A method for identifying cultivated areas using remote sensing vectors

Abstract

A kind of method for identifying cultivated area using remote sensing vector, it includes the following steps: step 1: obtain remote sensing image, and carry out image processing to obtain remote sensing image;Step 2: the feature extraction network and attention decoding module are trained with the prepared data set, and the segmentation model is obtained;Step 3: the segmentation model is trained, and the remote sensing image of the region to be identified is put into the trained segmentation model to obtain the segmentation result image;Step 4: the segmentation result image is spliced, and the area of cultivated land is calculated.The purpose of the present application is to solve the technical problems that the existing remote sensing image processing method for cultivated land area measurement has poor extraction effect and accuracy for the edge segmentation of cultivated land, and incomplete segmentation and excessive segmentation may occur at the edge of cultivated land, and to provide a method for identifying cultivated area using remote sensing vector, which can realize more accurate segmentation of cultivated area.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for identifying cultivated land areas using remote sensing vectors. Background Technology

[0002] In current technologies, the measurement and statistics of arable land area mainly rely on manual measurement. This method is outdated and inaccurate. The measurement method typically treats the shape of arable land as triangles and rectangles, but since many arable lands are irregularly shaped, the measured area differs from the actual area. Due to the development of remote sensing technology, the imaging quality of remote sensing images has improved, containing rich color and texture information. Many researchers have used machine learning methods to segment arable land areas on remote sensing images. However, the complex background of arable land areas in remote sensing images, containing a lot of noise, greatly interferes with the segmentation results, leading to less than ideal performance using machine learning methods. With the rise of deep learning technology, many research problems are being solved by incorporating deep learning techniques, including arable land extraction. Because of the vast amount of remote sensing imagery, which meets the need for large amounts of data for training deep learning models, deep learning has achieved good results in arable land extraction.

[0003] Although deep learning technology enables rapid segmentation and extraction of farmland remote sensing images, it still performs poorly in feature fusion and edge segmentation extraction. For example, in the paper "Research on Semantic Segmentation of Farmland Based on Multi-Temporal High-Resolution Remote Sensing Images using U-net", the segmentation of farmland areas was achieved by improving U-net. However, due to the direct fusion of low-level and high-level features in the network model, effective farmland information was destroyed, resulting in problems such as omission of farmland and incomplete edge segmentation in the segmentation results. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problems of existing remote sensing image processing methods for measuring cultivated land area, which have poor extraction effect and accuracy of cultivated land edge segmentation, and are prone to incomplete and redundant segmentation at the edge of cultivated land. The invention provides a method for identifying cultivated land areas using remote sensing vectors, which can achieve more accurate segmentation of cultivated land areas.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for identifying cultivated land areas using remote sensing vectors includes the following steps:

[0007] Step 1: Acquire remote sensing images and perform image processing to obtain remote sensing images;

[0008] Step 2: Train the feature extraction network and attention decoding module using the prepared dataset to obtain the segmentation model;

[0009] Step 3: Train the segmentation model by inputting the remote sensing image of the region to be identified into the trained segmentation model to obtain the segmentation result image;

[0010] Step 4: Stitch together the segmented images and calculate the area of ​​cultivated land.

[0011] In step 2, the feature extraction network includes a deep residual network and a feature fusion module. The input of the deep residual network is a remote sensing image, and the output of the deep residual network is connected to the input of the feature fusion module. The output of the feature fusion module outputs a high-level feature map.

[0012] The output of the feature fusion module is connected to the input of the attention decoding module. At the same time, the high-level feature map is input into the attention decoding module, and the output of the attention decoding module outputs the segmentation result.

[0013] The feature fusion module includes four parallel convolutional modules and pooling layer modules. The convolutional modules are the first convolutional module, the second convolutional module, the third convolutional module, and the fourth convolutional module. The feature map output by the deep residual network is input into the parallel convolutional modules and pooling layer modules, and the fused high-level feature map is obtained after parallel operation.

[0014] The attention decoding module includes two feature channel fusion modules, three convolution modules, a feature multiplication module, three upsampling modules, and a global average pooling module;

[0015] The feature channel fusion modules are a first feature channel fusion module and a second feature channel fusion module;

[0016] The convolutional modules are the fifth and sixth convolutional modules, respectively.

[0017] The upsampling modules are a first upsampling module, a second upsampling module, and a third upsampling module;

[0018] The output of the first upsampling module is connected to the input of the sixth convolutional block. The outputs of the fifth and sixth convolutional modules are connected to the input of the first feature channel fusion module. The output of the first feature channel fusion module is connected to the first input of the feature multiplication module.

[0019] The output of the second upsampling module is connected to the input of the global average pooling module, the output of the global average pooling module is connected to the input of the seventh convolution module, and the input of the seventh convolution module is connected to the second input of the feature multiplication module.

[0020] The outputs of the feature multiplication module and the third upsampling module are connected to the input of the second feature channel fusion module.

[0021] The remote sensing image is input into a deep residual network, which performs feature extraction to obtain feature maps F1, F2, F3, and F4. Feature map F4 is input into the feature fusion module, which outputs a high-level feature map.

[0022] Upsample feature map F2 to obtain a feature map of the same size as feature map F1. → Convolve this feature map and feature map F1 through the fifth convolution module to obtain feature maps F6 and F7 respectively. → Fuse feature maps F6 and F7 to obtain a low-level fused feature map F8. → Upsample the high-level feature map F5, then perform global average pooling and convolution operations, and multiply it with feature map F8 to obtain feature map F9. → Upsample feature map F5 and fuse it with feature channels to obtain feature map F9 to obtain feature map F10. → Perform two 3×3 convolution operations and one upsampling operation on feature map F10 to obtain the segmentation result.

[0023] Step 1 includes the following sub-steps:

[0024] 1-1: Use drones to photograph the target area and acquire remote sensing images;

[0025] 1-2: Select and label remote sensing images;

[0026] 1-3: Segment the remote sensing images and create a dataset;

[0027] 1-4: Perform data augmentation on the dataset by using random scaling to increase the number of images in the dataset.

[0028] In step 3, the loss function used for network training is defined as follows:

[0029]

[0030] Where M is the number of categories; N is the total number of pixels; y ic The sign function is 0 or 1; it takes the value 1 if the true class of sample i is equal to c, and 0 otherwise. ic Let be the predicted probability that observed sample i belongs to category c.

[0031] In step 3, the remote sensing image of the farmland area to be identified is segmented and input into the pre-trained model to obtain the segmented farmland image.

[0032] In step 4, the segmented images are stitched together, and the area of ​​the cultivated land is calculated based on the number of pixels and the pixel size of the stitched images.

[0033] The deep residual network is the ResNet-101 network.

[0034] Compared with the prior art, the present invention has the following technical effects:

[0035] 1) This invention achieves effective aggregation of multi-scale information of cultivated land by constructing a feature extraction network and an attention decoding module, and achieves more accurate segmentation of cultivated land areas. At the same time, it can achieve better fitting of cultivated land boundaries and faster acquisition of the cultivated land area of ​​the current region.

[0036] 2) The method provided by the invention can save a lot of manpower and material resources and is more efficient. Compared with the application of machine learning methods in farmland extraction, the recognition accuracy of the invention is higher.

[0037] 3) In the construction of the network model, this invention uses an attention decoding module to improve the segmentation accuracy of the model, making the segmentation effect of farmland boundaries more obvious. Existing networks for farmland segmentation lose edge information of the image as the number of network layers increases, resulting in poor segmentation effect at the edges of the image target. This invention strengthens the guidance for edge segmentation by weighting low-level features with high-level semantic information and fusing high-level semantic information and low-level features.

[0038] 4) This invention creates a feature fusion module, which uses convolutional kernels of different sizes to construct parallel feature extraction branches for the network to extract farmland features at different scales, and finally fuses the feature maps to obtain multi-scale semantic information. Attached Figure Description

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0040] Figure 1 This is a flowchart of the present invention;

[0041] Figure 2 This is a network structure diagram of the present invention;

[0042] Figure 3 This is a diagram of the attention decoding module of the present invention. Detailed Implementation

[0043] like Figure 1 As shown, a method for identifying cultivated land areas using remote sensing vectors includes the following steps.

[0044] Step 1: Acquire target images through remote sensing and perform image processing;

[0045] Step 2: Train the feature extraction network and attention decoding module using the prepared dataset to obtain the segmentation model;

[0046] Step 3: Input the image of the region to be identified into the trained model to obtain the segmentation result.

[0047] Step 4: Combine the segmented results and calculate the cultivated land area;

[0048] Step 1 specifically includes the following steps:

[0049] 1) Use drones to photograph the target area and acquire remote sensing images;

[0050] 2): Filter the images and annotate them;

[0051] 3): Segment the images and create a dataset;

[0052] 4) Data augmentation of the dataset: Random scaling is used to expand the number of images in the dataset. Dataset diversity includes scale diversity, lighting diversity, and target pose diversity. Among these, different scales for the same type of target are a common problem. Expanding the dataset using random scaling provides more direct and objective benefits to the model than random flipping, noise perturbation, or other similar manipulations.

[0053] Step 2 specifically includes the following steps:

[0054] 1): Construct a feature extraction network to extract target features;

[0055] 2): Construct an attention-based decoding model to predict segmentation results;

[0056] In step 1), the feature extraction network consists of ResNet-101 and a feature fusion module. ResNet-101 performs feature extraction and consists of four convolutional sets. When an image is input into ResNet-101, feature maps F1, F2, F3, and F4 are obtained sequentially. The obtained feature map F4 is input into the feature fusion module, which has four parallel convolutional block operations and one global average pooling operation. After parallel operations, a fused high-level feature map is obtained, and the number of channels in the high-level feature map is reduced.

[0057] In step 2), channel fusion is performed on the two low-level feature maps; the weights of different channels in the low-level feature maps are redistributed, giving higher weights to channels with higher quality features, while reducing interference information in the feature maps; the channel weights are obtained by a series of operations on the high-level feature maps. Specifically, this includes the following steps:

[0058] 2-1) Obtain two low-level feature maps F1 and F2 from ResNet-101. Upsample F2 to the same size as F1. Then perform a 1×1 convolution operation on F1 and F2, adjust the number of channels, and obtain feature maps F6 and F7 respectively. Perform feature fusion on F6 and F7 to obtain a low-level fused feature map F8.

[0059] 2-2) Upsample the high-level feature map, perform global pooling and 1×1 convolution, and then perform a × operation with the low-level feature map to obtain a weighted feature map.

[0060] 2-3) Perform feature fusion on the weighted feature map and the high-level feature map, and then perform 3×3 convolution and upsampling operations to obtain the prediction results.

[0061] In step 2, the loss function used for network training is defined as follows:

[0062]

[0063] Where M is the number of categories; N is the total number of pixels; y ic The sign function is 0 or 1; it takes the value 1 if the true class of sample i is equal to c, and 0 otherwise. ic Let be the predicted probability that observed sample i belongs to category c.

[0064] In step 3, the remote sensing image of the farmland area to be identified is segmented and input into the pre-trained model to obtain the segmented farmland image.

[0065] In step 4, the segmented images are stitched together, and the area of ​​the cultivated land is calculated based on the number of pixels and the pixel size of the stitched images.

[0066] Example:

[0067] In this embodiment, the method for extracting cultivated land areas from remote sensing images is performed as follows:

[0068] Step 1: The drone acquires remote sensing imagery. The imagery is then filtered, and suitable images are selected. ArcGIS software is used to label cultivated land areas on the images, with non-cultivated land areas serving as the background. After labeling, the remote sensing images and label maps are cut into 512×512 three-channel images to create a dataset. OpenCV is used to perform data augmentation on the dataset. Data augmentation methods mainly include the following:

[0069] 1) Random scaling: Randomly enlarging or reducing the image according to a proportional scale;

[0070] 2) Translation: Randomly translate the image in the horizontal or vertical direction;

[0071] 3) Rotation: Rotate the image around its center point, such as 90°, 180°, 270°;

[0072] 4) Noise perturbation: Add Gaussian noise to the image;

[0073] 5) Random Occlusion: Randomly using blank squares to occlude parts of the image;

[0074] 6) Contrast Transformation: Adjusts the grayscale values ​​of an image to make it darker or more variable.

[0075] This invention employs a method of random scaling to expand the number of original images and label images.

[0076] Step 2: Input the image into the feature extraction network. The image passes through the convolution set conv1, conv2, conv3, and conv4, resulting in feature maps of sizes 256×128×128 (F1), 512×64×64 (F2), 1024×32×32 (F3), and 2048×32×32 (F4), respectively. Input F4 into the feature fusion network, performing parallel convolutions with 1×1, 3×3, 3×3, 5×5, and 3×3 and 7×7 convolutions, and calculating using global average pooling. Then, fuse the five feature maps and reduce the number of channels in the fused feature map. The new feature map is F5.

[0077] Step 3: Perform a 1×1 convolution operation on F1, adjust the number of channels, and obtain feature map F6; upsample F2 to the same size as F1, perform a 1×1 convolution operation, adjust the number of channels, and obtain feature map F7; fuse F6 and F7 channels to obtain feature map F8; perform global average pooling on F5, perform a 1×1 convolution operation, adjust the number of channels to the number of channels of F8, and weight F8 to obtain feature map F9; upsample F5 to the same size as F9, and fuse F5 with F9 features to obtain F10;

[0078] Step 4: F10 undergoes two 3×3 convolutions and one upsampling to obtain the prediction result.

[0079] Step 5: Define the loss function:

[0080]

[0081] Where M is the number of categories; N is the total number of pixels; y ic The sign function is 0 or 1; it takes the value 1 if the true class of sample i is equal to c, and 0 otherwise. ic Let be the predicted probability that observed sample i belongs to category c.

[0082] Step 6: Segment the remote sensing image of the cultivated land area to be identified and input it into the pre-trained model to obtain the segmented image of the cultivated land. Stitch the segmented images together and calculate the area of ​​the cultivated land area based on the number and size of pixels in the stitched image.

Claims

1. A method for identifying cultivated land areas using remote sensing vectors, characterized in that, It includes the following steps: Step 1: Acquire remote sensing images and perform image processing to obtain remote sensing images; Step 2: Train the feature extraction network and attention decoding module using the prepared dataset to obtain the segmentation model; Step 3: Train the segmentation model by inputting the remote sensing image of the region to be identified into the trained segmentation model to obtain the segmentation result image; Step 4: Stitch together the segmented images and calculate the area of ​​cultivated land; In step 2, the attention decoding module (3) includes two feature channel fusion modules, three convolution modules, a feature multiplication module, three upsampling modules and a global average pooling module; The feature channel fusion modules are a first feature channel fusion module and a second feature channel fusion module; The convolutional modules are the fifth and sixth convolutional modules, respectively. The upsampling modules are a first upsampling module, a second upsampling module, and a third upsampling module; The output of the first upsampling module is connected to the input of the sixth convolutional block. The outputs of the fifth and sixth convolutional modules are connected to the input of the first feature channel fusion module. The output of the first feature channel fusion module is connected to the first input of the feature multiplication module. The output of the second upsampling module is connected to the input of the global average pooling module, the output of the global average pooling module is connected to the input of the seventh convolution module, and the input of the seventh convolution module is connected to the second input of the feature multiplication module. The outputs of the feature multiplication module and the third upsampling module are connected to the input of the second feature channel fusion module; The remote sensing image is input into the deep residual network (1), and the deep residual network (1) performs feature extraction to obtain feature maps F1, F2, F3, and F4; among them, feature map F4 is input into the feature fusion module (2), and the feature fusion module (2) outputs a high-level feature map; Upsample feature map F2 to obtain a feature map of the same size as feature map F1. → Perform convolution operation on feature map F2 and feature map F1 through the fifth convolution module to obtain feature map F6 and feature map F7 respectively. → Perform feature fusion on feature map F6 and feature map F7 to obtain low-level fused feature map F8. → Upsample high-level feature map F5, then perform global average pooling and convolution operations, and multiply it with feature map F8 to obtain feature map F9. → Upsample feature map F5 and perform feature channel fusion operation on feature map F9 to obtain feature map F10. → Perform convolution and upsampling operations on feature map F10 to obtain the segmentation result.

2. The method according to claim 1, characterized in that, In step 2, the feature extraction network includes a deep residual network (1) and a feature fusion module (2). The input end of the deep residual network (1) is a remote sensing image, the output end of the deep residual network (1) is connected to the input end of the feature fusion module (2), and the output end of the feature fusion module (2) outputs a high-level feature map. The output of the deep residual network (1) is connected to the input of the attention decoding module (3), and the high-level feature map is input into the attention decoding module (3). The output of the attention decoding module (3) outputs the segmentation result.

3. The method according to claim 2, characterized in that, The feature fusion module (2) includes four parallel convolutional modules and pooling layer modules. The convolutional modules are the first convolutional module, the second convolutional module, the third convolutional module, and the fourth convolutional module. The feature map output by the deep residual network (1) is input to the parallel convolutional modules and pooling layer modules, and the fused high-level feature map is obtained after parallel operation.

4. The method according to claim 1, characterized in that, Step 1 includes the following sub-steps: 1-1: Use drones to photograph the target area and acquire remote sensing images; 1-2: Select and label remote sensing images; 1-3: Segment the remote sensing images and create a dataset; 1-4: Perform data augmentation on the dataset by using random scaling to increase the number of images in the dataset.

5. The method according to claim 1, characterized in that, In step 3, the loss function used for network training is defined as follows: in, For the number of categories, N The total number of pixels; For the sign function, if the sample The true category equals If the value is 1, then the value is 0; otherwise, the value is 0. For observation samples Category The predicted probability.

6. The method according to claim 1, characterized in that, In step 3, the remote sensing image of the farmland area to be identified is segmented and input into the pre-trained model to obtain the segmented farmland image.

7. The method according to claim 1, characterized in that, In step 4, the segmented images are stitched together, and the area of ​​the cultivated land is calculated based on the number of pixels and the pixel size of the stitched images.

8. The method according to claim 1, 2, or 3, characterized in that, The deep residual network (1) is a ResNet-101 network.

Images