Multispectral remote sensing image super-resolution reconstruction method and system based on knowledge distillation

Abstract

The present invention discloses a method and system for super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation, and the method is as follows: Step 1: Introduce high-resolution bands in a non-reconstructed spectral domain as prior knowledge to construct a training set. Step 2: Construct a teacher-student offline distillation network, wherein the teacher network has high complexity but good reconstruction quality, and the student network is lightweight and has slightly lower reconstruction quality. The two network architectures are the same but the number of residual blocks is different; Step 3: The super-resolution remote sensing image output by the student network is respectively compared with the teacher network output and label calculation loss to form a teacher supervision loss and a data supervision loss, and the two are weighted to calculate the total loss. Step 4: Train the student network through back propagation of the total loss. Step 5: Use the trained student network model to upgrade the low-resolution band of the multispectral remote sensing image to the highest resolution band. The present invention can effectively reduce the computing cost and model complexity to better meet the needs of practical applications.

Description

Technical Field

[0001] The present invention belongs to the technical field of super-resolution reconstruction of remote sensing images, is used for fast super-resolution reconstruction of multispectral remote sensing images, relates to a model compression technology based on knowledge distillation, and specifically is a method and system for super-resolution reconstruction of multispectral remote sensing images. Background Art

[0002] As the data support and application basis of remote sensing technology, remote sensing images provide rich information for monitoring the earth's surface and are widely used in many fields such as ground object detection, disaster warning, environmental change detection, and military strikes. Compared with general images, remote sensing images are usually farther away from the imaging distance, and are prone to harsh environments and various jitters when imaging at high altitudes, which makes them more seriously affected by noise and blur. Therefore, remote sensing images generally have low resolution and poor quality. At present, the most widely used technology to improve the resolution of remote sensing images from the software aspect is super-resolution reconstruction technology. Among the many super-resolution reconstruction schemes, the technology of reconstructing high-resolution images using the feature information of low-resolution images using deep learning algorithms is mature and widely used, with the advantages of high feasibility, high availability, strong generalization, and low cost. However, this scheme also has its defects: in order to achieve a better reconstruction effect, the model depth and parameter volume are increasing, which brings disadvantages such as high resource consumption, long training time, and limited usage scenarios. To this end, it is necessary to study more mature design models and training schemes to solve this problem.

[0003] According to the existing literature, deep residual networks and generative adversarial networks are the most widely used methods for super-resolution reconstruction of multispectral remote sensing images based on deep learning. Among them, Charis Lanaras et al. published "Super-resolution of Sentinel-2 images: Learning a globally applicable deep neural network" in ISPRS JOURNAL in 2018, proposing a Dsen2 model for multispectral remote sensing images Sentinel-2. The model can effectively retain the texture details and spectral features of the original image by fusing spatial spectral information with a deep residual network for global regression training. However, with the increase in the number of network layers, the extraction of deeper features also faces the disadvantages of slow training speed, too many parameters, and narrow application range. At the same time, Frosti Palsson et al. published "Single Sensor Image Fusion Using A Deep Convolutional Generative Adversarial Network" in TIP in 2018, proposing a super-resolution reconstruction method for multispectral remote sensing images based on deep convolutional generative adversarial networks, and obtained better experimental results than traditional methods. However, both its generator and discriminator use deep convolutional design, so it also has disadvantages such as long convergence time and difficulty in deployment on devices with limited resources.

[0004] In summary, the field of super-resolution reconstruction of multispectral remote sensing images faces the problems of how to reduce model complexity, speed up training, reduce resource consumption, and improve generalization ability while ensuring the quality of image reconstruction. Therefore, it is necessary to seek new breakthroughs in model design and algorithm optimization to achieve more efficient and reliable super-resolution reconstruction technology for multispectral remote sensing images. Based on this, the present invention adopts the framework of the deep residual network model and uses the knowledge distillation method for model compression, and proposes a technical solution for super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation. Summary of the invention

[0005] The present invention aims at the defects of the current multi-spectral remote sensing image super-resolution reconstruction model based on deep learning, such as high resource consumption and long training time, and combines the characteristics of different spatial resolutions of each band of multi-spectral remote sensing images to provide a multi-spectral remote sensing image super-resolution reconstruction method and system based on knowledge distillation. The present invention makes the performance of the student network close to that of the teacher network with less time cost, and can realize efficient super-resolution reconstruction tasks for multi-spectral remote sensing images.

[0006] The present invention adopts the following technical scheme:

[0007] A multispectral remote sensing image super-resolution reconstruction method based on knowledge distillation, which is as follows:

[0008] Step 1: Acquire a multispectral remote sensing image and perform data preprocessing; preferably, according to the characteristics of the multispectral remote sensing image, take preprocessing steps such as band grouping, random cropping, down / up sampling, etc. to obtain a low-resolution image;

[0009] Step 2: Introduce the high-resolution band in the non-reconstructed spectral domain as prior knowledge and combine it with the low-resolution image to construct a super-resolution reconstruction training set;

[0010] Step 3: Construct a teacher-student offline distillation network based on multispectral remote sensing images, in which a pre-trained teacher network is used, which has high complexity and large number of parameters, but good reconstruction quality. A lightweight student network is constructed based on the idea of ​​network residual block cropping, which has the same architecture as the teacher network, but has low complexity, small number of parameters, and lower reconstruction quality than the teacher network;

[0011] Step 4: Construct a knowledge distillation framework for multispectral remote sensing images, input the image training set described in step 2 into the teacher network and student network described in step 3, output super-resolution remote sensing images respectively, calculate the loss band by band for the two and take the average as the teacher supervision loss (knowledge distillation loss), prompting the student network to learn the characteristics of the teacher network; calculate the loss band by band for the super-resolution remote sensing image output by the student network and the training set label and take the average as the data supervision loss, prompting the student network to learn the characteristics of the high-resolution label. The total loss is calculated by adding the two according to the weights;

[0012] Step 5: Use the training set described in step 2, introduce the total loss described in step 4 as the loss function for back propagation, and train the student network;

[0013] Step 6: Use the multispectral remote sensing image as input and use the student network model trained in step 5 to upgrade the low-resolution band to the highest resolution band in the spectral domain.

[0014] Furthermore, in step 1-2, the bands in the multispectral remote sensing image are grouped according to spatial resolution, and then randomly cropped to different sizes according to the magnification. The size needs to correspond to a receiving field of at least several hundred meters on the ground, so as to capture the local low-level texture and potential small semantic structures. Then, each group of bands is downsampled by Gaussian blur processing, and some (settable) low-resolution bands are upsampled by bilinear interpolation to make the low-resolution image consistent with the resolution of prior knowledge. Finally, since the teacher-student network training adds prior knowledge, transfers texture and spectral features from existing high-resolution bands, and guides super-resolution with high resolution, the constructed training set is not a traditional end-to-end low-resolution-high-resolution training set. The model training input is composed of the high-resolution prior band sampling and the low-resolution reconstructed band, and the label is the original resolution band.

[0015] Furthermore, in step 3, the pre-trained residual network Dsen2 is used as the teacher network. The main components of the teacher network structure are the convolution layer (Conv), the non-linear activation function (ReLU), and the residual neural network module (Resblock). The student network shares the same architecture with the teacher network, and the number of parameters is relatively small by reducing the number of residual blocks and channels.

[0016] Furthermore, in step 4, the teacher supervision loss and data supervision loss are calculated using the root mean square error RMSE, and the teacher supervision loss is used as the knowledge distillation loss to implement the knowledge distillation operation; the teacher supervision loss and the data supervision loss are added according to the weights to obtain the total loss, and the total loss is used as the network training error. After updating the iterative training, the final student network model is obtained, so that the student model can learn the characteristic knowledge of the teacher model to the greatest extent. The loss function is the root mean square error RMSE, and the model training is guided by pixel-level error back propagation, which is specifically expressed as:

[0017] L=αL DS +βL TS

[0018]

[0019] Among them, L is the total loss function, n is the number of iterations, L DS is the data supervision loss, L TS For teacher supervision losses, and are the student output image, teacher output image and true label image respectively, and α and β are the weight coefficients for balancing the losses in different aspects.

[0020] Furthermore, in step 5, the teacher-student network training assumes that the measurement relationship between bands with different spatial resolutions is self-similar within the relevant scale range, that is, scale invariance, which means that the s-fold super-resolution network model trained after uniformly downsampling all bands by n×s times can be applied to the s-fold super-resolution reconstruction of the original band, which is conducive to solving the problem that the high-resolution band has no ground truth value. Based on this, the total loss is used as the network training error, and the final student network model is obtained after updating the iterative training, which can achieve efficient super-resolution reconstruction.

[0021] Furthermore, in step 6, the converted lightweight student network model can be deployed in embedded devices with limited resources to complete corresponding engineering applications.

[0022] The present invention also discloses a multispectral remote sensing image super-resolution reconstruction system based on knowledge distillation, which includes the following modules based on the above method:

[0023] Image acquisition and preprocessing module, which acquires multispectral remote sensing images and performs data preprocessing;

[0024] The training set reconstruction module introduces the high-resolution bands in the non-reconstructed spectral domain as prior knowledge to construct a super-resolution reconstruction training set;

[0025] Teacher-student offline distillation network building module, builds a teacher-student offline distillation network based on multispectral remote sensing images, and designs a lightweight student network with the same architecture as the teacher network;

[0026] The total loss calculation module builds a knowledge distillation framework for multispectral remote sensing images. The training set in the training set reconstruction module is input into the teacher network and the student network in the teacher-student offline distillation network construction module. The super-resolution remote sensing images are output respectively, and the loss of the two is calculated as the teacher supervision loss, i.e., the knowledge distillation loss. The loss of the super-resolution remote sensing image output by the student network and the training set label is calculated as the data supervision loss. The total loss is calculated by adding the weights of the two.

[0027] The student network training module uses the training set in the training set reconstruction module and introduces the total loss in the total loss calculation module as the loss function for back propagation to train the student network.

[0028] The super-resolution reconstruction module takes the multispectral remote sensing image as input and uses the student network trained in the student network training module to upgrade the low-resolution band to the highest resolution band in the spectral domain.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The present invention uses residual connection to combine the feature maps before and after extraction, and extracts features on this basis. Residual convolution can free the network from learning only the resolution part of the coarse space of the input, and can focus on constructing the missing fine spatial details, thereby helping the network to converge faster and reach a deeper level, achieving a good reconstruction effect.

[0031] 2. The present invention introduces high-resolution bands as prior knowledge, combines high- and low-resolution bands and inputs them into network training to construct a non-traditional end-to-end low-resolution-high-resolution training set. It can transfer texture and spectral features from existing high-resolution bands, guide super-resolution with high resolution, retain spectral domain information, and achieve better reconstruction effects for multispectral remote sensing images.

[0032] 3. The knowledge distillation method proposed in the present invention can not only improve the training speed of the lightweight student network, but also enable the student network to achieve the reconstruction effect of the teacher network with a large number of parameters under the premise of a small number of parameters, making it more suitable for deployment in embedded devices with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the entire process of network design-training-deployment involved in the method for super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation in Example 1 of the present invention.

[0034] Figure 2 This is a flowchart of the network training for 2x super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation in Example 1 of the present invention.

[0035] Figure 3 This is a flowchart of the network training for 6x super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation in Example 1 of the present invention.

[0036] Figure 4 This is a schematic diagram of the internal structure of the residual block in the network of the multispectral remote sensing image super-resolution reconstruction method based on knowledge distillation in Example 1 of the present invention.

[0037] Figure 5 This is a block diagram of a multispectral remote sensing image super-resolution reconstruction system based on knowledge distillation according to Example 2 of the present invention. DETAILED DESCRIPTION

[0038] In order to more clearly illustrate the embodiments of the present invention, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings and other implementation methods can be obtained based on these accompanying drawings without creative work.

[0039] Embodiment 1:

[0040] like Figure 1 , 3 As shown in , this embodiment proposes a super-resolution reconstruction method for multispectral remote sensing images based on knowledge distillation. This embodiment takes the 6-fold super-resolution reconstruction of the L2A level image of Sentinel-2 / MSI as an example, and the specific steps are as follows:

[0041] Step 1: Log in to ESA to download Sentinel-2 image data and perform preprocessing. Specifically, according to the characteristics of Sentinel-2 remote sensing images, low-resolution images are obtained by taking steps such as band grouping (10m, 20m, 60m GSD), random cropping, down / up sampling, etc.;

[0042] Step 2: Introduce the high-resolution band in the non-reconstructed spectral domain as prior knowledge, combine it with the downsampled low-resolution image as training input, and use the original low-resolution band as a label (i.e., use the high-resolution band as prior knowledge to guide super-resolution reconstruction and construct an end-to-end low-resolution-high-resolution training set);

[0043] Step 3: Construct a teacher-student offline distillation network based on multispectral remote sensing images, in which a pre-trained teacher network is used, which has high complexity, large number of parameters, and good reconstruction quality. A lightweight student network is constructed based on the idea of ​​network residual block pruning, which has the same architecture as the teacher network, but has a small number of parameters and lower reconstruction quality than the teacher network.

[0044] Step 4: Construct a knowledge distillation framework for multispectral remote sensing images, input the image training set described in step 2 into the teacher network and student network described in step 3, output super-resolution remote sensing images respectively, calculate the loss band by band for the two and take the average as the teacher supervision loss (knowledge distillation loss), prompting the student network to learn the characteristics of the teacher network; calculate the loss band by band for the super-resolution remote sensing image output by the student network and the training set label and take the average as the data supervision loss, prompting the student network to learn the characteristics of the high-resolution label. The total loss is calculated by adding the two according to the weights;

[0045] Step 5: Use the training set described in step 2, introduce the total loss described in step 4 as the loss function for back propagation, and train the student network;

[0046] Step 6: Use the Sentinel-2 image as input and use the student network model trained in step 5 to upgrade the low-resolution band (60m GSD) to the highest resolution band (10m GSD) in the spectral domain.

[0047] In step 2 of this embodiment, it is necessary to obtain the L2A-level image of Sentinel-2 / MSI, which has three spatial resolutions of 10m, 20m, and 60m. Each band is divided into three groups according to the spatial resolution, and then randomly cropped according to the magnification. The three groups are uniformly Gaussian blurred and 6 times downsampled to 60m, 120m, and 360m spatial resolutions. The 120m and 360m bands are then bilinearly interpolated and upsampled to 60m spatial resolution, and then the training set and the validation set are divided. Since the teacher-student network training adds prior knowledge, the texture and spectral features are transferred from the existing high-resolution bands, and the super-resolution is guided by high resolution, the constructed training set is not a traditional end-to-end low-resolution-high-resolution training set, but introduces 60m and 120m upsampled to 60m high-resolution bands as prior knowledge. Therefore, the model training input is composed of the high-resolution band down / upsampled and combined with the low-resolution band, and the label is the original resolution band.

[0048] In step 3 of this embodiment, the pre-trained network Dsen2 (i.e., Dsen2 model) is used as the teacher network. The network is based on the idea of ​​band fusion, uses a residual neural network to learn the texture mapping relationship between low / high resolution images, and integrates spatial and spectral information to perform global regression training to achieve super-resolution reconstruction of Sentinel-2 images. The main components of the teacher network structure are convolutional layers (Conv), nonlinear activation functions (ReLU), and residual neural network modules (Resblock). Figure 4 As shown, including two layers of stacked convolution, skip links and residual scaling layers, the student network shares the same architecture as the teacher network, and the number of parameters is relatively small by reducing the number of residual blocks and channels.

[0049] In step 4 of this embodiment, the teacher supervision loss and the data supervision loss are added according to the weights to obtain the total loss, which is used as the network training error. After updating the iterative training, the final student network model is obtained, which forces the student model to learn the characteristic knowledge of the teacher model to the greatest extent. The loss function is RMSE (root mean square error), and the model training is guided by pixel-level error back propagation, which is specifically expressed as:

[0050] L=αL DS +βL TS

[0051]

[0052] Among them, L is the total loss function, n is the number of iterations, L DS is the data supervision loss, L TS For teacher supervision losses, and are the student output image, teacher output image and true label image respectively, and α and β are the weight coefficients for balancing the losses in different aspects.

[0053] In step 5 of this embodiment, the training is based on the scale invariance assumption, that is, the teacher-student network training assumes that the measurement relationship between bands with different spatial resolutions is self-similar within the relevant scale range, which means that the mapping between the resolutions 20m→10m and 40m→20m, 60m→10m and 360m→60m is basically the same. Based on this, since there is no ground truth for images with a resolution of 10m, in the 6x super-resolution reconstruction task, the 6x super-resolution reconstruction model of 360m→60m will be trained, which can be used for the super-resolution of 60m→10m. Finally, the pre-trained weights of the teacher network are imported, and the knowledge distillation method is introduced, that is, the teacher supervision loss (knowledge distillation loss) is added, back propagation, update iteration, and the student network is trained.

[0054] The specific implementation steps of this embodiment include the following steps based on logging in to the European Space Agency to download Sentinel-2 image data and performing preprocessing:

[0055] S1. Constructing a training set: Since Sentinel-2 images have three spatial resolutions of 10m, 20m, and 60m. Therefore, the remote sensing images are first cropped, using a pixel block of 96×96, and 500 patches are randomly sampled for each image. Each patch corresponds to a reception field of several hundred meters on the ground, which is sufficient to capture the local low-level texture and potential small semantic structures, such as a single building or a small water body, while ensuring the reconstruction effect. Avoid the inability to put the image into the GPU for training due to the large size. Then, each cropped image is divided into three groups according to the band: A = (B2, B3, B4, B8) (GSD = 10m), B = (B5, B6, B7, B8A, B11, B12) (GSD = 20m), and C = (B1, B9) (GSD = 60m). Gaussian blur processing is performed on each group to simulate its modulation transfer function, and y is obtained by averaging the 6×6 window to downsample. A (GSD = 60m), y B (GSD = 120m), y C (GSD=360m). B ,y C Then perform 2x and 6x bilinear interpolation upsampling respectively to obtain The training input of the model is The label is the original image band C (GSD = 60m). A , In fact, as a priori knowledge, the high-frequency information of C super-resolution comes from the multi-scale features coupled by A, B, and C. Finally, through a randomly generated index array containing Boolean values, 20% of the training data will be divided into a validation set by default.

[0056] S2. Design a lightweight student network: The student network is also based on a residual architecture. While using multiple residual blocks, by adopting "skip connections", the network can extract very deep features. Training data Bypassing part of the network and adding the prediction residual through the concatenation layer, the model only needs to learn the high-frequency residual between the high-resolution and low-resolution images. The network is mainly composed of convolutional layers (Conv), nonlinear activation functions (ReLU), and residual neural network modules (Resblock). By reducing the number of residual blocks and channels, the number of parameters of the student network is relatively reduced.

[0057] S3. Construct loss function: import pre-trained weights into the teacher network, input the training set into the teacher network, and output the super-resolution reconstructed image. At the same time, randomly initialize the student network, input the training set into the student network, and output the super-resolution reconstructed image. Calculate the RMSE loss of each band and take the average to get the teacher supervision loss L TS Then, the super-resolution image reconstructed by the student network and the label are subjected to the RMSE loss calculation for each band and averaged to obtain the data supervision loss L DS The two are added together according to the weights to get the total loss L of the model.

[0058] S4. Training model: The total loss is used as the training error of the classroom-student network, and the final student network model is obtained after updating the iterative training.

[0059] S5. Use the trained student network model to perform super-resolution reconstruction of multispectral remote sensing images, effectively improving the reconstruction efficiency.

[0060] Embodiment 2:

[0061] like Figure 5 As shown, this embodiment discloses a multispectral remote sensing image super-resolution reconstruction system based on knowledge distillation, which includes the following modules based on Embodiment 1:

[0062] Image acquisition and preprocessing module, which acquires multispectral remote sensing images and performs data preprocessing;

[0063] The training set reconstruction module introduces the high-resolution bands in the non-reconstructed spectral domain as prior knowledge to construct a super-resolution reconstruction training set;

[0064] Teacher-student offline distillation network building module, builds a teacher-student offline distillation network based on multispectral remote sensing images, and designs a lightweight student network with the same architecture as the teacher network;

[0065] The total loss calculation module builds a knowledge distillation framework for multispectral remote sensing images. The training set in the training set reconstruction module is input into the teacher network and the student network in the teacher-student offline distillation network construction module. The super-resolution remote sensing images are output respectively, and the loss of the two is calculated as the teacher supervision loss, i.e., the knowledge distillation loss. The loss of the super-resolution remote sensing image output by the student network and the training set label is calculated as the data supervision loss. The total loss is calculated by adding the weights of the two.

[0066] The student network training module uses the training set in the training set reconstruction module and introduces the total loss in the total loss calculation module as the loss function for back propagation to train the student network.

[0067] The super-resolution reconstruction module takes the multispectral remote sensing image as input and uses the student network trained in the student network training module to upgrade the low-resolution band to the highest resolution band in the spectral domain.

[0068] For other contents of this embodiment, please refer to Embodiment 1.

[0069] In summary, the present invention designs a multispectral remote sensing image super-resolution reconstruction method and system based on knowledge distillation, which is used to achieve high-efficiency and high-quality multispectral remote sensing image super-resolution reconstruction tasks, and can be applied to various engineering application scenarios with limited resources.

[0070] It should be noted that the above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, there will be changes in the specific implementation methods based on the ideas provided by the present invention, and these changes should also be regarded as the scope of protection of the present invention.

Claims

1. A multispectral remote sensing image super-resolution reconstruction method based on knowledge distillation, characterized by Follow these steps: Step 1: Acquire multispectral remote sensing images and perform data preprocessing; Step 2: Introduce the high-resolution bands in the non-reconstructed spectral domain as prior knowledge to construct a super-resolution reconstruction training set; Step 3: Build a teacher-student offline distillation network based on multispectral remote sensing images and design a lightweight student network with the same architecture as the teacher network; Step 4: Construct a knowledge distillation framework for multispectral remote sensing images, input the training set described in step 2 into the teacher network and the student network described in step 3, output super-resolution remote sensing images respectively, and calculate the loss of the two as the teacher supervision loss, i.e., the knowledge distillation loss; calculate the loss of the super-resolution remote sensing image output by the student network and the training set label as the data supervision loss, and calculate the total loss by adding the teacher supervision loss and the data supervision loss according to the weights; In step 4, the teacher supervision loss and data supervision loss are calculated using the root mean square error RMSE. The teacher supervision loss is used as the knowledge distillation loss to implement the knowledge distillation operation; the total loss is used as the network training error. After updating the iterative training, the final student network model is obtained, which is specifically expressed as: L=αL DS +βL TS Among them, L is the total loss function, n is the number of iterations, L DS is the data supervision loss, L TS For teacher supervision losses, and are the student output image, the teacher output image and the true label image respectively, and α and β are the weight coefficients for balancing the losses in different aspects; Step 5: Use the training set described in step 2, introduce the total loss described in step 4 as the loss function for back propagation, and train the student network; Step 6: Take the multispectral remote sensing image as input and use the student network trained in step 5 to upgrade the low-resolution band to the highest resolution band in the spectral domain.

2. The method for super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation according to claim 1, characterized in that: In step 1, the bands of the multispectral remote sensing image are grouped according to the spatial resolution, and the image blocks are randomly cropped. Then, each group of bands is downsampled by Gaussian blur processing, and some low-resolution bands are upsampled by bilinear interpolation.

3. The method for super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation as claimed in claim 1, characterized in that: In step 3, the pre-trained residual network Dsen2 is used as the teacher network, including convolutional layers, non-linear activation functions, and residual neural network modules.

4. The method for super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation as claimed in claim 1, characterized in that: In step 5, the training is based on the scale invariance assumption, that is, the teacher-student network training assumes that the measurement relationship between bands with different spatial resolutions is self-similar within the relevant scale range.

5. The method for super-resolution reconstruction of multispectral remote sensing images based on knowledge distillation according to any one of claims 1 to 4, characterized in that: In step 6, the obtained student network can be deployed on resource-limited embedded devices.

6. A multispectral remote sensing image super-resolution reconstruction system based on knowledge distillation, based on the method according to any one of claims 1 to 5, characterized in that Includes the following modules: Image acquisition and preprocessing module, which acquires multispectral remote sensing images and performs data preprocessing; The training set reconstruction module introduces the high-resolution bands in the non-reconstructed spectral domain as prior knowledge to construct a super-resolution reconstruction training set; Teacher-student offline distillation network building module, builds a teacher-student offline distillation network based on multispectral remote sensing images, and designs a lightweight student network with the same architecture as the teacher network; The total loss calculation module builds a knowledge distillation framework for multispectral remote sensing images. The training set in the training set reconstruction module is input into the teacher network and the student network in the teacher-student offline distillation network construction module. The super-resolution remote sensing images are output respectively, and the loss of the two is calculated as the teacher supervision loss, i.e., the knowledge distillation loss. The loss of the super-resolution remote sensing image output by the student network and the training set label is calculated as the data supervision loss. The total loss is calculated by adding the weights of the two. The student network training module uses the training set in the training set reconstruction module and introduces the total loss in the total loss calculation module as the loss function for back propagation to train the student network. The super-resolution reconstruction module takes the multispectral remote sensing image as input and uses the student network trained in the student network training module to upgrade the low-resolution band to the highest resolution band in the spectral domain.

Images