Deep learning-assisted agricultural greenhouse recognition method based on remote sensing images of megacities

Abstract

The present invention relates to a deep learning-assisted method for identifying agricultural greenhouses in megacities using remote sensing images. The method comprises the following steps: 1. obtaining vector annotation data of a greenhouse sample area; 2. setting corresponding category labels based on greenhouse classification to obtain classified vector annotation data; 3. rasterizing the classified vector annotation data, cropping it using a slider, and then expanding the sample to obtain a training sample; 4. training a target detection network to obtain a target detection model; 5. inputting processed actual satellite remote sensing images into the target detection model to obtain and optimize greenhouse recognition results; and 6. evaluating the optimized results to obtain a final recognition result, adding positive and negative samples to the initial sample from step 3 to perform model optimization. Compared with the prior art, the present invention has the advantages of improving the accuracy and efficiency of agricultural greenhouse recognition in megacities using remote sensing images.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent identification of agricultural greenhouses, and in particular to a method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning. Background Art

[0002] In agricultural land management, in order to meet the needs of refined management, it is necessary to timely monitor the types and numbers of greenhouses in areas at the municipal level or even larger levels. In order to improve the timeliness of the survey and solve the problem of high field research costs, it is necessary to use remote sensing images to conduct periodic monitoring and further improve monitoring efficiency through deep learning technology.

[0003] Currently, commonly used satellite remote sensing images are mainly updated on a quarterly basis. For city-wide greenhouse monitoring in megacities (taking Shanghai as an example), the method of fully manual visual interpretation is difficult to meet the timeliness requirements of monitoring. Traditional image classification models are limited by the setting and adjustment of thresholds, and it is difficult to meet the accuracy of large-scale image classification. Deep learning algorithms are often used to process single natural images, and the data volume of remote sensing images of megacities is very large. Commonly used deep learning networks are often used to process natural images and are difficult to directly use to process the recognition of remote sensing images of megacities. In addition, deep learning algorithms are often used to identify target types (such as pedestrians, animals, vehicles, etc.) in natural images, and are difficult to use directly for direct monitoring of agricultural greenhouses.

[0004] In summary, current target recognition methods are difficult to accurately and efficiently identify agricultural greenhouses through remote sensing images of megacities. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-assisted method for identifying agricultural greenhouses in remote sensing images of megacities in order to improve the accuracy and efficiency of identifying agricultural greenhouses in remote sensing images of megacities.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A deep learning-assisted method for identifying agricultural greenhouses in megacities using remote sensing images, comprising the following steps:

[0008] Step 1: Obtain satellite remote sensing images for training, select greenhouse sample areas on average within the target range of the image, outline the greenhouse sample vector boundaries within the greenhouse sample areas, and obtain vector annotation data of the greenhouse sample areas;

[0009] Step 2: Set the classification criteria for the agricultural greenhouses in the greenhouse sample area image according to their color and texture characteristics. Based on the classification criteria, obtain the classification results of three categories, including plastic greenhouses, black greenhouses, and multi-span greenhouses. Set corresponding color value labels for the agricultural greenhouses of the same category in the greenhouse sample area vector annotation data. The color value labels of agricultural greenhouses of different categories are different, and obtain the classified vector annotation data.

[0010] Step 3: rasterize the classified vector annotation data and crop the rasterized image using a slider. After cropping, there is a certain pixel overlap between every two adjacent sliders. The cropped image is used as the initial sample, and then the initial sample is expanded to obtain the training sample.

[0011] Step 4: Input the cropped images from the training samples into the target detection network for training to obtain a trained target detection model;

[0012] Step 5: Obtain an actual satellite remote sensing image, which is a pixel map. The actual satellite remote sensing image is first segmented, and the segmented image is traversed using a slider with the same characteristics as in step 3. The trained target detection model is input to obtain greenhouse recognition results, and the results are optimized.

[0013] Step 6: Perform manual accuracy evaluation on the optimized results to obtain the final recognition results. The correct recognition results and incorrect classification results are used as positive and negative samples respectively. The positive and negative samples are added to the initial samples in step 3 to optimize greenhouse recognition.

[0014] Furthermore, the specific steps of step 2 are:

[0015] Step 2.1: Set the classification criteria for agricultural greenhouses in the vector annotation data based on color and texture characteristics. Based on the classification criteria, the classification results are obtained, including three categories: plastic greenhouses, black greenhouses, and multi-span greenhouses;

[0016] Step 2.2: Use polygons to outline the agricultural greenhouses in the vector annotation data, and then set a color value label attribute column for each polygon;

[0017] Step 2.3: Set the same color value label for the color value label attribute column of the polygons of the same type of agricultural greenhouses. The color value labels of the polygons of different types of agricultural greenhouses are different. After completing the setting of all polygons, the classified vector annotation data is obtained.

[0018] Furthermore,

[0019] The specific steps of step 3 are:

[0020] Step 3.1: Rasterize the classified vector annotation data;

[0021] Step 3.2: Crop the rasterized image using sliders, with a certain amount of pixel overlap between sliders.

[0022] Step 3.3: The cropped image is used as the initial sample, and then the initial sample is expanded by rotation, mirroring, and color enhancement to obtain the training sample;

[0023] The specific steps of expanding by rotation are:

[0024] Rotate each initial sample by 90 and 180 degrees, and add the resulting image to the training sample;

[0025] The specific steps for expansion through mirroring are:

[0026] Call the left and right pixels of each initial sample and add the resulting image to the training sample;

[0027] The specific steps for augmentation through color enhancement are:

[0028] In both RGB and HSV color spaces, gamma transformation is performed on each initial sample, and the transformed image is added to the training sample.

[0029] Furthermore, the initial sample is transformed into:

[0030]

[0031] Among them, V in is the initial sample of RGB or HSV color space, V out is the image processed by gamma, A is a constant, γ is the gamma parameter, and when the gamma parameter changes, the grayscale of the gamma-processed image will be stretched and compressed.

[0032] Furthermore, the specific steps of step 4 are:

[0033] Step 4.1: Set the number of output types of the target detection network according to the number of classification results, and input the cropped images from the training samples into the target detection network for training;

[0034] Step 4.2: Store the trained object detection model in the pth file type.

[0035] Furthermore, the specific steps of step 5 are:

[0036] Step 5.1: Obtain an actual satellite remote sensing image, which is a pixel map. The actual satellite remote sensing image is first divided, and the divided image is traversed using a slider with the same characteristics as in step 3. The size of each image obtained after traversal is the same as the size of the slider in step 3, and there is a certain amount of pixel overlap between adjacent images.

[0037] Step 5.2: Input the traversed images into the target detection model in batches;

[0038] Step 5.3: For the overlapping parts of the traversed image, the two recognition results obtained after entering the target detection model are taken as the union. The obtained recognition result is the recognition result of the overlapping part. For the non-overlapping part, it is directly entered into the target detection model to obtain the corresponding recognition result. The two recognition results are combined as the greenhouse recognition result;

[0039] Step 5.4: Optimize the greenhouse recognition results through morphological processing.

[0040] Furthermore, morphological processing includes dilation and erosion. After morphological processing, the holes or extremely small area patches in the greenhouse recognition results are optimized.

[0041] Furthermore, the size of the slider is 400*400 pixels, and the overlap is 100 pixels.

[0042] Furthermore, the specific steps of step 6 are:

[0043] Step 6.1: Select a certain range of images from the divided images for manual recognition, compare the manual recognition results with the optimized greenhouse recognition results, and calculate the precision and recall rates;

[0044] Step 6.2: Vectorize the optimized greenhouse recognition results that meet the accuracy and recall rates to obtain the final recognition results;

[0045] Step 6.3: The optimized greenhouse recognition results that meet the accuracy and recall rates are used as positive samples, and the optimized greenhouse recognition results that do not meet the accuracy and recall rates are used as negative samples, and added to the training samples in step 3.

[0046] Furthermore, the image of a certain range selected in step 6.1 is 1km 2 The corresponding image.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] First, the sample range of greenhouses is outlined based on remote sensing images. Then, based on the different forms of greenhouse expression in the images, attribute classification is performed. Then, vector data is rasterized, and samples are cropped and enhanced. The network parameters are trained using the prepared greenhouse samples, and greenhouses are identified in remote sensing images of megacities using a slider mode. Finally, the recognition results are accuracy assessed and vectorized. For recognition results with high accuracy, they can be used as samples after manual optimization for sample optimization and supplementation. Accuracy assessment and sample expansion improve the accuracy of greenhouse recognition. At the same time, the slider cropping method is used to improve detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the present invention;

[0050] Figure 2 A schematic diagram of sample preparation and outline of an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of greenhouse classification according to an embodiment of the present invention;

[0052] Figure 4 A schematic diagram of a vector data rasterization tool according to an embodiment of the present invention;

[0053] Figure 5 A schematic diagram of sample cutting and sample strengthening according to an embodiment of the present invention;

[0054] Figure 6 2 is a schematic diagram of a slider mode test according to an embodiment of the present invention.

[0055] Figure 7 Schematic diagram of raster data vectorization results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0057] The present invention mainly solves the problem of the existing technology in monitoring agricultural greenhouses in megacities, and provides a deep learning-assisted method for intelligent recognition of agricultural greenhouses in megacities using remote sensing images. The flow chart of the method is as follows: Figure 1 The method comprises the following steps:

[0058] Step 1: Obtain satellite remote sensing images for training, select greenhouse sample areas on average within the target range of the image, outline the greenhouse sample vector boundaries within the greenhouse sample areas, and obtain vector annotation data of the greenhouse sample areas;

[0059] Step 2: Set the classification criteria for the agricultural greenhouses in the greenhouse sample area image according to their color and texture characteristics. Based on the classification criteria, obtain the classification results of three categories, including plastic greenhouses, black greenhouses, and multi-span greenhouses. Set corresponding color value labels for the agricultural greenhouses of the same category in the greenhouse sample area vector annotation data. The color value labels of agricultural greenhouses of different categories are different, and obtain the classified vector annotation data.

[0060] Step 3: rasterize the classified vector annotation data and crop the rasterized image using a slider. After cropping, there is a certain pixel overlap between every two adjacent sliders. The cropped image is used as the initial sample, and then the initial sample is expanded to obtain the training sample.

[0061] Step 4: Input the cropped images from the training samples into the target detection network for training to obtain a trained target detection model;

[0062] Step 5: Obtain an actual satellite remote sensing image, which is a pixel map. The actual satellite remote sensing image is first segmented, and the segmented image is traversed using a slider with the same characteristics as in step 3. The trained target detection model is input to obtain greenhouse recognition results, and the results are optimized.

[0063] Step 6: Perform manual accuracy evaluation on the optimized results to obtain the final recognition results. The correct recognition results and incorrect classification results are used as positive and negative samples respectively. The positive and negative samples are added to the initial samples in step 3 to optimize greenhouse recognition.

[0064] Sample preparation and sketching diagram Figure 2 In step 1, based on satellite remote sensing images, ArcMap software is used to evenly select greenhouse sample areas within the target range (taking a range of 1km*1km as an example, the present invention selects a total of 60 1km*1km range areas for sample preparation).

[0065] In step 2, based on the expressive features of agricultural greenhouses in the image and corresponding to the classification of agricultural greenhouses in the national standard, the greenhouses selected in step 1 are divided into plastic greenhouses (white-green in the image), black greenhouses, and multi-span greenhouses. Multi-span greenhouses appear in the image as greenhouses with a clear frame structure.

[0066] The specific steps of step 2 are:

[0067] Step 2.1: Set the classification criteria for agricultural greenhouses in the vector annotation data based on color and texture characteristics. Based on the classification criteria, the classification results are obtained, including three categories: plastic greenhouses, black greenhouses, and multi-span greenhouses;

[0068] Step 2.2: Use polygons to outline the agricultural greenhouses in the vector annotation data, and then set a color value label attribute column for each polygon;

[0069] Step 2.3: Set the same color value label for the color value label attribute column of the polygons of the same type of agricultural greenhouses. The color value labels of the polygons of different types of agricultural greenhouses are different. After completing the setting of all polygons, the classified vector annotation data is obtained.

[0070] Step 2.1: Set the classification criteria for the agricultural greenhouses in the greenhouse sample area image based on the color and texture characteristics. Based on the classification criteria, the classification results are obtained, including three categories: plastic greenhouses, black greenhouses, and multi-span greenhouses;

[0071] Step 2.2: Open the vector annotation data in ArcMap and use polygons to outline the agricultural greenhouses in the vector annotation data. To facilitate the subsequent rasterization steps, set the color value label attribute column fme_color for each polygon;

[0072] Step 2.3: Set the same color value label for the color value label attribute column of the polygons of the same type of agricultural greenhouses. The color value labels of the polygons of different types of agricultural greenhouses are different. After completing the setting of all polygons, the classified vector annotation data is obtained. For example, the color labels of plastic greenhouses ("0.004, 0.004, 0.004"), black greenhouses ("0.008, 0.008, 0.008"), and multi-span greenhouses ("0.012, 0.012, 0.012"). The schematic diagram of greenhouse classification is as follows. Figure 3 shown.

[0073] Schematic diagram of vector data rasterization tool Figure 4 As shown; the schematic diagram of sample cutting and sample strengthening (expansion) is as follows Figure 5 In step 3, the rasterized image is cropped using a slider to a uniform size (400*400 is used as an example in this patent).

[0074] The specific steps of step 3 are:

[0075] Step 3.1: Rasterize the classified vector annotation data;

[0076] Step 3.2: Crop the rasterized image using sliders, with a certain amount of pixel overlap between sliders.

[0077] Step 3.3: The cropped image is used as the initial sample, and then the initial sample is expanded by rotation, mirroring, and color enhancement to obtain the training sample;

[0078] The specific steps of expanding by rotation are:

[0079] Rotate each initial sample by 90 and 180 degrees, and add the resulting image to the training sample;

[0080] The specific steps for expansion through mirroring are:

[0081] Call the left and right pixels of each initial sample and add the resulting image to the training sample;

[0082] The specific steps for augmentation through color enhancement are:

[0083] In both RGB and HSV color spaces, gamma transformation is performed on each initial sample, and the transformed image is added to the training sample.

[0084] In step 3.1, first use FME tools to build a tool for vectorizing raster data, and use tools such as ImageRasterizer, RasterSelector, RasterBandRemover, Clipper, and RasterBandNodataSetter in FME tools to perform vector rasterization operations. In ImageRasterrize, the parameter values ​​of Xsize and Ysize need to be set according to the image pixels (for example, if the image has a resolution of 0.5 meters, both Xsize and Ysize are set to 0.5) for rasterization.

[0085] In step 3.2, since the greenhouse sample area has an image range of 1 km * 1 km, for a 0.5 m resolution satellite remote sensing image, a single image is 2000 * 2000 pixels, which is difficult to directly use as a sample for deep learning network training. Therefore, further cropping is required. The sample is cropped using a slider to crop the sample to a size of 400 * 400 pixels, with 100 pixels of overlap between the sliders.

[0086] In step 3.3, to expand the sample data, the samples need to be rotated, mirrored, and color enhanced. Rotation is to rotate each 400*400 pixel sample 90 and 180 degrees, mirroring is to swap the left and right pixels of each sample, and color enhancement refers to gamma transformation of the image in the RGB and HSV color spaces to reduce the data's dependence on colors such as saturation and brightness. This means that remote sensing images obtained under different lighting conditions can be well recognized. The gamma change formula is as follows:

[0087]

[0088] Where V in For image input in RGB or HSV color space, V outis the output after gamma processing, A is a constant, γ is the gamma parameter. By changing γ, the grayscale of the image can be stretched and compressed to achieve brightness control of the remote sensing image.

[0089] In step 4, the cropped images in the training samples are input into the target detection network for training.

[0090] The specific steps of step 4 are:

[0091] Step 4.1: Set the number of output types of the target detection network according to the number of classification results, and input the cropped images from the training samples into the target detection network for training;

[0092] Step 4.2: Store the trained object detection model in the pth file type.

[0093] In step 4.1, the target detection network is a combination of deeplabv3 and resnet101 networks, or other existing target detection networks such as yolov5 network are used.

[0094] Set the number of output types (NUM_OF_CLASS) according to the number of greenhouse sample types, set the GPU to accelerate training, and set the batch batch_size, learning rate learning_rate, and echo parameters according to the computer performance (this invention takes batch_size as 4, learning_rate as 10-4, and about 50,000 iterations as an example).

[0095] In step 5, a large range of satellite images are tested by using a slider, and the edge problem between sliders is handled and the results are optimized. Figure 6 The specific steps of step 5 are as follows:

[0096] Step 5.1: Obtain an actual satellite remote sensing image, which is a pixel map. The actual satellite remote sensing image is first divided, and the divided image is traversed using a slider with the same characteristics as in step 3. The size of each image obtained after traversal is the same as the size of the slider in step 3, and there is a certain amount of pixel overlap between adjacent images.

[0097] Step 5.2: Input the traversed images into the target detection model in batches;

[0098] Step 5.3: For the overlapping parts of the traversed image, the two recognition results obtained after entering the target detection model are taken as the union. The obtained recognition result is the recognition result of the overlapping part. For the non-overlapping part, it is directly entered into the target detection model to obtain the corresponding recognition result. The two recognition results are combined as the greenhouse recognition result;

[0099] Step 5.4: Optimize the greenhouse recognition results through morphological processing.

[0100] In step 5.1, since the amount of remote sensing image data at the city level is large, it is difficult for the computer to read and process such images in a single time. Therefore, it is necessary to grid the images and use FME software to cut them into processable images, while saving the spatial information file wld.

[0101] In step 5.2, since the single processing size of deep learning is 400*400 and the sample is trained at the original ratio, a 400*400 slider is required during the test process to check the remote sensing image after the division in step 5.1, and process it according to batch_size (this patent takes batch_size=4 as an example), that is, each batch_size slider is a batch of data, which is put into the network for testing.

[0102] In step 5.3, in order to ensure the integrity and accuracy of the data, there needs to be overlap between the sliders (the present invention takes 100 pixels of overlap as an example). Since there will be two greenhouse recognition results in the overlapping part, the union of the two recognition results is taken as the final recognition result.

[0103] In step 5.4, the results of automatic identification and detection may contain holes or very small patches, which will affect the statistical results of the number and area of ​​greenhouses. They need to be optimized through morphological processing such as dilation and erosion.

[0104] The formula for corrosion is:

[0105]

[0106] Where A is the eroded target, and B is an erosion structure similar to a convolution kernel. If B is completely contained in the overlapping area of ​​A at (x, y), the corresponding pixel (x, y) is output as 1, otherwise it is 0.

[0107] The formula for expansion is:

[0108]

[0109] Where A is the expanded target, and B is an expansion structure similar to a convolution kernel. If the intersection of B and A at (x, y) is not empty, the output of the corresponding pixel (x, y) is 1, otherwise it is 0.

[0110] In step 6, the accuracy of the test results in the non-sampling area is evaluated, and the results are vectorized for the entire city using the ArcMap tool to obtain the greenhouse detection results in shp format. The obtained greenhouse detection results are manually optimized, and the correctly classified and incorrectly classified ones are used as positive and negative samples. The operations of steps 1-4 are repeated to achieve effective sample expansion and continuously optimize the network learning effect. The schematic diagram of the raster data vectorization result is shown in the figure. Figure 7 shown.

[0111] The specific steps of step 6 are:

[0112] Step 6.1: Select a certain range of images from the divided images for manual recognition, compare the manual recognition results with the optimized greenhouse recognition results, and calculate the precision and recall rates;

[0113] Step 6.2: Vectorize the optimized greenhouse recognition results that meet the accuracy and recall rates to obtain the final recognition results;

[0114] Step 6.3: The optimized greenhouse recognition results that meet the accuracy and recall rates are used as positive samples, and the optimized greenhouse recognition results that do not meet the accuracy and recall rates are used as negative samples, and added to the training samples in step 3.

[0115] The range of a certain range of images is 1km 2 . Vectorization specifically involves vectorizing raster data through ArcMap.

[0116] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A deep learning-assisted method for identifying agricultural greenhouses in megacities using remote sensing images, characterized in that: The method comprises the following steps: Step 1: Obtain satellite remote sensing images for training, select greenhouse sample areas on average within the target range of the image, outline the greenhouse sample vector boundaries within the greenhouse sample areas, and obtain vector annotation data of the greenhouse sample areas; Step 2: Set the classification criteria for the agricultural greenhouses in the greenhouse sample area image according to their color and texture characteristics. Based on the classification criteria, obtain the classification results of three categories, including plastic greenhouses, black greenhouses, and multi-span greenhouses. Set corresponding color value labels for the agricultural greenhouses of the same category in the greenhouse sample area vector annotation data. The color value labels of agricultural greenhouses of different categories are different, and obtain the classified vector annotation data. Step 3: rasterize the classified vector annotation data and crop the rasterized image using a slider. After cropping, there is a certain pixel overlap between every two adjacent sliders. The cropped image is used as the initial sample, and then the initial sample is expanded to obtain the training sample. Step 4: Input the cropped images from the training samples into the target detection network for training to obtain a trained target detection model; Step 5: Obtain an actual satellite remote sensing image, which is a pixel map. The actual satellite remote sensing image is first segmented, and the segmented image is traversed using a slider with the same characteristics as in step 3. The trained target detection model is input to obtain greenhouse recognition results, and the results are optimized. Step 6: Perform manual accuracy evaluation on the optimized results to obtain the final recognition results. The correct recognition results and incorrect classification results are used as positive and negative samples respectively. The positive and negative samples are added to the initial samples in step 3 to optimize greenhouse recognition.

2. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 1, characterized in that: The specific steps of step 2 are: Step 2.1: Set the classification criteria for agricultural greenhouses in the vector annotation data based on color and texture characteristics. Based on the classification criteria, the classification results are obtained, including three categories: plastic greenhouses, black greenhouses, and multi-span greenhouses; Step 2.2: Use polygons to outline the agricultural greenhouses in the vector annotation data, and then set a color value label attribute column for each polygon; Step 2.3: Set the same color value label for the color value label attribute column of the polygons of the same type of agricultural greenhouses. The color value labels of the polygons of different types of agricultural greenhouses are different. After completing the setting of all polygons, the classified vector annotation data is obtained.

3. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 2, characterized in that: The specific steps of step 3 are: Step 3.1: Rasterize the classified vector annotation data; Step 3.2: Crop the rasterized image using sliders, with a certain amount of pixel overlap between sliders. Step 3.3: The cropped image is used as the initial sample, and then the initial sample is expanded by rotation, mirroring, and color enhancement to obtain the training sample; The specific steps of expanding by rotation are: Rotate each initial sample by 90 and 180 degrees, and add the resulting image to the training sample; The specific steps for expansion through mirroring are: Call the left and right pixels of each initial sample and add the resulting image to the training sample; The specific steps for augmentation through color enhancement are: In both RGB and HSV color spaces, gamma transformation is performed on each initial sample, and the transformed image is added to the training sample.

4. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 3, characterized in that: The initial sample is transformed into: Among them, V in is the initial sample of RGB or HSV color space, V out is the image processed by gamma, A is a constant, γ is the gamma parameter, and when the gamma parameter changes, the grayscale of the gamma-processed image will be stretched and compressed.

5. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 1, characterized in that: The specific steps of step 4 are: Step 4.1: Set the number of output types of the target detection network according to the number of classification results, and input the cropped images from the training samples into the target detection network for training; Step 4.2: Store the trained object detection model in the pth file type.

6. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 1, characterized in that: The specific steps of step 5 are: Step 5.1: Obtain an actual satellite remote sensing image, which is a pixel map. The actual satellite remote sensing image is first divided, and the divided image is traversed using a slider with the same characteristics as in step 3. The size of each image obtained after traversal is the same as the size of the slider in step 3, and there is a certain amount of pixel overlap between adjacent images. Step 5.2: Input the traversed images into the target detection model in batches; Step 5.3: For the overlapping parts of the traversed image, the two recognition results obtained after entering the target detection model are taken as the union. The obtained recognition result is the recognition result of the overlapping part. For the non-overlapping part, it is directly entered into the target detection model to obtain the corresponding recognition result. The two recognition results are combined as the greenhouse recognition result; Step 5.4: Optimize the greenhouse recognition results through morphological processing.

7. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 6, characterized in that: Morphological processing includes dilation and erosion. After morphological processing, the holes or extremely small areas in the greenhouse recognition results are optimized.

8. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 6, characterized in that: The size of the slider is 400*400 pixels, with an overlap of 100 pixels.

9. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 6, characterized in that: The specific steps of step 6 are: Step 6.1: Select a certain range of images from the divided images for manual recognition, compare the manual recognition results with the optimized greenhouse recognition results, and calculate the precision and recall rates; Step 6.2: Vectorize the optimized greenhouse recognition results that meet the accuracy and recall rates to obtain the final recognition results; Step 6.3: The optimized greenhouse recognition results that meet the accuracy and recall rates are used as positive samples, and the optimized greenhouse recognition results that do not meet the accuracy and recall rates are used as negative samples, and added to the training samples in step 3.

10. The method for identifying agricultural greenhouses in megacities using remote sensing images assisted by deep learning according to claim 9, characterized in that: The image of a certain range selected in step 6.1 is 1km 2 The corresponding image.

Images