Methods, systems, equipment and media for graded supervised super-resolution reconstruction of remote sensing images

Abstract

This invention discloses a hierarchical supervised super-resolution reconstruction method, system, device, and medium for remote sensing images. The method includes acquiring a high-resolution original remote sensing image, preprocessing it to obtain multiple sets of sub-pixel remote sensing images; constructing a sub-pixel remote sensing image reconstruction network, and pre-training it using a low-resolution remote sensing image and a set of sub-pixel remote sensing images as input and ground truth, respectively; constructing an upsampling-free remote sensing image super-resolution network, and optimizing it using multiple sets of sub-pixel images; constructing a multi-level supervised upsampling-free super-resolution network, and training it using an intermediate-resolution remote sensing image and a high-resolution original remote sensing image as ground truth, respectively; and performing super-resolution reconstruction based on the remote sensing image super-resolution reconstruction model, inputting the remote sensing image whose resolution needs to be improved, to obtain a high-resolution remote sensing image. This invention, by employing a hierarchical supervised and upsampling-free super-resolution reconstruction deep network, can improve the quality and detail of super-resolution reconstruction of remote sensing images.

Description

Technical Field

[0001] This invention relates to the field of image super-resolution reconstruction technology, and more specifically to a method, system, device, and medium for remote sensing image super-resolution reconstruction. Background Technology

[0002] High-resolution remote sensing images have wide applications in resource exploration, environmental monitoring, urban and rural planning, and change detection, providing an important data foundation for more accurate information extraction and analysis. However, limited by signal transmission bandwidth and imaging sensor capabilities, images acquired by remote sensing imaging equipment typically have low spatial resolution, which restricts the accuracy and effectiveness of remote sensing applications. Improving the resolution of remote sensing images at the hardware level is costly and time-consuming; therefore, image super-resolution reconstruction technology has significant research and application value.

[0003] Image super-resolution reconstruction technology uses one or more low-resolution images of the same scene as input, combined with prior knowledge of the images, to reconstruct a high-resolution image. This technology can effectively improve image resolution without changing existing hardware. Meanwhile, the development of deep learning has provided new technical means for remote sensing image super-resolution reconstruction. However, the upsampling techniques used in existing super-resolution reconstruction methods are prone to detail blurring.

[0004] Therefore, how to break away from existing upsampling techniques and construct an image super-resolution depth network to improve the quality and detail of low-resolution remote sensing images is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides an adaptive joint filtering scheme for removing band noise from hyperspectral remote sensing images.

[0006] To achieve the above objectives, this invention proposes a hierarchical supervised super-resolution reconstruction method for remote sensing images, comprising the following steps:

[0007] Acquire high-resolution raw remote sensing images, perform preprocessing, and obtain multiple sets of sub-pixel remote sensing images;

[0008] A subpixel remote sensing image reconstruction network was constructed, and pre-trained using a low-resolution remote sensing image and a set of subpixel remote sensing images as input and ground truth, respectively.

[0009] A super-resolution network for remote sensing images without upsampling is constructed. The network parameters for reconstructing subpixel remote sensing images obtained through pre-training are shared and optimized using multiple sets of subpixel images.

[0010] A multi-level supervised upsampling-free super-resolution network is constructed. The upsampling-free remote sensing image super-resolution network parameters obtained after optimization are shared. The network is trained using intermediate resolution remote sensing images and high-resolution original remote sensing images as ground values, respectively, to complete the construction of the remote sensing image super-resolution reconstruction model.

[0011] Based on the remote sensing image super-resolution reconstruction model, the input remote sensing image whose resolution needs to be improved is used for super-resolution reconstruction to obtain a high-resolution remote sensing image.

[0012] Furthermore, the preprocessing process yields multiple sets of sub-pixel remote sensing images, achieved by performing sub-pixel convolution on the high-resolution original remote sensing images.

[0013] Furthermore, a sub-pixel remote sensing image reconstruction network is constructed and pre-trained, as follows:

[0014] Constructing a subpixel remote sensing image reconstruction network;

[0015] The high-resolution original remote sensing image was downsampled twice to obtain the intermediate-resolution remote sensing image and the low-resolution remote sensing image.

[0016] The network is trained by using low-resolution remote sensing images as input and a set of subpixel remote sensing images as ground truth.

[0017] Furthermore, the non-upsampling remote sensing image super-resolution network includes multiple sets of sub-pixel remote sensing image reconstruction networks connected in parallel, and is connected in series at the end through convolutional layers to enhance the network's super-resolution reconstruction capability.

[0018] Moreover, the loss function of the non-upsampled remote sensing image super-resolution network consists of two levels of supervised loss, including the loss function of the feature extractor part and the loss function of the smoothing layer part.

[0019] Moreover, the multi-level supervised upsampling-free super-resolution network includes multiple upsampling-free remote sensing image super-resolution networks connected in series, and the network's loss function is composed of a weighted sum of the single-level loss functions of each level of the network.

[0020] Furthermore, the multi-level supervised upsampling-free super-resolution network includes a cascaded two-level upsampling-free remote sensing image super-resolution network.

[0021] On the other hand, the present invention also provides a hierarchical supervised super-resolution reconstruction system for remote sensing images, comprising the following modules:

[0022] The image acquisition module is used to acquire high-resolution raw remote sensing images, perform preprocessing, and obtain multiple sets of sub-pixel remote sensing images.

[0023] The subpixel remote sensing image reconstruction network construction module is used to construct the subpixel remote sensing image reconstruction network, and pre-trains it using a low-resolution remote sensing image and a set of subpixel remote sensing images as input and ground truth, respectively.

[0024] The module for constructing a super-resolution network for unsampling remote sensing images is used to build a super-resolution network for unsampling remote sensing images. It shares the network parameters for reconstructing subpixel remote sensing images obtained through pre-training and optimizes the network using multiple sets of subpixel images.

[0025] The multi-level supervised upsampling-free super-resolution network construction module is used to construct a multi-level supervised upsampling-free super-resolution network. It shares the upsampling-free remote sensing image super-resolution network parameters obtained after optimization, and uses intermediate resolution remote sensing images and high-resolution original remote sensing images as ground values ​​for training to complete the construction of the remote sensing image super-resolution reconstruction model.

[0026] The super-resolution reconstruction output module is used to perform super-resolution reconstruction based on the remote sensing image super-resolution reconstruction model. It takes the remote sensing image that needs to be improved as input and performs super-resolution reconstruction to obtain a high-resolution remote sensing image.

[0027] On the other hand, the present invention also provides an electronic device, including a memory and a processor, wherein the processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the remote sensing image hierarchical supervised super-resolution reconstruction method as described above by calling the program instructions.

[0028] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the remote sensing image hierarchical supervised super-resolution reconstruction method as described above.

[0029] This invention employs an upsampling-free approach to construct a deep network model, thereby improving the detail retention of the network model during super-resolution reconstruction. It also uses a multi-level supervision approach to construct and train the network, enhancing the network's ability to learn from the ground truth during training and improving the detail and quality of remote sensing image super-resolution reconstruction.

[0030] The present invention is simple and convenient to implement, highly practical, and solves the problems of low practicality and inconvenience in actual application of related technologies. It can improve user experience and has significant market value. Attached Figure Description

[0031] Figure 1 A flowchart of a hierarchical supervised super-resolution reconstruction method for remote sensing images provided in an embodiment of the present invention;

[0032] Figure 2This is a schematic diagram of sub-pixel shift decomposition provided in an embodiment of the present invention;

[0033] Figure 3 A schematic diagram of a subpixel remote sensing image reconstruction network provided in an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the structure of the non-upsampling remote sensing image super-resolution network provided in an embodiment of the present invention;

[0035] Figure 5 This is a schematic diagram of the structure of a multi-level supervised remote sensing image super-resolution network provided in an embodiment of the present invention;

[0036] Figure 6 This is a schematic diagram of a hierarchical supervised super-resolution reconstruction system for remote sensing images provided in an embodiment of the present invention. Detailed Implementation

[0037] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0038] In this invention, subpixel shift decomposition is performed on the acquired high-resolution original remote sensing images to obtain multiple sets of subpixel remote sensing image training sets, completing preprocessing. Based on the constructed subpixel remote sensing image reconstruction network, and using low-resolution remote sensing images and a set of subpixel remote sensing images as input and ground values ​​respectively for pre-training, high-resolution subpixel remote sensing image reconstruction based on low-resolution remote sensing images is achieved. Based on the constructed upsampling-free remote sensing image super-resolution network, the parameters of the pre-trained subpixel remote sensing image reconstruction network are shared, and the upsampling-free super-resolution network is optimized using multiple sets of subpixel images to achieve the construction of a single-level network. A multi-level supervised upsampling-free super-resolution network is constructed, and the upsampling-free remote sensing image super-resolution network parameters obtained after optimization are shared. The network is trained using intermediate-resolution remote sensing images and high-resolution original remote sensing images respectively as ground values ​​to complete the construction of the remote sensing image super-resolution reconstruction model. Based on the remote sensing image super-resolution reconstruction model, super-resolution reconstruction is performed on remote sensing images that need to be improved in resolution, which can effectively improve the quality and detail of the super-resolution reconstruction of remote sensing images.

[0039] like Figure 1 As shown in the figure, this invention discloses a method for super-resolution reconstruction of remote sensing images without upsampling, comprising the following steps:

[0040] Step 1: Acquire high-resolution raw remote sensing images, perform preprocessing to obtain multiple sets of sub-pixel remote sensing images, and use these multiple sets of sub-pixel remote sensing images as a training set.

[0041] The present invention further proposes that the specific implementation process of this step includes:

[0042] (1) Acquire high-resolution raw remote sensing images;

[0043] (2) Perform subpixel convolution on the high-resolution original remote sensing image to obtain multiple sets of subpixel remote sensing images and complete the preprocessing.

[0044] In the embodiment, subpixel shift decomposition is as follows: Figure 2 As shown, a high-resolution raw remote sensing image is decomposed into several groups using basic units of 2×2 pixels in length and width. The top-left, top-right, bottom-left, and bottom-right pixels of each basic unit (labeled 1, 2, 3, and 4 respectively) are then used to construct four sub-pixel images. The high-resolution raw remote sensing image has a resolution of 4n×4n, and the sub-pixel images have a resolution of 2n×2n. In other words, the acquired high-resolution raw remote sensing image undergoes sub-pixel shift decomposition to construct four sets of sub-pixel remote sensing image datasets, completing the preprocessing.

[0045] Step 2: Construct a subpixel remote sensing image reconstruction network, and pre-train it using a low-resolution remote sensing image and a set of subpixel remote sensing images as input and ground truth, respectively.

[0046] The present invention further proposes that the specific implementation process of this step includes:

[0047] (1) Construct a sub-pixel remote sensing image reconstruction network;

[0048] (2) The high-resolution original remote sensing image is downsampled twice to obtain the intermediate resolution remote sensing image and the low resolution remote sensing image.

[0049] (3) The sub-pixel remote sensing image reconstruction network is trained by taking low-resolution remote sensing images as input and a set of sub-pixel remote sensing images as ground values.

[0050] In this embodiment, a set of subpixel remote sensing image reconstruction networks is constructed and pre-trained;

[0051] Subpixel remote sensing image reconstruction network structure such as Figure 3 As shown in the figure, the network input is a low-resolution remote sensing image with a resolution of 2n×2n. After passing through the feature extractor, the network output is a sub-pixel remote sensing image.

[0052] The subpixel remote sensing image reconstruction network is pre-trained. The input for training is a low-resolution remote sensing image, which is obtained by downsampling the high-resolution original remote sensing image obtained in step 1 twice. A set of subpixel remote sensing images is used as the ground truth, which is any one of the four sets of subpixel remote sensing images obtained in step 1.

[0053] In this embodiment, there are no specific requirements for the subpixel remote sensing image reconstruction network, and the network infrastructure can be a convolutional neural network architecture or a Transformer network architecture.

[0054] Step 3: Construct a super-resolution network for remote sensing images without upsampling, share the network parameters for reconstructing subpixel remote sensing images obtained through pre-training, and optimize using multiple sets of subpixel images;

[0055] The present invention further proposes that the specific implementation process of this step includes:

[0056] (1) Construct a super-resolution network for remote sensing images without upsampling;

[0057] The unsampling remote sensing image super-resolution network consists of multiple sets of sub-pixel remote sensing image reconstruction networks connected in parallel and concatenated at the end by a smoothing layer;

[0058] (2) Use the pre-trained sub-pixel remote sensing image reconstruction network to share parameters within the super-resolution network of remote sensing images without upsampling.

[0059] (3) The super-resolution network for remote sensing images without upsampling is optimized by using multiple sets of subpixel images.

[0060] In this embodiment, a super-resolution network model for remote sensing images without upsampling is established and optimized based on a set of sub-pixel remote sensing image reconstruction networks constructed in step 2.

[0061] Unupsampled remote sensing image super-resolution network structure as follows Figure 4 As shown, the non-upsampled remote sensing image super-resolution network includes multiple sets of sub-pixel remote sensing image reconstruction networks in parallel (including feature extractors 1, 2, 3, and 4, respectively), and is cascaded at the end through convolutional layers to enhance the network's super-resolution reconstruction capability. The input image of the network is a low-resolution remote sensing image with a resolution of 2n×2n, which is obtained by downsampling the high-resolution original remote sensing image acquired in step 1 twice; the intermediate ground truth is four sets of sub-pixel remote sensing images with a resolution of 2n×2n; the output ground truth is a high-resolution original remote sensing image with a resolution of 4n×4n. The smoothing layer in this embodiment has no specific requirements and can be a convolutional layer.

[0062] During the model optimization phase, the pre-trained sub-pixel remote sensing image reconstruction network is first used as feature extractor 1. The parameters of feature extractor 1 are shared with feature extractors 2 through 4 to improve the network's convergence speed. The loss function of the un-upsampled remote sensing image super-resolution network consists of two levels of supervised loss. Specifically, the network's loss function is:

[0063] Loss = loss1 + loss2

[0064] Loss function for the feature extractor part: Where n is the number of feature extractors, j represents the feature extractor label, N is the number of network parameters in a single feature extractor, and i represents the feature extractor network parameter label. This represents the sub-pixel ground truth value of the remote sensing image under the i-th network parameter of the j-th feature extractor. f1() represents the input, and f1() represents the feature extraction mapping.

[0065] Loss function for the smoothing layer: Where M is the number of parameters in the smoothing layer, and n represents the index of the smoothing layer network parameter. Represents the true value of the original high-resolution remote sensing image. f2() represents the input to the smoothing layer, and f2() represents the smoothing layer mapping.

[0066] Step 4: Construct a multi-level supervised oversampling-free super-resolution network. Share the oversampling-free remote sensing image super-resolution network parameters obtained after optimization. Use intermediate resolution remote sensing images and high-resolution original remote sensing images as ground values ​​for training respectively. Once training is complete, the remote sensing image super-resolution reconstruction model is constructed.

[0067] The present invention further proposes that the specific implementation process of this step includes:

[0068] (1) Construct a multi-level supervised upsampling-free super-resolution network;

[0069] A multi-level supervised upsampling-free super-resolution network consists of multiple cascaded upsampling-free remote sensing image super-resolution networks. The network's loss function is a weighted sum of the single-level loss functions of each network level. In practice, the number of cascaded upsampling-free remote sensing image super-resolution networks, i.e., the number of levels, can be determined based on the specific task and experimental data.

[0070] (2) Use the optimized upsampling-free remote sensing image super-resolution network to share parameters within the multi-level supervised upsampling-free super-resolution network;

[0071] (3) The low-resolution remote sensing image is used as input, the intermediate-resolution remote sensing image is used as the intermediate ground truth, and the high-resolution original remote sensing image is used as the final ground truth to train the multi-level supervised unsampling super-resolution network.

[0072] In this embodiment, a multi-level supervised upsampling-free super-resolution network is established and trained based on the upsampling-free remote sensing image super-resolution network constructed in step 3.

[0073] A schematic diagram of a multi-level supervised, upsampling-free super-resolution network structure is shown below. Figure 5 In this embodiment of the invention, a two-stage supervised network is preferably employed. The input image to the network is a low-resolution remote sensing image, obtained by downsampling the high-resolution original remote sensing image acquired in step 1 twice. The first-stage output is a remote sensing image at twice the resolution, serving as an intermediate-resolution remote sensing image. The second-stage output is a remote sensing image at four times the resolution. The ground truth used for network training is the two-times-resolution remote sensing image and the four-times-resolution remote sensing image, respectively. The four-times-resolution remote sensing image is the high-resolution original remote sensing image obtained in step 1, and the two-times-resolution remote sensing image is obtained by downsampling the four-times-resolution remote sensing image once.

[0074] During the network training phase, the unupsampled remote sensing image super-resolution network parameters obtained in step 3 are first used as the first-level network parameters of the hierarchical supervised network and shared with the second-level network. Then, 2x and 4x resolution remote sensing images are used to train the multi-level supervised network, completing the construction of the multi-level supervised unupsampled remote sensing image super-resolution reconstruction model.

[0075] Step 5: Based on the remote sensing image super-resolution reconstruction model constructed in Step 4, input the remote sensing image whose resolution needs to be improved to perform super-resolution reconstruction and obtain a high-resolution remote sensing image.

[0076] In practice, the above process can be written into a program using Python, which has a fast processing speed and good effect. When the program is executed, it can realize the remote sensing image hierarchical supervised super-resolution reconstruction method as described above, thereby achieving image super-resolution.

[0077] In another possible embodiment, the present invention provides a hierarchical supervised super-resolution reconstruction system for remote sensing images, such as... Figure 6 As shown, it includes:

[0078] Image acquisition module: used to acquire high-resolution raw remote sensing images, perform preprocessing, and obtain multiple sets of sub-pixel remote sensing images;

[0079] Subpixel remote sensing image reconstruction network building module: used to build a subpixel remote sensing image reconstruction network, and pre-trained using a low-resolution remote sensing image and a set of subpixel remote sensing images as input and ground truth respectively;

[0080] Upsampling-free remote sensing image super-resolution network construction module: used to construct an upsampling-free remote sensing image super-resolution network, share the network parameters of the sub-pixel remote sensing images obtained through pre-training, and optimize them through multiple sets of sub-pixel images;

[0081] Multi-level supervised upsampling-free super-resolution network construction module: used to construct a multi-level supervised upsampling-free super-resolution network. It shares the upsampling-free remote sensing image super-resolution network parameters obtained after optimization, and uses intermediate resolution remote sensing images and high-resolution original remote sensing images as ground values ​​for training to complete the construction of the remote sensing image super-resolution reconstruction model.

[0082] Super-resolution reconstruction output module: Used to perform super-resolution reconstruction on remote sensing images that need to be improved based on the remote sensing image super-resolution reconstruction model, so as to obtain high-resolution remote sensing images.

[0083] The specific implementation process of each module in this system is the same as that in the above-described method embodiment, and will not be repeated here.

[0084] In another possible embodiment, the present invention also provides a computer device capable of super-resolution reconstruction of remote sensing images, which can be implemented by software and / or hardware. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, it implements the steps of the above-described hierarchical supervised super-resolution reconstruction method for remote sensing images.

[0085] Specifically, the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0086] In another possible embodiment, the present invention 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 remote sensing image hierarchical supervised super-resolution reconstruction method.

[0087] As a non-transitory computer-readable storage medium, memory can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program or instructions corresponding to the remote sensing image hierarchical supervision super-resolution reconstruction method in the above embodiments.

[0088] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function. The data storage area may store data created by the processor, etc.

[0089] In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device.

[0090] In some aspects, the memory may optionally include memory remotely located relative to the processor, which may be connected to the processor via a network.

[0091] Optionally, the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0092] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0093] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0094] The specific examples described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific examples or use similar methods to replace them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A hierarchical supervised super-resolution reconstruction method for remote sensing images, characterized in that, Includes the following steps: A high-resolution original remote sensing image is acquired and preprocessed to obtain multiple sets of sub-pixel remote sensing images; the preprocessing to obtain multiple sets of sub-pixel remote sensing images is achieved by performing sub-pixel convolution on the high-resolution original remote sensing image. A subpixel remote sensing image reconstruction network was constructed, and pre-trained using a low-resolution remote sensing image and a set of subpixel remote sensing images as input and ground truth, respectively. A non-upsampling remote sensing image super-resolution network is constructed, and the parameters of the sub-pixel remote sensing image reconstruction network obtained through pre-training are shared and optimized using multiple sets of sub-pixel images. The non-upsampling remote sensing image super-resolution network includes multiple sets of sub-pixel remote sensing image reconstruction networks connected in parallel, and concatenated at the end through convolutional layers to enhance the super-resolution reconstruction capability of the network. A multi-level supervised upsampling-free super-resolution network is constructed. The optimized upsampling-free remote sensing image super-resolution network parameters are shared. The network is trained using intermediate resolution remote sensing images and high-resolution original remote sensing images as ground truth, respectively, to complete the construction of the remote sensing image super-resolution reconstruction model. The multi-level supervised upsampling-free super-resolution network includes multiple upsampling-free remote sensing image super-resolution networks in series. The loss function of the network is composed of a weighted sum of the single-level loss functions of each level of the network. Based on the remote sensing image super-resolution reconstruction model, the input remote sensing image whose resolution needs to be improved is used for super-resolution reconstruction to obtain a high-resolution remote sensing image.

2. The hierarchical supervised super-resolution reconstruction method for remote sensing images according to claim 1, characterized in that: A subpixel remote sensing image reconstruction network is constructed and pre-trained, as follows: Constructing a subpixel remote sensing image reconstruction network; The high-resolution original remote sensing image was downsampled twice to obtain the intermediate-resolution remote sensing image and the low-resolution remote sensing image. The network is trained by using low-resolution remote sensing images as input and a set of subpixel remote sensing images as ground truth.

3. The hierarchical supervised super-resolution reconstruction method for remote sensing images according to claim 1, characterized in that: The loss function of the unupsampled remote sensing image super-resolution network consists of two levels of supervised loss, including the loss function of the feature extractor part and the loss function of the smoothing layer part.

4. The hierarchical supervised super-resolution reconstruction method for remote sensing images according to claim 1, characterized in that: The multi-level supervised upsampling-free super-resolution network consists of a cascaded two-level upsampling-free remote sensing image super-resolution network.

5. A hierarchical supervised super-resolution reconstruction system for remote sensing images, characterized in that, Includes the following modules, The image acquisition module is used to acquire high-resolution raw remote sensing images, perform preprocessing, and obtain multiple sets of sub-pixel remote sensing images; the preprocessing to obtain multiple sets of sub-pixel remote sensing images is achieved by performing sub-pixel convolution on the high-resolution raw remote sensing images. The subpixel remote sensing image reconstruction network construction module is used to construct the subpixel remote sensing image reconstruction network, and pre-trains it using a low-resolution remote sensing image and a set of subpixel remote sensing images as input and ground truth, respectively. The module for constructing a non-upsampled remote sensing image super-resolution network is used to construct a non-upsampled remote sensing image super-resolution network. It shares the pre-trained subpixel remote sensing image reconstruction network parameters and optimizes the network using multiple sets of subpixel images. The non-upsampled remote sensing image super-resolution network includes multiple sets of subpixel remote sensing image reconstruction networks connected in parallel and connected in series at the end through convolutional layers to enhance the network's super-resolution reconstruction capability. A multi-level supervised upsampling-free super-resolution network construction module is used to construct a multi-level supervised upsampling-free super-resolution network. The optimized upsampling-free remote sensing image super-resolution network parameters are shared, and the network is trained using intermediate-resolution remote sensing images and high-resolution original remote sensing images as ground truth, respectively, to complete the construction of the remote sensing image super-resolution reconstruction model. The multi-level supervised upsampling-free super-resolution network includes multiple upsampling-free remote sensing image super-resolution networks connected in series, and the network's loss function is a weighted sum of the single-level loss functions of each network level. The super-resolution reconstruction output module is used to perform super-resolution reconstruction based on the remote sensing image super-resolution reconstruction model. It takes the remote sensing image that needs to be improved as input and performs super-resolution reconstruction to obtain a high-resolution remote sensing image.

6. An electronic device, characterized in that: The system includes a memory and a processor, which communicate with each other via a bus. The memory stores program instructions that can be executed by the processor, and the processor can execute the remote sensing image hierarchical supervised super-resolution reconstruction method as described in any one of claims 1 to 4 by calling the program instructions.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the remote sensing image hierarchical supervised super-resolution reconstruction method as described in any one of claims 1 to 4.

Images