Method for realizing cultivated land planting purpose management and control based on remote sensing interpretation
By preprocessing and extracting features from remote sensing image data, and training a model using a convolutional neural network, the problem of insufficient resolution in remote sensing images was solved, enabling efficient and low-cost management of farmland planting uses and improving identification accuracy and management efficiency.
Patent Information
- Application Number
- CN202411088991.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-12
AI Technical Summary
The spatial resolution of remote sensing images is related to their recognition accuracy. High-resolution data is expensive, while insufficient resolution leads to poor recognition accuracy, making image annotation difficult, requiring high computing power, resulting in low accuracy of spectral model recognition, and making the management and maintenance system difficult.
By collecting spectral image data of the target area, radiometric correction, geometric correction, atmospheric correction, image cropping and image masking are performed to generate datasets and label sets. A spectral recognition model is trained using a convolutional neural network, and data fusion is performed with vector data to generate thematic maps and display them for management and control based on land parcels.
It achieved optimal recognition results at a spatial resolution of 10m, reduced data costs, improved recognition accuracy and management efficiency, and ensured the accuracy and reliability of farmland management and decision-making.
Smart Images

Figure SMS_1 
Figure SMS_2 
Figure SMS_3
Abstract
Description
Technical Field
[0001] The present application relates to the field of geographic information technology, and in particular to a method for controlling the use of cultivated land for planting based on remote sensing interpretation. Background Art
[0002] Remote sensing interpretation is the process of processing, analyzing, classifying and identifying remote sensing image data obtained by remote sensing technology. These remote sensing image data can be obtained through high-altitude satellites or low-altitude drones, including multispectral or hyperspectral image data. Nowadays, high-quality remote sensing data at home and abroad are becoming more and more abundant, and historical data are also very complete. In terms of crop spectral recognition, deep learning technology based on neural networks has already achieved a fairly high accuracy. Therefore, the establishment of a crop recognition model library can greatly reduce the collection of low-altitude remote sensing data and avoid repeated data collection.
[0003] In the era of big data, technologies such as multidimensional data visualization, vector data query and layer spatial analysis of geographic information have also been widely used. The basic data such as the third survey of cultivated land, contracted land and second-district demarcation are also very complete.
[0004] Driven by the above policies, technologies and data, we can gradually improve the archival information, farmland construction and full digital dynamic monitoring of high-standard farmland, and realize the precise supervision of "non-agriculturalization" and "non-grainization" of cultivated land. This will effectively improve the accuracy and efficiency of cultivated land planting use control and accelerate the process of establishing national cultivated land spatiotemporal big data.
[0005] However, the spatial resolution of remote sensing images is positively correlated with the accuracy of recognition, but the cost of high-resolution remote sensing image data is quite high, and insufficient resolution of remote sensing image data will lead to poor recognition accuracy. There are problems such as difficult image annotation, high computing power requirements, low spectral model recognition accuracy, and difficult management and maintenance of the control system. Summary of the invention
[0006] In order to solve at least one of the technical problems in the background technology, the embodiment of the present application provides a method for controlling the use of cultivated land based on remote sensing interpretation, including the following steps:
[0007] 1) Data collection: Collect spectral image data of the target area and sample data of scene types and ground object types appearing in the target area;
[0008] 2) Image preprocessing: perform radiation correction, geometric correction, atmospheric correction, image cropping and image mask processing on the collected spectral image data;
[0009] Through radiation correction, the digital values in the spectral image data are converted into reflectance or radiation brightness values to eliminate the influence of atmospheric disturbances at different times and locations;
[0010] Through geometric correction, spectral image data is geometrically transformed to eliminate geometric distortion and errors caused by image acquisition, transmission and processing;
[0011] Atmospheric correction converts the radiation values in spectral image data affected by atmospheric absorption and scattering into surface reflectance or radiation brightness values.
[0012] Image clipping is used to clip unnecessary parts of spectral image data to reduce data volume and improve processing efficiency;
[0013] By using masks, non-crop areas such as water bodies and buildings in spectral image data can be excluded to reduce noise interference and improve classification accuracy;
[0014] 3) Image annotation: extract features from spectral image data and sample data to generate data sets and label sets;
[0015] 4) Model training: Based on the data set and label set, a convolutional neural network is trained to generate a spectral recognition model with classification and recognition capabilities.
[0016] 5) Thematic map production: The crop identification in the target area is carried out through the spectral recognition model to obtain the crop distribution map containing the identification results. The noise of the crop distribution map is removed by binarization, and then the raster data is converted into vector data. Finally, a separate thematic layer file is output for each crop.
[0017] 6) Data fusion: The thematic map is integrated with the existing vector data including contracted land, the third survey of cultivated land and the second zone demarcation to form a thematic map with higher recognition accuracy;
[0018] 7) Generate a management and control map: Through the fusion of multiple layers, with plots as units, the map shows the types of crops, planting areas, planting systems, distribution areas and ownership relationships. Through on-site verification, data can be dynamically updated, and historical data can be queried, compared and analyzed in time series.
[0019] In one of the embodiments of the present application, in the method for controlling cultivated land planting purposes based on remote sensing interpretation, the spectral image data includes hyperspectral image data and multispectral image data, and the spectral image data uses images with a spatial resolution of 10m.
[0020] In one embodiment of the present application, in the method for implementing the control of cultivated land planting use based on remote sensing interpretation, the sample data includes historical sample data and real-time data;
[0021] The historical sample data adopts images with a spatial resolution of 1m, and the real-time data adopts spectral images or true color orthophotos collected by fixed-point photography of unmanned aerial vehicles.
[0022] In one embodiment of the present application, in the method for implementing the control of cultivated land planting use based on remote sensing interpretation, the number of sample data to be estimated is calculated by the formula:
[0023]
[0024] Among them, m is the number of samples to be estimated, n represents the total number, e represents the error level, and σ represents the standard deviation.
[0025] In one embodiment of the present application, the method for implementing cultivated land planting use control based on remote sensing interpretation, wherein the unnecessary parts of the spectral image data are clipped by image clipping to reduce the data volume and improve the processing efficiency, comprises:
[0026] By manually selecting areas, the unnecessary parts of the image can be cropped using GIS software to reduce the amount of data and improve processing efficiency;
[0027] Write a program to divide the image into multiple areas of equal size to match the machine computing power and processing time, avoid overall failure and recalculation, and save computing resources.
[0028] In one embodiment of the present application, the method for implementing cultivated land planting use control based on remote sensing interpretation, the image annotation: extracting features from spectral image data and sample data, generating a data set and a label set includes:
[0029] Open the spectral image data in the annotation software and associate three of the channels into R, G, and B channels to facilitate the identification of the landforms;
[0030] Select and create a labeled area. Select an area with multiple terrains and complex terrain. Try to include all the crops to be classified within each labeled area.
[0031] In the marked area, compare the sample data and use point, line and surface tools to draw the shape of each crop planting area, and use different numbers to identify its classification;
[0032] Merge objects of the same category so that there is only one object in each category, which is convenient for subsequent processing;
[0033] The spectral image data is cropped to make it the same size as the annotated image. After cropping, two image files are generated, which are used as the data set and label set for subsequent model training.
[0034] In one embodiment of the present application, the method for implementing the control of cultivated land planting purposes based on remote sensing interpretation, the model training: training through a convolutional neural network based on a data set and a label set to generate a spectral recognition model with classification and recognition capabilities includes:
[0035] Edge patching: Patching the edge pixels in the image to improve the resolution and quality of the image, remove or reduce the noise and interference signals in the image, and thus improve the clarity and recognizability of the image;
[0036] Tile processing: Converting each pixel of the spectral image data containing multiple bands into a 15*15*15 image block can better capture the spatial information in the image while retaining the spectral information, improving the accuracy of image classification, and avoiding excessive calculation or memory usage.
[0037] In one embodiment of the present application, the method for implementing cultivated land planting use control based on remote sensing interpretation, the data fusion: fusing the thematic map with the existing vector data including contracted land, three surveyed cultivated land and two zone demarcation to form a thematic map with higher recognition accuracy includes:
[0038] Based on the basic contracted land vector library, the spatial information, ownership relationship, and transfer information of each contracted land parcel are analyzed to roughly determine the planting situation of the contracted land parcel;
[0039] Taking the contracted land parcel as the basic object, query the third land survey vector database to determine the distribution, spatial coordinates, area, land use, and historical use changes of the contracted land parcel, and confirm the historical planting status of the contracted land parcel in one step;
[0040] Based on the two-zone demarcation vector library, it is determined whether the contracted land is a grain production functional zone or an agricultural product protection zone, and the planting purpose of the contracted land can be determined;
[0041] Combined with the crop planting status of the thematic map, the contracted plots are taken as the object, and the crops on the contracted plots are finally confirmed;
[0042] The results of the comprehensive analysis are saved again as thematic maps and compared with the sample data.
[0043] The implementation of this application can effectively solve the problems existing in the traditional methods of controlling the use of cultivated land for planting, and realize precise and automated control. Specific technical effects include:
[0044] 1. It solves the problem of achieving the best recognition results at a spatial resolution of 10m, improves the utilization rate of remote sensing data, makes it convenient to call historical data at any time, and significantly reduces the cost of data use.
[0045] 2. A feasible solution is proposed to solve the problem of image annotation, which effectively solves the problems of data annotation accuracy and verification of seasonal crop planting conditions on cultivated land in historical data.
[0046] 3. Utilize limited equipment resources to solve the problem of large data volume analysis and identification, improve the learning ability and recognition accuracy of the algorithm, and ensure the accuracy and reliability of farmland management and decision-making.
[0047] 4. Through geographic information system technology and system integration, analyze and present the crop planting structure, combine the cultivated land type and regional information, and generate a "one map" for the management and control of cultivated land planting use. Provide visual geographic information display and query functions, and improve the efficiency of managers' analysis and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0049] Figure 1 It is a flowchart of remote sensing interpretation and application;
[0050] Figure 2 It is a data fusion flow chart;
[0051] Figure 3 It is a flow chart of remote sensing interpretation data fusion;
[0052] Figure 4 are the training and validation accuracy plots;
[0053] Figure 5 is the training and validation loss rate plot;
[0054] Figure 6 is the original spectral pseudo RGB image;
[0055] Figure 7 This is the effect diagram after model recognition;
[0056] Figure 8 It is an entity relationship diagram. DETAILED DESCRIPTION
[0057] Remote sensing interpretation is the process of processing, analyzing, classifying and identifying remote sensing image data obtained by remote sensing technology. These remote sensing image data can be obtained through high-altitude satellites or low-altitude drones, including multispectral or hyperspectral data. Nowadays, high-quality remote sensing data at home and abroad are becoming more and more abundant, and historical data are also very complete. In terms of crop spectral recognition, deep learning technology based on neural networks has already achieved a fairly high accuracy. Therefore, the establishment of a crop recognition model library can greatly reduce the collection of low-altitude remote sensing data and avoid repeated data collection.
[0058] In the era of big data, technologies such as multidimensional data visualization, vector data query and layer spatial analysis of geographic information have also been widely used. The basic data such as the third survey of cultivated land, contracted land and second-district demarcation are also very complete.
[0059] Driven by the above policies, technologies and data, we can gradually improve the archival information, farmland construction and full digital dynamic monitoring of high-standard farmland, and realize the precise supervision of "non-agriculturalization" and "non-grainization" of cultivated land. This will effectively improve the accuracy and efficiency of cultivated land planting use control and accelerate the process of establishing national cultivated land spatiotemporal big data.
[0060] However, the spatial resolution of remote sensing images is positively correlated with the accuracy of recognition, but the cost of high-resolution remote sensing image data is quite high, and insufficient resolution of remote sensing image data will lead to poor recognition accuracy. There are problems such as difficult image annotation, high computing power requirements, low spectral model recognition accuracy, and difficult management and maintenance of the control system.
[0061] The technical problems to be solved by this application are:
[0062] 1. High resolution requirements for remote sensing data: The resolution here refers to the spatial resolution of remote sensing images, which is positively correlated with the accuracy of recognition. This brings about a cost issue. The cost difference between sub-meter and 30-meter resolution data purchased from commercial companies is quite large. This patent can achieve the best recognition results under general resolution conditions.
[0063] 2. Difficulty in image annotation: Remote sensing data is often historical data. For crops on cultivated land that are already highly seasonal, it is difficult to manually verify the planting situation on site. Even with a resolution of ten meters, it is difficult for the naked eye to distinguish specific plots of land and crops. This patent proposes a practical solution to solve this problem.
[0064] 3. High computing power requirements: The amount of hyperspectral image data is huge, and the memory, GPGPU and other hardware attributes required for training are relatively high. This patent solves the problem of how to use limited device resources to solve the problem of large-scale data analysis and recognition.
[0065] 4. Low spectral model recognition accuracy: The spectral recognition model established based on remote sensing image data of general resolution does not meet the application requirements in terms of model prediction accuracy. This requires the algorithm to fully learn and understand the relationship between spectral information, texture information and spatial information, and to combine the data of the third national land survey, the data of the two districts demarcation, and the data of rural land contract management rights to perform layer and patch fusion, thereby increasing the recognition accuracy to more than 95%.
[0066] 5. The management and control system is difficult to manage and maintain: This method integrates geographic information system technology with the system to achieve functions such as precise positioning of plots, terrain analysis, and soil type classification. It designs an intuitive and friendly user interface and visualizes complex agricultural data and information through charts, images, and maps, making it easier for managers and users to understand and analyze.
[0067] This application involves the fields of spectral recognition, image annotation, deep learning, data analysis, and geographic information systems. By extracting features from remote sensing data and establishing a recognition model based on a deep convolutional neural network, it is possible to quickly realize automatic recognition of large-scale plots and crops. At the same time, by using vector data in the management and control system for layer fusion and spatial analysis, accurate and standardized management and control of cultivated land planting purposes based on geographic information is achieved, providing solid support for agricultural production.
[0068] 1. Design Method and Process
[0069] See also Figure 1 As shown, to achieve the control of cultivated land planting use based on remote sensing interpretation technology, the following four processes are included.
[0070] 1. Data collection: including sample data and high / multi-spectral data. Using high-altitude satellites or low-altitude drones equipped with spectral cameras, spectral image data of cultivated land can be obtained. After interpretation, these data can be used for crop identification. Sample data can be high / multi-spectral data or color images taken by ordinary cameras. They are mainly used for feature extraction and verification, and have certain requirements for resolution.
[0071] 2. Image preprocessing: including radiation correction, geometric correction, atmospheric correction, image clipping, and image masking. Image masking will exclude non-crop areas such as water bodies and buildings to reduce noise interference and improve classification accuracy.
[0072] 3. Image interpretation: Extract features from data sets and label sets through convolutional neural networks to generate models with classification and recognition capabilities. This includes image annotation, model training, testing and optimization, image recognition, thematic map production, and data fusion.
[0073] 4. Manage and control one map: Through the fusion of multiple layers, with plots as units, the map displays the types of crops, planting areas, planting systems, distribution areas and ownership relationships. Through on-site verification, data can be dynamically updated. Historical data can be queried, compared and analyzed in time series.
[0074] 2. Key Technology Analysis
[0075] 1. Spectral data
[0076] Spectral data includes hyperspectral and multispectral data. The main data sources are obtained by applying for relevant licenses or fees, such as GEE Sentinel-1 / 2, Gaofen-1, etc. Through multiple comparative analyses of costs and results, it is determined that the use of 10m resolution images has the best cost-effectiveness in terms of economy and accuracy after interpretation (see Table 1). After later data fusion, the accuracy can reach more than 95%, which can fully meet the requirements of production.
[0077] Table 1 Comparison of spectral image data with different spatial resolutions
[0078] Serial number Spatial resolution Procurement costs Preliminary interpretation accuracy 1 30m Low 70%~80% 2 10m middle 80%~90% 3 5m Medium to high 85%~90% 4 2m high 88%~95% 5 1m Very high 90%~95% 6 0.5m Very high More than 95%
[0079] 2. Sample data
[0080] When establishing a classification and recognition model for spectral images, training data and test data are provided for feature extraction and result verification respectively, which need to meet the following requirements:
[0081] ① Data source: Sample data should come from areas related to the target mission and cover possible scenes and types of objects in the target area. For historical samples, images with a spatial resolution of about 1m can be purchased; for real-time data, high-resolution spectral images or true color orthophotos can be collected by drones. Because the sample data is small, the purchase cost is lower. In order to eliminate human errors, app surveys and other methods are not allowed to collect data.
[0082] ②Data quality: Sample data should be of high quality and high precision, avoiding problems such as noise, blur and artifacts.
[0083] ③ Data quantity: The sample data should be rich enough, containing enough images of different categories and scenes, so that the algorithm can learn various features in the data set. It can be calculated according to the following Cochran formula, where m is the number of samples to be estimated, n represents the total number, e represents the error level, and σ represents the standard deviation:
[0084]
[0085] For example, if a county covers an area of 1,500 square kilometers, the error level is 0.1, that is, the accuracy is controlled at 90%, and the standard deviation is 0.2, it can be obtained that the sample size in the county is about 4 square kilometers.
[0086] ④Data distribution: Sample data should cover different regions and seasons so that the algorithm can learn the features of different regions and seasons. Taking an ordinary camera with 20 million pixels and a focal length of 35mm as an example, the best flight altitude is about 100 meters. The camera can capture an area of about 1 square kilometer with a spatial resolution of 1m. Choose 4 places with different features such as forests, rivers and lakes, dry land, and gardens. For seasons, such as rice, it can be collected 3 times a year according to the needs of early, mid-, and late rice areas.
[0087] ⑤Data format: Sample data should be stored in a standard format, such as GeoTIFF, JPG, PNG, CVS, etc., so that the algorithm can read and process the data.
[0088] 3. Image cropping
[0089] ① The first step is to manually select the area and use GIS software to crop the unnecessary parts of the image to reduce the amount of data and improve processing efficiency.
[0090] ② The second step is to write a program to split the image into multiple areas of the same size to match the machine computing power and processing time, avoid overall failure and recalculation, and save computing resources. For example, a cluster of 30 RTX 2060s has a single-precision computing power of about 190TFLOPS. For a 40000*40000 pixel 4-channel multispectral image, the computing time is about 2 hours. The vertical segmentation code is as follows:
[0091]
[0092] 4. Image Annotation
[0093] This step combines the original spectral image data and sample data to extract features and provide a data set and label set for training the spectral recognition model. The quality of the annotation has a great impact on the accuracy of the model. The scope and quantity of the annotation correspond to the sample data one by one. The annotation process is as follows:
[0094] ① Select labeling software: You can choose labelme, GeoLabel, ArcMap and other software. ArcMap is recommended as it has powerful functions.
[0095] ② Open the original remote sensing image to be annotated in arcmap, such as tiff or img format. Associate three channels into R, G, and B channels to facilitate the identification of landforms.
[0096] ③ Select and create a labeled area on the original image. Generally, select an area with multiple terrains and complex terrain. The number of labeled ranges must meet the requirements mentioned above. Try to include all the crops to be classified in each labeled range. For example, in a remote sensing image, there are rice, corn, and wheat. All three crops must be included in the created labeled area. For example, on the original spectral image of 40000*40000, an area of about 1000*1000 pixels is intercepted for labeling.
[0097] ④ Manually mark the type of features. In the marked area, use point, line, and surface tools to draw the shape of each crop planting range in comparison with the sample data, and use different numbers to identify its classification. For example, if there are rapeseed, wheat, ridges, plastic greenhouses, grasslands, etc. in this area, mark their corresponding categories in the attribute table as 1, 2, 3, 4, 5, and 6, a total of six categories, of which 6 represents all other unlabeled features.
[0098] ⑤ After marking, merge and save. Merge objects of the same category so that there is only one object in each category, which is convenient for subsequent processing. Then convert the marked graphics into a tif file for saving.
[0099] ⑥ Annotation cropping. Crop the original image to make it the same size as the annotated image. After cropping, two tif image files will be obtained, which will be used as the data set and label set for subsequent model training.
[0100] 5. Model training
[0101] As shown in Table 2, the model training was carried out based on the convolutional neural network structure of deep learning. After repeated tests, parameter adjustment and algorithm optimization, the test accuracy was finally improved to about 90%. The following are some key points:
[0102] ①Parameter adjustment:
[0103] Test set: The training set is set to 0.3;
[0104] Number of training rounds: 20 times to reach 90%, 30 times to reach 95%;
[0105] Activation function: relu and softmax;
[0106] Batch size: related to video memory, can be set to 128;
[0107] Tile size: 9*9*9 or 15*15*15 is optimal;
[0108] Learning efficiency: 1e-03;
[0109] Learning rate decay: 1e-06
[0110] ② Edge patching: Patching the edge pixels in the image to improve the resolution and quality of the image, remove or reduce the noise and interference signals in the image, and thus improve the clarity and recognizability of the image. The patch size is half the size of the tile.
[0111] ③ Tile processing: Converting each pixel in the hyperspectral image (containing multiple bands) into a 15*15*15 image block can better capture the spatial information in the image while retaining the spectral information, improving the accuracy of image classification and avoiding excessive computation or memory usage.
[0112] ④ Optimize the convolutional neural network structure and recognition algorithm: Limited by the accuracy and availability of remote sensing images, we first extract and call various features in remote sensing images by optimizing object-oriented segmentation parameters. Then, we perform feature space optimization on spectral features, vegetation index features, texture features, and geometric features. Finally, we set up multiple feature optimization tests and integrate multiple classification algorithms for classification comparison to obtain the optimal classification model.
[0113] Table 2 Network structure parameters
[0114] Network structure Output size Convolution kernel, filter, step size conv layer1 13×13×13 [3×3×3,16,1] conv laver2 9×9×9 [5×5×5,32,1] conv layer3 5×5×5 [5×5×5,64,1] conv layer4 3×3×3 [3×3×3,128,1] conv laver5 1×1×1 softmax Total Param 5,488,243 FLOPS <![CDATA[1.73×10 8 ]]>
[0115] 6. Thematic map production
[0116] After crop identification through the spectral recognition model, a crop distribution map of the identification results will be obtained. First, use binarization to remove image noise, then convert raster data into vector data, and finally output a thematic layer file in shapefile format for each crop. The relevant code is as follows:
[0117]
[0118] 7. Data Fusion
[0119] The prediction accuracy of the spectral model after forming the thematic map is 90%. At this time, it is necessary to combine the existing vector data such as contracted land, the third survey of cultivated land and the second zoning, and further perform layer data fusion to improve the accuracy to more than 96%.
[0120] See also Figure 2 , the specific fusion steps are as follows:
[0121] ① Based on the basic contracted land vector library, the spatial information, ownership relationship, and transfer information of each plot of land are analyzed to roughly determine the planting situation of the plot of land;
[0122] ② Taking the plot as the basic object, query the third land survey vector database to determine the distribution, spatial coordinates, area, land use, and historical use changes of the plot, and further confirm the historical planting status of the plot;
[0123] ③ Based on the two-zone demarcation vector library, determine whether the plot is a grain production functional zone or an agricultural product protection zone, which can basically determine the planting purpose of the plot;
[0124] ④ Combined with the crop planting status of the thematic map, take the plot as the object and finally confirm the crops of the plot.
[0125] ⑤ Save the comprehensive analysis results as thematic maps and compare them with sample data. After multiple adjustments to the algorithm, the actual recognition accuracy reached more than 96%.
[0126] 8. Query analysis.
[0127] Import various vector data into the spatial database postgreSQL, and use postGIS to perform complex spatial queries, statistics, overlays, interpolation, clustering, tracing, and association analysis; use WEB GI S technology and WMS protocol, use Tiandi Map as the base map, dynamically generate various thematic layers through the WMS protocol, and overlay them on the base map for display. Realize the integrated display of multiple data sources and improve the visualization effect and information expression ability of map data.
[0128] Figure 3 It explains the process of fully integrating the thematic maps after remote sensing interpretation with the existing survey data, significantly improving the accuracy of crop identification, and ultimately forming a single map for management and control. Specific embodiment:
[0130] The following uses rapeseed spectral image recognition in Qiyang County as an example to introduce the process of farmland use control:
[0131] 1. Purchase spectral remote sensing satellite data with a spatial resolution of 10m. The satellite sources can be LAND SAT series, Environmental series, ALOS series, TERRA series, Sustainable Development Science Satellite series, Sentinel series, Tianxian series, Luojia No. 1 series, AQUA series, etc. Take Qiyang County as an example, the area is 2547 square kilometers. According to the price in the table below, it is less than one scene, and the price is between 2500 and 5000 yuan, which is very cost-effective.
[0132] Table 3 Price list of domestic high-resolution satellite image series (sold by scene)
[0133]
[0134] 2. Image preprocessing, radiation calibration, geometric correction and atmospheric correction are performed to eliminate interference factors.
[0135] 3. Then select the band or area of interest and ignore other bands or areas to obtain a mask.
[0136] 4. Finally, in order to meet the computing power requirements, the image is cropped according to the specified pixel width and pixel height.
[0137] 5. Use ArcMap software to mark the crops on different types of plots to obtain data sets and label sets.
[0138] 6. Use the convolutional neural network structure to build a spectral recognition model, divide the data set and label set into training set and test set, and perform model training.
[0139] 7. Use the test set to verify the trained model, repeatedly adjust the model parameters, optimize the algorithm, and increase the model prediction accuracy to more than 90%. The accuracy and loss rate of the training and optimization process are as follows: Figure 4 and Figure 5 shown.
[0140] 8. Save the trained spectral model as a .h5 file, and then perform crop recognition on the preprocessed image.
[0141] 9. Save the crop types as thematic shapefiles.
[0142] 10. Combined with the contracted land, the third round of cultivated land survey, and the second zone demarcation data of the Ministry of Agriculture and Rural Affairs, the crop thematic map data will be integrated, and the recognition accuracy will eventually be increased to more than 95%. Figure 7 This is the final recognition result of rapeseed in Qiyang County. Figure 6 is the original spectral pseudo-RGB image, Figure 7 This is the effect diagram after model recognition.
[0143] 11. After data extraction, conversion and cleaning, the identification results are uniformly stored in the spatial database PostgreSQL to establish a land parcel information file. Figure 8 , including plot information and planting information.
[0144] 12. Dynamic maintenance of plot planting information, including new data entry, historical data update, and data release.
[0145] ① New data entry: Continuously update the current status of cultivated land planting. The process is detailed in the previous steps.
[0146] ② Historical data update: For those who have completed spectral analysis and whose data has been stored, mobile app software can be used to verify the location, area, attributes and other information of each plot on the cultivated land by village, and conduct additional investigations on crop types, planting systems, sowing areas, and abandoned land. Verification and investigation shall be carried out on a county basis to find out the current status of cultivated land planting and use. After the verification and supplementary investigation information is reviewed by the provincial agricultural and rural departments, the historical results database shall be updated in a timely manner.
[0147] ③Data release: Build a map server GeoServer, and provide county-level cultivated land planting use status layer and cultivated land planting use verification layer through WMS and HTTPS protocols. The layer is composed of crop thematic maps analyzed in the early stage.
[0148] 13. Use WEB GIS technology to form a management and control map for query and analysis.
[0149] ① Using the sky and earth image map as the base map, various thematic layers can be displayed in a tree-like hierarchical structure according to business categories. It should include but not be limited to: national land space planning map, land use status map, cultivated land planting use control map, and "two zones" demarcation map.
[0150] ②Each layer can be displayed in superposition, and the layers can be dragged, enlarged and rotated at will.
[0151] ③ Click to view any plot number, land type, slope, property nature, ownership unit, administrative division and other information.
[0152] ④ Check the current status of cultivated land planting, including data analysis results such as crop type, planting system, and yield.
[0153] ⑤ Statistics are carried out by administrative divisions, time series, nature of plots, etc., supporting pie charts, cube charts, line charts, etc.
[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A method for controlling the use of cultivated land based on remote sensing interpretation. It is characterized in that The steps include: 1) Data collection: Collect spectral image data of the target area and sample data of scene types and ground object types appearing in the target area; 2) Image preprocessing: perform radiation correction, geometric correction, atmospheric correction, image cropping and image mask processing on the collected spectral image data; Through radiation correction, the digital values in the spectral image data are converted into reflectance or radiation brightness values to eliminate the influence of atmospheric disturbances at different times and locations; Through geometric correction, spectral image data is geometrically transformed to eliminate geometric distortion and errors caused by image acquisition, transmission and processing; Atmospheric correction converts the radiation values in spectral image data affected by atmospheric absorption and scattering into surface reflectance or radiation brightness values. Image clipping is used to clip unnecessary parts of spectral image data to reduce data volume and improve processing efficiency; By using masks, non-crop areas such as water bodies and buildings in spectral image data can be excluded to reduce noise interference and improve classification accuracy; 3) Image annotation: extract features from spectral image data and sample data to generate data sets and label sets; 4) Model training: Based on the data set and label set, a convolutional neural network is trained to generate a spectral recognition model with classification and recognition capabilities; 5) Thematic map production: The crop identification in the target area is carried out through the spectral recognition model to obtain the crop distribution map containing the identification results. The noise of the crop distribution map is removed by binarization, and then the raster data is converted into vector data. Finally, a separate thematic layer file is output for each crop. 6) Data fusion: The thematic map is integrated with the existing vector data including contracted land, the third survey of cultivated land and the second zone demarcation to form a thematic map with higher recognition accuracy; 7) Generate a management and control map: Through the fusion of multiple layers, with plots as units, the map shows the types of crops, planting areas, planting systems, distribution areas and ownership relationships. Through on-site verification, data can be dynamically updated, and historical data can be queried, compared and analyzed in time series.
2. The method for controlling the use of cultivated land based on remote sensing interpretation according to claim 1, It is characterized in that The spectral image data includes hyperspectral image data and multispectral image data, and the spectral image data adopts images with a spatial resolution of 10m.
3. The method for controlling the use of cultivated land based on remote sensing interpretation according to claim 1, It is characterized in that The sample data includes historical sample data and real-time data; The historical sample data adopts images with a spatial resolution of 1m, and the real-time data adopts spectral images or true color orthophotos collected by fixed-point photography of unmanned aerial vehicles.
4. The method for controlling the use of cultivated land based on remote sensing interpretation according to claim 3, It is characterized in that The sample data to be estimated is calculated by the formula: Among them, m is the number of samples to be estimated, n represents the total number, e represents the error level, and σ represents the standard deviation.
5. The method for controlling the use of cultivated land based on remote sensing interpretation according to claim 1, It is characterized in that The method of clipping unnecessary parts of the spectral image data to reduce the data volume and improve the processing efficiency includes: By manually selecting areas, the unnecessary parts of the image can be cropped using GIS software to reduce the amount of data and improve processing efficiency; Write a program to divide the image into multiple areas of equal size to match the machine computing power and processing time, avoid overall failure and recalculation, and save computing resources.
6. The method for controlling the use of cultivated land based on remote sensing interpretation according to claim 1, It is characterized in that The image annotation: extracting features from spectral image data and sample data to generate a data set and a label set includes: Open the spectral image data in the annotation software and associate three of the channels into R, G, and B channels to facilitate the identification of the landforms; Select and create a labeled area. Select an area with multiple terrains and complex terrain. Try to include all the crops to be classified within each labeled area. In the marked area, compare the sample data and use point, line and surface tools to draw the shape of each crop planting area, and use different numbers to identify its classification; Merge objects of the same category so that there is only one object in each category, which is convenient for subsequent processing; The spectral image data is cropped to make it the same size as the annotated image. After cropping, two image files are generated, which are used as the data set and label set for subsequent model training.
7. The method for controlling the use of cultivated land based on remote sensing interpretation according to claim 1, It is characterized in that The model training: training through a convolutional neural network based on a data set and a label set to generate a spectral recognition model with classification and recognition capabilities includes: Edge patching: Patching the edge pixels in the image to improve the resolution and quality of the image, remove or reduce the noise and interference signals in the image, and thus improve the clarity and recognizability of the image; Tile processing: Converting each pixel of the spectral image data containing multiple bands into a 15*15*15 image block can better capture the spatial information in the image while retaining the spectral information, improving the accuracy of image classification, and avoiding excessive calculation or memory usage.
8. The method for controlling the use of cultivated land based on remote sensing interpretation according to claim 7, It is characterized in that The data fusion: fusing the thematic map with the existing vector data including contracted land, the third survey of cultivated land and the second zone demarcation to form a thematic map with higher recognition accuracy, including: Based on the basic contracted land vector library, the spatial information, ownership relationship, and transfer information of each contracted land parcel are analyzed to roughly determine the planting situation of the contracted land parcel; Taking the contracted land parcel as the basic object, query the third land survey vector database to determine the distribution, spatial coordinates, area, land use, and historical use changes of the contracted land parcel, and confirm the historical planting status of the contracted land parcel in one step; Based on the two-zone demarcation vector library, it is determined whether the contracted land is a grain production functional zone or an agricultural product protection zone, and the planting purpose of the contracted land can be determined; Combined with the crop planting status of the thematic map, the contracted plots are taken as the object, and the crops on the contracted plots are finally confirmed; The results of the comprehensive analysis are saved again as thematic maps and compared with the sample data.
Citation Information
Cited By
Unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing
CN120870015A
