Rare earth mining area land utilization refined classification method for hyperspectral remote sensing images

Abstract

The invention discloses a rare earth mining area land utilization fine classification method for hyperspectral remote sensing images, and belongs to the field of remote sensing. Based on a hyperspectral remote sensing image of a rare earth mining area, an object-oriented feature extraction module and an object-oriented one-dimensional convolutional neural network classification model based on a multilayer attention mechanism are designed, and a rare earth mining area land utilization classification task is completed through the two modules. By combining the object-oriented and convolutional neural network, the classification precision of different land utilization categories is enhanced, and the phenomenon of salt and pepper noise in the classification result is avoided at the same time.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing, and specifically relates to hyperspectral remote sensing, image processing and computer vision, mining area monitoring and environmental protection. Background Art

[0002] As the country pays more and more attention to environmental restoration measures such as rare earth mine ecosystem reconstruction and land reclamation, it is urgent to strengthen the development and management of rare earths and the monitoring and judgment of the ecological environment recovery status of mining areas. Land use survey is an important part of mine environmental monitoring. Classifying remote sensing images through different classification algorithms to obtain land use results is an important means of conducting land use surveys in large-scale mining areas. The development of remote sensing technology has provided a rich data source for land use surveys in mining areas. High-resolution remote sensing images have great application potential in land use monitoring in mining areas, but their cost is too high and the image processing workload is also large. Rare earth mines are small and scattered. Although medium-resolution multispectral images such as Landsat have a large monitoring range and are economical and labor-saving, they cannot reflect more detailed land use information in rare earth mining areas. Hyperspectral remote sensing images not only have a wide image width and can be used for large-scale macro-monitoring, but also have high spectral resolution, a large amount of information, and contain multiple narrow spectral bands. They are very sensitive to tiny spectral details and can obtain detailed spectral characteristics of land types and their differences, showing great potential in the fine classification of land use.

[0003] With the continuous deepening of research on automatic extraction technology of remote sensing information, the classification methods used for extracting land use information in mining areas have gradually become richer. Compared with pixel-based classification methods, object-oriented classification methods use pixels with the same characteristics as homogeneous objects as classification units, and the classification results can avoid the "salt and pepper" phenomenon. At the same time, object-oriented classification methods can fully consider the spectral and spatial characteristics of the objects, and the overall classification effect is better. However, due to the complex composition of surface features in rare earth mining areas (mining working faces, surface reclamation vegetation, etc.), it is difficult to apply traditional object-oriented classification methods to the fine classification of land use in rare earth mining areas. With the development of artificial intelligence technology in image recognition and processing, convolutional neural network algorithms with efficient dimensionality reduction and high-level feature extraction capabilities have been initially applied to the classification of objects and land use information extraction in open-pit mining areas. Although the convolutional neural network classification model has achieved good classification accuracy, its classification results still cannot avoid the appearance of "salt and pepper" noise. Therefore, in the context of massive amounts of hyperspectral remote sensing data, how to better combine the advantages of object-oriented classification and convolutional neural network classification models to achieve more efficient and automated data analysis and scene interpretation is one of the most pressing issues to be addressed. This will provide intelligent technical support for relevant regulatory authorities to monitor the mining and production and reclamation and restoration of rare earth mining areas. Summary of the invention

[0004] Therefore, the present invention combines object-oriented and convolutional neural networks according to the land feature types and special environments of the mining areas, and constructs an object-oriented-convolutional neural network method for fine classification of land use in rare earth mining areas using hyperspectral remote sensing images. The method makes full use of the advantages of hyperspectral remote sensing image data to improve the model's target recognition accuracy, speed and other performances. It can quickly and accurately identify various typical land feature targets in the mining area, obtain accurate land use classification information, and provide intelligent technical support for relevant regulatory authorities to monitor the mining production and reclamation and restoration of rare earth mining areas.

[0005] To this end, the present invention provides a rare earth mining area land use fine classification method for hyperspectral remote sensing images, the implementation steps of which include:

[0006] Step 1: Determine the classification system of rare earth mining areas based on current national standards and existing relevant research, combined with field survey information and hyperspectral remote sensing images of rare earth mining areas;

[0007] Step 2: Using the object-oriented feature extraction module, which is mainly responsible for extracting image features from hyperspectral remote sensing images;

[0008] Step 3: Construct an object-oriented one-dimensional convolutional neural network classification model (Object Convolutional on Transformer and CBAM, OCTC) (hereinafter referred to as the OCTC model) based on a multi-layer attention mechanism, complete the deep image feature learning and the high-light remote sensing image classification through the convolutional neural network, and then complete the land use classification task of the rare earth mining area.

[0009] Preferably, the object-oriented feature extraction module has the following specific implementation steps: the first step is to determine the optimal segmentation scale of the hyperspectral remote sensing image according to the rare earth mining area classification system using a scale parameter estimation model; the second step is to apply the optimal segmentation scale to the segmentation of the hyperspectral remote sensing image to obtain the hyperspectral remote sensing image segmentation result; the third step is to extract the image features of the hyperspectral remote sensing image from the hyperspectral remote sensing image segmentation result.

[0010] Preferably, the image features of the hyperspectral remote sensing image include spectral features and spatial features.

[0011] Preferably, the spectral characteristics include band characteristics and index characteristics.

[0012] Preferably, the band characteristics include 32 means and 32 standard deviations of each band.

[0013] Preferably, the index feature selects the reflectance of 32 original bands of the high-light remote sensing image as input parameters to construct 8 index features, including chlorophyll vegetation index, normalized vegetation index, ratio vegetation coefficient, green normalized vegetation index, difference vegetation index, 2 normalized difference red edge indices, and terrestrial chlorophyll index.

[0014] Preferably, the spatial features include geometric features and texture features.

[0015] Preferably, the texture feature is constructed by the synergy, contrast, dissimilarity, information entropy, mean, variance, correlation, and second-order moment of the hyperspectral remote sensing image segmentation result based on the gray level co-occurrence matrix.

[0016] Preferably, the geometric features introduce the idea of ​​object-oriented classification, and based on the segmentation results of the hyperspectral remote sensing image, the asymmetry, boundary index, compactness, density, ellipse fitting, radius of the largest closed ellipse, radius of the smallest closed ellipse, rectangle fitting, roundness, shape index, area, boundary length, length, aspect ratio, number of pixels, volume, and width of the hyperspectral remote sensing image are calculated.

[0017] Preferably, the OCTC model has the following specific implementation steps: the first step is to construct a one-dimensional convolutional neural network, input the image features of the hyperspectral remote sensing image, and extract its deep image features; the one-dimensional convolutional neural network includes a convolutional layer, an activation function, a pooling layer, and a fully connected layer; the second step is to introduce a Transformer module in the feature extraction process, which enables the model to dynamically focus on local image features and global image features, and make full use of the global context, thereby improving classification accuracy; the third step is to use the convolutional attention module to extract significant feature information in the network, which is equivalent to an implicit optimization of the image features, which is conducive to mining distinguishability in the rare earth mining area land use classification task; the fourth step is to classify the significant image feature information output by the one-dimensional convolutional neural network to obtain land use classification information.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. The present invention proposes a method for fine-grained classification of land use in rare earth mining areas based on hyperspectral remote sensing images, and establishes a framework for fine-grained classification of land use in rare earth mining areas based on hyperspectral remote sensing images. This method framework can achieve higher accuracy in rare earth mining area ground object detection.

[0020] 2. The present invention introduces the concept of object-oriented classification, which can obtain richer spectral and spatial features to reduce the interference of complex objects in rare earth mining scenes and emphasize the characteristic differences between categories. At the same time, it solves the problem of noise in the classification results and achieves the goal of extracting more accurate land use information in complex scenes of rare earth mining areas.

[0021] 3. The present invention constructs an object-oriented one-dimensional convolutional neural network classification model based on a multi-layer attention mechanism. For an object at a given scale, the model reshapes all its corresponding pixel-level features into a one-dimensional feature vector, performs deep feature learning and prediction, and completes the final rare earth mining area land use classification task. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A framework diagram of the implementation steps of the method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images provided by the present invention;

[0023] Figure 2 The classification results of the classification method of the present invention are compared with those of other classification methods. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical scheme and advantages of the invention more clear, the present invention will be further described below in conjunction with embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all 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.

[0025] This embodiment is based on the hyperspectral remote sensing image of the Lingbei rare earth mining area in southern Jiangxi.

[0026] A refined classification method for land use in rare earth mining areas based on hyperspectral remote sensing images is provided. The implementation steps are shown in the following figure. Figure 1 As shown, the implementation steps of the method include:

[0027] Step 1: Determine the classification system of rare earth mining areas based on current national standards and existing relevant research, combined with field survey information and hyperspectral remote sensing images of the Lingbei rare earth mining area;

[0028] Step 2: Use the object-oriented feature extraction module, which is mainly responsible for extracting image features from hyperspectral remote sensing images; Figure 1 The hyperspectral image in the object-oriented feature extraction module shown is obtained in step 1;

[0029] Step 3: Construct the OCTC model, complete the deep image feature learning and the high-light remote sensing image classification through the convolutional neural network, and then complete the land use classification task of the rare earth mining area.

[0030] Preferably, the hyperspectral remote sensing image feature extraction module is as follows: Figure 1 (Object-oriented feature extraction module), the specific implementation steps are: the first step is to determine the optimal segmentation scale of the hyperspectral remote sensing image according to the rare earth mining area classification system using a scale parameter estimation model; the second step is to apply the optimal segmentation scale to the segmentation of the hyperspectral remote sensing image to obtain a hyperspectral remote sensing image segmentation result; the third step is to extract the image features of the hyperspectral remote sensing image from the hyperspectral remote sensing image segmentation result.

[0031] Preferably, the image features of the hyperspectral remote sensing image include spectral features and spatial features.

[0032] Preferably, the spectral features include band features and index features; the spatial features include geometric features and texture features.

[0033] Preferably, the spectral characteristics include 32 means and 32 standard deviations of each band.

[0034] Preferably, the index features are constructed using the reflectance of 32 original bands of the high-light remote sensing image as input parameters, and the index features include chlorophyll vegetation index, normalized vegetation index, ratio vegetation coefficient, green normalized vegetation index, difference vegetation index, 2 normalized difference red edge indices, and terrestrial chlorophyll index.

[0035] Preferably, the texture feature is constructed by the synergy, contrast, dissimilarity, information entropy, mean, variance, correlation, and second-order moment of the hyperspectral remote sensing image segmentation result based on the gray level co-occurrence matrix.

[0036] Preferably, the geometric features introduce the idea of ​​object-oriented classification, and based on the segmentation results of the hyperspectral remote sensing image, calculate the asymmetry, boundary index, compactness, density, ellipse fitting, radius of the largest closed ellipse, radius of the smallest closed ellipse, rectangle fitting, roundness, shape index, area, boundary length, length, aspect ratio, number of pixels, volume, and width of the hyperspectral remote sensing image.

[0037] Specific relevant information of the image features is shown in Table 1:

[0038] Table 1. Specific information table of features

[0039]

[0040]

[0041] The OCTC model, Figure 1 As shown in the (OCTC model), the specific implementation steps are: the first step is to construct a one-dimensional convolutional neural network to extract deep features from the original input features; the one-dimensional convolutional neural network includes a convolutional layer, an activation function, a pooling layer and a fully connected layer; the second step is to introduce the Transformer module in the feature extraction process, which allows the model to dynamically focus on local features and global features, and make full use of the global context to improve the classification accuracy; the third step is to use the convolutional attention module to extract the significant feature information in the network, which is equivalent to an implicit feature optimization, which is conducive to mining distinguishability in classification tasks; the fourth step is to classify the significant feature information output by the one-dimensional convolutional neural network to obtain land use classification information.

[0042] Compared with the existing technology, the present invention does not require cumbersome operations, and also has very good performance in classification efficiency and classification accuracy. The original image of this embodiment is classified into land use types using two typical machine learning methods, a one-dimensional convolutional neural network classification method, and the classification method of the present invention. The classification accuracy obtained is shown in Table 2. The typical machine learning methods are K-nearest neighbor classification method and random forest classification method. The classification results are shown in Table 2. Figure 2As shown, Figure 2 (a) is the original image of the rare earth mining area in Lingbei Mining Area. Figure 2 (b) is the result of K-nearest neighbor classification method. Figure 2 (c) is the result of the random forest classification method. Figure 2 (d) is the result of the one-dimensional convolutional neural network classification method. Figure 2 (e) is the result diagram of the classification method of the present invention. According to the classification category of rare earth mining area, the classification results are compared to show that on the basis of object-oriented, the results of each classification method effectively avoid the "salt and pepper" phenomenon. However, in the classification results of the machine learning algorithm, the classification effect of the K-nearest neighbor classification method is also the worst. From the comparison of its classification result diagram and the original image, it can be seen that the K-nearest neighbor classification method is difficult to accurately distinguish between cultivated land and gardens, and there are many omissions for transportation land. The results of the random forest classification method are generally good, but in some details, especially in the multi-land coupling area, there are widespread misclassification phenomena, such as the coupling area of ​​the mining area and the reclaimed vegetation, the extraction effect of the reclaimed vegetation is not good, and at the same time, there are more confusion and misclassification between the reclaimed vegetation and the garden. This is because the classification characteristics between the reclaimed vegetation and the garden are relatively similar, and there are only some small differences. The random forest classification method has a poor effect on distinguishing the two. The classification method of the present invention has the best classification effect. It not only has the highest accuracy, but also can be seen from its result graph that it has more advantages than traditional machine learning algorithms in processing areas with small and fragmented land types. The phenomenon of misclassification and omission is greatly reduced, and the extraction accuracy of fragmented land objects is effectively improved. For land objects with similar classification characteristics, the classification accuracy of the classification method of the present invention is also higher, and it can effectively distinguish between reclaimed vegetation and gardens. In the task of environmental monitoring of rare earth mining areas, the classification method of the present invention can accurately and efficiently identify typical land objects in the mining area, and can provide technical means and support for the information-based supervision of rare earth mining areas.

[0043] Table 2 Comparison of classification accuracy of different methods

[0044]

[0045] The above description is only a preferred embodiment of the present invention and is 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 scope of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for fine classification of land use in rare earth mining areas based on hyperspectral remote sensing images, characterized by: The implementation steps of this method are: Step 1: Determine the classification system of rare earth mining areas based on current national standards and existing relevant research, combined with field survey information and hyperspectral remote sensing images of rare earth mining areas; Step 2: extracting image features from the hyperspectral remote sensing image using an object-oriented feature extraction module; Step 3: Construct an object-oriented one-dimensional convolutional neural network classification model based on a multi-layer attention mechanism, complete deep feature learning and the classification of the high-light remote sensing image through the convolutional neural network, and then complete the land use classification task of the rare earth mining area.

2. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images according to claim 1, characterized in that: The object-oriented feature extraction module has the following specific implementation steps: the first step is to determine the optimal segmentation scale of the hyperspectral remote sensing image based on the rare earth mining area classification system using a scale parameter estimation model; the second step is to apply the optimal segmentation scale to the segmentation of the hyperspectral remote sensing image to obtain the hyperspectral remote sensing image segmentation result; the third step is to extract the image features of the hyperspectral remote sensing image from the hyperspectral remote sensing image segmentation result.

3. The method for fine classification of land use in rare earth mining areas based on hyperspectral remote sensing images as claimed in claim 2, characterized in that: The image features of the hyperspectral remote sensing image include spectral features and spatial features.

4. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images as claimed in claim 3, characterized in that: The spectral characteristics include band characteristics and index characteristics.

5. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images as claimed in claim 4, characterized in that: The band characteristics are 32 means and 32 standard deviations of each band.

6. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images as claimed in claim 4, characterized in that: The index features are constructed using the reflectance of 32 original bands of the high-light remote sensing image as input parameters, and the index features are chlorophyll vegetation index, normalized vegetation index, ratio vegetation coefficient, green normalized vegetation index, difference vegetation index, 2 normalized difference red edge indices, and terrestrial chlorophyll index.

7. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images as claimed in claim 3, characterized in that: The spatial features include geometric features and texture features.

8. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images as claimed in claim 7, characterized in that: The texture feature is constructed by the synergy, contrast, dissimilarity, information entropy, mean, variance, correlation, and second-order moment of the hyperspectral remote sensing image segmentation result based on the gray level co-occurrence matrix.

9. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images as claimed in claim 7, characterized in that: The geometric features introduce the idea of ​​object-oriented classification. Based on the segmentation results of the hyperspectral remote sensing image, the asymmetry, boundary index, compactness, density, ellipse fitting, radius of the largest closed ellipse, radius of the smallest closed ellipse, rectangle fitting, roundness, shape index, area, boundary length, length, aspect ratio, number of pixels, volume, and width of the hyperspectral remote sensing image are calculated.

10. The method for fine classification of rare earth mining area land use based on hyperspectral remote sensing images as claimed in claim 1, characterized in that: The object-oriented one-dimensional convolutional neural network classification model based on a multi-layer attention mechanism has the following specific implementation steps: the first step is to construct a one-dimensional convolutional neural network, and input the image features of the hyperspectral remote sensing image to extract its deep image features; the one-dimensional convolutional neural network includes a convolutional layer, an activation function, a pooling layer and a fully connected layer; the second step is to introduce a Transformer module into the one-dimensional convolutional neural network to dynamically focus on local image features and global image features; the third step is to use the convolutional attention module to extract significant image feature information in the one-dimensional convolutional neural network, perform image feature optimization, and explore the distinguishability of the land use classification task; the fourth step is to classify the significant image feature information output by the one-dimensional convolutional neural network to obtain land use classification information.

Images