A rare earth mine area scene typical feature identification method for multispectral remote sensing images

Abstract

The application discloses a rare earth mine area scene typical object recognition method for multispectral remote sensing images, and model construction is carried out from five aspects of constructing a multispectral network input head, suppressing background noise, constructing a class attention constraint mechanism, constructing a multiscale fusion module and designing a loss function, so that the advantages of high-resolution remote sensing image data are fully utilized to improve the target recognition accuracy, speed and other performances of the model, the detection efficiency can be effectively improved, the complexity and real-time requirement of rare earth mine area management can be better met, and the sustainable management of the rare earth mine area and the realization of the ecological environment protection target are facilitated.

Description

Technical Field

[0001] This invention belongs to the field of image processing, specifically relating to a method for identifying typical features in rare earth mining areas based on multispectral remote sensing images. Background Technology

[0002] Ion-adsorption rare earth elements from southern China are widely used in high-tech fields, becoming a key strategic resource. However, the long-term global supply of rare earths has led to overconsumption and ecological damage. To protect resources and promote sustainable development, China has implemented rare earth mining quota controls and emphasized mine reclamation and restoration. Against this backdrop, the development of high-resolution remote sensing technology and artificial intelligence has provided new opportunities for precise detection and vegetation monitoring in rare earth mining areas, and offers an important pathway to realizing intelligent remote sensing monitoring methods to promote information-based management of rare earth mining areas.

[0003] However, ion-adsorption rare earth deposits are characterized by low grades, and mining sites are typically scattered and small-scale, often located in remote mountainous areas with complex terrain. Manual surveys are costly and difficult to conduct in a timely manner, failing to meet the needs of real-time monitoring in mining areas. The process-oriented nature of rare earth mining, especially the dispersed and small-scale nature of mining sites, makes traditional manual survey methods difficult and inefficient. In this context, the need for accurate detection of rare earth mining activities and monitoring of vegetation growth during mine reclamation has become increasingly apparent. Therefore, the demand for automated and intelligent monitoring methods in the field of rare earth mining area monitoring has become urgent.

[0004] With the rapid development of remote sensing technology, space-based Earth observation techniques, especially remote sensing, have made significant progress. The spatial resolution of remote sensing images has reached the sub-meter level, revealing richer spatial structures and surface textures of ground objects. This development holds immense potential for refined ground object identification in complex scenarios such as rare earth mining areas. However, it's important to note that while improved spatial resolution of remote sensing images provides more detail, it doesn't necessarily translate to increased accuracy in remote sensing data analysis and interpretation. In fact, deep learning-based target identification methods in high-resolution images suffer from several common drawbacks. First, the accuracy of existing techniques remains relatively low due to background noise in remote sensing images. Second, there are issues of heterospectral similarity and heterogeneous similarity. In high-resolution images, the same ground object may exhibit different spectral responses, leading to heterospectral similarity; conversely, different ground objects may have similar spectral responses, resulting in heterogeneous similarity. These problems make traditional spectral information-based classification methods difficult. Finally, there are issues arising from differences in spatial scale. High-resolution images may contain targets at multiple scales, posing a challenge to algorithms based on fixed scales and necessitating multi-scale processing. These issues are all pressing problems that need to be addressed in current deep learning target recognition research, as resolving them can further improve the model's recognition accuracy. Therefore, in scenarios involving massive amounts of high-resolution remote sensing data, how to perform data analysis and scene interpretation more efficiently and automatically has become one of the most pressing issues for achieving high-precision and intelligent mine supervision. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for identifying typical features in rare earth mining areas based on multispectral remote sensing images. By constructing an algorithm model applicable to the identification of typical features throughout the entire process from mining to reclamation in rare earth mining areas, the method can achieve monitoring of the mining process and the vegetation in the reclamation area, thus providing technical support for the intelligent and information-based mining of future mining areas.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention includes the following steps:

[0007] A method for identifying typical ground features in rare earth mining areas based on multispectral remote sensing imagery, specifically:

[0008] The multispectral channel features an input head module (MBH) design:

[0009] Unlike traditional RGB images, remote sensing multispectral images possess richer spectral dimensional features. Obtaining more effective feature maps from the original spectral information is a prerequisite for subsequent networks to capture more significant class features. Therefore, the MBH module established in this paper, after inputting multispectral channel information, obtains more feature differences in both spatial and channel dimensions through convolutional operations in different branches, ultimately retaining more discriminative initial feature maps from the original data. First, the original image is simultaneously input into two branches for parallel feature extraction. One branch, while expanding the receptive field through a 3x3 convolution to extract features, primarily aims to expand the combined spectral channel features, increasing the number of input channels from 4 to 64. The other branch uses a depthwise seperable convolution with a kernel size of 3x3. This branch separates the spatial and channel dimensional correlations, focusing more on spatially perceptual features on each feature map. It is important to note that both branches use the same receptive field but different convolutional methods for feature extraction, and both maintain the same image size as the input. Then, the outputs of the two branches are concatenated using a cat operation along the channel dimension to obtain a feature map of the same size as the original image but with more channels. The channel dimension is a stack of spectral combined features and spatially perceptual features of each channel. Next, average pooling and max pooling are used in channel attention to calculate the statistical features of the feature maps in each channel, and these are fed into a shared fully connected layer. After completing the channel feature mapping, channel attention is used to perform weighted fusion with the feature maps, improving the interdependence between multi-channel features. Finally, an 11-convolution is added to capture the differences between the features extracted by the two branches, while also further compressing the number of channels to reduce the number of network model parameters.

[0010] Feature extraction structure design:

[0011] We chose the CSP Block as the basic module for feature extraction in the network, such as Figure 3 As shown, this is used to further compress the feature maps output by the network head module and extract high-level semantic features. Our backbone network has four layers. Each layer first uses a convolutional kernel with a stride of 2 to achieve a downsampling effect. The subsequent CSP Block module contains three bottleneck residual units forming a large residual block, which to some extent prevents network degradation. The feature map size is reduced by half after each layer, and the ratio of the output feature map to the original image size is {1 / 2, 1 / 4, 1 / 8, 1 / 16}. Furthermore, the SPP structure is used in the highest semantic layer to further expand the receptive field.

[0012] Class-constrained attention module design:

[0013] To address the issues of overlooked features and semantic confusion between categories in high-resolution remote sensing imagery of rare earth mining areas, a class-constrained attention module suitable for target detection was established. This module enhances category semantic features, thereby mitigating semantic confusion in complex mining scenes. The class-constrained attention module consists of three parts: First, the Class Mapping Generation (CMGM) submodule performs a category mapping process from the feature maps of the backbone network to generate category feature maps. Then, the Class Feature Library (CFLM) submodule further generates a class feature library based on the category feature maps to obtain intra- and inter-class semantic feature sets. Finally, the Class Feature Library (CAM) submodule combines global scene context information and the class feature library to establish an attention mechanism with class information constraints.

[0014] Design of multi-scale feature enhancement and fusion module:

[0015] The high-precision detection of multi-scale ground objects in rare earth mining areas relies on accurate deep semantic features and refined shallow localization features. To fully integrate deep semantic and shallow localization features for more accurate target feature representation, this study designs a multi-scale feature enhancement and fusion module. This module extracts multi-scale feature maps with semantic features ranging from shallow to deep through the backbone network. This study selects the remaining three layers (excluding the shallowest layer) for further feature enhancement and fusion. Before fusion, the multi-scale feature maps pass through a CCA module to obtain multi-scale class-constrained attention. Then, the class-constrained attention and the feature maps are multiplied using an element-wise multiplication strategy to impose stronger class constraints on the multi-scale feature maps, thereby enhancing and balancing the multi-scale features of the target. To fully integrate multi-level features and improve the network's multi-scale target detection capability, a path aggregation network (PANet), a feature pyramid structure with bidirectional fusion paths from top to bottom and bottom to top, is used for the fusion of deep semantic features and shallow localization features. Ultimately, this multi-scale feature with strong class constraints further strengthens the class feature representation during the fusion process.

[0016] Multi-task loss function design:

[0017] In order to integrate the process of constructing class constraints with the tasks of target localization and classification into a unified end-to-end network training, this study redesigned a novel multi-task loss, MCCA-Loss, which includes traditional loss functions such as bounding box classification, regression, and confidence in the target detection task, as well as a class mapping loss function to assist in the process of constructing class information constraints.

[0018] Furthermore, in the design of the input head module (MBH), branch a1 expands the spectral channel combination features while extracting features through a 33-convolution to expand the receptive field, increasing the number of input channels from 4 to 64; branch a2 uses a depthwise separable convolution with a kernel size of 33 for feature extraction, which separates the correlation between spatial dimension and channel dimension.

[0019] Among them, branches a1 and a2 use the same receptive field and different convolution methods to perform feature extraction operations, and both maintain the same output and input image size.

[0020] Furthermore, the channel dimension in the input header module (MBH) design is a stack of spectral combination features and spatially perceptual features of each channel.

[0021] The beneficial effects of this invention are as follows:

[0022] The target recognition method constructed in this invention has higher accuracy than currently popular methods, and can accurately identify typical features in rare earth mining areas. Compared with existing technologies, the accuracy of the deep learning target recognition results in this invention is already sufficient, and no further post-processing is needed to improve the model's recognition accuracy. Furthermore, in the process of technical optimization, this invention employs several optimization methods specifically for the rare earth mining area scenario. First, to address the interference caused by environmental shadows in mining areas, the spectral features of remote sensing images are fully utilized. A novel network input head structure supporting multispectral channels is constructed to introduce richer spectral dimension features, and a branching structure is used to obtain more discriminative features, mitigating the interference caused by environmental shadows. Second, to address the problem that small features in mining areas are not significant and are easily affected by complex background noise, thus exacerbating semantic confusion between geographic entities, a class-constrained attention mechanism is constructed to suppress complex background noise in remote sensing images and emphasize the semantic feature differences between categories, achieving more accurate discovery of potential small targets in complex mining areas. Thirdly, a multi-scale fusion module was constructed, which fuses multi-scale feature maps with strong class constraints through bidirectional feature fusion such as top-down and bottom-up, for the fusion of deep semantic features and shallow localization features, so as to achieve more accurate identification of target individuals. The method constructed by this invention can effectively capture the subtle features of rare earth mining area features, solve the problem of small and densely distributed features, and thus improve the accuracy of identification.

[0023] Compared with existing technologies, this invention does not require cumbersome operations and also performs exceptionally well in terms of recognition efficiency and accuracy. In the process monitoring of rare earth mining, it can accurately and efficiently identify typical features within the mining area, providing excellent technical means and support for the information-based supervision of rare earth mines. Attached Figure Description

[0024] Figure 1 The method framework constructed for this invention.

[0025] Figure 2 A visual comparison chart of class-constrained attention.

[0026] Figure 3 The results show the identification of typical features in rare earth mining areas using existing technologies and the method of this invention.

[0027] Figure 4 The results of vegetation reclamation identification are based on existing technologies and the method of this invention. Implementation

[0028] The present invention will be further described in detail below with reference to the embodiments. Example

[0029] This embodiment identifies multispectral remote sensing images of the same mining area using recognition methods with and without class-constrained attention modules, respectively. The results are shown in [the table / document / reference]. Figure 2 .

[0030] from Figure 2 Lines 1-2 show that without class-based attention, while there is good activation of the main foreground targets, this activation is not significant, and small objects are not effectively activated. However, after adding class-based attention, the network shows more significant activation of all foreground targets, while this pixel-level attention still provides good activation of small objects. In mining scenes with complex backgrounds, such as... Figure 2 As shown in the second column of rows 3-4, false activations occurred, mainly due to confusion caused by the dark features generated by shadows. However, the more accurate class features constructed by incorporating multispectral information into the CA module were used to apply stronger class supervision, enhancing foreground target features and suppressing complex background information, resulting in more accurate activation of ground targets, while irrelevant background information was not falsely activated. Example

[0031] This embodiment uses existing technology and the land cover identification method constructed in this invention to identify typical land covers in multispectral remote sensing images of the same rare earth mining area. The identification results are shown in [the provided text]. Figure 3 .

[0032] Figure 3The first two rows show the identification results of densely distributed sedimentation ponds in rare earth mining areas. Compared with existing technologies, the method constructed in this invention has higher accuracy in the identification results. It can not only accurately identify the number of densely distributed ground features, but also correctly classify them. The third to ninth rows show the ground feature identification results of various methods in the face of complex terrain environment in rare earth mining areas. As can be seen from the figure, the identification results of this method are superior to other methods in both mountainous and hilly areas, and the effect of distinguishing sedimentation ponds from abandoned sedimentation ponds is particularly obvious. Example

[0033] This embodiment uses existing technology and the land cover identification method constructed in this invention to identify typical land cover in multispectral remote sensing images of the same reclaimed vegetation area. The identification results are shown in [the provided text]. Figure 4 .

[0034] from Figure 4 As can be seen in lines 1-3, the method in this invention has stronger feature representation capabilities and can reduce the influence of factors such as surface weeds and shadows, thereby avoiding false detection of ground features. In lines 4-6, some existing methods have the problem of multiple clustered plants being falsely detected as a single plant in the identification results. However, the method of this invention achieves more accurate individual plant detection by implementing pixel-level category supervision on multi-scale feature maps.

[0035] In summary, compared with existing technologies, the target identification method constructed in this invention has higher accuracy. During the rare earth mining process, deep learning methods are used to accurately identify leaching sedimentation ponds within the mining area, thereby clearly understanding the mining scale and situation. It can also be applied to the monitoring of illegal rare earth mining, accurately and efficiently identifying typical features within the mining area, providing technical means and support for the information-based supervision of rare earth mining areas. Furthermore, deep learning methods are used to accurately identify reclaimed vegetation planted after rare earth mining, allowing for monitoring of the growth of reclaimed vegetation and providing a reference for ecological environment assessment of the mining area. Moreover, this invention requires no cumbersome operation and is easy to use.

Claims

1. A method for identifying typical ground features in rare earth mining areas based on multispectral remote sensing imagery, characterized in that: It includes the following five modules: Input Header Module Design: After inputting multispectral channel information, the input header module obtains more feature differences from the spatial and channel dimensions through convolution operations of different branches, and finally retains a more discriminative initial feature map from the original data. Feature extraction structure design: used to further compress the feature maps output by the network head module and extract high-level semantic features; Class-constrained attention module design: Enhance category semantic features by establishing a class-constrained attention module; Design of Multi-Scale Feature Enhancement and Fusion Module: By designing a multi-scale feature enhancement and fusion module, we can fully integrate deep semantic and shallow localization features to obtain a more accurate target feature representation; Multi-task loss function design: Integrate the process of constructing class constraints with target localization and classification tasks into a unified end-to-end network training; Specifically, the input header module is designed as follows: After acquiring the original image, the original image is input into a multispectral channel with an input head module design. The input head module design has two branches, a1 and a2. After passing through the two branches, the output results are stitched together in the channel dimension to obtain a feature map with the same size as the original image but with more channels. Then, average pooling and max pooling are used in channel attention to calculate the statistical features of the feature map in each channel, and they are sent to a shared fully connected layer. After completing the channel feature mapping, the feature map is weighted and fused through channel attention. Finally, an I1 convolution is added to capture the difference features between the features extracted by the two branches. Specifically, the feature extraction structure design is as follows: The CSP Block is chosen as the basic module for feature extraction in the network. The backbone network has four layers. Each layer first uses a convolutional kernel with a stride of 2 to achieve a downsampling effect. The subsequent CSP Block module contains three bottleneck residual units, which can form a large residual block. The size of the feature map is reduced by half after each layer. The ratio of the output feature map of the four layers to the original image size is {1 / 2, 1 / 4, 1 / 8, 1 / 16}. The SPP structure is used in the highest semantic layer to further expand the receptive field. Specifically, the class-constrained attention module is designed as follows: The class-constrained attention module is divided into three parts. First, the class mapping generation submodule completes the class mapping process from the feature map of the backbone network to generate a class feature map. Then, the class feature library submodule further generates a class feature library based on the class feature map to obtain the semantic feature set within and between classes. Finally, the class feature library submodule combines the global context information of the scene and the class feature library to establish an attention with class information constraints. Specifically, the design of the multi-scale feature enhancement and fusion module is as follows: Before fusion, the multi-scale feature maps first pass through the CCA module to obtain multi-scale class-constrained attention. Then, the class-constrained attention and the feature maps are multiplied by an element-wise strategy to apply stronger class constraints to the multi-scale feature maps, so as to enhance and balance the multi-scale features of the target. In order to fully fuse multi-level features and improve the multi-scale target detection capability of the network, the path aggregation network PANet, a feature pyramid structure with bidirectional fusion paths from top to bottom and bottom to top, is used to fuse deep semantic features and shallow localization features. Finally, this multi-scale feature with strong class constraints further strengthens the class feature representation during the fusion process. Specifically, the design of the multi-task loss function is as follows: A multi-task loss function, MCCA-Loss, incorporates traditional loss functions for bounding box classification, regression, and confidence in object detection tasks, as well as a class mapping loss function used to assist in constructing class information constraints.

2. The method for identifying typical ground features in rare earth mining areas based on multispectral remote sensing imagery as described in claim 1, characterized in that: Branch a1 expands the receptive field and extracts features through a 33-convolution while increasing the spectral channel combination features, increasing the number of input channels from 4 to 64; Branch a2 uses a depthwise separable convolution with a kernel size of 33 for feature extraction, which separates the correlation between spatial dimension and channel dimension. Among them, branches a1 and a2 use the same receptive field and different convolution methods to perform feature extraction operations, and both maintain the same output and input image size.

3. The method for identifying typical ground features in rare earth mining areas based on multispectral remote sensing imagery as described in claim 1, characterized in that: The channel dimension is a stack of spectral combination features and spatial perception features of each channel.

Images