A method for identifying sand-buried road sections using remote sensing images

Through high-resolution remote sensing imaging and deep learning technology, the spectral and texture characteristics of sand buried sections are extracted, and the depth autoencoder model is constructed, which solves the problems of high detection cost and small coverage of sand buried sections in the existing technology, achieving high Accurate and fast identification of sand-burned road sections.

CN114663767BActive Publication Date: 2024-11-19CHINA COMM SPACE INFORMATION TECH (BEIJING)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210352114.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-03
Publication Date
2024-11-19
Estimated Expiration
2042-04-03

AI Technical Summary

Technical Problem

The existing sand-buried road section detection methods consume a lot of time and economic costs, and it is difficult to detect and identify sand-buried road sections in large areas.

Method used

High-resolution remote sensing images combined with deep learning methods are used to extract spectral features and texture features, and a depth autoencoder model is constructed to achieve high-precision recognition of sand-buried road sections.

Benefits of technology

It realizes high-precision identification of sand-buried road sections, with wide coverage, high spatial resolution, fast and convenient effects, and reduces detection costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114663767B_ABST
    Figure CN114663767B_ABST
Patent Text Reader

Abstract

The present invention provides a remote sensing image sand-buried road section identification method, comprising: acquiring remote sensing images; performing image processing on the remote sensing images, and producing detection road section pixels and associated pixel labels according to whether the road section is buried in sand; extracting spectral features and texture features from the detection road section pixels, and performing vector splicing of the spectral features and texture features to obtain spectral-texture feature vectors; dividing the spectral-texture feature vectors and associated pixel labels into a training set and a test set; inputting the training set into a deep autoencoder for training, and constructing a deep autoencoder model; inputting the test set into the deep autoencoder model, and obtaining classification and identification results of sand-buried road sections and non-sand-buried road sections. The present invention combines the existing traditional remote sensing image classification technology with deep learning methods, and uses high-resolution remote sensing images, which can retain important information in the data to the maximum extent, realize high-precision sand-buried road section identification, and has the effects of wide coverage, high spatial resolution, and quick and convenient.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image information extraction, and in particular to a remote sensing image sand-buried road section recognition method. Background Art

[0002] Sand burial is the main type of sandstorm hazard to desert roads. The existence of sand-buried road sections affects transportation and endangers road traffic safety. Real-time acquisition of sand-buried road section information in the region can be used to judge road commuting conditions and provide auxiliary information for decision-making.

[0003] The existing sand-buried road section detection methods mainly include on-site measurement, device detection and other methods. The on-site measurement method can be carried out by on-site measurement, observation, taking photos and other methods. The device detection method realizes the detection of sand-buried roads by setting up sand-buried detection devices on fixed sections of the road. However, the existing sand-buried road section positioning detection method consumes a lot of time and economic costs, and it is difficult to realize the detection and identification of sand-buried road sections in a large area. Therefore, a method for identifying sand-buried road sections is urgently needed. Summary of the invention

[0004] In order to solve the above-mentioned technical problems, the present invention provides a method for identifying sand-buried road sections using remote sensing images. By combining the existing traditional remote sensing image classification technology with the deep learning method, and adopting high-resolution remote sensing images, it can retain the important information in the data to the maximum extent and realize high-precision identification of sand-buried road sections, with the effects of wide coverage, high spatial resolution, and quick and convenient operation.

[0005] The present invention provides a method for identifying sand-buried road sections using remote sensing images, comprising:

[0006] Acquire remote sensing images; perform image processing on the remote sensing images, and generate detection section pixels and associated pixel labels according to whether the road section is buried in sand; extract spectral features and texture features from the detection section pixels, and perform vector concatenation of the spectral features and texture features to obtain spectral-texture feature vectors; divide the spectral-texture feature vectors and associated pixel labels into a training set and a test set; input the training set into a deep autoencoder for training and construct a deep autoencoder model; input the test set into the deep autoencoder model to obtain classification and recognition results of sand-buried sections and non-sand-buried sections.

[0007] Furthermore, the road section information is extracted based on the vector road network data and road binary images of the remote sensing images. After orthorectification and image fusion, the spectral curve differences of the road sections are analyzed, and the pixels of the detected road sections and the pixel labels associated with them are determined based on whether they are buried in sand.

[0008] Furthermore, the pixel values ​​of the red band, green band, blue band and near-infrared band of the detection section pixels are counted, and the difference index, ratio index and normalized index are obtained by band combination to extract the spectral characteristics of the detection section pixels.

[0009] Furthermore, the gray level co-occurrence matrix is ​​used to statistically process the gray level information contained in the pixels of the detected road section, and the statistics of contrast, entropy, energy and correlation are obtained as feature parameters to describe the texture features;

[0010] Furthermore, the characteristic parameters also include: local binary pattern, Markov random field and fractal dimension.

[0011] Furthermore, the training set is input into the deep autoencoder for training so that the network weights and bias values ​​of the deep autoencoder reach the optimal value, and it has the functions of feature vector splicing and fusion, deep feature extraction and classification and recognition of sand-buried roads. A deep learning network model for classification and recognition of sand-buried sections and non-sand-buried sections is constructed to form a deep autoencoder model.

[0012] Furthermore, before inputting the training set into the deep autoencoder for training, the deep autoencoder is pre-trained using the training set.

[0013] Furthermore, fine-tuning training is performed after pre-training, and the reference value of the deep autoencoder is adjusted using error direction propagation to optimize the network weights and bias values.

[0014] Furthermore, the network structure of the deep autoencoder is divided into an encoding part and a decoding part, which is composed of an input layer, a hidden layer, and an output layer; the hidden layer is set in multiple layers.

[0015] Furthermore, the deep autoencoder also includes a classifier;

[0016] The classifier is set between the last hidden layer and the output layer;

[0017] The classifier sets two output classification points to output the possibility probability values ​​of the sand-buried section and the non-sand-buried section respectively to obtain the classification recognition results.

[0018] The present invention uses a variety of feature extraction methods to extract the features of sand-buried sections and non-sand-buried sections, and combines deep learning methods to perform feature dimension reduction and pixel classification, thereby achieving high-precision recognition of sand-buried sections in high-resolution remote sensing images.

[0019] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other features, advantages and aspects of the embodiments of the present invention will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:

[0021] Figure 1 is a technical flow chart of the sand-buried road section identification method of this embodiment;

[0022] Figure 2 is a block diagram of sand-buried road section recognition of the deep autoencoder model of this embodiment;

[0023] Figure 3 This is an example 1 of the recognition result of the sand-buried road section recognition method of this embodiment;

[0024] Figure 4 This is the second example of the recognition result of the sand-buried road section recognition method of this embodiment;

[0025] Figure 5 This is the third example of the recognition result of the sand-buried road section recognition method of this embodiment;

[0026] in, Figure 1-5 The corresponding relationship between the reference numerals and component names in the figure is:

[0027] 1 Non-buried section, 2 buried section. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] In addition, the term "and / or" in this article is only a description of the association relationship between the associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0030] Refer to the following Figure 1To describe a remote sensing image sand-buried section recognition method provided by the present invention, high-resolution remote sensing image data is used, and the rich spectral information and texture information of target objects in high-resolution remote sensing images are fully utilized to recognize sand-buried sections; through the data set production step, the data set required for deep learning network training and testing is produced; through the feature extraction step, multiple spectral and texture features of sand-buried sections and non-sand-buried sections in remote sensing images are extracted; through the model construction and training step, a deep autoencoder model is constructed based on an open source deep learning framework, and model training is completed on the sand-buried section data set; through the sand-buried section recognition step, the trained model is used to realize the recognition of sand-buried sections in high-resolution remote sensing images. The specific recognition method is as follows:

[0031] S1: Dataset preparation; obtaining remote sensing images; performing image processing on remote sensing images, and producing detection road section pixels and associated pixel labels according to whether the road section is buried in sand;

[0032] S2: Feature extraction: extract spectral features and texture features from the pixels of the detected road section, and perform vector concatenation of the spectral features and texture features to obtain a spectral-texture feature vector; divide the spectral-texture feature vectors and pixel labels of the pixels of the detected road section into a training set and a test set;

[0033] S3: Model construction and training;

[0034] Build a deep learning network model, which is a deep autoencoder network model that includes a classifier;

[0035] Input the training set into the deep autoencoder for pre-training;

[0036] After pre-training, fine-tuning training is performed to adjust the reference value of the deep autoencoder using error direction propagation to optimize the network weights and bias values;

[0037] Input the training set into the deep autoencoder for training and build a deep autoencoder model;

[0038] S4: Identification of sand-buried road sections: The test set is input into the deep autoencoder model to obtain the classification and identification results of sand-buried road sections and non-sand-buried road sections.

[0039] In the above implementation, the 4-band GF-2 remote sensing image is used as an example to identify the buried road section:

[0040] S1: Dataset preparation: To produce high-quality samples, regional road section information is extracted based on vector road network data and road binary images of remote sensing images. After image preprocessing such as orthorectification and image fusion, the spectral curve differences of the road sections are analyzed. Visual interpretation is mainly used to determine the pixels of the detected road sections and the pixel labels associated with them according to whether they are buried in sand (pixels of non-buried sections can be marked as 0, and pixels of buried sections can be marked as 1, and 0 or 1 is the pixel label), and the pixels of the detected road sections and the pixel labels associated with them are divided into training set and test set.

[0041] S2: Feature extraction: High-resolution multispectral images contain relatively rich spectral information and texture information. Different objects have different pixel reflectances in the same / different bands, and the textures of sand and roads are not consistent. Therefore, the spectrum and texture information of the objects can be used for category judgment. However, the information contained in a single spectral feature or texture feature is difficult to use to distinguish between sand-buried and non-sand-buried sections. In order to deeply mine image information to achieve more accurate automatic identification of sand-buried sections, a method of combining multiple spectral features and texture features to construct a spectral-texture feature vector is used to achieve automatic identification. The specific feature extraction is as follows:

[0042] (1) Spectral feature extraction: For each pixel label corresponding to the remote sensing image detection road section pixel, the pixel values ​​of the four bands of red band, green band, blue band, and near-infrared band are counted, and spectral indexes such as difference index, ratio index, and normalization index are obtained by band combination to extract the spectral features of the detection road section pixel;

[0043] (2) Texture feature extraction: For each pixel label corresponding to the remote sensing image detection road segment pixel, the gray level co-occurrence matrix is ​​used to perform statistical processing on the gray level information contained therein. In order to better describe the texture information of the image road segment pixels, the obtained gray level co-occurrence matrix is ​​not used directly, but statistical processing is performed to obtain four secondary statistics of contrast, entropy, energy and correlation as the feature parameters of the image to describe the gray level co-occurrence matrix texture features of the image road segment pixels. In addition, in order to obtain better texture information of the image road segment pixels and facilitate the subsequent construction and training of the deep autoencoder dimensionality reduction classification model and road segment recognition, local binary patterns (Local Binary Pattern), Markov random fields (MRF), and fractal dimensions are also used to statistically describe the texture features of the image road segment pixels.

[0044] (3) Constructing the spectral-texture feature vector: quantize the extracted spectral features and texture features, concatenate the spectral feature vector and the texture feature vector to obtain the spectral-texture feature vector of the pixel of the detected road section. The specific information is shown in Table 1. The spectral-texture feature vector of the pixel of the detected road section is added to the training set and test set corresponding to the pixel label associated with it.

[0045] Table 1 Spectral-texture feature vector information

[0046]

[0047]

[0048] S3: Model building and training:

[0049] (1) Construction of deep autoencoder network model: A deep learning network model for classification and identification of buried sand sections is constructed. The network structure of the model is divided into an encoding part and a decoding part, and is composed of an input layer, a hidden layer, and an output layer; the hidden layer is set in multiple layers. The deep autoencoder also includes a Softmax classifier, which is set between the last hidden layer and the output layer. In this embodiment, a deep autoencoder including 10 hidden layers is constructed. The number of neurons in each hidden layer of the encoder part decreases successively to achieve feature dimensionality reduction. The network structure of the deep autoencoder used is shown in Table 2:

[0050] Table 2 Deep autoencoder network structure

[0051]

[0052] The number of hidden layer neural units in the encoding part is 32, 29, 27, 25, 24, 22, 20, 18, 16, and 14, respectively, and a Softmax classifier is added after the last hidden layer of the encoding part. Since the detection road section pixels in this implementation can be divided into two categories: sand-buried road sections and non-sand-buried road sections, the Softmax classifier used has two output classification points, and a probability value is assigned to each output classification result, indicating the possibility of belonging to each category, and the category with the highest probability is taken as the final classification result.

[0053] (2) Deep learning model training: Through the pre-training and training process of the deep autoencoder, the weights and bias values ​​of the entire deep autoencoder network can be optimized, and it has the functions of feature fusion, deep feature extraction and sand-buried road recognition and classification. A deep learning network model for classification and recognition of sand-buried and non-sand-buried road sections is constructed to form a deep autoencoder model. The process of deep autoencoder encoding and Softmax classification is as follows: Figure 2 shown.

[0054] Among them, pre-training is to input the spectral-texture feature vector data in the training set into the deep autoencoder for pre-training; after pre-training, fine-tuning training is also required to train the entire deep autoencoder, and use error direction propagation to adjust the reference value of the deep autoencoder to make the network weights and bias values ​​reach the optimal value; the optimal standard is selected according to the loss function corresponding to the deep learning training.

[0055] Taking into account the GPU memory, input image size, algorithm performance and execution efficiency, a set of hyperparameter combinations to be used in the training process were selected in the model pre-training and fine-tuning training stages, as shown in Tables 3 and 4.

[0056] Table 3 Pre-training experiment hyperparameter settings

[0057]

[0058] Table 4 Hyperparameter settings for model fine-tuning training experiments

[0059]

[0060] The training process is to input the spectrum-texture feature vector data in the training set into the deep autoencoder for training, use the deep autoencoder to extract deep features, and obtain a trained deep autoencoder model;

[0061] The specific model training process is as follows:

[0062] 1) Train the first hidden layer of the deep autoencoder using unsupervised training;

[0063] 2) The output of the previous hidden layer training is regarded as the input of the next hidden layer, and then trained in an unsupervised manner;

[0064] 3) Repeat step 2) until all hidden layers are trained.

[0065] 4) After the pre-training is completed, the entire deep autoencoder is trained, and the parameter values ​​of the entire system are adjusted using error back propagation so that both the weights and bias values ​​are optimal.

[0066] S4: Sand-buried road section recognition: For multi-band high-resolution remote sensing images and road binary images, the feature extraction operation mentioned in S2 is used to extract the spectral features and texture features of the target objects in the remote sensing images. The spectral-texture feature vector data in the test set is input into the trained deep autoencoder model to obtain the classification and recognition results of the sand-buried road sections in the image. For specific recognition information, please refer to Figures 3 to 5 As shown, it can be seen that the technical solution of the present invention can clearly identify the sand-buried road section and its structure is accurate.

[0067] When adopting the above implementation method, firstly, this implementation method makes full use of the rich spectral information and texture information of the target objects in the high-resolution remote sensing images; in addition, this implementation method introduces deep learning technology and uses a deep autoencoder for deep feature extraction and fusion, which can retain important information in the data to the maximum extent and realize high-precision identification of sand-buried road sections, with wide coverage, high spatial resolution, and quick and convenient effects.

[0068] In the description of the present invention, the terms "connection", "installation", "fixation" and the like should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0069] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0070] The above is only a preferred implementation of the present invention and is not intended to limit the present invention. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying sand-buried road sections using remote sensing images, characterized in that: include: Acquisition of remote sensing images; Perform image processing on remote sensing images, and generate detection road section pixels and associated pixel labels based on whether the road section is buried in sand; The road section information is extracted based on the vector road network data and the road binary image of the remote sensing image, and after orthorectification and image fusion, the spectral curve difference of the road section is analyzed, and the pixel of the detected road section and the pixel label associated therewith are determined based on whether it is buried in sand; Extracting spectral features and texture features from the detected road section pixels, and performing vector splicing of the spectral features and the texture features to obtain a spectral-texture feature vector; Dividing the spectral-texture feature vectors and the associated pixel labels into a training set and a test set; Inputting the training set into a deep autoencoder for training to construct a deep autoencoder model; The test set is input into a deep autoencoder model to obtain classification and recognition results of sand-buried road sections and non-sand-buried road sections.

2. The method for identifying a sand-buried road section according to claim 1, characterized in that: The pixel values ​​of the red band, green band, blue band and near-infrared band of the detection section pixels are counted, and the difference index, ratio index and normalized index are obtained by band combination to extract the spectral characteristics of the detection section pixels.

3. The method for identifying a sand-buried road section according to claim 1, characterized in that: The grayscale information contained in the pixels of the detected road section is statistically processed using a grayscale co-occurrence matrix to obtain statistics of contrast, entropy, energy and correlation as feature parameters to describe the texture features.

4. The method for identifying a sand-buried road section according to claim 3, characterized in that: The characteristic parameters also include: local binary pattern, Markov random field and fractal dimension.

5. The method for identifying a sand-buried road section according to claim 1, characterized in that: The training set is input into the deep autoencoder for training, so that the network weights and bias values ​​of the deep autoencoder are optimized, and the deep autoencoder has the functions of feature vector splicing and fusion, deep feature extraction and classification and recognition of sand-buried roads, and a deep learning network model for classification and recognition of sand-buried sections and non-sand-buried sections is constructed to form the deep autoencoder model.

6. The method for identifying a sand-buried road section according to claim 5, characterized in that: Before inputting the training set into the deep autoencoder for training, the deep autoencoder is pre-trained using the training set.

7. The method for identifying a sand-buried road section according to claim 6, characterized in that: Fine-tuning training is performed after the pre-training, and error direction propagation is used to adjust the reference value of the deep autoencoder so that the network weights and the bias values ​​are optimized.

8. The method for identifying a sand-buried road section according to claim 1, characterized in that: The network structure of the deep autoencoder is divided into an encoding part and a decoding part, and is composed of an input layer, a hidden layer, and an output layer; the hidden layer is arranged in multiple layers.

9. The method for identifying a sand-buried road section according to claim 8, characterized in that: The deep autoencoder also includes a classifier; The classifier is arranged between the last hidden layer and the output layer; The classifier sets two output classification points to respectively output the possibility probability values ​​of the sand-buried road section and the non-sand-buried road section to obtain the classification recognition result.