Ploughing greenhouse remote sensing extraction method and system based on improved target detection

Abstract

The invention discloses a cultivated land greenhouse remote sensing extraction method and system based on improved target detection, and the method comprises the steps: S1, obtaining a high-resolution remote sensing image of a time sequence, carrying out the remote sensing preprocessing, and making a sample data set; s2, constructing and improving a YOLOv8-OBB deep learning target detection algorithm model; s3, performing iterative training on the YOLOv8-OBB target detection algorithm model, setting various hyper-parameters, and performing comparison to obtain an optimal model after iterative training; and S4, target detection result post-processing: based on the YOLOv8-OBB target detection algorithm model of the optimal parameters, carrying out identification extraction on buildings and greenhouses in the region, obtaining a target detection result, and outputting the target detection result as a geographic vector data file. According to the method, certain success is achieved in building and greenhouse room identification and extraction, the buildings and the greenhouse rooms in the cultivated land range can be accurately and efficiently identified and extracted, accurate point location data is provided for related departments, and the efficiency of dynamic cultivated land monitoring and supervision is improved.

Description

Technical Field

[0001] The invention relates to the field of agricultural remote sensing monitoring, and in particular to a remote sensing extraction method and system for greenhouse houses on cultivated land based on improved target detection. Background Art

[0002] Cultivated land is the foundation for the survival and development of human civilization, and it is related to the national economy and social development. In recent years, due to the rapid economic development and the accelerating urbanization process in my country, the non-agriculturalization of cultivated land has occurred from time to time, and even in some areas, the non-agriculturalization of cultivated land has continued to increase, seriously threatening my country's food security. Therefore, how to quickly, accurately and automatically identify and extract the non-agriculturalization of cultivated land is of great significance to strictly adhere to the red line of cultivated land in my country and ensure the sustainable development of social economy.

[0003] High-resolution remote sensing images have the characteristics of high spatial resolution, large coverage, short revisit period, rich texture features and geometric features, etc. They have been widely used in natural resource monitoring and provide strong data support for timely and accurate acquisition of cultivated land change information. At present, the identification, extraction and monitoring of high-frequency and large-scale cultivated land non-agriculturalization are mainly divided into two categories: one is the traditional method and the other is the deep learning method.

[0004] The traditional method of extracting farmland non-agriculturalization is to construct a highly interpretable farmland non-agriculturalization identification and extraction model based on the spectral and spatial characteristics of buildings and greenhouses on remote sensing images, thereby extracting buildings and greenhouses. However, the traditional extraction and monitoring method requires artificial design of feature extraction algorithm models based on prior knowledge, so the algorithm model has a certain degree of subjectivity. At the same time, remote sensing images taken in different regions, at different times, and at different angles are easily affected by external factors such as region, lighting, and sensors, which makes the algorithm model's generalization ability for monitoring large-scale farmland non-agriculturalization behavior relatively poor, and the extraction accuracy of buildings and greenhouses and the extraction accuracy of contour edges are relatively poor.

[0005] The method of extracting farmland from non-agricultural areas based on deep learning generally adopts a semantic segmentation algorithm model, which uses samples to automatically learn the deep features of remote sensing images, such as the texture features, geometric features, and spectral features of buildings and greenhouses in remote sensing images, and then builds a farmland non-agricultural identification and extraction model. Although the deep learning method has a high degree of automation, the existing deep learning identification and extraction model still faces the problem of unreliable automatic interpretation results under remote sensing images with a wide and complex range, which is mainly manifested in problems such as inaccurate extraction accuracy and contour edge extraction.

[0006] In summary, traditional methods usually need to rely on rich prior knowledge to design algorithms for the image features of cultivated land non-agriculturalization. However, due to the obvious differences in remote sensing image features in different regions and at different times, the selection of features needs to be adjusted repeatedly, resulting in poor generalization ability of the algorithm model and poor accuracy in contour edge extraction. Although deep learning methods have a strong ability to automatically learn image features and extract universal features by learning a large number of samples, they still ignore the representation of image features of cultivated land non-agriculturalization in different time and space, resulting in insufficient model targeting and relatively low accuracy in contour edge extraction. Summary of the invention

[0007] The present invention provides a remote sensing extraction method and system for greenhouse houses on cultivated land based on improved target detection. The method and system are based on the YOLOv8-OBB target detection algorithm, an improved model algorithm, a remote sensing data processing module (sliding window, band combination) and an inference module, and a directional frame extraction method for identification and extraction. The method solves the technical problems of low extraction accuracy of buildings and greenhouse houses, poor edge extraction accuracy, weak model generalization ability, and low reasoning and prediction efficiency.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A remote sensing extraction method for greenhouse houses on cultivated land based on improved target detection comprises the following steps:

[0010] S1, obtain time series high-resolution remote sensing images and perform remote sensing preprocessing to produce sample data sets;

[0011] S2. Build and improve the YOLOv8-OBB deep learning target detection algorithm model;

[0012] S3. Iteratively train the YOLOv8-OBB target detection algorithm model, set various hyperparameters, and compare them to obtain the optimal model after iterative training;

[0013] S4. Post-processing of target detection results: Based on the YOLOv8-OBB target detection algorithm model with optimal parameters, the buildings and greenhouses in the area are identified and extracted, the target detection results are obtained, and output as geographic vector data files.

[0014] As a preferred embodiment of the above solution, step S1 specifically includes:

[0015] S101. Screening sample areas that meet target conditions: Based on the acquired time-series high-resolution remote sensing images, the remote sensing images are pre-processed using ENVI 5.3 remote sensing processing software;

[0016] S102, drawing vector patches: Based on the high-resolution remote sensing image processed in step S101, a rectangular vector patch is drawn using the rectangle drawing tool of the vector editing module through QGIS 3.4.2 geographic information processing software. The patch contains a single or two targets such as buildings and greenhouses, but all are marked with FN as labels;

[0017] S103, clipping of raster data: Based on the high-resolution remote sensing image processed by S101 and the vector patches containing the target drawn by S102, the raster clipping tool by mask layer in QGIS 3.4.2 geographic information processing software is used to clip the remote sensing image one by one using the vector patches, and the clipped raster image is saved as image data;

[0018] S104, preparation of oriented frame OBB sample data set: based on the raster data clipped and saved in S103, the roLabelImg tool is used to draw and adjust the direction and size of the rectangular frame according to the direction and size of the building and greenhouse, and save it;

[0019] S105, oriented frame sample data set label format conversion: convert the data saved in S104 into a format suitable for YOLOv8-OBB target detection algorithm model label input.

[0020] As a preferred embodiment of the above scheme, remote sensing preprocessing includes: radiation calibration, atmospheric correction, geometric correction, image fusion, geographic registration and image uniform color mosaic operation.

[0021] As a preferred embodiment of the above solution, step S2 specifically includes:

[0022] S201. Based on the Windows operating system, the anaconda deep learning package installation manager is used to build the YOLOv8-OBB target detection algorithm model based on the YOLO v8 official website document through the PyTorch deep learning framework;

[0023] S202. Based on the YOLOv8-OBB algorithm model, the main part is improved, the Conv and C2f in Backone are replaced by SCConv space and channel reconstruction convolution, and the input mode and output mode of the prediction module are modified to make it suitable for reading and outputting remote sensing images;

[0024] S203, using a sliding window method to obtain remote sensing image data, recombining the bands, and automatically outputting vector data through coordinate conversion.

[0025] As a preferred embodiment of the above scheme, in step S201: according to the environment required by YOLO v8, a YOLOv8-OBB virtual environment is constructed, in which a GPU version of the PyTorch deep learning framework is installed, according to the YOLO v8 official website, the environment library required by YOLO v8 is installed, and according to the geographic information requirements, a remote sensing geographic information management package is installed.

[0026] As a preferred embodiment of the above solution, step S3 specifically includes:

[0027] S301. During iterative model training, download the official pre-trained model, load it into the YOLOv8-OBB target detection algorithm model, input your own sample training set, and set iterative training hyperparameters;

[0028] S302. By debugging different hyperparameters, the influencing factors of the target detection, recognition and extraction precision and accuracy are determined, and then the optimal training parameters are obtained to determine the optimal model.

[0029] As a preferred embodiment of the above solution, step S4 specifically includes:

[0030] S401, based on the sliding window, the entire scene remote sensing image is continuously input into the optimal YOLOv8-OBB target detection algorithm model for prediction, and the YOLOv8-OBB target detection result is obtained, which mainly includes confidence, category information and pixel coordinates of 4 points;

[0031] S402, reading the geographic reference information and projection information of the remote sensing image captured by the sliding window, and converting the pixel coordinates of the four points into a geographic coordinate system through the latitude and longitude formula;

[0032] S403, constructing a vector file through the gdal library, and writing the confidence level, category information, the four converted geographic coordinate points, and the longitude and latitude of the center point of each vector spot into the vector file.

[0033] A remote sensing extraction system for greenhouse houses on cultivated land based on improved target detection, comprising:

[0034] A remote sensing image preprocessing module is used to obtain time-series high-resolution remote sensing images and perform remote sensing image preprocessing operations on the remote sensing images;

[0035] Vector drawing and cropping module, used for sample drawing and cropping of pre-processed remote sensing image data;

[0036] YOLOv8-OBB target detection algorithm model training module, which is used to build the model main structure module, training module, testing module and prediction module of the YOLOv8-OBB detection algorithm model;

[0037] YOLOv8-OBB target detection result post-processing module, used to identify and extract buildings and greenhouses in the area, obtain target detection results, and output them as geographic vector data files

[0038] Due to the above structure, the beneficial effects of the present invention are:

[0039] The present invention is based on the YOLOv8-OBB target detection algorithm, improves the model algorithm, and adopts the directional frame extraction method for identification and extraction based on the remote sensing data processing module (sliding window, band combination) reasoning module (threshold method). It has achieved certain success in the identification and extraction of buildings and greenhouses. It can accurately and efficiently identify and extract buildings and greenhouses within the scope of cultivated land, provide accurate point data for relevant departments, and improve the efficiency of dynamic monitoring and supervision of cultivated land. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below.

[0041] Figure 1 is a flow chart of the method of the present invention;

[0042] Figure 2 To obtain the time series images, the remote sensing images are pre-processed by remote sensing;

[0043] Figure 3 Based on the pre-processed remote sensing image, a rectangular vector file diagram of buildings and greenhouses within the cultivated land area is drawn by using vector tools;

[0044] Figure 4 This is a schematic diagram of a cropped remote sensing image based on a pre-processed remote sensing image using the Crop Raster by Mask Layer tool;

[0045] Figure 5 This is a schematic diagram of obb data labeled based on the roLabelImg tool;

[0046] Figure 6 This is a schematic diagram of the storage format of the original XML format label data annotated by roLabelImg;

[0047] Figure 7 This is a schematic diagram of the storage format of the original XML format labels annotated by roLabelImg into XML format label data of the DOTA dataset;

[0048] Figure 8 This is a diagram of the conversion of the XML tags of the DOTA dataset into the TXT format tag data storage format of the DOTA dataset;

[0049] Fig. 9This is a schematic diagram of the storage format of the txt format label data used for converting the txt format label of the DOTA dataset to obb;

[0050] Fig.10 Schematic diagram of SCConv structure;

[0051] Fig.11 Replace the improved diagram based on YOLOv8-OBB;

[0052] Fig.12 Schematic diagram of the original remote sensing image to be predicted;

[0053] Fig.13 Schematic diagram of the original remote sensing image to be predicted and the corresponding vector result overlay diagram;

[0054] Fig.14 Schematic diagram of the original remote sensing image to be predicted;

[0055] Fig.15 Schematic diagram of the original remote sensing image to be predicted and the corresponding vector result overlay diagram;

[0056] Fig.16 Schematic diagram of the original GF-2 image to be predicted;

[0057] Fig.17 Schematic diagram of the vector result overlay corresponding to the original GF-2 image to be predicted;

[0058] Fig.18 Schematic diagram of the original GF-2 image to be predicted;

[0059] Fig.19 Schematic diagram of the superposition of the corresponding vector results of the GF-2 original image to be predicted DETAILED DESCRIPTION

[0060] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.

[0061] See also Figures 1 to 19 .

[0062] This embodiment provides a remote sensing extraction method for greenhouse houses on cultivated land based on improved target detection, comprising the following steps:

[0063] S1. Obtain time series of high-resolution remote sensing images. Based on the scope of the study area, obtain time series of high-resolution remote sensing images through a third-party remote sensing mall. Specific remote sensing index parameters include but are not limited to: 4-band multispectral, spatial resolution better than 0.75 meters, cloud cover less than 10%, key areas must be cloudless, and temporal resolution reaches one period per month. The remote sensing images used in the present invention are 2 periods in total, mid-April and mid-May, respectively, with cloud cover less than 5%, and the spatial resolution of cloudless key areas is better than 0.75 meters for 4-band remote sensing image data. According to the target areas of buildings and greenhouses, 240 raster images with a pixel value size of 640×640 are randomly cropped to facilitate the production of a directional box target detection algorithm model training data set. Specifically including:

[0064] S101, screening sample areas that meet the target conditions. Based on the acquired time-series high-resolution remote sensing images, the remote sensing images are preprocessed using ENVI 5.3 remote sensing processing software, mainly including: radiation calibration, atmospheric correction, geometric correction, image fusion, geographic registration, and image uniform color mosaic operations.

[0065] S102, drawing vector patches. Based on the high-resolution remote sensing image processed in step S101, a rectangular vector patch with a pixel value of 640×640 is drawn using the rectangle drawing tool of the vector editing module through QGIS 3.4.2 geographic information processing software. The patch contains a single or two targets such as buildings and greenhouses, but they are all labeled with FN.

[0066] S103, clipping of raster data. Based on the high-resolution remote sensing image processed by S101 and the vector patches containing the target drawn by S102, the high-resolution remote sensing image processed by S101 is clipped one by one using the vector patches containing the target drawn by S102 through the clipping raster tool by mask layer in QGIS 3.4.2 geographic information processing software, and the clipping of the raster image after S101 is saved as image data in .tif format.

[0067] S104. Production of oriented frame (obb) sample data set. Based on the raster data clipped and saved in S103, use the roLabelImg tool, open the tool, click Open Dir, open the raster data path (the path should not contain Chinese characters, otherwise there will be an error), and the first raster image will be automatically displayed. Then, visually check the location of the building and greenhouse, click CreateRotated RBox, draw a rectangular box, and adjust the direction and size of the rectangular box according to the direction and size of the building and greenhouse. Use the ZXCV shortcut key to adjust the direction, and adjust the size by pressing the inflection point with the left mouse button. After adjustment, click Save.

[0068] S105, oriented box (obb) sample data set label format conversion. In step S104, when drawing labels, the saved format is xml format. However, the input format of the YOLOv8-OBB target detection algorithm model label is txt text format. Therefore, it is necessary to convert the xml format to a txt format available to YOLOv8-OBB. In this process, due to the different storage forms of the training sample data, three conversions are required. First, the original xml format data drawn by the roLabelImg tool is converted to xml format data in the form of a DOTA data set; secondly, the xml format data of the DOTA data set is converted to the txt format of the DOTA data set; finally, the txt format of the DOTA data set is converted to a txt format available to YOLOv8-OBB. The data storage forms of the four formats are as follows: (1) The original xml format data drawn by the roLabelImg tool is stored in the following form: category: name, the coordinates of the center point of the rectangle: cx, cy, the width and height of the rectangle w, h, and the rotation angle of the rectangle relative to the horizontal: angle; (2) The xml format data storage form of the DOTA dataset is stored in the following form: category: name, the coordinates of the four points of the rectangle are: x0, y0, x1, y1, x2, y2, x3, y3; (4) The data storage format of the DOTA dataset in xml format is: the coordinates of the four points of the rectangular box: x0, y0, x1, y1, x2, y2, x3, y3, category number: 0, category: cls, separated by spaces, and rows are separated by line breaks; (5) The data storage format of the YOLOv8-OBB available in txt format is: category number: 0, the coordinates of the four points of the rectangular box: x0, y0, x1, y1, x2, y2, x3, y3, separated by spaces, and rows are separated by line breaks.

[0069] Based on the above S101-105, the acquisition, preprocessing, sample vector spot drawing, sample image cropping, sample image label drawing, and sample label format conversion of time series images are completed. Figure 2 In order to obtain time series remote sensing data, the preprocessed remote sensing images are Figure 3 Draw a schematic diagram for the sample vector patch, Figure 4 This is a schematic diagram of the sample image after cropping. Figure 5 This is a schematic diagram of the labels of the sample image annotated by the roLabelImg tool. Figure 6-Figure 9 Schematic diagram of different storage formats of sample labels.

[0070] S2. Build and improve the YOLOv8-OBB deep learning target detection algorithm model, and improve the reasoning prediction module. The experiment of the present invention uses a laptop computer based on the Windows operating system as the hardware environment, uses the anaconda deep learning software package installation manager, and uses the PyTorch deep learning framework to build the YOLOv8-OBB target detection algorithm model based on the YOLO v8 official website document. First, based on the YOLOv8-OBB algorithm model, improve the main part (Backone), and use SCConv (spatial and channel reconstruction convolution, structure see Fig.10 Replace Conv and C2f in Backone for specific locations. Fig.11 As shown. The efficient convolution module of SCConv (spatial and channel reconstruction convolution) is used to reduce the spatial and channel redundancy in the convolutional neural network, optimize the feature extraction process, reduce the consumption of computing resources and improve the network performance, thereby improving the algorithm extraction speed and extraction accuracy; secondly, the model input method is modified to make it suitable for the remote sensing image raster data format, which is convenient for iterative training of the YOLOv8-OBB target detection algorithm model; finally, the output method of the YOLOv8-OBB target detection algorithm model is modified to output the predicted results as geographic vector data files, which is convenient for statistical analysis of the results. Specifically include:

[0071] S201, build YOLOv8-OBB deep learning target detection algorithm model, and improve the reasoning prediction module. The present invention is based on the windows operating system, installs the anaconda deep learning software management package, and builds a virtual environment named YOLOv8-OBB according to the environment required by YOLO v8. In the YOLOv8-OBB virtual environment, install the GPU version of the PyTorch deep learning framework, and according to the YOLO v8 official website, install the environment library required by YOLO v8, such as the ultralytics library, sahi library, matplotlib library, etc. According to the geographic information requirements, install the remote sensing geographic information management package: gdal remote sensing database, tifffile library. At this point, the relevant environment is configured.

[0072] S202, algorithm improvement, first, based on the YOLOv8-OBB algorithm model, improve the main part (Backone), use SCConv (spatial and channel reconstruction convolution, structure see Fig.10 Replace Conv and C2f in Backone for specific locations. Fig.11As shown in the figure. The efficient convolution module of SCConv (spatial and channel reconstruction convolution) is used to reduce the spatial and channel redundancy in the convolutional neural network, optimize the feature extraction process, reduce the consumption of computing resources and improve the network performance; secondly, the model input method is modified to make it suitable for the remote sensing image raster data format, which is convenient for iterative training of the YOLOv8-OBB target detection algorithm model; finally, the output method of the YOLOv8-OBB target detection algorithm model is modified to output the predicted results as geographic vector data files, which is convenient for statistical analysis of the results.

[0073] S203, improve the reasoning prediction module, based on the characteristics of remote sensing image data, that is, the number of channels is 4 channels, namely R, G, B, NIR, and at the same time, the remote sensing image data is large, and the width and height are usually tens of thousands of pixels. Therefore, in the reasoning prediction module, the present invention adopts a sliding window method to obtain remote sensing image data, recombines the bands to make it suitable for the reasoning model input method, and then obtains the reasoning prediction results. At the same time, after coordinate conversion, vector data is automatically output to improve the automation of remote sensing image recognition and extraction.

[0074] S3, iterative training of YOLOv8-OBB target detection algorithm model. The above-mentioned directional box sample data set is input into the YOLOv8-OBB target detection algorithm model for iterative training until the performance of the YOLOv8-OBB target detection algorithm model meets the preset conditions. Through experiments, iterative training is performed based on a small sample training data set (240 images of Jilin-1), and the model accuracy is 95.88%, and the prediction accuracy is 94.587%, which reflects the accuracy of the model. Inference and prediction are performed on the GF-2 images of different times, different regions, and different sensors, and there is still an 80% extraction accuracy, which shows the good generalization of the model. By performing inference and prediction on the remote sensing images of the Yangtze River New Area (610 square kilometers) in different periods, the overall time spent is 16 minutes, which reflects the high efficiency of the algorithm. Specifically including:

[0075] S301. Since the sample data produced by the present invention is 240 in total, it is a small sample data set. Therefore, during the iterative training of the model, download the official pre-trained model, load it into the YOLOv8-OBB target detection algorithm model, input your own sample training set, and set the iterative training hyperparameters;

[0076] S302. By debugging different hyperparameters, determine the influencing factors of target detection and recognition extraction precision and accuracy, and then obtain the optimal training parameters and determine the optimal model. Through comparative analysis, the specific hyperparameters of this experiment are as follows: lr0:0.001, lrf:0.01, epochs:100, patience:100, batch:8, imgsz:640, cache:False, device:0, workers:8, optimizer:auto, single_cls:True, close_mosaic:10, and other parameters are default values. Through iterative training of the model, the model extraction accuracy reached 93.38%, the recall rate reached 90.63%, the maAP50 reached 92.58%, and the mAP50-90 reached 71.72%.

[0077] S4. Post-processing of target detection results. Based on the YOLOv8-OBB target detection algorithm model with optimal parameters, identify and extract buildings and greenhouses in the area to obtain target detection results. By modifying the output result format, it can be output as a geographic vector data file. Through vector operations, the latitude and longitude coordinates, area, and confidence level of identification and extraction of each vector spot are obtained to facilitate statistical analysis of the results and on-site inspection. Specifically include:

[0078] S401, post-processing of YOLOv8-OBB target detection results, based on step S2, the YOLOv8-OBB target detection algorithm model with the best parameters can be obtained. Based on the sliding window, the entire scene remote sensing image is continuously input into the optimal YOLOv8-OBB target detection algorithm model for prediction to obtain the YOLOv8-OBB target detection results, which mainly include confidence, category information and pixel coordinates of 4 points;

[0079] S402: Since the four point coordinates are based on the pixel coordinate values ​​of the image, they need to be converted into a geographic coordinate system. By reading the projection information and geographic reference information of the remote sensing image, the obtained point coordinates are converted into a geographic coordinate system, i.e., longitude and latitude, according to a formula.

[0080] Longitude calculation formula:

[0081] geo_x=top_left_x+x*geo[1]+y*geo[2]

[0082] Latitude calculation formula:

[0083] geo_y=top_left_y+x*geo[4]+y*geo[5]

[0084] Where: top_left_x: x-axis coordinate of the upper left corner, top_left_y: y-axis coordinate of the upper left corner, x: x-axis coordinate of the four points, y: y-axis coordinate of the four points, geo[1]: horizontal spatial resolution of the remote sensing image, geo[2]: row rotation, geo[4]: column rotation, geo[5]: vertical spatial resolution of the remote sensing image. After obtaining the geographic vector data file, you can use vector tools to operate it and convert it into the common file format of the KML map, which is convenient for relevant personnel to conduct on-site inspections and reduce manpower, material resources and time.

[0085] S403, constructing a vector file through the gdal library, and writing the confidence level, category information, the four converted geographic coordinate points, and the longitude and latitude of the center point of each vector spot into the vector file.

[0086] Based on S2-S4 above, the optimal YOLO v8obb target detection algorithm model can be obtained, and then the test image is input to obtain the recognition and extraction results, which are converted into geographic vector files to facilitate subsequent statistical analysis and field investigations. Fig.10 , 11 , 12, and 13 are schematic diagrams of the superposition of the original test image and the corresponding vector results. Fig.14 , 15 , 12, and 13 are schematic diagrams of inference prediction on GF-2 images to test the generalization ability of the model. The feasibility of the present invention can be directly proved by comparing the result vector diagram with the original image.

[0087] This embodiment also provides a farmland greenhouse remote sensing extraction system based on improved target detection, including:

[0088] Remote sensing image preprocessing module. This module is used to obtain time-series high-resolution remote sensing images, and perform remote sensing image preprocessing operations on the remote sensing images, such as radiometric calibration, atmospheric correction, geometric correction, image fusion, geographic registration, and image uniform color mosaicking, so as to obtain preprocessed remote sensing image data.

[0089] Vector drawing and cropping module. This module is mainly used for sample drawing and cropping of pre-processed remote sensing image data. Based on this module, select buildings and greenhouse areas, draw vector patches that meet the input size of the algorithm model, and then crop each vector patch to obtain sample image data.

[0090] YOLOv8-OBB target detection algorithm model training module. This module mainly constructs the YOLOv8-OBB detection algorithm model, including the model main structure module, training module, test module and prediction module. This module mainly uses the training module and the prediction module. The training module determines the influencing factors of the target detection and identification extraction precision and accuracy by debugging different hyperparameters, and then obtains the optimal training parameters and determines the optimal model. For the prediction module, the remote sensing image sliding window is improved, and the 640-size window is continuously slid. The 640-size image is input into the optimal model for prediction, and its prediction results are obtained. According to the latitude and longitude calculation formula, the pixel coordinate point is converted into a geographic coordinate system and saved as a geographic vector data file to facilitate the statistical analysis of subsequent results.

[0091] YOLOv8-OBB target detection result post-processing module. Based on the sliding window, this module divides the entire remote sensing image into 640×640 images, continuously inputs them into the optimal model, and obtains the YOLOv8-OBB target detection results, mainly including confidence, category information, and pixel coordinates of 4 points. At the same time, according to the remote sensing image intercepted by the sliding window, the geographic reference information and projection information of the image are read. Then, the pixel coordinates of the 4 points are converted into the geographic coordinate system through the longitude and latitude formula. According to the read geographic reference information and projection information, a vector file is constructed through the gdal library, and the confidence, category information and the converted 4 geographic coordinate points are written into the vector file. At the same time, according to the geographic coordinate system, the longitude and latitude of the center point of each vector spot are calculated and written into the vector file, which is convenient for the statistical analysis and field investigation of subsequent results.

[0092] For the specific processes of the remote sensing image preprocessing module, vector drawing and cropping module, YOLOv8-OBB target detection algorithm model training module, and YOLOv8-OBB target detection result post-processing module in this embodiment, refer to S1-S4 in the embodiment.

[0093] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A remote sensing extraction method for greenhouse houses on cultivated land based on improved target detection, characterized in that: The following steps are involved: S1, obtain time-series high-resolution remote sensing images and perform remote sensing preprocessing to produce a sample data set; S2. Build and improve the YOLOv8-OBB deep learning target detection algorithm model; S3. Iteratively train the YOLOv8-OBB target detection algorithm model, set various hyperparameters, and compare them to obtain the optimal model after iterative training; S4. Post-processing of target detection results: Based on the YOLOv8-OBB target detection algorithm model with optimal parameters, the buildings and greenhouses in the area are identified and extracted, the target detection results are obtained, and output as geographic vector data files.

2. According to the method of claim 1, the remote sensing extraction method of greenhouse houses on cultivated land based on improved target detection is characterized in that: Step S1 specifically includes: S101, screening sample areas that meet the target conditions: based on the acquired time-series high-resolution remote sensing images, the remote sensing images are pre-processed using ENVI5.3 remote sensing processing software; S102, drawing vector patches: Based on the high-resolution remote sensing image processed in step S101, a rectangular vector patch is drawn using the rectangle drawing tool of the vector editing module through QGIS 3.4.2 geographic information processing software. The patch contains a single or two targets such as buildings and greenhouses, but all are marked with FN as labels; S103, clipping of raster data: Based on the high-resolution remote sensing image processed by S101 and the vector patches containing the target drawn by S102, the raster clipping tool by mask layer in QGIS 3.4.2 geographic information processing software is used to clip the remote sensing image one by one using the vector patches, and the clipped raster image is saved as image data; S104, preparation of oriented frame OBB sample data set: based on the raster data clipped and saved in S103, the roLabelImg tool is used to draw and adjust the direction and size of the rectangular frame according to the direction and size of the building and greenhouse, and save it; S105, oriented frame sample data set label format conversion: convert the data saved in S104 into a format suitable for YOLOv8-OBB target detection algorithm model label input.

3. A remote sensing extraction method for greenhouse houses on cultivated land based on improved target detection according to claim 1 or 2, characterized in that: Remote sensing preprocessing includes: radiation calibration, atmospheric correction, geometric correction, image fusion, geographic registration and image uniform color mosaic operations.

4. The remote sensing extraction method for greenhouse houses on cultivated land based on improved target detection according to claim 1 is characterized in that: Step S2 specifically includes: S201. Based on the Windows operating system, the anaconda deep learning package installation manager is used to build the YOLOv8-OBB target detection algorithm model based on the YOLO v8 official website document through the PyTorch deep learning framework; S202. Based on the YOLOv8-OBB algorithm model, the main part is improved, the Conv and C2f in Backone are replaced by SCConv space and channel reconstruction convolution, and the input mode and output mode of the prediction module are modified to make it suitable for reading and outputting remote sensing images; S203, using a sliding window method to obtain remote sensing image data, recombining the bands, and automatically outputting vector data through coordinate conversion.

5. The remote sensing extraction method of cultivated land greenhouse house based on improved target detection according to claim 4 is characterized in that: In step S201: according to the environment required by YOLO v8, a YOLOv8-OBB virtual environment is constructed, in which the GPU version of the PyTorch deep learning framework is installed, according to the YOLO v8 official website, the environment library required by YOLO v8 is installed, and according to the geographic information requirements, the remote sensing geographic information management package is installed.

6. The remote sensing extraction method of cultivated land greenhouse house based on improved target detection according to claim 1 is characterized in that: Step S3 specifically includes: S301. During iterative model training, download the official pre-trained model, load it into the YOLOv8-OBB target detection algorithm model, input your own sample training set, and set iterative training hyperparameters; S302. By debugging different hyperparameters, the influencing factors of the target detection, recognition and extraction precision and accuracy are determined, and then the optimal training parameters are obtained to determine the optimal model.

7. The remote sensing extraction method of cultivated land greenhouse house based on improved target detection according to claim 1 is characterized in that: Step S4 specifically includes: S401, based on the sliding window, the entire scene remote sensing image is continuously input into the optimal YOLOv8-OBB target detection algorithm model for prediction, and the YOLOv8-OBB target detection result is obtained, which mainly includes confidence, category information and pixel coordinates of 4 points; S402, reading the geographic reference information and projection information of the remote sensing image captured by the sliding window, and converting the pixel coordinates of the four points into a geographic coordinate system through the latitude and longitude formula; S403, constructing a vector file through the gdal library, and writing the confidence level, category information, the four converted geographic coordinate points, and the longitude and latitude of the center point of each vector spot into the vector file.

8. The remote sensing extraction system for greenhouse houses on cultivated land based on improved target detection according to claim 1 is characterized in that: include: A remote sensing image preprocessing module is used to obtain time-series high-resolution remote sensing images and perform remote sensing image preprocessing operations on the remote sensing images; Vector drawing and cropping module, used for sample drawing and cropping of pre-processed remote sensing image data; YOLOv8-OBB target detection algorithm model training module, which is used to build the model main structure module, training module, testing module and prediction module of the YOLOv8-OBB detection algorithm model; The YOLOv8-OBB target detection result post-processing module is used to identify and extract buildings and greenhouses in the area, obtain target detection results, and output them as geographic vector data files.

Images