Cultivated land information high-precision extraction method based on high-resolution remote sensing image

Abstract

The invention belongs to the technical field of remote sensing information extraction, and particularly relates to a cultivated land information high-precision extraction method based on a high-resolution remote sensing image. According to the method, the remote sensing cultivated land extraction precision can be improved based on feature dense superposition fusion and information homogeneous enhancement. Firstly, existing remote sensing cultivated land extraction samples are collected to construct a sample library, and on this basis, extraction and dense fusion of information of different scales are rapidly realized by applying a feature dense superposition fusion module, and expression of the model on global consistency and local features is improved. Information interaction between a bottom layer and a high layer is realized by using an information homogeneous enhancement module, fusion of feature information of different branches is improved, consistent expression of internal features of cultivated land and difference enhancement of edge features are realized, and the precision of remote sensing cultivated land extraction is improved. In application, a remote sensing cultivated land extraction technology with practical value is obtained, and application and development of ground feature extraction in a remote sensing image are expected to be practically promoted.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing information extraction, and in particular relates to a method for high-precision extraction of cultivated land information based on high-resolution remote sensing images. Background Art

[0002] With the development of urbanization and the increase in population, a large amount of cultivated land has been used for urban construction or industrial parks, resulting in a reduction in cultivated land area, which poses a serious threat to food security and people's basic living needs. Rapid, timely and effective understanding of cultivated land utilization is crucial to alleviating the pressure caused by the reduction of cultivated land. Dynamic cultivated land monitoring provides important data support for cultivated land protection and helps to solve food security issues. High-resolution remote sensing technology can monitor a large area of ​​cultivated land, provide real-time information, and promptly detect and respond to cultivated land changes. However, due to the differences in crop types and their growth conditions in cultivated land, as well as the problem of narrow cultivated land edges, the extraction accuracy is low when using high-resolution remote sensing technology, and the cultivated land edges are often incomplete.

[0003] The development of farmland segmentation technology for remote sensing images is based on computer vision image segmentation technology, which can be divided into traditional algorithms and deep learning algorithms. Traditional algorithms usually use low-level features of images, such as color, texture, and edges, to perform segmentation through methods such as clustering and region growing. However, traditional algorithms have limitations in processing complex scenes and irregular shapes, especially in extracting high-level semantic information, which leads to poor performance in farmland segmentation tasks.

[0004] Among deep learning algorithms, deep learning models (such as convolutional neural networks, CNN) have the ability to extract semantic features from large data sets and can be effectively applied to remote sensing cultivated land extraction. Researchers extracted cultivated land using methods such as DeepLab v3+, UNet, SegNet, and multi-task models. These methods have achieved good results in extracting cultivated land from remote sensing high-resolution images, but due to the differences in crops grown in cultivated land and the narrow ridges between plots, these methods still have some limitations in terms of the integrity and boundary accuracy of cultivated land.

[0005] In view of this, the inventors hope to provide a high-precision method for extracting cultivated land information based on high-resolution remote sensing images. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned problems existing in traditional technologies, provide a high-precision method for extracting cultivated land information based on high-resolution remote sensing images, improve the accuracy of cultivated land extraction from remote sensing images, and achieve high-precision cultivated land extraction.

[0007] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0008] The present invention provides a method for extracting cultivated land information with high precision based on high-resolution remote sensing images, comprising the following steps:

[0009] Step 1: Collect remote sensing image farmland extraction samples;

[0010] Step 2: Constructing a remote sensing image farmland extraction model based on feature-intensive paper fusion and homogeneous information enhancement;

[0011] Step 3: Model implementation and training;

[0012] Step 4: Model testing and evaluation;

[0013] Furthermore, the specific operations of step one are: first, download various public data sets of remote sensing images from the Internet; second, organize the large amount of collected image data and remove damaged or poor quality files; then perform data preprocessing, which includes image cropping, scaling, and rotation operations to remove irrelevant content and adjust the image to an appropriate size to complete the construction of the sample library.

[0014] Furthermore, the specific operation of step 2 is as follows: first, the image is quickly downsampled at different scales through the feature dense overlapping fusion module (FFM) to achieve multi-scale expression of the same target, and the features of different scales are transferred and spliced ​​to each other to improve the consistency of context information, strengthen the local differences of cultivated land edge information, and enhance the model's ability to utilize feature information; then, a multi-scale feature transmission network from bottom to top and from left to right is established through the homogeneous information enhancement module, which makes full use of the information interaction between low-level and high-level features and optimizes the transmission of inter-layer information. This structural design enhances the integrity of cultivated land extraction plots, while showing excellent performance in processing edge details, and improves the accuracy and reliability of cultivated land extraction results.

[0015] Furthermore, in step 2, the model mainly includes a feature-intensive overlapping fusion module and a homogeneous information enhancement module.

[0016] Furthermore, the feature-intensive overlay fusion module mainly realizes the extraction and fusion of feature information; the model continuously generates new feature layers through the feature-intensive overlay fusion module to build a multi-branch model; the core of the feature-intensive overlay fusion module lies in the combination of long connections and short connections to realize the fusion and transmission of multi-level features. Long connections span multiple feature layers, transfer shallow information to deep layers, prevent information loss in the deep network, and maintain the consistency of global features; short connections extract detail features through local convolution operations for local information extraction and transmission; through the local features of input and output, the representation of local details is strengthened, which is especially suitable for capturing the differences in farmland edge information and local features; the feature-intensive overlay fusion module improves the model's comprehensive processing capabilities for global and local features through a multi-level feature fusion mechanism, effectively enhancing the generalization and stability of the model.

[0017] Furthermore, each branch in the homogeneous information enhancement module contains a convolution layer, a residual connection and an upsampling operation; after the feature information is processed by the convolution operation, the residual connection fuses the convolution output with the input; this structure allows the network to be designed deeper while ensuring that information is effectively propagated within the deep network, alleviating the gradient vanishing problem, and improving the training stability and efficiency of the network; after the last convolution layer of each branch, an upsampling operation is performed; the features obtained by upsampling are used as the input of the next layer of convolution, residual concatenation and upsampling operations; through upsampling, the upper layer feature information is expanded to a size that matches the resolution of the feature map of the next layer, realizing the fusion of features of different scales at the same level; finally, the output feature information of each layer is upsampled and additively fused, so that the model can consider local and global information and enhance the network's ability to understand and generate images; by realizing multi-scale feature fusion, the model can better adapt to input data of different types and complexities, thereby improving the generalization ability of the network and effectively processing various data with different resolutions and levels of detail.

[0018] Furthermore, in step 3, the software environment configuration of the experiment was performed on the Windows operating system. The network model and the model used in the comparative experiment were based on PyTorch as the deep learning framework. The deep learning environment for model training was configured by Anaconda. The corresponding Python version used was 3.9. The final development platform of the experiment was Pycharm. Numpy and OpenCV dependency libraries were mainly used in the training and testing of the network model. At the same time, in order to increase the running speed of the model, GPU acceleration technology was adopted, and the parallel programming model CUDA was used to accelerate the parallel computing of the GPU to speed up the processing of complex image processing tasks.

[0019] In terms of experimental hardware environment configuration, all experiments were run on a high-performance computer equipped with an NVIDIA GeForce RTX4060Ti graphics card and 16GB of memory, and a Kingston 2T solid-state drive and a Seagate 4T mechanical hard drive;

[0020] In the training part of the model, the model is trained on the GF-2 and JL-1 datasets. First, the dataset is cropped, and the original image is cropped to 512×512 size, the step size is 256, and the pixel overlap between two adjacent blocks is 256; momentum and Adam optimizers are used for training, the initial learning rate is set to 0.001, the weight decay is 0.0025, and the total batch size is 4.

[0021] Furthermore, in step 4, the model is tested using the GF-2 test set. The Gaofen-2 satellite is equipped with two high-resolution cameras: a 1-meter panchromatic camera and a 4-meter multispectral camera, both of which have sub-meter spatial resolution. This resolution enables accurate annotation of farmland plots. The labeled vector data is converted into a binary image format, where 1 and 0 in the binary image represent farmland plots and other areas, respectively; the labeled data is cropped into samples of 512×512 pixels, forming a total of 3764 samples, which are further randomly divided into training and test sets at a ratio of 4:1. Finally, precision (P), recall (R), intersection over union (IOU), F1 score, and overall accuracy (OA) are selected as evaluation indicators, and the average is taken as the final accuracy.

[0022] The beneficial effects of the present invention are:

[0023] In view of the differences in crop types and growth conditions in cultivated land, as well as the narrow ridges between plots, the present invention designs a remote sensing high-resolution cultivated land extraction method based on dense feature superposition fusion and information homogeneity enhancement to improve the accuracy of remote sensing cultivated land extraction. The method first collects existing remote sensing cultivated land extraction samples to build a sample library, and on this basis, by applying a dense feature superposition fusion module, quickly realizes the extraction and dense fusion of information of different scales, and improves the model's expression of global consistency and local features. And the information homogeneity enhancement module is used to interact the information between the bottom layer and the upper layer, improve the fusion of feature information of different branches, realize the consistent expression of the internal features of the cultivated land and the differential enhancement of the edge features, and improve the accuracy of remote sensing cultivated land extraction. In terms of application, the present invention obtains a remote sensing cultivated land extraction technology with practical value, in order to effectively promote the application and development of ground object extraction in remote sensing images.

[0024] Of course, any product implementing the present invention does not necessarily need to achieve all of the above advantages at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0026] Figure 1 This is the main technical flow chart of the present invention;

[0027] Figure 2 It is a feature dense overlay fusion module diagram;

[0028] Figure 3 It is the information homogeneity reinforcement module diagram;

[0029] Figure 4 It is a model map of high-resolution cultivated land extraction from remote sensing based on dense feature overlay fusion and homogeneous information enhancement;

[0030] Among them, Conv: convolution; BN: normalization processing; Relu: activation function; Dropout: Dropout layer prevents overfitting; Avgpool: average pooling; RC: residual connection; Up: upsampling operation;

[0031] Figure 5 This is the extraction result diagram of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The present invention proposes a remote sensing high-resolution farmland extraction method based on feature-intensive overlapping fusion and homogeneous information enhancement. The method proposes a feature-intensive overlapping fusion module to extract feature information, uses a feature-intensive overlapping fusion module to extract features at different scales, constructs feature layers with different branches, and then fuses these feature layers. In the deep network, the feature layer of the previous branch is upsampled after convolution residual connection, and is fused with different feature layers of the next branch, making full use of global and local information and reducing the loss of details in the transmission of feature information. Compared with other semantic segmentation methods, this model achieves high-precision farmland extraction while ensuring the accuracy of edge details, thereby making the farmland boundary clearer.

[0034] The specific embodiments of the present invention are as follows:

[0035] This embodiment provides a method for extracting farmland information with high precision based on high-resolution remote sensing images, comprising the following steps:

[0036] Step 1: Collect farmland extraction samples from remote sensing images

[0037] Create a sample library based on remote sensing images. First, download various public remote sensing image data sets from the Internet. Second, organize the large amount of image data collected and remove damaged or poor quality files. Then perform data preprocessing, which includes operations such as cropping, scaling, and rotating images to remove irrelevant content and adjust the images to a suitable size to complete the construction of the sample library.

[0038] Step 2: Construction of remote sensing high-resolution cultivated land extraction model based on feature-intensive overlay fusion and homogeneous information enhancement

[0039] In order to solve the problems of differences in crop types and growth conditions in cultivated land, as well as narrow ridges between plots, this embodiment uses a feature-intensive overlay fusion module (FFM) to quickly downsample images at different scales to achieve multi-scale expression of the same target, and transfers and splices features of different scales to each other, improves the consistency of contextual information, strengthens the local differences in cultivated land edge information, and enhances the model's ability to utilize feature information. Then, a multi-scale feature transmission network from bottom to top and from left to right is established through a homogeneous information enhancement module, which makes full use of the information interaction between low-level and high-level features and optimizes the transmission of inter-layer information. This structural design enhances the integrity of cultivated land extraction plots, while showing excellent performance in processing edge details, and improves the accuracy and reliability of cultivated land extraction results.

[0040] The main structure and components of the network model are introduced in detail below. The model mainly includes: feature superposition fusion module and homogeneous information enhancement module.

[0041] Feature Overlay Fusion Module: New feature layers are continuously generated through FFM to build a multi-branch model. The core of the feature dense overlay fusion module lies in the combination of long connections and short connections, which realizes the fusion transmission of multi-level features. Long connections span multiple feature layers, transfer shallow information to deep layers, prevent information loss in deep networks, and maintain the consistency of global features. Short connections extract detail features through local convolution operations for local information extraction and transmission. Through the local features of input and output, the representation of local details is strengthened, which is particularly suitable for capturing the differences in farmland edge information and local features. FFM improves the model's comprehensive processing capabilities for global and local features through a multi-level feature fusion mechanism, effectively enhancing the generalization and stability of the model.

[0042] Homogeneous Information Enhancement Module: To achieve effective multi-scale feature fusion and maximize the flow of feature information between different scales, we adopt a homogeneous information enhancement module in the deep network. Each branch contains a convolution layer, a residual connection, and an upsampling operation. After the feature information is processed by the convolution operation, the residual connection fuses the convolution output with the input. This structure allows the network to be designed deeper while ensuring effective information propagation within the deep network, alleviating the gradient vanishing problem, and improving the training stability and efficiency of the network. After the last convolution layer of each branch, an upsampling operation is performed. The features obtained by upsampling are used as the input of the next layer of convolution, residual concatenation, and upsampling operations. Through upsampling, the upper layer feature information is expanded to a size that matches the resolution of the feature map of the next layer, achieving the fusion of features of different scales at the same level. Finally, the output feature information of each layer is upsampled and additively fused, allowing the model to consider local and global information and enhance the network's ability to understand and generate images. By achieving multi-scale feature fusion, the model better adapts to input data of different types and complexities, thereby improving the generalization ability of the network and effectively processing various data with different resolutions and levels of detail.

[0043] Step 3: Model implementation and training

[0044] In terms of the experimental software environment configuration, the experiments in this embodiment are all operated on the Windows operating system. The network model and the models used in the comparative experiment are based on PyTorch as the deep learning framework. The deep learning environment for model training is configured by Anaconda. The corresponding Python version used is 3.9. The final development platform of the experiment is Pycharm. In the training and testing process of the network model, Numpy, OpenCV and other dependent libraries are mainly used. At the same time, in order to improve the running speed of the model, GPU acceleration technology is adopted, and the parallel programming model CUDA is used to accelerate the parallel computing of the GPU to speed up the processing of complex image processing tasks.

[0045] In terms of experimental hardware environment configuration, all experiments were run on a machine equipped with an NVIDIA GeForce RTX4060Ti graphics card and 16GB of memory, and a high-performance configuration computer equipped with a Kingston 2T solid-state drive and a Seagate 4T mechanical hard drive.

[0046] In the training part of the model, the proposed model is trained on the GF-2 dataset and the JL-1 dataset. First, the dataset is cropped, and the original image is cropped to 512×512 size, the step size is 256, and the pixel overlap between two adjacent blocks is 256. The momentum and Adam optimizers are used for training, the initial learning rate is set to 0.001, the weight decay is 0.0025, and the total batch size is 4.

[0047] Step 4: Model testing and evaluation

[0048] The model was tested using the test set of GF-2. The Gaofen-2 satellite is equipped with two high-resolution cameras: a 1-meter panchromatic camera and a 4-meter multispectral camera, both of which have sub-meter spatial resolution. This resolution enables accurate annotation of farmland plots. The labeled vector data was converted into a binary image format, where 1 and 0 in the binary image represent farmland plots and other areas, respectively. The labeled data was cropped into samples of 512×512 pixels, forming a total of 3764 samples, which were further randomly divided into training and test sets at a ratio of 4:1. Finally, precision (P), recall (R), intersection over union (IOU), F1 score, and overall accuracy (OA) were selected as evaluation indicators, and the average was taken as the final accuracy.

[0049] A specific application of this embodiment is:

[0050] Using the remote sensing cultivated land extraction datasets GF-2 and JL-1, the original image size of the dataset was cropped to 512×512, the step size was 256, and the pixel overlap between two adjacent blocks was 256. The multi-scale model was trained for 80 epochs on the GF-2 and JL-1 datasets, using momentum and Adam optimizers for training, with the initial learning rate set to 0.001 and the weight decay to 0.0025. The NVIDIA GeForce RTX 4060Ti graphics card was used for training, with a total batch size of 4. At the same time, other commonly used models were selected for cultivated land extraction and compared with the results of the method in this embodiment. As shown in Table 1, the remote sensing cultivated land extraction method proposed in this embodiment achieves state-of-the-art performance.

[0051] Table 1

[0052]

[0053] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A high-precision method for extracting cultivated land information based on high-resolution remote sensing images, characterized in that: The following steps are involved: Step 1: Collect remote sensing image farmland extraction samples; Step 2: Constructing a remote sensing image farmland extraction model based on feature-intensive paper fusion and homogeneous information enhancement; Step 3: Model implementation and training; Step 4: Model testing and evaluation.

2. The high-precision method for extracting cultivated land information based on high-resolution remote sensing images according to claim 1 is characterized in that: The specific operations of step one are: first, download various public remote sensing image data sets from the Internet; second, organize the large amount of collected image data and remove damaged or poor quality files; Then the data is preprocessed, which includes cropping, scaling, and rotating the image to remove irrelevant content and adjust the image to an appropriate size to complete the construction of the sample library.

3. The high-precision method for extracting cultivated land information based on high-resolution remote sensing images according to claim 2 is characterized in that: The specific operations of step 2 are as follows: first, the image is quickly downsampled at different scales through the feature dense overlay fusion module to achieve multi-scale expression of the same target, and the features of different scales are transferred and spliced ​​to each other to improve the consistency of contextual information, strengthen the local differences of cultivated land edge information, and enhance the model's ability to utilize feature information; then, a multi-scale feature transmission network is established from bottom to top and from left to right through the homogeneous information enhancement module, which makes full use of the information interaction between low-level and high-level features and optimizes the transmission of inter-layer information.

4. The method for high-precision extraction of cultivated land information based on high-resolution remote sensing images according to claim 3 is characterized in that: In step 2, the model mainly includes a feature-intensive overlapping fusion module and a homogeneous information enhancement module.

5. The method for high-precision extraction of cultivated land information based on high-resolution remote sensing images according to claim 4 is characterized in that: The feature-intensive overlapping fusion module mainly realizes the extraction and fusion of feature information; the model continuously generates new feature layers through the feature-intensive overlapping fusion module to build a multi-branch model; the core of the feature-intensive overlapping fusion module lies in the combination of long connections and short connections to realize the fusion transmission of multi-level features; Long connections span multiple feature layers, transferring shallow information to deep layers, preventing information loss in deep networks and maintaining the consistency of global features; short connections extract detail features through local convolution operations for local information extraction and transmission; and strengthen the representation of local details through local features of input and output.

6. The method for high-precision extraction of cultivated land information based on high-resolution remote sensing images according to claim 5 is characterized in that: Each branch in the homogeneous information enhancement module contains a convolution layer, a residual connection and an upsampling operation. After the feature information is processed by the convolution operation, the residual connection fuses the convolution output with the input. This structure allows the network to be designed deeper while ensuring that information is effectively propagated within the deep network, alleviating the gradient vanishing problem and improving the training stability and efficiency of the network. After the last convolution layer of each branch, an upsampling operation is performed. The features obtained by upsampling serve as the input of the next layer of convolution, residual concatenation and upsampling operations. Through upsampling, the upper layer feature information is expanded to a size that matches the resolution of the feature map of the next layer, realizing the fusion of features of different scales at the same level. Finally, the output feature information of each layer is upsampled and additively fused, so that the model can consider local and global information and enhance the network's ability to understand and generate images.

7. The high-precision method for extracting cultivated land information based on high-resolution remote sensing images according to claim 3 is characterized in that: In step 3, the software environment configuration of the experiment was performed on the Windows operating system. The network model and the model used in the comparative experiment were based on PyTorch as the deep learning framework. The deep learning environment for model training was configured by Anaconda. The corresponding Python version used was 3.

9. The final development platform of the experiment was Pycharm. Numpy and OpenCV dependency libraries were mainly used in the training and testing of the network model. At the same time, in order to increase the running speed of the model, GPU acceleration technology was adopted, and the parallel programming model CUDA was used to accelerate the parallel computing of the GPU to speed up the processing of complex image processing tasks. In terms of experimental hardware environment configuration, all experiments were run on a high-performance computer equipped with an NVIDIA GeForce RTX 4060Ti graphics card and 16GB of memory, and a Kingston 2T solid-state drive and a Seagate 4T mechanical hard drive; In the training part of the model, the model is trained on the GF-2 and JL-1 datasets. First, the dataset is cropped, and the original image is cropped to 512×512 size, the step size is 256, and the pixel overlap between two adjacent blocks is 256; momentum and Adam optimizers are used for training, the initial learning rate is set to 0.001, the weight decay is 0.0025, and the total batch size is 4.

8. The method for high-precision extraction of cultivated land information based on high-resolution remote sensing images according to claim 7 is characterized in that: In step 4, the model was tested using the GF-2 test set. The labeled vector data was converted into a binary image format, where 1 and 0 in the binary image represented farmland and other areas, respectively. The labeled data was cropped into samples of 512 × 512 pixels, forming a total of 3764 samples, which were randomly divided into training and test sets at a ratio of 4:

1. Finally, precision, recall, intersection-over-union, F1 score, and overall accuracy were selected as evaluation indicators, and the average value was taken as the final accuracy.

Images