A multi-source remote sensing image enhancement method, device, equipment and storage medium

Abstract

This application discloses a method, apparatus, device, and storage medium for multi-source remote sensing image enhancement. The method includes: acquiring multi-source remote sensing images and preprocessing the images to construct an initial image sample library; performing distortion removal processing on a first-resolution image in the initial image sample library and correcting processing on a second-resolution image to obtain a paired image sample library; constructing a deep learning resolution enhancement network using a stacked architecture combined with an attention mechanism; training the deep learning resolution enhancement network using the target image sample library using a progressive training strategy; and inputting the remote sensing image to be enhanced into the fully trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images. This invention, through sample library construction, distortion removal, network design, and training strategies, enables the generated images to simultaneously achieve high spatial and temporal resolution, enhancing the model's generalization ability and enhancement effect.

Description

Technical Field

[0001] This application relates to the field of image enhancement technology, and more specifically, to a method, apparatus, device, and storage medium for enhancing multi-source remote sensing images. Background Technology

[0002] Remote sensing imagery is a crucial means of acquiring geographic information. With the rapid development of remote sensing technology, multi-source remote sensing imagery data is becoming increasingly abundant, providing vital data support for ground feature identification and dynamic monitoring. However, due to limitations in satellite sensor hardware performance, a single remote sensing data source cannot simultaneously meet the dual requirements of high spatial resolution and high temporal resolution. How to effectively integrate the advantages of multi-source remote sensing imagery to achieve a synergistic improvement in spatial detail and timeliness has become a key issue that urgently needs to be addressed in the field of remote sensing image processing.

[0003] In existing technologies, the following conventional approach is typically used for resolution enhancement of multi-source remote sensing images: First, remote sensing images of the same area at different resolutions are acquired, and training sample pairs are constructed through manual selection or simple registration; second, low-resolution images are directly used as input, and a super-resolution reconstruction model based on convolutional neural networks is used for feature extraction and upsampling; finally, the network is trained end-to-end using a single pixel-level loss function to output the enhanced high-resolution image.

[0004] However, the aforementioned existing technologies have some shortcomings in practical applications. Regarding sample construction, conventional methods lack automated screening mechanisms for sample quality and fail to effectively correct for radiometric differences between images from different sensors, thus limiting the model's generalization ability and enhancement effect. In the preprocessing stage, existing technologies directly use low-resolution images containing inherent mosaic distortion as input, without designing specific distortion removal schemes for the imaging mechanisms of specific sensors, resulting in jagged edges and geometric distortions in the enhanced images. In terms of network structure design, traditional convolutional neural networks struggle to capture the long-range dependencies widely present in remote sensing images, while directly introducing the Transformer model faces excessive computational complexity and lacks targeted enhancement mechanisms for multi-scale detail information. Regarding training strategies, the constraints of a single loss function often lead to overly smooth enhancement results that lose realistic texture, while directly using generative adversarial networks for training is prone to mode collapse and gradient explosion, making it difficult to guarantee training stability and success rate. Summary of the Invention

[0005] To address at least one deficiency or improvement need in the existing technology, the present invention provides a multi-source remote sensing image enhancement method, apparatus, device, and storage medium, which solves the problem that the existing technology is unable to simultaneously achieve the synergistic enhancement of high spatial resolution and high temporal resolution, and has deficiencies in sample library construction, distortion preprocessing, network structure design, and training strategies, resulting in severe geometric distortion of the enhanced image, insufficient detail fidelity, weak model generalization ability, and poor training stability.

[0006] To achieve the above objectives, according to a first aspect of the present invention, a multi-source remote sensing image enhancement method is provided, comprising: Collect multi-source remote sensing images and preprocess them to construct an initial image sample library; Distortion removal processing is performed on the first-resolution images in the initial image sample library, and correction processing is performed on the second-resolution images to obtain a paired image sample library. Construct a deep learning resolution enhancement network that employs a stacked architecture combined with an attention mechanism; A progressive training strategy is adopted, and the deep learning resolution enhancement network is trained using a target image sample library. The remote sensing images to be enhanced are input into a well-trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images.

[0007] In one possible implementation, acquiring and preprocessing multi-source remote sensing images to construct an initial image sample library further includes: Multi-source remote sensing images are collected from the same geographical area within a preset collection time range and with cloud cover meeting the requirements at different resolutions. Geometric registration and radiometric consistency correction were performed on the acquired multi-source remote sensing images of different resolutions. The sliding window method is used to crop the corrected multi-source remote sensing images to generate multiple original image pairs of preset sizes. The original image pairs are automatically screened for quality, and image pairs that meet the quality standards are retained to form an initial image sample library.

[0008] In one possible implementation, distortion removal processing is performed on the first-resolution images in the initial image sample library, and correction processing is performed on the second-resolution images to obtain a paired image sample library, which further includes: Identify the first-resolution image from the first sensor in the initial image sample library. The first-resolution image has inherent mosaic distortion due to the imaging process. A gridded local average filter model was established to simulate the formation mechanism of mosaic distortion, and an adaptive guided filtering algorithm was used to dynamically adjust the filtering parameters of the first resolution image to eliminate grid-like artifacts and preserve the boundary information of ground features. Radiometric consistency correction is performed on the second-resolution images from the second sensor in the initial image sample library to ensure that the second-resolution images are consistent with the first-resolution images after distortion removal in terms of radiometric characteristics. High-frequency residual information is extracted between the first-resolution image after adaptive guided filtering and the original first-resolution image. The high-frequency residual information is then correlated with the second-resolution image after radiometric correction to form a paired image sample library containing the first-resolution image after distortion removal and the second-resolution image after correction.

[0009] In one possible implementation, constructing a deep learning resolution enhancement network employing a stacked architecture combined with an attention mechanism also includes: Construct a network backbone consisting of multiple codecs stacked sequentially, where each codec contains sequentially connected encoding and decoding stages; In the encoding stage of each codec, the feature map resolution is gradually reduced through multiple downsampling operations, forming multiple scale levels with resolution from high to low; In the decoding stage of each codec, the feature map resolution is gradually restored through multiple upsampling operations, forming multiple reconstruction scales corresponding to the scale levels of the encoding stage; Within each codec, the upsampled outputs of each level in the decoding stage are fused with the original feature maps of the corresponding levels in the encoding stage across levels. In a multi-codec architecture, the output feature map of the previous codec is used as the input of the next codec.

[0010] In one possible implementation, within each codec, the upsampled outputs of each level in the decoding stage are fused across levels with the original feature maps of the corresponding levels in the encoding stage, and the process further includes: At each scale level of the codec, the original feature map output from the encoding stage is fused with the feature map obtained after upsampling in the decoding stage to form a fused feature map. Features at different spatial locations and in different feature channels in the fused feature map are recalibrated separately. The feature maps with different recalibrations are fused to output a cross-layer fused feature map.

[0011] In one possible implementation, a progressive training strategy is employed, using a target image sample library to train the deep learning resolution enhancement network, and further includes: In the first training phase, a first comprehensive loss function is constructed by combining loss functions of multiple constraint types to optimize the deep learning resolution enhancement network and learn the basic structural mapping relationship of images at different resolutions. In the second training phase, a discriminator network is introduced for adversarial optimization. At least one loss function from the first phase is retained as a constraint term, which together with the adversarial loss constitutes the second comprehensive loss function for optimization.

[0012] In one possible implementation, during the first training phase, a first comprehensive loss function is constructed by combining loss functions of multiple constraint types, and also includes: Obtain the enhanced image output by the deep learning resolution enhancement network and the corresponding resolution labeled image from the target image sample library; Calculate the difference between the enhanced image and the labeled image in pixel space, and construct a pixel-level spatial loss; Semantic features are extracted from both the enhanced and labeled images, and the difference between them in the feature space is calculated to construct a perceptual loss. The enhanced image and the labeled image are respectively converted to the frequency domain space, and the difference between the two in the frequency domain representation is calculated to construct the frequency domain loss; The pixel-level spatial loss, perceptual loss, and frequency domain loss are weighted and summed according to preset weights to form the first comprehensive loss function.

[0013] According to a second aspect of the present invention, a multi-source remote sensing image enhancement device is also provided, comprising: The initial sample module is configured to acquire multi-source remote sensing images and preprocess them to build an initial image sample library. The sample optimization module is configured to perform distortion removal processing on the first resolution image in the initial image sample library and correction processing on the second resolution image to obtain a paired image sample library. The network building module is configured to build deep learning resolution-enhancing networks that employ a stacked architecture combined with an attention mechanism. The network training module is configured to use a progressive training strategy to train the deep learning resolution enhancement network using a target image sample library. The image enhancement module is configured to input the remote sensing images to be enhanced into a well-trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images.

[0014] According to a third aspect of the present invention, a multi-source remote sensing image enhancement device is also provided, comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of any of the above-described multi-source remote sensing image enhancement methods.

[0015] According to a fourth aspect of the present invention, a storage medium is also provided, which stores a computer program executable by a multi-source remote sensing image enhancement device, which, when run on the multi-source remote sensing image enhancement device, causes the multi-source remote sensing image enhancement device to perform the steps of any of the above-described multi-source remote sensing image enhancement methods.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention provides a multi-source remote sensing image enhancement method. In terms of sample library construction, it constructs an initial image sample library by acquiring and preprocessing multi-source remote sensing images. This ensures that the training sample data covers various land cover types, avoiding the problem of insufficient generalization ability caused by single-source training data or class imbalance. Furthermore, by performing specialized distortion removal processing on the first-resolution images in the initial sample library and correction processing on the second-resolution images, a pairwise image sample library with artifacts removed is obtained. This removes the grid-like distortion caused by sensor pixel aggregation technology from the low-resolution images input to the enhancement network, preventing the network from mistaking mosaic grids for real textures during the learning process and thus preventing jagged edges and regular artifacts in the enhanced images. In terms of network structure design, a deep learning resolution enhancement network using a stacked architecture combined with an attention mechanism is constructed. This allows for repeated feature refinement and correction during multiple encoding and decoding processes. Simultaneously, the attention mechanism selectively enhances features from both spatial and channel dimensions, effectively capturing long-distance dependencies in remote sensing images. In terms of training strategy, the network is optimized in stages using a target image sample library. In the first stage, the focus is on learning accurate structural mapping relationships and laying a stable training foundation. In the second stage, an adversarial mechanism is introduced to improve visual realism. At the same time, the loss function from the previous stage is retained as a constraint to prevent deviation, thus ensuring stable convergence of the training process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an embodiment of the multi-source remote sensing image enhancement method provided by the present invention; Figure 2 A schematic diagram of a structure of an embodiment of the multi-source remote sensing image enhancement device provided by the present invention; Figure 3 This is a schematic diagram of the structure of a multi-source remote sensing image enhancement device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0021] This invention provides a method, apparatus, device, and storage medium for enhancing multi-source remote sensing images, which will be described below.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the multi-source remote sensing image enhancement method provided by the present invention. In a specific embodiment of the present invention, a multi-source remote sensing image enhancement method is disclosed, comprising: S101. Acquire multi-source remote sensing images and preprocess them to construct an initial image sample library; S102. Distortion removal processing is performed on the first resolution image in the initial image sample library, and correction processing is performed on the second resolution image to obtain a paired image sample library. S103. Construct a deep learning resolution enhancement network using a stacked architecture combined with an attention mechanism; S104. A progressive training strategy is adopted, and the deep learning resolution enhancement network is trained using the target image sample library. S105. Input the remote sensing image to be enhanced into a well-trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images.

[0023] In the above embodiments, based on the principle of spatiotemporal consistency, remote sensing images from different sources and at different resolutions are acquired for the same geographical area, including high spatial resolution images and their corresponding high temporal resolution images. During the acquisition process, the acquisition time window for image pairs needs to be controlled to ensure that changes in ground features are within an acceptable range, while image pairs with cloud cover below a threshold and no obvious sensor anomalies are selected. Subsequently, geometric registration processing is performed on the acquired multi-source images, and a high-precision registration algorithm is used to achieve sub-pixel-level alignment, ensuring that the position of the same ground feature corresponds accurately on different images.

[0024] Based on this, radiometric consistency correction is performed on the registered images to eliminate radiometric deviations caused by differences in atmospheric conditions or sensor response characteristics, making images from different sources comparable in radiometric features. After correction, a sliding window cropping method is used to generate multiple image patch pairs of preset sizes from the image pairs, and an automated quality assessment is performed on each image patch pair to select samples with high clarity and low noise levels. Finally, land cover category identification and statistics are performed on the selected samples, and the number of samples for different types of land cover is adjusted through oversampling or undersampling techniques to form an initial image sample library covering multiple typical land cover types such as water bodies, woodlands, farmland, buildings, and roads, with a balanced distribution of samples in each category.

[0025] The initial sample library is composed of remote sensing images from specific sensors (such as Sentinel-2). These images inherently exhibit grid-like mosaic distortion due to the pixel-gathering imaging process. For these images, a gridded local averaging filter model is established to simulate the distortion formation mechanism and clarify the distribution and manifestation of distortion on the image. Subsequently, an adaptive guided filtering algorithm is used to process the distorted images. The filtering intensity is dynamically adjusted based on the gradient characteristics of local image regions. Stronger smoothing is applied to uniform areas within the image to eliminate the mosaic effect, while weaker filtering is maintained at the edges of ground features to preserve true boundary information. After filtering, high-frequency residual information between the original distorted image and the filtered image is further extracted. This residual information contains potential details that were smoothed out by the filter. After normalization, the high-frequency residual information is used as an auxiliary guiding channel and fused with the filtered image to form enhanced input data containing original structural information and potential detail cues. Finally, the target image sample library after distortion removal is obtained.

[0026] A network backbone is constructed by stacking multiple codecs, each containing sequentially connected encoding and decoding stages. In the encoding stage of each codec, the feature map resolution is progressively reduced through multiple downsampling operations, forming multiple scale levels with varying resolutions, thereby extracting semantic information and structural features of the image at different scales. In the decoding stage of each codec, the feature map resolution is progressively restored through multiple upsampling operations, forming multiple reconstruction scales corresponding to the scale levels of the encoding stage. Within each codec, the upsampled outputs of each level of the decoding stage are fused with the original feature maps of the corresponding levels of the encoding stage across levels, allowing for full interaction between high-level semantic information and low-level detail information.

[0027] Simultaneously, a detail enhancement module consisting of parallel spatial attention units and channel attention units is introduced onto the fused feature map: the spatial attention unit recalibrates the importance of different spatial locations in the feature map, enhancing the feature representation of key regions; the channel attention unit recalibrates the importance of different feature channels, enhancing the channel response sensitive to specific land cover types. The feature map enhanced by the dual attention mechanism is used as the input to subsequent layers. Between multiple codecs, the output feature map of the previous codec is used as the input to the next codec, forming a cascaded refined structure, enabling the network to gradually correct reconstruction errors and refine the enhancement results during multiple iterations.

[0028] The training process is divided into two phases. In the first phase, a first comprehensive loss function is constructed by combining pixel-level spatial loss, perceptual loss, and frequency domain loss to optimize the network. Pixel-level spatial loss constrains the difference in pixel values ​​between the enhanced image and the ground truth label, ensuring basic reconstruction accuracy. Perceptual loss calculates the distance between the enhanced image and the ground truth label in a high-level feature space by inputting them into a pre-trained remote sensing feature extraction network, ensuring semantic consistency between the enhanced image and the ground truth label. Frequency domain loss transforms the enhanced image and the ground truth label into the frequency domain, constraining the difference in high-frequency components and guiding the network to focus on reconstructing detailed textures. Through the joint constraints of multiple loss types, the network initially focuses on learning the accurate structural mapping relationship from low-resolution to high-resolution images.

[0029] After completing the first training phase, the second training phase begins. A multi-scale discriminator network is introduced to distinguish between genuine and fake augmented images. A generative adversarial loss function is used to adversarially optimize the network, improving the visual realism and texture naturalness of the augmented images. Simultaneously, the perceptual loss from the first phase is retained as a regularization term in the second training phase, forming a second comprehensive loss function together with the adversarial loss to prevent the learned accurate structure from being destroyed during adversarial optimization. Through two phases of progressive training, the network gradually enhances visual realism while maintaining structural accuracy, ultimately achieving a balance between pixel precision and visual quality.

[0030] Specifically, the first step is to determine the source sensor type of the low-resolution remote sensing image to be processed. This image can originate from the Sentinel-2 satellite, the Gaofen series satellites, or other types of remote sensing satellites. Based on the source sensor type, a deep learning resolution enhancement network model that has completed two-stage training and converged is selected to match this data source. Because different sensor images have different imaging mechanisms, spectral response characteristics, and distortion patterns, enhancement network models trained for different data sources cannot be used interchangeably. It is essential to ensure that the low-resolution image used in the inference stage and the low-resolution image used in the training stage originate from the same sensor or have highly consistent imaging characteristics.

[0031] The image to be enhanced is input into a fully trained network model based on the corresponding data source. During the forward propagation of the network, the image first undergoes progressive feature extraction and refinement through multiple stacked encoder-decoder subnetworks. In each subnetwork, global semantic features and local structural features are extracted through multi-scale downsampling in the encoding stage, and spatial resolution is restored through multi-scale upsampling in the decoding stage. At each scale level, key details are enhanced through cross-level fusion and dual attention mechanisms. After multiple iterations of refinement, the final layer of the network outputs a high-resolution image reconstructed through sub-pixel convolutional layer upsampling. The final high-resolution image has the same spatial size as the target resolution, with clear boundaries of ground features, rich texture details, and a natural and realistic overall visual effect. It can be directly used for subsequent applications such as remote sensing interpretation, ground feature recognition, or dynamic monitoring.

[0032] In each codec, global semantic features and local structural features are extracted through multi-scale downsampling during the encoding stage, and spatial resolution is restored through multi-scale upsampling during the decoding stage. Furthermore, key detail representation is enhanced at each scale level through cross-level fusion and dual attention mechanisms. After multiple iterations of refinement, the final layer of the network outputs a high-resolution image reconstructed through sub-pixel convolutional layer upsampling. The resulting high-resolution image has the same spatial dimensions as the target resolution, with clear boundaries of ground features, rich texture details, and a natural and realistic overall visual effect. It can be directly used for applications such as remote sensing interpretation, ground feature identification, or dynamic monitoring.

[0033] Compared with existing technologies, this embodiment provides a multi-source remote sensing image enhancement method. In terms of sample library construction, it collects and preprocesses multi-source remote sensing images to build an initial image sample library. This ensures that the training sample data covers multiple land cover types, avoiding the problem of insufficient generalization ability caused by single training data or class imbalance. Furthermore, by performing specialized distortion removal processing on remote sensing images containing inherent mosaic distortion in the initial sample library, a target image sample library with artifacts removed is obtained. This removes the grid-like distortion caused by sensor pixel aggregation technology from the low-resolution images input to the enhancement network, preventing the network from mistaking the mosaic grid for real texture during the learning process and thus preventing jagged edges and regular artifacts in the enhanced image. In terms of network structure design, a deep learning resolution enhancement network using a stacked architecture combined with an attention mechanism is constructed. This allows for repeated refinement and correction of features during multiple encoding and decoding processes. Simultaneously, the attention mechanism selectively enhances features from both spatial and channel dimensions, effectively capturing long-distance dependencies in remote sensing images. In terms of training strategy, the network is optimized in stages using a target image sample library. In the first stage, the focus is on learning accurate structural mapping relationships and laying a stable training foundation. In the second stage, an adversarial mechanism is introduced to improve visual realism. At the same time, the loss function from the previous stage is retained as a constraint to prevent deviation, thus ensuring stable convergence of the training process.

[0034] In some embodiments of the present invention, acquiring multi-source remote sensing images and preprocessing the multi-source remote sensing images to construct an initial image sample library further includes: Multi-source remote sensing images are collected from the same geographical area within a preset collection time range and with cloud cover meeting the requirements at different resolutions. Geometric registration and radiometric consistency correction were performed on the acquired multi-source remote sensing images of different resolutions. The sliding window method is used to crop the corrected multi-source remote sensing images to generate multiple original image pairs of preset sizes. The original image pairs are automatically screened for quality, and image pairs that meet the quality standards are retained to form an initial image sample library.

[0035] In the above embodiments, image downloads are performed for the same geographical area based on spatiotemporal consistency. For images with precise timestamps, such as Sentinel-2 and Gaofen series images, the acquisition time interval is controlled within a preset time, such as within 14 days. For quarterly images, such as Jilin series images, the acquisition time of the corresponding low-resolution images is controlled within a time difference of 7 days in the middle of the quarter, while ensuring that cloud cover is ≤10% and there are no obvious sensor anomalies.

[0036] In a preferred embodiment, during the geometric registration stage, this invention employs the RoMa correction algorithm to achieve sub-pixel-level fine registration, with the final registration error controlled within one pixel (based on high-resolution imagery). For radiometric consistency processing, the 6S atmospheric correction model is first applied to eliminate atmospheric effects, and then histogram matching technology is used to perform radiometric adjustments on each band of the multi-source remote sensing imagery, referencing high-resolution images from the same period and region.

[0037] In the image cropping stage, an adaptive gridded sliding window method is employed to generate image blocks of uniform size, while a 20% overlap rate is set to increase sample diversity. The quality screening stage introduces an automated evaluation system, retaining only high-quality image pairs with a signal-to-noise ratio ≥30dB based on multi-dimensional standards such as signal-to-noise ratio. A large database containing 180,000 standardized samples is constructed, covering at least eight typical land cover categories, including water bodies, woodlands, farmland, buildings, and roads. Oversampling technology ensures a balanced proportion of samples across each category. The entire process achieves complete automation from data acquisition to sample generation, significantly improving sample construction efficiency and quality consistency.

[0038] It should be noted that Sentinel-2 imagery refers to remote sensing imagery data acquired by the European Space Agency's Copernicus program Earth observation satellites. Its core advantage lies in its short revisit cycle; with the two satellites working together, a complete coverage image of the same area can be completed every 5 days, enabling high-frequency dynamic monitoring. While Sentinel-2 imagery has a spatial resolution of 10 meters in the visible light band, which can meet the needs of rapid inspection over large areas, the pixel aggregation technology used in its imaging process results in a grid-like mosaic distortion in the original images, and the relatively insufficient spatial resolution makes it difficult to clearly present the fine boundaries and textures of ground features.

[0039] The Gaofen series of images consists of image data acquired by civilian remote sensing satellites under the "High-Resolution Earth Observation System Major Project," primarily including Gaofen-2 and Gaofen-7. Gaofen-2, launched in 2014, boasts a panchromatic resolution of 0.8 meters and a multispectral resolution of 4 meters. Gaofen-7, launched in 2019, is a space-based stereo mapping satellite with a panchromatic resolution better than 0.8 meters. The spatial resolution of the Gaofen series reaches sub-meter levels, clearly revealing the morphology and texture of detailed targets such as water body boundaries, illegal structures, and sand dredgers. It provides intuitive color and outline data for high-precision interpretation; however, due to limitations in satellite revisit cycles, high-frequency dynamic monitoring is difficult to achieve.

[0040] The Jilin series of images mainly refers to commercial remote sensing image data acquired by the "Jilin-1" satellite constellation. This constellation, independently developed by Changguang Satellite Technology Co., Ltd., is currently the world's largest sub-meter resolution commercial remote sensing satellite constellation. Jilin-1 imagery boasts a spatial resolution of up to 0.5 meters, classifying it as sub-meter resolution high-resolution remote sensing imagery, capable of providing precise services for agriculture, forestry, water conservancy, environmental protection, and other fields. In the technical solution, Jilin-1 imagery appears as a high-resolution image source, working together with the Gaofen series of imagery to form a real-world label for enhanced targets.

[0041] High-resolution (HMR) and low-resolution (LMR) imagery are relative classifications based on their spatial resolution. HMR imagery typically refers to remote sensing images with a spatial resolution below 5 meters, even reaching sub-meter levels. Its characteristics include extremely rich detail and clear ground feature boundaries, making it suitable for refined target identification and high-precision interpretation. However, it has higher data acquisition costs and longer revisit cycles, making it difficult to meet the needs of high-frequency monitoring. Low-resolution (LMR) imagery typically refers to remote sensing images with a spatial resolution of 10 meters or higher. Its characteristics include wide coverage, short revisit cycles, and low data acquisition costs, enabling high-frequency dynamic monitoring. However, insufficient spatial resolution leads to blurred ground feature boundaries and unclear textures. In this technical solution, LMR imagery serves as the input to the augmentation network, while HMR imagery serves as the reference for the augmented target's ground truth label. By learning the mapping relationship between the two, a synergistic enhancement of high temporal and spatial resolution is ultimately achieved.

[0042] In some embodiments of the present invention, distortion removal processing is performed on the first-resolution images in the initial image sample library, and correction processing is performed on the second-resolution images to obtain a paired image sample library, which further includes: Identify the first-resolution image from the first sensor in the initial image sample library. The first-resolution image has inherent mosaic distortion due to the imaging process. A gridded local average filter model was established to simulate the formation mechanism of mosaic distortion, and an adaptive guided filtering algorithm was used to dynamically adjust the filtering parameters of the first resolution image to eliminate grid-like artifacts and preserve the boundary information of ground features. Radiometric consistency correction is performed on the second-resolution images from the second sensor in the initial image sample library to ensure that the second-resolution images are consistent with the first-resolution images after distortion removal in terms of radiometric characteristics. High-frequency residual information is extracted between the first-resolution image after adaptive guided filtering and the original first-resolution image. The high-frequency residual information is then correlated with the second-resolution image after radiometric correction to form a paired image sample library containing the first-resolution image after distortion removal and the second-resolution image after correction.

[0043] In the above embodiments, this invention proposes a specialized preprocessing method to address the mosaic distortion problem unique to Sentinel-2 imagery. First, the distortion mechanism is analyzed in depth, and a gridded local averaging filter model is established to accurately simulate the pixel aggregation process. Based on this, an adaptive guided filtering algorithm is designed. This algorithm dynamically adjusts the filtering parameters according to the local gradient characteristics of the image, performing strong smoothing operations in uniform areas within the image to eliminate the mosaic block effect, while using weak filtering in the edge areas of ground features to maintain the integrity of true boundaries.

[0044] To further enhance detail recovery capabilities, this invention introduces a detail guidance mechanism. The high-frequency residual between the original low-resolution image and the preprocessed image is calculated, normalized, and used as auxiliary channel information. This information is then input into the subsequent enhancement network along with the main image, preventing the model from learning distortion artifacts and providing guidance information for the reconstruction of true details.

[0045] It should be noted that the first sensor refers to a remote sensing satellite sensor that uses pixel aggregation and gridding technology during the imaging process, resulting in inherent grid-like mosaic distortion in the output image. In a specific embodiment, the first sensor is represented by the multispectral imager of the Sentinel-2 satellite. Its imaging principle determines that the output low-resolution image contains blocky artifacts, requiring specialized distortion removal processing.

[0046] First-resolution imagery refers to the raw remote sensing imagery acquired and output by the first sensor. Its spatial resolution is relatively low. In a preferred embodiment, images with a resolution of 10 meters or higher are defined as first-resolution. This type of imagery is the input for resolution enhancement tasks. It is characterized by high temporal resolution and short revisit periods, meeting the needs of high-frequency dynamic monitoring. However, due to limitations in sensor imaging technology, it suffers from mosaic distortion, blurred ground feature boundaries, and unclear textures, making it unsuitable for direct high-precision interpretation.

[0047] The second sensor refers to a remote sensing satellite sensor with mature imaging technology capable of acquiring high spatial resolution images. Second sensors are represented by the imaging payloads of satellites such as Gaofen-2, Gaofen-7, and Jilin-1, whose output images achieve sub-meter spatial resolution (0.5 to 1 meter), displaying rich details and clear boundaries of ground features, serving as real-world labels for resolution enhancement tasks.

[0048] Second-resolution imagery refers to the raw remote sensing imagery acquired and output by a second sensor. It has a relatively high spatial resolution. In a preferred embodiment, images with a resolution of 1 meter or less are defined as second-resolution images. These images are the learning targets for resolution enhancement tasks. They are characterized by extremely rich spatial details and clear textures, making them suitable for refined target identification and high-precision interpretation. However, due to limitations in revisit cycles, they are difficult to monitor dynamically at high frequencies.

[0049] In some embodiments of the present invention, constructing a deep learning resolution enhancement network employing a stacked architecture combined with an attention mechanism further includes: Construct a network backbone consisting of multiple codecs stacked sequentially, where each codec contains sequentially connected encoding and decoding stages; In the encoding stage of each codec, the feature map resolution is gradually reduced through multiple downsampling operations, forming multiple scale levels with resolution from high to low; In the decoding stage of each codec, the feature map resolution is gradually restored through multiple upsampling operations, forming multiple reconstruction scales corresponding to the scale levels of the encoding stage; Within each codec, the upsampled outputs of each level in the decoding stage are fused with the original feature maps of the corresponding levels in the encoding stage across levels. In a multi-codec architecture, the output feature map of the previous codec is used as the input of the next codec.

[0050] In some embodiments of the present invention, within each codec, the upsampled outputs of each level of the decoding stage are fused across levels with the original feature maps of the corresponding levels of the encoding stage, further comprising: At each scale level of the codec, the original feature map output from the encoding stage is fused with the feature map obtained after upsampling in the decoding stage to form a fused feature map. Features at different spatial locations and in different feature channels in the fused feature map are recalibrated separately. The feature maps with different recalibrations are fused to output a cross-layer fused feature map.

[0051] In the above embodiments, the network architecture Hourglass-DetailTransformer (HG-DT) is adopted, which combines the hourglass stacking architecture and the dual-branch aggregation Transformer module design. The stacked fusion module adopts an encoder-decoder structure and realizes multi-level stacking.

[0052] The encoding stage constructs a multi-scale feature pyramid through progressive downsampling, fusing global and local features at different scales. The decoding stage performs upsampling reconstruction for each scale, introducing a bidirectional feature fusion mechanism. Each decoding layer not only receives upsampled features from the previous layer, but also fuses the original features of the same scale from the encoding stage through skip connections, while embedding a dedicated detail enhancement module (Detail Block).

[0053] Inspired by the Dual Aggregation Transformer (DAT), the detail enhancement module consists of residual modules composed of attention mechanisms and lightweight convolutions. It selectively enhances the most important high-frequency details at the current scale. The Dual Channel-attention Transformer Block (DCTB) and Dual Spatial-attention Transformer Block (DSTB) utilize spatial and channel self-attention Transformer blocks respectively, extracting features from two dimensions. For features at different scales, the network uses a three-layer residual convolutional module for modeling.

[0054] In some embodiments of the present invention, a progressive training strategy is employed, using a target image sample library to train the deep learning resolution enhancement network, and the method further includes: In the first training phase, a first comprehensive loss function is constructed by combining loss functions of multiple constraint types to optimize the deep learning resolution enhancement network and learn the basic structural mapping relationship of images at different resolutions. In the second training phase, a discriminator network is introduced for adversarial optimization. At least one loss function from the first phase is retained as a constraint term, which together with the adversarial loss constitutes the second comprehensive loss function for optimization.

[0055] In some embodiments of the present invention, the first training phase, which combines loss functions of multiple constraint types to form a first comprehensive loss function, further includes: Obtain the enhanced image output by the deep learning resolution enhancement network and the corresponding resolution labeled image from the target image sample library; Calculate the difference between the enhanced image and the labeled image in pixel space, and construct a pixel-level spatial loss; Semantic features are extracted from both the enhanced and labeled images, and the difference between them in the feature space is calculated to construct a perceptual loss. The enhanced image and the labeled image are respectively converted to the frequency domain space, and the difference between the two in the frequency domain representation is calculated to construct the frequency domain loss; The pixel-level spatial loss, perceptual loss, and frequency domain loss are weighted and summed according to preset weights to form the first comprehensive loss function.

[0056] In the above embodiments, a two-stage progressive training strategy is adopted to ensure the stability of the training process and the final model performance.

[0057] The first phase focuses on structural accuracy training. This phase does not introduce adversarial training, allowing the network to concentrate on learning accurate structural mappings and establishing a stable foundation. This phase accounts for approximately 70% of the total training epochs. Overall loss As shown in formula (1):

[0058] in Loss can be uniformly distributed across a noise region; perceived loss By extracting high-level semantic features through pre-trained convolutional neural networks, the limitations of pixel-level loss, which only focuses on pixel value differences, are avoided, and the enhanced image can maintain semantic consistency with the reference image.

[0059] To address the characteristics of remote sensing images, a VGG model fine-tuned with a large-scale remote sensing dataset was selected instead of a model pre-trained on natural images. The fine-tuned model is better able to identify ground features unique to remote sensing. Specifically, the output of the fourth convolutional layer of the VGG network was chosen as the feature map. The features at this layer have deviated from the original pixel information and focus on the structural and category attributes of ground features, which can effectively avoid over-smoothing of small target ground features during the optimization process.

[0060]

[0061] in, For the fine-tuned VGG Feature extraction function of layer, These represent the width, height, and number of channels of the feature map in this layer, respectively. The pixel coordinates and channel indexes of the corresponding feature map. To enhance the image, These are the true labels in the sample pairs.

[0062] Frequency loss transforms the image to the frequency domain, directly constraining the differences in high-frequency components and specifically optimizing detail recovery. Considering the characteristic of remote sensing images often exhibiting blocky landform structures (such as large areas of farmland or regular building areas), Discrete Cosine Transform (DCT) is chosen instead of Fourier Transform. By using L1 norm constraints to enhance the difference between the DCT-transformed image and the reference image, the model is forced to learn the generation rules of high-frequency details, avoiding detail loss or distortion. The formula is as follows:

[0063] in, It is a two-dimensional discrete cosine transform. The L1 norm can effectively suppress abnormal deviations in high-frequency components and improve the stability of detail recovery.

[0064] The second stage introduces adversarial training on a stable foundation to enhance visual realism. A multi-scale PatchGAN discriminator architecture is employed, which distinguishes between real and fake image patches at different scales, providing multi-layered adversarial supervision for the generator. Least Squares GAN (LSGAN) loss is used instead of standard GAN loss in the discriminator loss to effectively prevent mode collapse and gradient explosion. The capability balance between the generator and discriminator is monitored in real time during training; if one becomes too dominant, the learning rate is dynamically adjusted or updates are paused to ensure co-evolution. Simultaneously, the perceptual loss from the first stage is retained as a regularization term to prevent adversarial training from corrupting the learned accurate structure. Its overall expression is:

[0065] The learning rate is scheduled using a cosine annealing strategy and is reset to a smaller value at the beginning of the second phase for fine-tuning, accounting for approximately 30% of the total training rounds. The coefficients in formulas (1) and (4) are set as follows: =0.2、 =0.5、 =0.3、 =0.1、 =0.5, this training strategy achieves a good balance between pixel accuracy and visual quality while ensuring training stability.

[0066] To better implement the multi-source remote sensing image enhancement method in the embodiments of the present invention, based on the multi-source remote sensing image enhancement method, please refer to the corresponding... Figure 2 , Figure 2 This is a schematic diagram of an embodiment of the multi-source remote sensing image enhancement device provided by the present invention. The embodiment of the present invention provides a multi-source remote sensing image enhancement device 200, comprising: The initial sample module 210 is configured to acquire multi-source remote sensing images and preprocess the multi-source remote sensing images to build an initial image sample library. The sample optimization module 220 is configured to perform distortion removal processing on remote sensing images containing inherent mosaic distortion in the initial image sample library to obtain the target image sample library. Network building module 230 is configured to build a deep learning resolution enhancement network that employs a stacked architecture combined with an attention mechanism; The network training module 240 is configured to train the deep learning resolution enhancement network using a progressive training strategy and a target image sample library. The image enhancement module 250 is configured to input the remote sensing image to be enhanced into a well-trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images.

[0067] It should be noted that the device 200 provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be found in the corresponding content in the above method embodiments, and will not be repeated here.

[0068] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a multi-source remote sensing image enhancement device provided in an embodiment of the present invention. Based on the above-described multi-source remote sensing image enhancement method, the present invention also provides a multi-source remote sensing image enhancement device, which can be a computing device such as a mobile terminal, desktop computer, laptop, handheld computer, or server. The multi-source remote sensing image enhancement device 300 includes a processor 310, a memory 320, and a display 330. Figure 3 Only some components of the multi-source remote sensing image enhancement device are shown. However, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0069] In some embodiments, the memory 320 may be an internal storage unit of the multi-source remote sensing image enhancement device 300, such as a hard disk or memory of the multi-source remote sensing image enhancement device 300. In other embodiments, the memory 320 may be an external storage device of the multi-source remote sensing image enhancement device 300, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the multi-source remote sensing image enhancement device 300. Furthermore, the memory 320 may include both internal storage units and external storage devices of the multi-source remote sensing image enhancement device 300. The memory 320 is used to store application software and various types of data installed on the multi-source remote sensing image enhancement device 300, such as the program code of the multi-source remote sensing image enhancement device 300. The memory 320 may also be used to temporarily store data that has been output or will be output. In one embodiment, a multi-source remote sensing image enhancement program 340 is stored in the memory 320. This multi-source remote sensing image enhancement program 340 can be executed by the processor 310 to implement the multi-source remote sensing image enhancement method of the various embodiments of this application.

[0070] In some embodiments, processor 310 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in memory 320 or process data, such as executing multi-source remote sensing image enhancement methods.

[0071] In some embodiments, display 330 may be an LED display, a liquid crystal display, a touch-screen liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 330 is used to display information from the multi-source remote sensing image enhancement device 300 and to display a user interface for visualization. Components 310-330 of the multi-source remote sensing image enhancement device 300 communicate with each other via a system bus.

[0072] In one embodiment, when the processor 310 executes the multi-source remote sensing image enhancement program 340 in the memory 320, it implements the steps in the multi-source remote sensing image enhancement method described above.

[0073] This embodiment also provides a computer-readable storage medium storing a multi-source remote sensing image enhancement program, which, when executed by a processor, performs the following steps: Collect multi-source remote sensing images and preprocess them to construct an initial image sample library; Distortion removal processing is performed on remote sensing images containing inherent mosaic distortion in the initial image sample library to obtain the target image sample library; Construct a deep learning resolution enhancement network that employs a stacked architecture combined with an attention mechanism; A progressive training strategy is adopted, and the deep learning resolution enhancement network is trained using a target image sample library. The remote sensing images to be enhanced are input into a well-trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images.

[0074] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0075] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0076] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0078] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0079] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0080] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0081] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0082] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-source remote sensing image enhancement method, characterized in that, include: Collect multi-source remote sensing images and preprocess the multi-source remote sensing images to construct an initial image sample library; Distortion removal processing is performed on the first resolution image in the initial image sample library, and correction processing is performed on the second resolution image to obtain a paired image sample library. Construct a deep learning resolution enhancement network that employs a stacked architecture combined with an attention mechanism; A progressive training strategy is adopted, and the deep learning resolution enhancement network is trained using the paired image sample library. The remote sensing images to be enhanced are input into a well-trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images. The construction of the deep learning resolution enhancement network, which employs a stacked architecture combined with an attention mechanism, includes: Construct a network backbone consisting of multiple codecs stacked sequentially, where each codec contains sequentially connected encoding and decoding stages; In the encoding stage of each codec, the feature map resolution is gradually reduced through multiple downsampling operations, forming multiple scale levels with resolution from high to low; In the decoding stage of each codec, the feature map resolution is gradually restored through multiple upsampling operations, forming multiple reconstruction scales corresponding to the scale levels of the encoding stage; Within each codec, the upsampled outputs of each level in the decoding stage are fused with the original feature maps of the corresponding levels in the encoding stage across levels. Among multiple codecs, the output feature map of the previous codec is used as the input of the next codec; The upsampled outputs from each level of the decoding stage are fused across levels with the original feature maps from the corresponding levels of the encoding stage, including: At each scale level of the codec, the original feature map output from the encoding stage is fused with the feature map obtained after upsampling in the decoding stage to form a fused feature map. Features at different spatial locations and in different feature channels in the fused feature map are recalibrated separately. The feature maps with different recalibrations are fused to output a cross-layer fused feature map. The progressive training strategy, which trains the deep learning resolution enhancement network using the paired image sample library, includes: In the first training phase, a first comprehensive loss function is constructed by combining loss functions of multiple constraint types to optimize the deep learning resolution enhancement network and learn the basic structural mapping relationship of images at different resolutions. In the second training phase, a discriminator network is introduced for adversarial optimization. At least one loss function from the first phase is retained as a constraint term, which together with the adversarial loss constitutes the second comprehensive loss function for optimization.

2. The multi-source remote sensing image enhancement method of claim 1, wherein, The process of acquiring multi-source remote sensing images and preprocessing them to construct an initial image sample library includes: Multi-source remote sensing images are collected from the same geographical area within a preset collection time range and with cloud cover meeting the requirements at different resolutions. Geometric registration and radiometric consistency correction were performed on the acquired multi-source remote sensing images of different resolutions. The sliding window method is used to crop the corrected multi-source remote sensing images to generate multiple original image pairs of preset sizes. The original image pairs are subjected to automated quality screening, and image pairs that meet the quality standards are retained to form an initial image sample library.

3. The multi-source remote sensing image enhancement method of claim 1, wherein, The process involves performing distortion removal processing on the first-resolution images in the initial image sample library and correction processing on the second-resolution images to obtain a paired image sample library, including: Identify the first resolution image from the first sensor in the initial image sample library. The first resolution image has inherent mosaic distortion due to the imaging process. A gridded local average filter model is established to simulate the formation mechanism of the mosaic distortion, and an adaptive guided filtering algorithm is used to dynamically adjust the filtering parameters of the first resolution image to eliminate grid-like artifacts and preserve the boundary information of ground features. Radiometric consistency correction processing is performed on the second resolution images from the second sensor in the initial image sample library to make the second resolution images consistent with the first resolution images after distortion removal processing in terms of radiometric characteristics. High-frequency residual information is extracted between the first-resolution image after adaptive guided filtering and the original first-resolution image. The high-frequency residual information is then associated with the second-resolution image after radiometric correction to form a paired image sample library containing the first-resolution image after distortion removal and the second-resolution image after correction.

4. The multi-source remote sensing image enhancement method of claim 1, wherein, In the first training phase, a first comprehensive loss function is constructed by combining loss functions of multiple constraint types, including: Obtain the enhanced image output by the deep learning resolution enhancement network and the corresponding resolution labeled image from the paired image sample library; Calculate the difference between the enhanced image and the labeled image in pixel space, and construct a pixel-level spatial loss; Semantic features are extracted from the enhanced image and the labeled image respectively, and the difference between the two in the feature space is calculated to construct the perceptual loss; The enhanced image and the labeled image are respectively converted to the frequency domain space, and the difference between their frequency domain representations is calculated to construct the frequency domain loss; The pixel-level spatial loss, the perceptual loss, and the frequency domain loss are weighted and summed according to preset weights to form a first comprehensive loss function.

5. A multi-source remote sensing image enhancement device for implementing the multi-source remote sensing image enhancement method according to any one of claims 1 to 4, characterized in that, include: The initial sample module is configured to acquire multi-source remote sensing images and preprocess the multi-source remote sensing images to construct an initial image sample library. The sample optimization module is configured to perform distortion removal processing on the first resolution image in the initial image sample library and correction processing on the second resolution image to obtain a paired image sample library. The network building module is configured to build deep learning resolution-enhancing networks that employ a stacked architecture combined with an attention mechanism. The network training module is configured to train the deep learning resolution enhancement network using the paired image sample library with a progressive training strategy. The image enhancement module is configured to input the remote sensing images to be enhanced into a well-trained deep learning resolution enhancement network to generate enhanced multi-source remote sensing images.

6. A multi-source remote sensing image enhancement device, characterized in that, It includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of the multi-source remote sensing image enhancement method according to any one of claims 1 to 4.

7. A storage medium, characterized by It stores a computer program executable by a multi-source remote sensing image enhancement device, which, when run on the multi-source remote sensing image enhancement device, causes the multi-source remote sensing image enhancement device to perform the steps of the multi-source remote sensing image enhancement method according to any one of claims 1 to 4.

Images