High-precision lake extraction method based on unsupervised remote sensing super-resolution

Abstract

The invention discloses a high-precision lake extraction method based on unsupervised remote sensing super-resolution. The method comprises the following steps: 1) collecting a low-resolution remote sensing image of a target area; 2) constructing a multi-channel unsupervised depth super-division model of a fusion gradient; 3) performing model training, and determining optimized model parameters; 4) performing high-spatial-resolution image reconstruction on the acquired low-resolution remote sensing image of the target area by using the depth super-resolution model obtained by training to obtain a high-resolution remote sensing image; 5) performing image feature enhancement based on the original low-resolution remote sensing image and the high-resolution remote sensing image obtained in the step 4) to obtain a final high-resolution remote sensing image; 6) calculating the NDWI index of the super-resolution remote sensing image according to the final high-resolution remote sensing image, and 7) extracting the lake region according to the NDWI index.The lake extraction method is used for extracting the lake, and the lake region with clearer boundary, more accurate area and higher spatial resolution can be obtained.

Description

Technical Field

[0001] The invention relates to remote sensing information technology, and in particular to a high-precision lake extraction method based on unsupervised remote sensing super-resolution. Background Art

[0002] Lakes are dynamic systems that support huge biodiversity and provide important food supply and cultural ecosystem services to people everywhere. Because changes in lakes are closely related to the impact of climate and human activities, lakes have been identified as important indicators of climate change, and more detailed lake area can better reflect climate change.

[0003] Remote sensing technology makes it possible to monitor lakes over a long time series and with wide coverage. Many studies have used remote sensing data to extract lake areas. However, since remote sensing monitoring technology cannot simultaneously obtain images with high temporal resolution, high spatial resolution, and high spectral resolution, the fine extraction of small lakes is very challenging. Super-resolution technology aims to reconstruct higher-resolution images from their original low-resolution remote sensing images. Image super-resolution technology provides the possibility of improving the spatial resolution of remote sensing images and the accuracy of lake extraction. With the development of deep learning, many super-resolution methods have been developed, which can be divided into two categories: supervised super-resolution methods and unsupervised super-resolution methods. Since supervised super-resolution methods require the construction of sufficient sample sets to complete the model training process, and for remote sensing images, the remote sensing images generated by different sensors are quite different, the sample construction process is complicated, and the generalization ability of the supervised model is limited, it is difficult to guarantee the accuracy of the model. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a high-precision lake extraction method based on unsupervised remote sensing super-resolution in view of the defects in the prior art. An unsupervised super-resolution method constructed by the present invention can complete the super-resolution process of the target image without using additional data to construct a sample data set.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a high-precision lake extraction method based on unsupervised remote sensing super-resolution, comprising the following steps: Step 1) collecting a remote sensing image of the target area; the remote sensing image has a Green band and a NIR band; in, Represents the green band of remote sensing images, Represents the near-infrared band of remote sensing images; Step 2) constructing a multi-channel unsupervised deep super-resolution model that fuses gradients; the multi-channel unsupervised deep super-resolution model is used to reconstruct high spatial resolution images of the collected remote sensing images of the target area; The model includes a gradient information extraction module, an information fusion module, a deep feature extraction module, an upsampling module and a reconstruction module; The gradient information extraction module is used to calculate the gradient map of the low-resolution image of the original image; the low-resolution image is obtained by downsampling the remote sensing image collected in step 1); The information fusion module is used to perform cascade fusion on a feature map obtained by performing information extraction on a low-resolution image of the original image and a feature map obtained by performing feature extraction on the obtained gradient map to obtain fusion information; The deep feature extraction module is used to use multiple residual network structures to perform deep feature modeling and extraction on the fusion information; The up-sampling module is used to perform an up-sampling operation on the information; The reconstruction module is used to reconstruct a high-resolution image using a convolution kernel; Step 3) Train the model and obtain the optimized model parameters; Step 4) Reconstruct the collected remote sensing images of the target area with high spatial resolution using the trained multi-channel unsupervised deep super-resolution model to obtain high-resolution remote sensing images; Step 5) enhancing the image features based on the original remote sensing image and the high-resolution remote sensing image obtained in step 4) to obtain a final high-resolution remote sensing image; 5.1) High-resolution image I HR' Downsample to the same resolution as the original image I, and calculate the difference Diff between the original image I and the estimated downsampled image L , then Diff L Upsample to high resolution image I HR' The same resolution gets the image difference Diff H ; 5.2) By taking the high-resolution image I HR' Add image difference Diff H As a high-resolution image I HR' Updates; 5.3) Iteratively update high-resolution image I HR' , until the image difference Diff H When the value is less than the set threshold or reaches the set number of iterations n, the final updated high-resolution image is obtained.

[0006] Step 6) Calculate the NDWI index of the super-resolution remote sensing image based on the final high-resolution remote sensing image; Step 7) Extract lake area according to NDWI index; The land features are classified according to the NDWI index, where the NDWI values ​​in the range of 0-1 are lakes, and the lake areas are extracted.

[0007] According to the above scheme, the Sobel operator is used to calculate the gradient map of the input low-resolution image, as follows:

[0008] Among them, GM represents the gradient map, and Sob represents the sobel operator.

[0009] According to the above scheme, the training process of the model in step 3) is as follows: 3.1) The target image According to the super-resolution Downsample to get the corresponding low-resolution image ; 3.2) Based on low resolution -image Yes, perform adaptive extraction of training samples; 3.3) Use image augmentation algorithms including rotation, flipping, cropping, translation, shearing, and scaling to expand the training samples to obtain an expanded sample data set; 3.4) Determine the network activation function as ReLU, the optimization algorithm as Adam, the learning rate reduction strategy and the batch size, train the unsupervised deep super-resolution network, and obtain the model optimization parameters.

[0010] According to the above scheme, in step 3), the loss function used in model training is the L1 loss function.

[0011] Among them, M represents the number of all samples, represents the sample image obtained by using the image augmentation algorithm based on the target image, Represents the corresponding high-resolution image reconstructed by the model.

[0012] According to the above scheme, the NDWI index of the image is calculated as follows: ; in, Represents the green band of remote sensing images, Represents the near-infrared band of remote sensing images.

[0013] The beneficial effects produced by the present invention are: 1. The present invention constructs an unsupervised deep super-resolution network. The network can construct training samples based on the target image itself without the need for other auxiliary images, and can improve the spatial resolution of remote sensing images of different image sizes, resolutions, and numbers of channels, thereby adaptively extracting high-precision lake areas.

[0014] 2. The present invention improves the quality of remote sensing images through secondary processing. First, an unsupervised deep super-resolution model is constructed to improve the resolution of remote sensing images. Then, based on the original remote sensing images and the high-resolution remote sensing images obtained by the model, the image characteristics of the high-resolution remote sensing images obtained by the model are enhanced to obtain the final high-resolution remote sensing images. On this basis, the NDWI index is used to extract lakes in high-resolution images, thereby obtaining lake areas with clearer boundaries, more accurate areas, and higher spatial resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is a method flow chart of an embodiment of the present invention; Figure 2 Schematic diagram of the principle of the unsupervised super-resolution model of an embodiment of the present invention; Figure 3 Schematic diagram of the network structure of the unsupervised super-resolution model of an embodiment of the present invention; Figure 4 Schematic diagram of the unsupervised super-resolution model training process according to an embodiment of the present invention; Figure 5 is a schematic diagram comparing lake extraction results according to an embodiment of the present invention; Figure 6 3 is a schematic diagram comparing lake extraction results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0017] like Figure 1 As shown, the present invention takes lake extraction from MODIS remote sensing images as an example and provides a high-precision lake extraction method based on remote sensing super-resolution, comprising the following steps: 1) Collect remote sensing images containing lakes; 2) Construct a multi-channel unsupervised deep super-resolution model that integrates gradients. The multi-channel unsupervised deep super-resolution model is used to reconstruct high spatial resolution images of the collected remote sensing images of the target area, such as Figure 2 ;include: Construct the overall framework of the unsupervised super-resolution model, which includes gradient information extraction module, information fusion module, deep feature extraction module, upsampling module and reconstruction module; The network structure of each module in the model is as follows Figure 3 As shown; A gradient information extraction module is used to calculate the gradient map of the collected low-resolution image; the low-resolution image is obtained by downsampling the remote sensing image collected in step 1); When calculating the gradient map of the input low-resolution image, the Sobel operator is used, as follows:

[0018] Among them, GM represents the gradient map, and Sob represents the sobel operator.

[0019] An information fusion module is used to cascade fuse a feature map obtained by extracting information from the original image and a feature map obtained by extracting features from the obtained gradient map; The information fusion module first establishes a multi-branch CNN model. Branch one extracts information from the original low-resolution image, and branch two extracts features from the obtained gradient map. On this basis, the feature maps obtained from the two branches are cascaded and fused.

[0020] A deep feature extraction module is used to model and extract deep features of information using multiple residual network structures; An upsampling module is used to perform an upsampling operation on information; the upsampling module uses multiple pixel-shuffle layers to perform an upsampling operation on information; The reconstruction module uses convolution kernels to reconstruct high-resolution images; 3) Perform adaptive training and testing on the model to determine the model optimization parameters, as follows: Figure 4 As shown: 3.1) The target image According to the super-resolution Downsample to get the corresponding low-resolution image ; 3.2) Based on low resolution -image Yes, perform adaptive extraction of training samples; 3.3) Use image augmentation algorithms including rotation, flipping, cropping, translation, shearing, and scaling to expand the training samples to obtain sufficient sample data sets; 3.4) Determine the network activation function as ReLU, the optimization algorithm as Adam, the learning rate reduction strategy and the batch size, train the unsupervised deep super-resolution network, and obtain the model optimization parameters; The loss function used in model training is L1 loss. ; Among them, M represents the number of all samples, represents the sample image obtained by using the image augmentation algorithm based on the target image, Represents the corresponding high-resolution image reconstructed by the model.

[0021] 4) Use the trained deep super-resolution model to improve the spatial resolution of the target remote sensing image to obtain a high-resolution remote sensing image.

[0022] 5) Based on the original low-resolution remote sensing image and the obtained high-resolution remote sensing image, image features are enhanced to obtain the final high-resolution remote sensing image, as follows: 5.1) The high-resolution image I HR' Downsample to the same resolution as the original image I, and calculate the difference Diff between the original image I and the estimated downsampled image L , then Diff L Upsample to high resolution image I HR' The same resolution gets the image difference Diff H ; The downsampling method is Bicubic interpolation.

[0023] 5.2) By taking the high-resolution image I HR' Add image difference Diff H As a high-resolution image I HR' Updates; 5.3) Iteratively update high-resolution image I HR' , until the image difference Diff H When the value is less than the set threshold or reaches the set number of iterations n, the final updated high-resolution image is obtained.

[0024] 6) Calculate the NDWI index of super-resolution remote sensing images: NDWI=(Green-NIR) / (Green+NIR) Among them, Green represents the green band of the remote sensing image, and NIR represents the near infrared band of the remote sensing image.

[0025] 7) Extract lake areas based on the NDWI index, and lakes are those with NDWI between 0 and 1.

[0026] 8) Evaluate the lake extraction results. The evaluation indicators include overall classification accuracy (OA), Kappa coefficient (Kappa), average producer accuracy (APA) and average user accuracy (AUA), peak signal-to-noise ratio (PNSR) and structural similarity (SSIM).

[0027] Comparison results of 8x super-resolution algorithms:

[0028] Through the calculation results, we can know that the comprehensive performance evaluation index of the present invention is better than the existing methods.

[0029] Using the method of the present invention, Figure 5 As shown, lakes are directly extracted from the original remote sensing image. The accuracy of lake extraction is limited by the spatial resolution of the remote sensing image. The edge of the lake extracted by the method of the present invention is clearer. Figure 6 The following is a schematic diagram of the extraction of three typical lakes, and the area of ​​the lakes extracted by the images is calculated. The leftmost column uses the original MODIS image (spatial resolution of 250m) for lake extraction; the middle column uses the super-resolution algorithm of the present invention to improve the MODIS low-resolution image for lake extraction; the rightmost column uses the Landset image for lake extraction (spatial resolution of 30m, which we use as a high-resolution reference here). It can be seen that for the third small lake, it is directly lost in the MODIS image, but after being improved by the super-resolution algorithm, the lake can be extracted; for the two lakes of Pongyin Co and Timachaka, after the spatial resolution is improved by the super-resolution algorithm, the extracted lake edges are clearer and the area is more accurate.

[0030] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A high-precision lake extraction method based on unsupervised remote sensing super-resolution, characterized in that: The following steps are involved: Step 1) collecting a remote sensing image of the target area; the remote sensing image has a Green band and a NIR band; in, Represents the green band of remote sensing images, Represents the near-infrared band of remote sensing images; Step 2) constructing a multi-channel unsupervised deep super-resolution model that fuses gradients; the multi-channel unsupervised deep super-resolution model is used to reconstruct high spatial resolution images of the collected remote sensing images of the target area; The model includes a gradient information extraction module, an information fusion module, a deep feature extraction module, an upsampling module and a reconstruction module; The gradient information extraction module is used to calculate the gradient map of the low-resolution image of the original image; the low-resolution image is obtained by downsampling the remote sensing image collected in step 1); The information fusion module is used to perform cascade fusion on a feature map obtained by performing information extraction on a low-resolution image of the original image and a feature map obtained by performing feature extraction on the obtained gradient map to obtain fusion information; The deep feature extraction module is used to use multiple residual network structures to perform deep feature modeling and extraction on the fusion information; The up-sampling module is used to perform an up-sampling operation on the information; The reconstruction module is used to reconstruct a high-resolution image using a convolution kernel; Step 3) Train the model and obtain the optimized model parameters; Step 4) Reconstruct the collected remote sensing images of the target area with high spatial resolution using the trained multi-channel unsupervised deep super-resolution model to obtain high-resolution remote sensing images; Step 5) performing image enhancement based on the original remote sensing image and the high-resolution remote sensing image obtained in step 4) to obtain a final high-resolution remote sensing image; 5.1) The high-resolution image I HR' Downsample to the same resolution as the original image I, and calculate the difference Diff between the original image I and the downsampled image L , then Diff L Upsample to high resolution image I HR' The same resolution gets the image difference Diff H ; 5.2) By taking the high-resolution image I HR' Add image difference Diff H As a high-resolution image I HR' Updates; 5.3) Iteratively update high-resolution image I HR' , until the image difference Diff H When the value is less than the set threshold or reaches the set number of iterations n, the final updated high-resolution image is obtained. Step 6) Calculate the NDWI index of the super-resolution remote sensing image based on the final high-resolution remote sensing image; Step 7) Extract lake areas based on the NDWI index.

2. The high-precision lake extraction method based on unsupervised remote sensing super-resolution according to claim 1 is characterized in that: In step 2), when calculating the gradient map of the input low-resolution image, the Sobel operator is used, as follows: Among them, GM represents the gradient map, and Sob represents the sobel operator.

3. The high-precision lake extraction method based on unsupervised remote sensing super-resolution according to claim 1 is characterized in that: The training process of the model in step 3) is as follows: 3.1) The target image According to the super-resolution Downsample to get the corresponding low-resolution image ; 3.2) Based on low resolution -image Yes, perform adaptive extraction of training samples; 3.3) Use image augmentation algorithms including rotation, flipping, cropping, translation, shearing, and scaling to expand the training samples to obtain sufficient sample data sets; 3.4) Determine the network activation function as ReLU, the optimization algorithm as Adam, the learning rate reduction strategy and the batch size, train the unsupervised deep super-resolution network, and obtain the model optimization parameters.

4. The high-precision lake extraction method based on unsupervised remote sensing super-resolution according to claim 3 is characterized in that: In model training, the loss function used is L1 loss; ； Among them, M represents the number of all samples, represents the sample image obtained according to the target image, Represents the corresponding high-resolution image reconstructed by the model.

5. The high-precision lake extraction method based on unsupervised remote sensing super-resolution according to claim 1 is characterized in that: In step 6), the NDWI index of the image is calculated as follows: in, Represents the green band of remote sensing images, Represents the near-infrared band of remote sensing images.

6. The high-precision lake extraction method based on unsupervised remote sensing super-resolution according to claim 1 is characterized in that: In the step 7), the ground objects are classified according to the NDWI index, wherein the NDWI values ​​in the range of 0-1 are lakes, and the lake area is extracted.

7. An electronic device, characterized in that: include: one or more processors; as well as a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images