calculate low-frequency normalized cross-correlation similarity : ， in, Indicates spatial location index, and Low-frequency brightness temperature images and Mean pixel intensity in the spatial domain For constant terms; 3) Introduce the Softmax operator to normalize the cross-correlation similarity of low frequencies. Mapping to the probability space yields the temporal weights. .

6. The multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature as described in claim 5, characterized in that, The time-series weights , is represented as: ， in, This is a temperature coefficient used to adjust the discriminative power or smoothness of the weight distribution.

7. The multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature as described in claim 3, characterized in that, S26 introduces a temporal-spatial joint attention feature enhancement mechanism to improve the initial local fusion features. Enhancement is performed to obtain local temporal enhancement features, specifically including: 1) Temporal correlation attention modeling: Preliminary local fusion features corresponding to the central reference frame Using temporal anchors, features are mapped to a low-dimensional embedding space through a convolutional embedding operator with shared parameters, and inter-frame correlations are calculated, represented as: , ,in, and It is a convolution mapping operator; The embedded features are element-wise multiplied and summed along the channel dimension to obtain the temporal correlation response between adjacent frames and the center frame. , is represented as: ,in, This represents the c-th channel; Using the Sigmoid activation function to correlate responses Mapped to normalized temporal attention weights ; The obtained normalized temporal attention weights The corresponding frame features are applied pixel-by-pixel. After temporal correlation attention weighting, all temporal frame features are concatenated along the channel dimension and a channel compression operator is used to obtain the fused features. , is represented as: ,in, for Convolution operator, Represents element-wise product; 2) Multi-scale spatial attention modeling: Fusion features After pooling and convolution operations to generate a coarse-scale spatial response, it is upsampled stepwise and fused with shallow spatial attention features to obtain a spatial attention map. , is represented as: ,in, This represents a spatial attention generation operator consisting of multiple convolutional, pooling, and upsampling operations. 3) Joint attention feature enhancement output: Utilizing the obtained spatial attention Mapping pairs of fused features Element-wise modulation is performed, and a residual enhancement term is introduced to obtain local temporal enhancement features. , is represented as: ,in, This is the convolution mapping operator.

8. The multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature as described in claim 1, characterized in that, The S3 section introduces a bidirectional temporal propagation mechanism to enhance the local temporal features. Perform global modeling to obtain the feature representation after global temporal modeling. Specifically, it includes: 1) Backward time propagation: Starting from the end of the sequence, the time-reverse cluster update propagation features are as follows: ,in, This represents the backpropagation feature update operator, which internally consists of... Implemented by combining convolutional layers and gating units; 2) Forward temporal propagation: Based on the fusion of backward propagation features, the propagation features are updated from the temporal start frame to the last frame as follows: ,in, This represents the forward propagation feature update operator, which internally consists of... Implemented by combining convolutional layers and gating units; 3) By fusing forward ship features and backpropagation features, a feature representation after global temporal modeling is obtained: ， in, for Convolutional layers are used to compress channels and fuse bidirectional spatiotemporal information.

9. A multi-image super-resolution reconstruction method for Antarctica based on passive microwave brightness temperature as described in claim 3, characterized in that, The S4 step, based on a residual learning mechanism, performs super-resolution reconstruction of the brightness temperature image, specifically including: For low-resolution brightness temperature images Bicubic interpolation was performed to obtain a low-frequency brightness temperature substrate. ; Using convolutional neural networks, feature representations after global temporal modeling High-frequency residual information is obtained from the prediction. ; Superimposed high-frequency residual information With low-frequency brightness temperature substrate High-resolution brightness temperature reconstruction results were obtained. .

10. A multi-image super-resolution reconstruction method for Antarctic passive microwave brightness temperature as described in claim 1, characterized in that, An end-to-end optimization is performed by constructing a joint loss function that incorporates physical constraints. The joint loss function specifically includes: ， in, This is the pixel reconstruction loss term, used to constrain the consistency of the super-resolution results with the true high-resolution brightness temperature at the pixel level. These are high-frequency constraint terms based on real observations; Indicates the weighting coefficient; The pixel reconstruction loss term , is represented as: ， in, The super-resolution predicted image and the actual super-resolution image are respectively in the 1st... The pixel value at each pixel; This indicates the total number of pixels in each frame of the image; The time frame number of the brightness temperature sequence; Represents the stability constant; The high-frequency constraint terms based on real observations , is represented as: ， in: For network prediction results The corresponding high-frequency components, For true high-resolution images The corresponding high-frequency components are represented by the difference between the high-frequency components and the low-frequency components extracted after two-dimensional Gaussian low-pass filtering of the brightness temperature image and alignment.

Images