Abandoned land identification method and system based on remote sensing satellite image

Abstract

The invention discloses an abandoned land identification method and system based on a remote sensing satellite image. The method comprises the following steps: acquiring and preprocessing a remote sensing image to obtain an input image; the input image is input into a TransU-Net + model to obtain a recognition result, and the TransU-Net + model comprises a down-sampling module for extracting multi-scale features, a Transform module for extracting global correlation features, an up-sampling module for introducing a channel attention network for feature fusion and feature recovery, and an output layer for introducing an FC-CRF module; and the identification result is visualized. According to the abandoned land recognition and segmentation method, global pixel interaction modeling can be carried out on the feature map, so that the perception capability of the model is improved, and the abandoned land is recognized and segmented more accurately.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer graphics processing, and particularly relates to a method and system for abandoned land recognition based on remote sensing satellite images. BACKGROUND

[0002] In the process of rural and agricultural development, automatic recognition and control of abandoned land is an extremely important part. However, due to the diversity and complexity of remote sensing images, such as occlusion, shadow and seasonal changes of crops, it still presents great challenges to accurately recognize abandoned land. In the field of remote sensing image semantic segmentation, convolutional neural networks (CNN), especially fully convolutional neural networks (FCN), have always played a crucial role. Although they perform well in image feature extraction, due to the local characteristics of convolution operation, CNN-based methods often have difficulty in fully understanding the long-distance context information in the image. Therefore, when the existing methods are applied to the segmentation of abandoned land in remote sensing images, there are often great challenges. Therefore, a method specifically for abandoned land recognition in remote sensing images is needed to cope with the above challenges. SUMMARY

[0003] In order to solve the above technical problems, the present application proposes a method and system for abandoned land recognition based on remote sensing satellite images, which can more accurately recognize and segment abandoned land.

[0004] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0005] The method for abandoned land recognition based on remote sensing satellite images comprises the following steps:

[0006] Obtaining a remote sensing image and pre-processing to obtain an input image;

[0007] Inputting the input image into a TransU-Net+ model to obtain a recognition result, wherein the TransU-Net+ model comprises a down-sampling module for extracting multi-scale features, a Transformer module for extracting global correlation features,

[0008]

[0008] an up-sampling module introducing a channel attention network for feature fusion and feature restoration, and an output layer introducing an FC-CRF module;

[0009] Visualizing the recognition result.

[0010] Preferably, the pre-processing comprises padding and cropping.

[0011] Preferably, the downsampling module includes a 3×3 convolutional block and four downsampling convolutional blocks connected in sequence. The 3×3 convolutional block includes a 3×3 convolutional calculation with a step size of 1, a BatchNorm layer, and a ReLU activation function. The downsampling convolutional block includes a 2×2 Maxpool layer and a 3×3 convolutional block.

[0012] Preferably, the Transformer module processing includes: the feature map processed by the last downsampling convolutional block is transformed into a dimensionality transformation and then input into a 6-layer stacked Transformer layer; the output features of the 6-layer stacked Transformer layer are transformed into a dimensionality transformation and then input into a 3×3 convolutional block to obtain a global correlation feature map; the global correlation feature map is then input into the upsampling module.

[0013] Preferably, the upsampling module includes four sets of upsampling convolutional blocks, a channel attention network, and a 3×3 convolutional block. The channel attention network fuses the features after the upsampling convolutional block operation with the features after the corresponding downsampling convolutional block operation and inputs them into the 3×3 convolutional block to obtain a fused feature map.

[0014] Preferably, the calculation formula for the FC CRF module is as follows:

[0015]

[0016] θ i =-logp(x i )

[0017] Where p(x) i The probability of the class at pixel i is calculated by the TransU-Net+ network.

[0018]

[0019] Where, when x i =x j When μ(x) i ,x j ) = 0, otherwise 1; p i Let p be the class probability of the i-th pixel. j Let I represent the class probability of the j-th pixel, where I represents the RGB color of the pixel; w and σ represent the weights and hyperparameters, respectively.

[0020] Preferably, the recognition result is visualized, specifically including the following steps:

[0021] The Argmax function is used to obtain the category map for each pixel from the recognition results;

[0022] The category map of each pixel is overlaid on the original remote sensing image and visualized.

[0023] Based on the above, the present invention also discloses an abandoned land identification system based on remote sensing satellite imagery, comprising:

[0024] The preprocessing module is used to acquire remote sensing images and perform preprocessing to obtain the input image;

[0025] The recognition module is used to input the input image into the TransU-Net+ model to obtain the recognition result. The TransU-Net+ model includes a downsampling module for extracting multi-scale features, a Transformer module for extracting global correlation features, an upsampling module that introduces a channel attention network for feature fusion and feature recovery, and an output layer that introduces an FC-CRF module.

[0026] The visualization module is used to visualize the recognition results.

[0027] Based on the above technical solution, the beneficial effects of this invention are as follows: Firstly, it improves the learning and representation capabilities of the feature extraction module by leveraging the global context information encoding capability of the Transformer. Secondly, it enhances the model's perceptual ability by introducing a channel attention network into the feature fusion layer to automatically focus on the most relevant information. Thirdly, it enhances the global consistency of the results by introducing a fully connected CRF network for smoothing. This allows the invention to retain the high-resolution spatial information and global context information of the traditional TransU-Net while incorporating a channel attention network and an FC CRF module to improve the model's feature perception and enhance the global consistency of the results, thus enabling more accurate semantic segmentation of abandoned land. Experimental results using remote sensing satellite data from Wenchang City, Hainan Province in June and August of 2024 show that compared to the traditional U-Net network, Attention U-Net network, and TransU-Net network, the accuracy of this invention is improved by 3.5%-6.12%, the recall by 3.74%-6.54%, and the precision by 4.18%-7.63%. Attached Figure Description

[0028] Figure 1 This is a flowchart of an implementation method for identifying abandoned land based on remote sensing satellite images;

[0029] Figure 2 This is a schematic diagram of the TransU-Net+ model processing flow in one embodiment. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0031] Symbol explanation: R a×b×c Represents the shape or size of data, such as: A∈R a×b×c Let A be a 3-dimensional data point with a shape or size of [a, b, c].

[0032] like Figure 1 , 2 As shown, this embodiment provides a method for identifying abandoned land based on remote sensing satellite imagery, specifically including the following steps:

[0033] Step 1: Data acquisition and processing, preprocessing the input remote sensing image data.

[0034] Step 1.1: Process the length (H) and width (W) of the original remote sensing satellite image to multiples of 224, and fill any insufficient parts with 0;

[0035] Step 1.2: Crop the image processed in Step 1.1 into several smaller 224×224 images, which will be used as input to the TransU-Net+ model.

[0036] Step 2: The lightweight feature extraction module extracts the feature map.

[0037] Step 2.1: Define a 3×3 convolutional block and input the image obtained in Step 1 into this block, outputting a level 1 feature map C1∈R. 16×224×224 The 3×3 convolutional block contains a 3×3 convolutional computation with a stride of 1, a BatchNorm layer, and a ReLU activation function;

[0038] Step 2.2: Define a downsampling convolutional block, input C1 into the downsampling convolutional block to obtain a level 2 feature map C2∈R. 64×112×112 The downsampling module contains a 2×2 Maxpool layer and a 3×3 convolutional block;

[0039] Step 2.3: Repeat step 2.2 three times to obtain the three-level feature map C3∈R. 128×56×56 Level 4 feature map C4∈R 256×28×28 and the 5th level feature map C5∈R 512×14×14 .

[0040] Step 3: The Transformer module constructs a global correlation feature map.

[0041] Step 3.1: Change the shape of the level 5 feature map C5 obtained in step 2 from [512, 14, 14] to [196, 512];

[0042] Step 3.2: Input the level 5 feature map C5 from Step 3.1 into a Transformer module consisting of 6 standard Transformer Encoder layers, and output the global correlation feature map T∈R. 196×512 ;

[0043] Step 3.3: Restore the shape of the global correlation feature map from 3.2 back to [512,14,14].

[0044] Step 4: Input the global correlation feature map into the 3×3 convolutional module defined in Step 2 to obtain the global correlation feature map T2∈R. 512×14×14 .

[0045] Step 5: The attention fusion and upsampling module calculates the class probability for each pixel.

[0046] Step 5.1: Input the global correlation feature map T2 from Step 4 into a 3×3 convolutional block, and input the output of the convolutional block into an upsampling convolution (ConvTranspose) to obtain the first-level upsampling feature U1∈R. 256×28×28 ;

[0047] Step 5.2: Define a channel attention network and input the level 4 feature map C4 and the level 1 upsampled feature map U1 obtained in Step 2 into the channel attention network to obtain the level 1 fused feature map F1∈R. 512×28×28 The channel attention network is defined as follows:

[0048] F(X,Y)=sigmoid(MLP(AvgPool(cat(X,Y))))×cat(X,Y)

[0049] Where cat() represents vector concatenation, sigmoid() is the activation function, and AvgPool() represents global average pooling;

[0050] Step 5.3: Similarly, obtain the second-level upsampled feature U2∈R from the first-level fused feature map F1 using the steps in 5.1. 128 ×56×56 ;

[0051] Step 5.4: Input the level 2 upsampled feature U2 and the level 3 feature map C3 obtained in step 2 into the channel attention network to obtain the level 2 fused feature map F2∈R. 128×56×56 ;

[0052] Step 5.5: Obtain the level 3 upsampled feature U3∈R from the level 2 fused feature map F2 as described in step 5.1. 64 ×112×112 ;

[0053] Step 5.6: Input the level 3 upsampled feature U3 and the level 2 feature map C2 obtained in step 2 into the channel attention network to obtain the level 3 fused feature map F3∈R. 64×112×112 ;

[0054] Step 5.7: Obtain the 4th-level upsampled feature U4∈R from the 3-level fused feature map F3 as in Step 5.1. 16 ×224×224 ;

[0055] Step 5.8: Input the level 4 upsampled feature U4 and the level 1 feature map C1 obtained in step 2 into the channel attention network to obtain the level 4 fused feature map F4∈R. 32×224×224 ;

[0056] Step 5.9: Input the 4-level fused feature map F4 into a 3×3 convolutional block and a 1×1 convolution to calculate a pixel class probability map P∈R. 2×224×224 .

[0057] Step 6: Input the pixel category probability map P into the FC CRF module for smoothing. The FC CRF module is a fully connected CRF function, and its calculation method is as follows:

[0058]

[0059] θ i =-logp(x i )

[0060] Where p(x) i The probability of the class at pixel i is calculated by the TransU-Net+ network.

[0061]

[0062] Where, when x i =x j When μ(x) i ,x j ) = 0, otherwise 1; p i Let p be the class probability of the i-th pixel. j Let I represent the class probability of the j-th pixel, and let I represent the RGB color of the pixel. w and σ represent the weights and hyperparameters, respectively, and in this embodiment, they are set to: w1 = 5, w2 = 3, σ α =60,σ β =5,σ γ =3.

[0063] Step 7: Visualize the results.

[0064] Step 7.1: Use the Argmax() function to obtain the class map of each pixel from the smoothed probability map output in Step 6;

[0065] Step 7.2: Overlay the category map of each pixel with the original image and visualize it.

[0066] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0067] One embodiment provides a system for identifying abandoned land based on remote sensing satellite imagery, comprising:

[0068] The preprocessing module is used to acquire remote sensing images and perform preprocessing to obtain the input image;

[0069] The recognition module is used to input the input image into the TransU-Net+ model to obtain the recognition result. The TransU-Net+ model includes a downsampling module for extracting multi-scale features, a Transformer module for extracting global correlation features, an upsampling module that introduces a channel attention network for feature fusion and feature recovery, and an output layer that introduces an FC-CRF module.

[0070] The visualization module is used to visualize the recognition results.

[0071] The above description is merely a preferred embodiment of the method and system for identifying abandoned land based on remote sensing satellite imagery disclosed in this invention, and is not intended to limit the scope of protection of the embodiments in this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments in this specification should be included within the scope of protection of the embodiments in this specification.

Claims

1. A method for identifying abandoned land based on remote sensing satellite imagery, characterized in that, Includes the following steps: Acquire remote sensing images and perform preprocessing to obtain the input image; The input image is fed into the TransU-Net+ model to obtain the recognition result. The TransU-Net+ model includes a downsampling module for extracting multi-scale features, a Transformer module for extracting global correlation features, an upsampling module that introduces a channel attention network for feature fusion and feature recovery, and an output layer that introduces an FC-CRF module. The recognition results are visualized.

2. The method for identifying abandoned land based on remote sensing satellite imagery according to claim 1, characterized in that, The preprocessing includes filling and trimming.

3. The method for identifying abandoned land based on remote sensing satellite imagery according to claim 1, characterized in that, The downsampling module includes a 3×3 convolutional block and four downsampling convolutional blocks connected in sequence. The 3×3 convolutional block includes a 3×3 convolution calculation with a step size of 1, a BatchNorm layer, and a ReLU activation function. The downsampling convolutional block includes a 2×2 Maxpool layer and a 3×3 convolutional block.

4. The method for identifying abandoned land based on remote sensing satellite imagery according to claim 3, characterized in that, The Transformer module processing includes: the feature map processed by the last downsampling convolutional block is transformed into a dimension and then input into a 6-layer stacked Transformer layer; the output features of the 6-layer stacked Transformer layer are transformed into a dimension and then input into a 3×3 convolutional block to obtain a global correlation feature map; the global correlation feature map is then input into the upsampling module.

5. The method for identifying abandoned land based on remote sensing satellite imagery according to claim 4, characterized in that, The upsampling module includes four sets of upsampling convolutional blocks, a channel attention network, and a 3×3 convolutional block. The channel attention network fuses the features after the upsampling convolutional block operation with the features after the corresponding downsampling convolutional block operation and inputs them into the 3×3 convolutional block to obtain a fused feature map.

6. The method for identifying abandoned land based on remote sensing satellite imagery according to claim 1, characterized in that, The calculation formula for the FCCRF module is as follows: Where p(x) i The probability of the class at pixel i is calculated by the TransU-Net+ network. Where, when x i =x j When μ(x) i ,x j ) = 0, otherwise 1; p i Let p be the class probability of the i-th pixel. j Let I represent the class probability of the j-th pixel, where I represents the RGB color of the pixel; w and σ represent the weights and hyperparameters, respectively.

7. The method for identifying abandoned land based on remote sensing satellite imagery according to claim 1, characterized in that, The visualization of the recognition results specifically includes the following steps: The Argmax function is used to obtain the category map for each pixel from the recognition results; The category map of each pixel is overlaid on the original remote sensing image and visualized.

8. A system for identifying abandoned land based on remote sensing satellite imagery, characterized in that, include: The preprocessing module is used to acquire remote sensing images and perform preprocessing to obtain the input image; The recognition module is used to input the input image into the TransU-Net+ model to obtain the recognition result. The TransU-Net+ model includes a downsampling module for extracting multi-scale features, a Transformer module for extracting global correlation features, an upsampling module that introduces a channel attention network for feature fusion and feature recovery, and an output layer that introduces an FC-CRF module. The visualization module is used to visualize the recognition results.

Images