Land utilization mapping method combining AI Earth and U-Net network
By combining the land use mapping methods of AI Earth and U-Net networks, the problem of land use mapping in the prior art relying on manual, time-consuming and inconsistent results is solved, and automated land use segmentation is achieved, reducing costs and improving the consistency and repeatability of results.
Patent Information
- Application Number
- CN202510008343.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-03
- Publication Date
- 2025-05-23
AI Technical Summary
Existing land use mapping methods rely on manual interpretation, which is time-consuming and has limited consistency and repeatability of results. The deep learning model has a high demand for computing resources and labeled samples, resulting in high costs.
Combining the land use mapping methods of AI Earth and U-Net networks, automated land use segmentation is achieved by pre-processing of satellite remote sensing images, sample set annotation, U-Net network training and mapping accuracy evaluation.
Reduces the dependence of land use mapping on labor, reduces costs, and improves the consistency and repeatability of results.
Smart Images
Figure CN120032069A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of land management, and in particular to a land use mapping method combining AI Earth and a U-Net network. Background Art
[0002] Land use mapping is an important means of understanding and managing land resources. It plays a vital role in planning urban development, maintaining the ecological environment, and formulating relevant policies. With the acceleration of population growth and urbanization, land resources are facing the threat of over-exploitation and environmental degradation. Remote sensing technology is a key tool for obtaining land use information. Its high resolution, high spectral and high temporal resolution make large-scale, rapid and periodic surface monitoring possible.
[0003] Traditional remote sensing mapping methods rely on professional software and manual interpretation, but are often time-consuming and subject to personal experience and subjective judgment, resulting in limited consistency and repeatability of results. Machine learning algorithms improve the efficiency and objectivity of mapping through automated feature learning and pattern recognition, but they usually require a large amount of training data and features to optimize the model. Although deep learning models perform well in processing complex images and fine segmentation tasks, they have high requirements for computing resources and usually require a large number of labeled samples to train the model. The production of labeled samples often requires manual work, which is not only time-consuming but also costly when the amount of data is large or the samples are difficult to obtain.
[0004] Therefore, how to reduce the reliance on manual labor in land use mapping, reduce costs, and improve the consistency and repeatability of results has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention
[0005] In view of the above-mentioned defects of the prior art, the present invention provides a land use mapping method combining AI Earth and U-Net network, which aims to reduce the dependence of land use mapping on manual work, reduce costs, and improve the consistency and repeatability of results, which has become a technical problem that technical personnel in this field urgently need to solve.
[0006] To achieve the above object, the present invention discloses a land use mapping method combining AI Earth and U-Net network, comprising the following steps:
[0007] Step 1: Preprocess satellite remote sensing images;
[0008] Step 2: Sample set annotation;
[0009] Step 3: U-Net network training;
[0010] Step 4: Cartographic accuracy assessment.
[0011] Preferably, the satellite remote sensing image is a multispectral image, and the preprocessing performed includes radiation correction, atmospheric correction and geometric correction;
[0012] The radiation correction is used to eliminate the error of the sensor itself;
[0013] The atmospheric correction is used to eliminate the influence of the atmosphere on the image;
[0014] The geometric correction is to align the satellite remote sensing image with the geographic coordinate system.
[0015] Preferably, step 2 is as follows:
[0016] Step 2.1, inputting the pre-processed satellite remote sensing image into the AI Earth platform for segmentation to obtain an initial land use spatial distribution map;
[0017] Step 2.2, load the initial land use segmentation results and images into ArcGIS software, and cut them into multiple sub-patches using the SplitRaster tool in the toolbox;
[0018] Step 2.3, randomly selecting more than one fifth of the sub-patches from all the sub-patches as training samples;
[0019] In ArcGIS, each of the selected sub-patch is converted into a vector format and modified to obtain an accurately labeled sample;
[0020] Then, the vector format files of each sub-patch after the modification to obtain the accurate labeled samples are converted into raster files to form an accurate land use labeled sample set;
[0021] The precise land use annotation sample set is used as a training set.
[0022] More preferably, in step 2.2, the sub-patches are cropped into 3600 pieces of 128×128 and 128×128×4 sizes by the SplitRaster tool; and in step 2.3, 800 pieces of the sub-patches are randomly selected as training samples.
[0023] More preferably, each land use category in the training set includes cultivated land, forest land, shrubs, water area, buildings and bare land.
[0024] More preferably, step 3 is as follows:
[0025] Step 3.1: Build a U-Net network. The encoder is used to capture the features of the image and gradually reduce the spatial dimension of the feature map. The decoder is used to restore the spatial dimension of the image and output the final segmentation map.
[0026] The encoder comprises a plurality of convolutional blocks;
[0027] Each of the convolution blocks includes a plurality of convolution operations, followed by a maximum pooling operation;
[0028] Each of the convolution operations is followed by a ReLU activation function;
[0029] The decoder gradually enlarges the size of the feature map by transposing the convolution layer or the deconvolution layer;
[0030] Step 3.2: Use the training set to train the U-Net model. The trained network is applied to the entire remote sensing image to automatically segment land use, thereby obtaining the final land use distribution map.
[0031] More preferably, each of the convolution blocks includes two 3x3 convolution operations.
[0032] More preferably, each of the maximum pooling operations is a 2x2 maximum pooling operation with a step size of 2.
[0033] More preferably, during the upsampling process of step 3, the decoder will be spliced with the feature map of the corresponding layer in the encoder. After splicing, the decoder will also include two 3x3 convolution operations; each 3x3 convolution operation uses a ReLU activation function.
[0034] More preferably, step 4 is to use the verification sample to perform accuracy test on the segmentation result, and save and output the result after completion.
[0035] Beneficial effects of the present invention:
[0036] The application of the present invention can reduce the dependence of land use mapping on manual labor, reduce costs, and improve the consistency and repeatability of results, which has become a technical problem that technical personnel in this field urgently need to solve.
[0037] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 An embodiment of the present invention is shown
[0039] Figure 2 The Sentinel-2 multispectral image disclosed by the European Space Agency in one embodiment of the present invention is shown.
[0040] Figure 3 A distribution diagram of training samples in one embodiment of the present invention is shown.
[0041] Figure 4 Shows the land use spatial distribution map in an embodiment of the present invention. Detailed implementation manners
[0042] Embodiment
[0043] As Figure 1 shown, the land use mapping method combining AI Earth and U-Net network includes the following steps:
[0044] Step 1, preprocess the satellite remote sensing image;
[0045] Step 2, label the sample set;
[0046] Step 3, train the U-Net network;
[0047] Step 4, evaluate the mapping accuracy.
[0048] In some embodiments, the satellite remote sensing image is a multi-spectral image, and the preprocessing performed includes radiometric correction, atmospheric correction, and geometric correction;
[0049] Radiometric correction is used to eliminate the errors of the sensor itself;
[0050] Atmospheric correction is used to eliminate the influence of the atmosphere on the image;
[0051] Geometric correction is to align the satellite remote sensing image with the geographic coordinate system.
[0052] In practical applications, it is a key step to obtain high-quality remote sensing data;
[0053] Among them, radiometric correction can ensure the comparability of images acquired at different times and by different sensors;
[0054] Atmospheric correction can improve the spectral purity of the image;
[0055] Geometric correction can ensure the accuracy of the image.
[0056] As Figure 2 shown, the Sentinel-2 multi-spectral image publicly available from the European Space Agency.
[0057] In some embodiments, Step 2 is specifically as follows:
[0058] Step 2.1, input the preprocessed satellite remote sensing image into the AI Earth platform for segmentation to obtain an initial land use spatial distribution map;
[0059] Step 2.2, load the initial land use segmentation result and the image into the ArcGIS software, and crop them into multiple sub-patches through the SplitRaster tool in the toolbox;
[0060] Step 2.3, randomly select more than one fifth of all sub-patches as training samples;
[0061] Each selected sub-patch was converted into vector format in ArcGIS and modified to obtain accurate labeled samples;
[0062] Then, the vector format files of each sub-patch after the modification to obtain the accurate labeled samples are converted into raster files to form an accurate land use labeled sample set;
[0063] The precise land use labeling sample set is used as the training set.
[0064] In some embodiments, in step 2.2, the image is cropped into 3600 sub-patches of sizes of 128×128 and 128×128×4 using the SplitRaster tool; and in step 2.3, 800 sub-patches are randomly selected as training samples.
[0065] In some embodiments, the land use categories in the training set include cultivated land, forest land, shrubs, water areas, buildings, and bare land.
[0066] For example Figure 2 After executing step 2, the multispectral image shown in Figure 3 The training sample distribution diagram is shown in the figure, where the statistics of each land use category in the training samples are shown in the following table:
[0067] category serial number Number of pixels Proportion (%) arable land 1 3603103 6.11 woodland 2 7192537 12.19 shrub 3 237132 0.40 Waters 4 422039 0.72 building 5 1539782 2.61 Bare Land 6 112607 0.19
[0068] In some embodiments, step 3 is as follows:
[0069] Step 3.1: Build a U-Net network. The encoder is used to capture the features of the image and gradually reduce the spatial dimension of the feature map. The decoder is used to restore the spatial dimension of the image and output the final segmentation map.
[0070] The encoder includes multiple convolutional blocks;
[0071] Each convolution block includes multiple convolution operations, followed by a maximum pooling operation;
[0072] Each convolution operation is followed by a ReLU activation function;
[0073] The decoder gradually enlarges the size of the feature map by transposing the convolution layer or deconvolution layer;
[0074] Step 3.2: Use the training set to train the U-Net model. The trained network is applied to the entire remote sensing image to automatically segment land use, thereby obtaining the final land use distribution map.
[0075] In some embodiments, each convolution block includes two 3x3 convolution operations.
[0076] In some embodiments, each maximum pooling operation is a 2x2 maximum pooling operation with a stride of 2.
[0077] In practical applications, each convolution block is followed by a 2x2 maximum pooling operation for downsampling, and the stride is set to 2 to reduce the size of the feature map.
[0078] In some embodiments, during the upsampling process in step 3, the decoder is concatenated with the feature map of the corresponding layer in the encoder. After concatenation, the decoder also includes two 3x3 convolution operations; each 3x3 convolution operation uses a ReLU activation function.
[0079] The above technical means can further process the feature map and generate the final segmentation result.
[0080] In such Figure 3 Based on the training sample distribution diagram, step 3 is performed to obtain Figure 4 The spatial distribution map of land use is shown.
[0081] In some embodiments, step 4 is to use the verification sample to perform accuracy check on the segmentation result, and save and output the result after completion.
[0082] In practical applications, Figure 4 The inspection of the land use spatial distribution map of a city shown in the figure is shown in the following table:
[0083]
[0084] The present invention uses the AIE-SEG remote sensing general basic segmentation model in the AI Earth platform to perform preliminary segmentation on the image and obtain a preliminary land use spatial distribution map; the SplitRaster tool in ArcGIS is used to crop the image and the preliminary segmentation results into sub-patches of a certain size, and a number of sub-patches are randomly selected as a sample set, and converted into a vector file to remove small areas, and the vector file is modified in combination with the remote sensing image. The modified vector file is then converted into a raster file to obtain an accurate land use annotation sample set; a U-Net network is constructed and trained, which mainly includes an encoder and a decoder. Among them, the encoder is used to extract image features and gradually reduce the spatial size of the feature map, and the decoder is used to restore the spatial size of the image and output the segmentation result; the trained network is used for the segmentation of the entire image to obtain a land use distribution map, and the segmentation result is tested for accuracy using a verification sample, and the result is output and saved.
[0085] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A land use mapping method combining AI Earth and U-Net network; characterized in that: The steps include: Step 1: Preprocess satellite remote sensing images; Step 2: Sample set annotation; Step 3: U-Net network training; Step 4: Cartographic accuracy assessment.
2. The land use mapping method combining AI Earth and U-Net network according to claim 1, characterized in that: The satellite remote sensing image is a multispectral image, and the preprocessing performed includes radiation correction, atmospheric correction and geometric correction; The radiation correction is used to eliminate the error of the sensor itself; The atmospheric correction is used to eliminate the influence of the atmosphere on the image; The geometric correction is to align the satellite remote sensing image with the geographic coordinate system.
3. The land use mapping method combining AI Earth and U-Net network according to claim 1, characterized in that: Step 2 is as follows: Step 2.1, inputting the pre-processed satellite remote sensing image into the AIEarth platform for segmentation to obtain an initial land use spatial distribution map; Step 2.2, load the initial land use segmentation results and images into ArcGIS software, and cut them into multiple sub-patches using the SplitRaster tool in the toolbox; Step 2.3, randomly selecting more than one fifth of the sub-patches from all the sub-patches as training samples; In ArcGIS, each of the selected sub-patch is converted into a vector format and modified to obtain an accurately labeled sample; Then, the vector format files of each sub-patch after the modification to obtain the accurate labeled samples are converted into raster files to form an accurate land use labeled sample set; The precise land use annotation sample set is used as a training set.
4. The land use mapping method combining AI Earth and U-Net network according to claim 3 is characterized in that: In step 2.2, the sub-patches are cut into 3600 pieces of 128×128 and 128×128×4 sizes by the SplitRaster tool; in step 2.3, 800 pieces of the sub-patches are randomly selected as training samples.
5. The land use mapping method combining AI Earth and U-Net network according to claim 3, characterized in that: The land use categories in the training set include cultivated land, forest land, shrubs, water areas, buildings and bare land.
6. The land use mapping method combining AI Earth and U-Net network according to claim 3, characterized in that: Step 3 is as follows: Step 3.1: Build a U-Net network. The encoder is used to capture the features of the image and gradually reduce the spatial dimension of the feature map. The decoder is used to restore the spatial dimension of the image and output the final segmentation map. The encoder comprises a plurality of convolutional blocks; Each of the convolution blocks includes a plurality of convolution operations, followed by a maximum pooling operation; Each of the convolution operations is followed by a ReLU activation function; The decoder gradually enlarges the size of the feature map by transposing the convolution layer or the deconvolution layer; Step 3.2: Use the training set to train the U-Net model. The trained network is applied to the entire remote sensing image to automatically segment land use, thereby obtaining the final land use distribution map.
7. The land use mapping method combining AI Earth and U-Net network according to claim 6, characterized in that: Each of the convolution blocks includes two 3x3 convolution operations.
8. The land use mapping method combining AI Earth and U-Net network according to claim 3, characterized in that: Each of the maximum pooling operations is a 2x2 maximum pooling operation with a step size of 2.
9. The land use mapping method combining AI Earth and U-Net network according to claim 3, characterized in that: During the upsampling process in step 3, the decoder will be spliced with the feature map of the corresponding layer in the encoder. After splicing, the decoder will also include two 3x3 convolution operations; each 3x3 convolution operation uses the ReLU activation function.
10. The land use mapping method combining AI Earth and U-Net network according to claim 3, characterized in that: Step 4 is to use the verification sample to test the accuracy of the segmentation results, and save and output the results after completion.