An ESRGAN-based single-channel super-resolution reconstruction method for FY-4B satellite remote sensing

Abstract

The application provides a FY4B remote sensing single-channel super-resolution reconstruction method based on ESRGAN, which comprises the following steps: analyzing, uniformly cutting, scale constraining and quality detecting the resolution data of FY4B and FY3D to obtain pretreatment data; generating a maximum effective pixel intersection mask according to the pretreatment data; calculating the effective pixel ratio and performing patch screening according to the maximum effective pixel intersection mask to obtain training data; constructing an ESRGAN network structure and performing data reconstruction on the training data to obtain a predicted image. Through the establishment of paired samples of the same day, the uniform cutting alignment and scale constraint, the patch screening and mask loss calculation based on the maximum intersection of effective pixels, and the reconstruction strategy combining pixels, high frequencies and adversarial learning, the FY4B single-channel data is realized from low resolution to high resolution in detail enhancement and structure reconstruction.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing image processing and super-resolution reconstruction technology, and in particular to a single-channel super-resolution reconstruction method based on ESRGAN from the Fengyun-4B satellite. Background Technology

[0002] Against the backdrop of continuously increasing demands for "high timeliness, precision, and quantifiability" in meteorological operations, the Fengyun-4B (FY4B) geostationary meteorological satellite, due to its wide coverage and high observation frequency, is widely used in scenarios such as cloud monitoring, severe convection identification, and monitoring of haze and dust storms. However, limited by the imaging system and observation conditions, the remote sensing data from Fengyun-4B (FY4B) has a relatively low resolution (most channels have a resolution of 4000m) under current technologies, exhibiting significant smoothing and loss of detail in information such as cloud boundaries, convective structures, and fine-scale textures. Because fine structures are weakened or aliased during imaging and downsampling, directly using traditional interpolation upsampling (bilinear / bicubic) can only amplify the pixel size and cannot recover true high-frequency details, resulting in enhancement results that are still blurry and structurally unclear, making it difficult to meet the needs of precision monitoring. Meanwhile, existing general super-resolution methods are mostly geared towards natural images, and their training often relies on stable, clean, high-resolution ground truth data and consistent imaging distribution. However, in meteorological remote sensing scenarios, cross-satellite (FY4 and FY3) observations present problems such as imaging differences, missing / invalid pixels (NaN), and block-level missing data. Without strict pairing, effective pixel selection, and mask constraints, the training process will be affected by missing pixels and degraded samples, leading to unstable model convergence, overly smoothed output, or the generation of pseudo-textures, making it difficult to achieve usable single-channel super-resolution reconstruction results. Summary of the Invention

[0003] The purpose of this invention is to provide a single-channel super-resolution reconstruction method for FY4B satellite remote sensing based on ESRGAN, which can significantly improve the edge sharpness and detail expression of FY4B single-channel images without relying on simple interpolation.

[0004] To achieve the above objectives, the present invention provides the following solution: A single-channel super-resolution reconstruction method based on ESRGAN from Fengyun-4B satellite remote sensing includes the following steps: The resolution data of Fengyun-4B and Fengyun-3D satellites were analyzed to obtain a pairing index table; The paired index table is uniformly pruned, scaled, and quality checked to obtain preprocessed data. Generate the maximum effective pixel intersection mask based on the pairing index table and preprocessed data; The effective pixel ratio is calculated based on the maximum effective pixel intersection mask, and patch filtering is performed to obtain training data; An ESRGAN network structure is constructed, and the training data is reconstructed using the ESRGAN network structure to obtain the predicted image.

[0005] Optionally, the paired index table is uniformly pruned, scaled, and quality checked to obtain preprocessed data, including: The paired index table is cropped at a fixed resolution to obtain FY3D single-channel image patches; Calculate the total number of pixels, the number of valid pixels, and the number of invalid pixels in a single-channel image block of FY3D; The invalid pixel ratio is calculated based on the total number of pixels, the number of valid pixels, and the number of invalid pixels; Outlier filtering is performed on FY3D single-channel image blocks based on the proportion of invalid pixels to obtain preprocessed data.

[0006] Optionally, a maximum effective pixel intersection mask is generated based on the pairing index table and preprocessed data, including: The paired index table is cropped at a fixed resolution to obtain the FY4B low-resolution image; Low-resolution and high-resolution masks are generated based on the FY4B low-resolution image and preprocessed data, respectively. The low-resolution mask is mapped to the high-resolution network by nearest neighbor upsampling to obtain the mapped mask; Calculate the maximum effective pixel intersection mask based on the mapping mask and the high-resolution mask.

[0007] Optionally, the effective pixel ratio is calculated based on the maximum effective pixel intersection mask, and patch filtering is performed to obtain training data, including: Training patches are generated on the maximum effective pixel intersection mask using a sliding window; the training patches include: high-resolution patches and low-resolution patches. The effective pixel percentage of the training patch is calculated using the integral image. The training patches are filtered based on the effective pixel ratio to obtain the training data.

[0008] Optionally, the effective pixel percentage of the training patch is calculated using the integral image, including: Zero-padding is applied to the training patches to obtain the integral image. Calculate the number of effective pixels and the total number of pixels based on the integral image; The effective pixel count is normalized by matching the lower and upper bounds of the channel data in the index table. The percentage of effective pixels is calculated based on the normalized number of effective pixels and the total number of pixels.

[0009] Optionally, an ESRGAN network structure is constructed, and the training data is reconstructed using the ESRGAN network structure to obtain the predicted image, including: Initial features of the training data are extracted using a two-dimensional convolution with a kernel size of 3×3, a stride of 1, and padding of 1. Deep features are extracted from the initial features through the backbone network; the backbone network consists of multiple RRDBs, and each RRDB consists of multiple Residual Dense Blocks. The deep features are connected to the initial features via a global residual concatenation to obtain the fused features; The fused features are subjected to double upsampling and convolution to obtain the predicted image.

[0010] Optionally, the ESRGAN network structure includes a generator and a discriminator; the loss function of the discriminator is a binary cross-entropy with logits, and the loss function of the generator is: L G =w pix ·L1+w perc ·L perc +w grad ·L grad +w freq ·L freq +w stat ·L stat +lambda adv ·L adv ;where w pix w perc w grad w freq w stat and lambda adv All are weighting coefficients, L1 is the pixel reconstruction loss, L perc To perceive loss, L grad For gradient loss, L freq For frequency domain loss, L stat To calculate the loss, L adv To combat the losses.

[0011] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a single-channel super-resolution reconstruction method for FY4B satellite remote sensing based on ESRGAN. This method includes: parsing the resolution data of FY4B and FY3D satellites to obtain a pairing index table; performing unified cropping, scale constraints, and quality checks on the pairing index table to obtain preprocessed data; generating a maximum effective pixel intersection mask based on the pairing index table and preprocessed data; calculating the effective pixel ratio based on the maximum effective pixel intersection mask and performing patch selection to obtain training data; constructing an ESRGAN network structure and reconstructing the training data using the ESRGAN network structure to obtain a predicted image. This method achieves detail enhancement and structural reconstruction of FY4B single-channel data from low resolution to high resolution by establishing paired samples of spectral channels with similar dates and times, unified cropping and alignment with scale constraints, patch selection and mask loss calculation based on the maximum intersection of effective pixels, and combining pixel, high-frequency, and adversarial learning reconstruction strategies. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the Fengyun-4B satellite remote sensing single-channel super-resolution reconstruction method based on ESRGAN, as described in this invention. Figure 2 This is a schematic diagram illustrating the working principle of the ESRGAN framework in an embodiment of the present invention. Detailed Implementation

[0014] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1 As shown, this embodiment of the invention provides a method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN, including the following steps: Step 100: Analyze the resolution data of Fengyun-4B and Fengyun-3D satellites to obtain a pairing index table; Step 200: Perform uniform pruning, scale constraints, and quality checks on the paired index table to obtain preprocessed data; Step 300: Generate the maximum effective pixel intersection mask based on the pairing index table and preprocessed data; Step 400: Calculate the percentage of effective pixels based on the maximum effective pixel intersection mask and perform patch filtering to obtain training data; Step 500: Construct the ESRGAN network structure and use the ESRGAN network structure to reconstruct the training data to obtain the predicted image.

[0017] In the specific implementation process, step 100 involves collecting 4000m single-channel low-resolution data from the AGRI of Fengyun-4B (FY4B) satellite and 1000m high-resolution data from the MERSI of Fengyun-3D (FY3D) satellite by traversing the data root directory. Fields are parsed according to file naming rules and used as unique keys to create an index. Subsequently, the intersection of the key sets of FY3D and FY4B is taken to generate a paired index table, which serves as the sole data source for subsequent training and evaluation.

[0018] In the specific implementation process, step 200, to ensure pixel-level supervised alignment, performs uniform cropping and scale constraints on each pair of paired samples: setting the magnification factor scale=4, cropping FY3D to a fixed high-resolution target size hr_target_hw, and cropping FY4B to a fixed low-resolution target size lr_target_hw, satisfying hr_target_hw=lr_target_hw×scale; the cropping method supports fixed modes (such as top-left corner alignment or center alignment) to eliminate misalignment caused by redundant pixels at the boundary. For missing values ​​and outliers, the proportion of invalid pixels (NaN / Inf ratio) is calculated for FY3D single-channel image blocks to determine whether the data block can be used for statistics, training, or inversion. Specifically, let the FY3D image block be a two-dimensional array X with a size of H×W, where X[i,j] is the pixel value in the i-th row and j-th column. Define the effective cell mask M: If X[i,j] is a finite number (referring to a normal real value in floating-point representation, not undefined or overflowing, i.e., not NaN, representing "not a number", usually generated during invalid operations, indicating that the cell value is undefined, and not +Inf / -Inf, where Inf represents an infinitely large or infinitely small value generated by numerical overflow or division by zero, belonging to abnormal cells), then M[i,j]=1; otherwise, M[i,j]=0. Total number of cells N=HW; Effective number of cells N val =sum(M[i,j]) (sums over all i=1..H, j=1..W); invalid pixels Ninv =NN val Invalid pixel ratio r inv =N inv / N=1-N val / (HW). Let the threshold τ (in this embodiment, τ = 0.10, representing 10%), when r inv When the value is greater than τ, the key (date, block, channel) is considered to have severe missing data. "date" represents the observation date, corresponding to the specific observation time; "block" represents the spatial block number, used to identify a fixed spatial sub-region within the same satellite image; and "channel" represents the remote sensing band number (channel number), corresponding to different physical observation bands. This key is skipped to avoid contaminating the training with missing samples, thus obtaining preprocessed data.

[0019] In the specific implementation process, step 300 generates effective pixel masks (isfinite) for both FY3D and FY4B. First, it determines whether each pixel is finite, generating a 0 / 1 binary mask to mask the impact of NaN / Inf on training and statistics. Let X be the high-resolution image after cropping from FY3D. hr The low-resolution image after FY4B cropping is X. lr The magnification s=4, and after cropping, H is satisfied. hr =s·H lr W hr =s·W lr H hr ×W hr H represents the size of the high-resolution image. lr ×W lr Let M be the size of the low-resolution image. Construct masks separately: M hr [i,j]=1 (if X= hr [i,j] are valid) otherwise 0; M lr [u,v]=1 (if X) lr [u,v] valid) otherwise 0. Set the low-resolution mask M. lr Upsampling to a high-resolution grid using the nearest neighbor method yields... The mapping relationship is as follows: , This represents the floor function (i.e., the largest integer not greater than x). This represents mapping the i-th row in the high-resolution image back to the corresponding row index in the low-resolution image. Since the high-resolution image is obtained by expanding the low-resolution image by a factor of s, every s rows of high-resolution pixels correspond to one row of pixels in the low-resolution image; rounding down ensures that the mapping is to the correct low-resolution pixel index. Then, the maximum effective pixel intersection mask is calculated: The AND operator represents a logical AND operation (the result is 1 if both values ​​are 1, otherwise it is 0). This intersection mask is used to ensure that subsequent training is performed only on pixels that are simultaneously valid in both FY3D and FY4B.

[0020] In the specific implementation process, step 400 uses a sliding window to generate training patches within the intersection area of ​​the maximum effective pixel intersection mask. Among these, HR... patch =256×512, indicating that the high-resolution patch is 256 pixels high and 512 pixels wide; stride=128×256, indicating that the sliding window moves 128 pixels high and 256 pixels wide each time (i.e., the stride). Since the magnification s=4, the corresponding low-resolution patch size and stride are scaled down proportionally to: LR patch =64×128, stride=32×64. Where HR (HighResolution) represents high resolution (FY3D), LR (LowResolution) represents low resolution (FY4B), and satisfies coordinate mapping: if the top left corner of the high resolution patch is (y,x), then the top left corner of the low resolution patch is (floor(y / s),floor(x / s)).

[0021] Furthermore, to quickly calculate the proportion of effective pixels within each candidate window, this embodiment uses an integral image (also known as a summed-areatable) to accumulate and sum the intersection mask. First, let B be the 0 / 1 array of the intersection mask (True→1, False→0), and construct an integral image S (with zero-padding at the boundaries), defined as: S(0,j)=0, S(i,0)=0; for any i>=1, j>=1, we have S(i,j)=B(i,j)+S(i-1,j)+S(i,j-1)-S(i-1,j-1). For any window starting at the top left corner (y,x) and of size h×w (h=256, w=512 in this embodiment), the number of effective pixels within the window is: N val (y,x) = S(y+h,x+w) - S(y,x+w) - S(y+h,x) + S(y,x). The total number of pixels in the window is N. win =h·w, then the effective pixel ratio r val (y,x)=N val (y,x) / (h·w). Let the patch filtering threshold be η (η=0.98 in this embodiment), when r valWhen (y,x)>=η, the HR or LR patch coordinates and path information corresponding to the window are written to the patchlist. The patchlist (patch list / patch index table) is a structured patch list file (CSV format in this example) used for training and validation. Each row describes a reproducible training sample patch and includes at least: date, block, channel, and fy3. path fy4 path hr yx ,lr yx hr hw ,lr hw ,crop offse Fields such as 't' are included. Using patchlists, samples are deterministically loaded according to the list during the training phase, avoiding the unreproducibility caused by random segmentation and facilitating auditing and traceability of training data quality.

[0022] Furthermore, the numerical range of effective FY3D pixels for each channel is statistically analyzed and used for normalization. Here, lo and hi represent the "lower bound" and "upper bound" of the data for that channel, respectively, used to clip pixel values ​​and linearly map them to a uniform numerical range. The physical meaning of lo / hi can be understood as follows: among the effective observed pixels of that channel, lo corresponds to a smaller effective observed value, and hi corresponds to a larger effective observed value; during normalization, values ​​below lo are treated as lo, and values ​​above hi are treated as hi, thus avoiding extreme values ​​from causing instability in training. In this embodiment, full-range statistics are used, i.e., min / max: let the set of all FY3D image patches participating in the statistics for each channel be {X}. k}, where X k It is a two-dimensional array; the effective cell set is defined as Val={x|x is all X k Let $x$ be the number of pixels that satisfy $isfinite(x) = True$. Then $lo = min(Val)$, $hi = max(Val)$, where $lo$ is the minimum effective pixel value and $hi$ is the maximum effective pixel value; $isfinite(x)$ means that $x$ is neither NaN nor ±1$. If it is necessary to pre-screen samples with severe missing data, the proportion of invalid pixels for each sample can be calculated and removed before calculating $lo / hi$.

[0023] Specifically, in the normalized mapping, "input / target uniform clipping mapping to [-1,1]" means that the input tensor of the training samples and the target tensor of the supervision samples are both learned in the same numerical domain. Here, "input" refers to the FY4B single-channel low-resolution patch (LR) received by the model. patchAfter necessary cropping and alignment, the extracted data is normalized and used as the generator input; "target" refers to the FY3D single-channel high-resolution patch (HR) with the same key (date, block, channel) as this LR patch. targetpatch After cropping and alignment, the value is extracted and normalized to serve as the supervisory ground truth. The normalization process is as follows: for any pixel value v, first crop v... clip =min(max(v,lo),hi), then perform a linear mapping v norm =2*(v clip -lo) / (hi-lo)-1, making it fall within [-1,1]. Here, lo / hi represents the upper and lower bounds obtained from the channel statistics; when hi=lo, v can be... norm Set to 0 to avoid division by zero. The above mapping is applied to both the input (FY4BLR patch) and the target (FY3DHR target patch) to ensure consistent numerical scale during training.

[0024] like Figure 2 As shown, in the specific implementation process, step 500 uses the PyTorch framework to construct the ESRGAN network structure, including a generator G and a discriminator D. The generator uses RRDB (Residual-in-Residual Dense Block) as the backbone structure to extract deep texture features; the input is an FY4B single-channel LR patch (after normalization), and the output is a single-channel SR result. Specifically, the generator first extracts initial features through shallow convolution, denoted as I. LR (shape is 1×H) lr ×W lr The first convolutional layer is F0 = Conv3 × 3(I LR ), where Conv3×3 represents a 2D convolution with a kernel size of 3×3, a stride of 1, and padding of 1, and the number of output channels is, for example, nf=64; then F0 is input into a backbone network composed of nb RRDBs to obtain the deep feature F. trunk =RRDB stack (F0), where RRDB stack This indicates that nb RRDBs are concatenated (nb=16 in this example), each RRDB consists of several Residual Dense Blocks, and residual scaling is used to stabilize training; the backbone output is then fused by a convolutional layer and globally residually connected with shallow features, expressed as: F=F0+Conv3×3(F trunk Next, the upsampling module UpsampleBlock is entered, which uses "interpolation upsampling + convolution" to amplify the sample twice in succession: the first upsampling yields F. up1=Conv3×3(Interp(F,scale=2)), the second upsampling yields F. up2 =Conv3×3(Interp(F_up1,scale=2)), where Interp represents interpolation upsampling (bilinear operation in this example), and the total magnification is 4. Finally, a single-channel result is output through reconstruction convolution, expressed as: I SR =Conv3×3(F up2 The output shape is 1×H. hr ×W hr Here, "SR result" refers to "super-resolution reconstruction result / super-resolution output image patch," which is the predicted image after the model restores the low-resolution input to a high-resolution scale.

[0025] Furthermore, during network training, _SR Patch I with normalized HR target HR The loss is calculated within the effective pixel mask area, and parameters are updated via backpropagation. The discriminator D employs a relative discriminative structure. Its basic idea is not to judge whether an image is real or fake, but rather to judge whether the real sample is more real than the generated sample. The main body of the discriminator uses a patch discriminative structure, with output logits (discrimination scores) obtained through multi-layer convolution and progressive downsampling. Let the original output of the discriminator for the input image x be C(x) (logits, which can be a scalar or feature map; in implementation, it is often a patchlogits map). The expression for the relative discriminative output is: D re l(real,fake)=C(real)-mean(C(fake)),D rel (fake, real) = C(fake) - mean(C(real)), where mean(·) represents the mean over the batch dimension (and optionally the spatial dimension), used to construct a "relative comparison," and C(·) is the discriminator backbone output logits. The adversarial target is calculated in the form of "relatively real / relatively generated": the discriminator aims to achieve D during training. rel The larger (real, fake) is, the better (it should be judged as true), D rel The smaller the (fake, real) ratio, the better (it should be judged as fake). Specifically, BCEWithLogitsLoss (binary cross-entropy with logits) is used as the loss function, with the expression: L D =BCEWithLogits(D rel (real,fake),1)+BCEWithLogits(D rel (fake, real), 0); Label 1 uses smoothing to improve stability. The generator's adversarial loss has the opposite objective: to generate samples that are more "realistic" than actual samples, expressed as: =BCEWithLogits(D rel (fake, real), 1); in, BCEWithLogits() is the adversarial loss for the generator, and BCEWithLogits() is the cross-entropy loss with logits.

[0026] It should be noted that by mapping the input and target to [-1,1] with lo / hi as the boundary, generating effective pixel masks for HR target patches and replacing and zeroing invalid pixels, using RRDB backbone + two-level ×2 interpolation upsampling to achieve ×4 super-resolution reconstruction, and using relative discriminative adversarial targets for training, this invention can improve the expression of texture details while maintaining training stability in numerical remote sensing single-channel super-resolution scenarios, making the output closer to the high-resolution observation form of FY3D.

[0027] Furthermore, the network structure training employs an alternating update strategy: in each iteration, the discriminator is updated first, followed by the generator. The discriminator loss uses BCEWithLogits, and the ground truth labels are smoothed (real=0.9) to suppress oscillations caused by an overly strong discriminator. The specific construction process is as follows: the discriminator outputs logits for the ground truth high-resolution patch HR, denoted as pred. real =D(HR), where pred is the output logits of the super-patch SR (with gradients disconnected on SR) of the generator output. fake =D(SR detach Let the true label be y. real =0.9, false label y fake If the value is 0.0, then the expression for the discriminator loss is: L D =mean(BCEWithLogits(pred real ,y real ))+mean(BCEWithLogits(pred fake ,y fake )); BCEWithLogits(z,y)=-[y·log(sigmoid(z))+(1-y)·log(1-sigmoid(z))]; sigmoid(z) = 1 / (1 + exp(-z); Here, mean(·) represents averaging over the batch dimension and the spatial dimension, and label smoothing is used to prevent the discriminator from tending to extremes too quickly, thereby causing adversarial training oscillations.

[0028] The generator loss is a combined loss, including: pixel reconstruction loss (L1, used to maintain overall structure and numerical consistency), perceptual loss (obtained by copying a single channel into a three-channel input VGG19 to extract multi-layer features and calculate L1 differences, used to enhance the texture morphology "like FY3D"), gradient loss (L1 based on horizontal / vertical first-order differences, used to enhance edges and structural transitions), frequency domain loss (obtained by performing a two-dimensional FFT on SR and HR, used to constrain amplitude spectrum differences to match high-frequency energy distribution), statistical loss (used to constrain mean, standard deviation, and extreme value differences to stabilize the overall distribution), and adversarial loss (used to make SR closer to the real HR on the discriminator). Its specific expression and construction process are as follows: the generator outputs SR=G(LR), and various losses are calculated within the effective pixel mask. The expression for the pixel reconstruction loss is: L1=sum(|SR-HR|·mask) / (sum(mask)+eps); Where eps is a minimal constant to prevent division by zero. Perceptual loss L perc After copying SR and HR into three channels and performing VGG input normalization, extract several layers of features Phi_l(·), expressed as: ; Where Σ_I represents the summation of selected multi-layer feature sets, and I is a pre-defined index of several feature layers in the VGG19 network. Gradient loss L grad Define first-order differences Gx(X) = X[:,:,:,1:] - X[:,:,:,:-1] and Gy(X) = X[:,:,1:,:] - X[:,:,:-1,:], and obtain the function expressions as follows: ; Frequency domain loss L freq The expression is: F SR =FFT2(SR); F HR =FFT2(HR); A SR =abs(F SR ); A HR =abs(F HR ); L freq =mean(|A SR -A HR |); Where A is the amplitude spectrum. Statistical loss L stat : Calculate mu on the effective pixels of mask SR 、mu HR std SR std HR min SR ,max SR min HR ,max HR ,mu SR mu represents the mean value of the SR image within the effective pixel area. HR std represents the mean value of the HR image within the effective pixel region. SR std represents the standard deviation of an SR image within the effective pixel area. HR The minimum value represents the standard deviation of the HR image within the effective pixel area. SR Max represents the minimum value within the effective pixel area of ​​the SR image. SR min represents the maximum value of the SR image within the effective pixel area. HR The maximum value represents the minimum value within the effective pixel area of ​​the HR image. HR The maximum value of the HR image within the effective pixel region is expressed as: L stat =|mu SR -mu HR |+|std SR -std HR |+|min SR -min HR |+|max SR -max HR |; Combat loss L adv : Retrieve when generator is updated And let y g =1.0 (or 0.9, consistent with the actual label; one must be fixed), then the expression is: ; The final expression for the total loss of the generator is: L G =w pix ·L1+w perc ·L perc +w grad ·L grad +w freq ·L freq +w stat ·L stat +lambda adv ·L adv ; Where wpix w perc w grad w freq w stat With lambda adv These are the weighting coefficients. During training, the L1 metric is periodically calculated on the validation set, and the optimal model (G) is saved. best The validation set (.pt) is obtained by splitting the patch list during the generation phase according to the date dimension: all sample keys are grouped by date and the training date set and validation date set are divided in a fixed ratio of 8:2, generating training and validation date sets respectively. patches .csv and val patches .csv, training only uses train patches .csv, validation only uses val patches .csv; During verification, the SR is obtained by forward processing batch by batch for each val patch and calculated within the mask: L1 val-batch =sum(|SR-HR|·mask) / (sum(mask)+eps), taking the mean of the entire validation set to obtain the L1 gradation. val If the current L1 val If the generator is better than the historical best, its weight G is saved. best .pt. At the same time, a fixed output visualization mesh (LR / SR / HR / differential 16-bit PNG mosaic) is used to quickly determine "whether it is closer to FY3D, whether it is blurry or has false textures".

[0029] Simultaneously, during the inference phase, arbitrary FY4B single-channel LR data undergoes the same cropping and normalization as during training (using the same channel lo / hi). The cropped LR is then fed into the generator G, directly outputting an SR result consistent with the low-resolution target size (internal ×4 upsampling is performed). Subsequently, the effective mask of the LR is upsampled to the SR size and used for invalid region processing in the output result (i.e., setting the invalid region to NaN or maintaining a consistent fill value). The SR is then denormalized back to the physical quantity numerical domain and saved as a computable super-resolution product. At the same time, a 16-bit image is output for quick visual verification by the business side, thereby improving the spatial resolution and detail representation of FY4B single channel, making the output texture and shape closer to the FY3D observation effect.

[0030] It should be noted that this invention does not simply enlarge FY4B pixels, but rather utilizes high-resolution FY3D data for supervised learning. An ESRGAN generator performs nonlinear mapping on the low-resolution input, thereby reconstructing high-frequency details and edge structures that interpolation methods cannot produce, making the result visually closer to FY3D. Furthermore, to address missing data and outliers in remote sensing data, an effective pixel mask and a "maximum effective pixel intersection" patch selection mechanism are introduced, along with global quantile normalization. This avoids invalid pixels interfering with training and improves convergence stability and generalization ability. Simultaneously, through relative discriminant analysis and multiple combined losses (pixel / gradient / frequency domain / statistical / adversarial), the invention enhances detail representation while maintaining a generally reasonable overall structure, significantly outperforming the enhancement effect of conventional interpolation methods, which "can only enlarge but remain blurry."

[0031] The beneficial effects of this invention are as follows: 1) Using the ESRGAN structure, super-resolution reconstruction of single-channel low-resolution remote sensing images is performed through an improved generator and discriminator. The generator learns the residuals of high-frequency details and low-frequency structures, which solves the problem of high-frequency texture and detail that is difficult to recover by traditional interpolation methods. 2) By calculating the effective pixel ratio and integral image for efficient screening, it is ensured that the input patches in the network training have sufficient effective pixel areas, avoiding the impact of overly smoothed low-quality samples on network convergence; 3) A combined loss function was designed, which includes L1 pixel reconstruction loss, perceptual loss, frequency domain loss and adversarial loss, improving the super-resolution results in terms of spatial resolution, texture structure and frequency domain characteristics. 4) Through a sophisticated data processing and evaluation mechanism, the data quality during the training process and the accuracy of the super-resolution results during the inference stage are ensured; 5) Smoothing labels and adversarial training optimization methods were adopted in the training process, which not only improved the stability of the training process, but also effectively suppressed the generation of pseudo high frequencies, further improving the naturalness and realism of the images. 6) To address the low resolution and lack of high-frequency details in the remote sensing data from Fengyun-4B (FY4B) satellite, a highly adaptable super-resolution method is proposed. This method is customized for different meteorological information in the remote sensing images, overcoming the limitation of existing super-resolution methods that are difficult to adapt to the special characteristics of remote sensing data.

[0032] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0033] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN, characterized in that, Includes the following steps: The resolution data of Fengyun-4B and Fengyun-3D satellites were analyzed to obtain a pairing index table; The paired index table is uniformly pruned, scaled, and quality checked to obtain preprocessed data; Generate the maximum effective pixel intersection mask based on the pairing index table and the preprocessed data; The effective pixel ratio is calculated based on the maximum effective pixel intersection mask, and patch filtering is performed to obtain training data; An ESRGAN network structure is constructed, and the training data is reconstructed using the ESRGAN network structure to obtain the predicted image.

2. The method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN according to claim 1, characterized in that, The paired index table is uniformly pruned, scaled, and quality checked to obtain preprocessed data, including: The paired index table is cropped at a fixed resolution to obtain FY3D single-channel image blocks; Calculate the total number of pixels, the number of valid pixels, and the number of invalid pixels in the FY3D single-channel image block; The invalid pixel ratio is calculated based on the total number of pixels, the number of valid pixels, and the number of invalid pixels. The FY3D single-channel image block is filtered for outliers based on the proportion of invalid pixels to obtain the preprocessed data.

3. The method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN according to claim 1, characterized in that, Generating a maximum effective pixel intersection mask based on the pairing index table and the preprocessed data includes: The paired index table is cropped at a fixed resolution to obtain a low-resolution FY4B image; Low-resolution mask and high-resolution mask are generated based on the FY4B low-resolution image and the preprocessed data, respectively; The low-resolution mask is mapped to the high-resolution network by nearest neighbor upsampling to obtain the mapped mask; The maximum effective pixel intersection mask is calculated based on the mapping mask and the high-resolution mask.

4. The method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN according to claim 1, characterized in that, The effective pixel ratio is calculated based on the maximum effective pixel intersection mask, and patch filtering is performed to obtain training data, including: Training patches are generated on the maximum effective pixel intersection mask using a sliding window; the training patches include: high-resolution patches and low-resolution patches; The effective pixel percentage of the training patch is calculated using the integral image. The training patches are filtered based on the effective pixel ratio to obtain the training data.

5. The method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN according to claim 4, characterized in that, The effective pixel percentage of the training patch is calculated using the integral image, including: The training patch is zero-padded at the boundaries to obtain the integral image; Calculate the number of effective pixels and the total number of pixels based on the integral image; The effective pixel count is normalized and mapped using the lower and upper bounds of the channel data in the pairing index table. The percentage of effective pixels is calculated based on the normalized number of effective pixels and the total number of pixels.

6. The method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN according to claim 1, characterized in that, Constructing an ESRGAN network structure and reconstructing the training data using the ESRGAN network structure to obtain a predicted image includes: The initial features of the training data are extracted by a two-dimensional convolution with a kernel size of 3×3, a stride of 1, and padding of 1. Deep features of the initial features are extracted through a backbone network; the backbone network consists of multiple RRDBs, and each RRDB consists of multiple Residual Dense Blocks. The deep features are connected to the initial features via a global residual concatenation to obtain the fused features; The fused features are subjected to double upsampling and convolution processing to obtain the predicted image.

7. The method for single-channel super-resolution reconstruction of Fengyun-4B satellite based on ESRGAN according to claim 1, characterized in that, The ESRGAN network structure includes a generator and a discriminator; the loss function of the discriminator is a binary cross-entropy with logits, and the loss function of the generator is: L G =w pix ·L1+w perc ·L perc +w grad ·L grad +w freq ·L freq +w stat ·L stat +lambda adv ·L adv ;where w pix w perc w grad w freq w stat and lambda adv All are weighting coefficients, L1 is the pixel reconstruction loss, L perc To perceive loss, L grad For gradient loss, L freq For frequency domain loss, L stat To calculate the loss, L adv To combat the losses.

Images