A method for extracting vegetation information from high-resolution urban remote sensing images

Abstract

The present invention discloses a method for extracting vegetation information from high-resolution urban remote sensing images, comprising the following steps: 1) data enhancement; 2) training; 3) obtaining a sample image set; 4) obtaining a block extraction result map; 5) obtaining a superpixel sub-region block; 6) obtaining a block sample image; 7) merging. This method can improve the generalization ability of the deep model, reduce splicing traces, and has high extraction accuracy.

Description

Technical Field

[0001] The invention relates to the field of high-resolution urban remote sensing image information extraction, in particular to a method for extracting vegetation information from a high-resolution urban remote sensing image. Background Art

[0002] Urban vegetation ecosystems play an irreplaceable role in maintaining atmospheric carbon-oxygen balance, reducing heat island effects, purifying and beautifying the urban environment. With the continuous development of remote sensing satellite technology, the use of high-resolution urban remote sensing images to extract and analyze vegetation information can save a lot of manpower, financial resources and material resources, and can help plan and manage vegetation in urban areas more efficiently.

[0003] The existing methods for extracting vegetation information from urban remote sensing images are mainly divided into the following two types: one is the traditional machine learning method represented by support vector machines. This type of method mainly uses spectral features to process pixels individually, ignoring the spatial structure and rich texture information contained in high-resolution remote sensing images. The spectral characteristics of vegetation are affected by many factors such as environmental conditions and atmospheric conditions, and have certain regionality and effectiveness, which often makes the extraction results have serious "salt and pepper phenomenon"; the other is the object-oriented vegetation extraction method, but this method has a low degree of automation, the extraction effect is extremely dependent on the manually set threshold, and the performance is not stable enough. Summary of the invention

[0004] The purpose of the present invention is to provide a method for extracting vegetation information from high-resolution urban remote sensing images in view of the shortcomings of the prior art. This method can improve the generalization ability of the deep model, reduce splicing traces, and has high extraction accuracy.

[0005] The technical solution for achieving the purpose of the present invention is:

[0006] A method for extracting vegetation information from high-resolution urban remote sensing images comprises the following steps:

[0007] 1) Data enhancement: Randomly crop the dataset and rotate the images in the cropped dataset, such as 90°, 180°, or 270° counterclockwise or clockwise, flip the image, mirror, blur, add noise, such as Gaussian noise, salt and pepper noise data enhancement operations;

[0008] 2) Training: Using the UNet++ network structure, select 80% of the images in the dataset after the data augmentation operation in step 1) as the training set, and then input them into the UNet++ network for training. The remaining 20% ​​of the dataset images are used as the validation set to verify the model effect;

[0009] 3) Obtaining a sample image set: Select any image as a sample image for vegetation extraction, perform dilation and block processing on the sample image, and obtain a sample image set;

[0010] 4) Obtaining a block extraction result map: loading the best model trained in step 2), using the model to perform information extraction processing on the sample image set obtained in step 3), and obtaining a block extraction result map;

[0011] 5) Obtaining super-pixel sub-region blocks: performing SLIC segmentation on the image in the sample image set of step 3) to obtain super-pixel sub-region blocks;

[0012] 6) Obtaining the block sample image: Fuse the block extraction result image obtained in step 4) with the superpixel sub-region block obtained in step 5), and count the predicted label values ​​of all pixels in the sub-region. Calculate the entropy value of the sub-region, where H(S) is the entropy value of the sub-region S, Nc represents the total number of categories, l(i) represents the label category of the i-th pixel, and M represents the total number of pixels in the sub-region S. Set the threshold T for judgment. If the entropy value of the sub-region is greater than the threshold T, further segmentation and calculation are performed until the entropy value of each sub-region is lower than the threshold T. Finally, the predicted label value containing the maximum number of pixels in the sub-region is assigned to the sub-region to obtain a block sample image.

[0013] 7) Merge: Merge the divided sample images in step 6) to obtain the final extraction result.

[0014] Since the remote sensing image dataset has few samples, data augmentation operations are performed to increase the sample size and improve the generalization ability of the deep model; the UNet++ network structure embeds a 1-4-layer U-net structure, and the redesigned jump path reduces the semantic gap between the encoder and decoder sub-paths. The dense jump connection improves the mapping of semantically similar features, and deep supervision provides pruning operations, making the model more flexible and general; the expansion block method can avoid obvious splicing traces in the extracted result image; the SLIC segmentation algorithm performs better than other superpixel segmentation methods in terms of running speed, compactness of generated superpixels, and contour preservation, laying the foundation for the accuracy of subsequent sub-region judgment; by re-judging the sub-regions of the fused samples, the influence of the "salt and pepper phenomenon" can be effectively removed and the overall extraction accuracy can be improved.

[0015] This method can improve the generalization ability of deep models, reduce splicing traces, and achieve high extraction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the process of an embodiment;

[0017] Figure 2Schematic diagram of extraction comparison of the embodiment. DETAILED DESCRIPTION

[0018] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, but the present invention is not limited thereto.

[0019] Example:

[0020] Reference Figure 1 , a method for extracting vegetation information from high-resolution urban remote sensing images, comprising the following steps:

[0021] 1) Data enhancement: Randomly crop the dataset and rotate the images in the cropped dataset, such as rotating 90°, 180° or 270° counterclockwise or clockwise, flipping the image, mirroring, blurring, adding noise, such as Gaussian noise, salt and pepper noise data enhancement operations. This example uses the public experimental dataset of Potsdam, Germany provided by the International Society for Photogrammetry (ISPRS). The dataset contains 38 high-resolution urban remote sensing images of 6000×6000 pixels and 38 corresponding label images of the same size. 34 samples are selected for network model training and verification. The 34 images are expanded into a set of 10,000 images of 256×256 through data enhancement operations, and divided into a training set of 8,000 and a verification set of 2,000. The remaining four are used to verify the effect of the overall experiment;

[0022] 2) Training: Using the UNet++ network structure, select 80% of the images in the dataset after the data augmentation operation in step 1) as the training set, and then input them into the UNet++ network for training. The remaining 20% ​​of the dataset images are used as the validation set to verify the model effect;

[0023] 3) Obtaining a sample image set: Select any image as a sample image for vegetation extraction, perform dilation and block processing on the sample image, and obtain a sample image set. In this example, a sample image with a size of 6000×6000 is selected as a sample image for vegetation extraction;

[0024] 4) Obtaining a block extraction result map: loading the best model trained in step 2), using the model to perform information extraction processing on the sample image set obtained in step 3), and obtaining a block extraction result map;

[0025] 5) Obtaining super-pixel sub-region blocks: performing SLIC segmentation on the image in the sample image set of step 3) to obtain super-pixel sub-region blocks;

[0026] 6) Obtaining the block sample image: Fuse the block extraction result image obtained in step 4) with the superpixel sub-region block obtained in step 5), and count the predicted label values ​​of all pixels in the sub-region. Calculate the entropy value of the sub-region, where H(S) is the entropy value of the sub-region S, Nc represents the total number of categories, l(i) represents the label category of the i-th pixel, and M represents the total number of pixels in the sub-region S. Set the threshold T for judgment. If the entropy value of the sub-region is greater than the threshold T, further segmentation and calculation are performed until the entropy value of each sub-region is lower than the threshold T. Finally, the predicted label value containing the maximum number of pixels in the sub-region is assigned to the sub-region to obtain a block sample image.

[0027] 7) Merge: Merge the divided sample images in step 6) to obtain the final extraction result, such as Figure 2 shown.

Claims

1. A method for extracting vegetation information from high-resolution urban remote sensing images, characterized in that: The steps include: 1) Data enhancement: Randomly crop the dataset and perform image rotation, image flipping, blurring, and noise enhancement operations on the images in the cropped dataset; 2) Training: Using the UNet++ network structure, select 80% of the images in the dataset after the data augmentation operation in step 1) as the training set, and then input them into the UNet++ network for training. The remaining 20% ​​of the dataset images are used as the validation set to verify the model effect; 3) Obtaining a sample image set: Select any image as a sample image for vegetation extraction, perform dilation and block processing on the sample image, and obtain a sample image set; 4) Obtaining a block extraction result map: loading the best model trained in step 2), using the model to perform information extraction processing on the sample image set obtained in step 3), and obtaining a block extraction result map; 5) Obtaining super-pixel sub-region blocks: performing SLIC segmentation on the image in the sample image set of step 3) to obtain super-pixel sub-region blocks; 6) Obtaining the block sample image: Fuse the block extraction result image obtained in step 4) with the superpixel sub-region block obtained in step 5), and count the predicted label values ​​of all pixels in the sub-region. Calculate the entropy value of the sub-region, where H(S) is the entropy value of the sub-region S, Nc represents the total number of categories, l(i) represents the label category of the i-th pixel, and M represents the total number of pixels in the sub-region S. Set the threshold T for judgment. If the entropy value of the sub-region is greater than the threshold T, further segmentation and calculation are performed until the entropy value of each sub-region is lower than the threshold T. Finally, the predicted label value containing the maximum number of pixels in the sub-region is assigned to the sub-region to obtain a block sample image. 7) Merge: Merge the divided sample images in step 6) to obtain the final extraction result.

Images