A method, device, equipment and medium for target identification in an open pit mine

Abstract

The application discloses an open-pit mine target identification method and device, equipment and medium, relates to the technical field of remote sensing intelligent interpretation and open-pit mine target identification, and the method comprises the steps of obtaining unmanned aerial vehicle remote sensing images and laser radar point cloud data of an open-pit mine area to be identified; obtaining a digital orthographic image based on the unmanned aerial vehicle remote sensing images; obtaining slope data based on the laser radar point cloud data; obtaining an open-pit mine semantic segmentation mask based on the digital orthographic image and the slope data by using an open-pit mine semantic segmentation model; and obtaining an open-pit mine target identification result based on the open-pit mine semantic segmentation mask. The application can improve the geometric accuracy and stability of the open-pit mine target identification result.

Description

Technical Field

[0001] This application relates to the fields of remote sensing intelligent interpretation and open-pit mine target recognition technology, and in particular to a method, apparatus, equipment and medium for open-pit mine target recognition. Background Technology

[0002] Open-pit mining is common in mineral resource development, and the resulting open-pit mines represent a significant type of anthropogenically disturbed surface. These surface units often exhibit morphological characteristics such as large spatial scale, steep slopes, clearly defined terraces, and stripped or severely disturbed surface cover. Because mining directly alters the original topography and surface material composition, it can potentially lead to ecological damage, increased soil erosion, and a heightened risk of localized geological disasters, while also posing a potential threat to the stability and safety of surrounding infrastructure such as roads and pipelines. Therefore, to meet regulatory and governance needs, there is an urgent need to establish a technical method capable of quickly identifying open-pit mine targets, providing early warnings of suspected illegal mining activities, and accurately delineating their operational boundaries and disturbance areas. This would provide a basis for continuous monitoring, risk assessment, and law enforcement evidence collection.

[0003] Traditional regulatory methods rely heavily on personnel patrols, on-site surveys, location tracking, and evidence collection. Constrained by transportation conditions and equipment, these methods generally suffer from high costs, low efficiency, and discontinuous coverage. With improved Earth observation capabilities, high spatial resolution remote sensing imagery provides a feasible path for acquiring information on surface changes over large areas and periodically, laying the foundation for large-scale resource verification, ecological supervision, and soil and water conservation monitoring. However, for a long time, remote sensing interpretation has primarily relied on manual visual interpretation and experience-based inference. Facing operational scenarios with high timeliness requirements, large regional spans, and a large number of images, stable and repeatable high-precision interpretation is difficult to achieve. In recent years, artificial intelligence methods have developed rapidly in the field of remote sensing information extraction. Automatic identification and interpretation based on machine learning and deep learning have gradually become an important direction for improving efficiency. Traditional machine learning algorithms such as support vector machines, decision trees, and maximum likelihood classification have been used for target identification and classification in mines, improving automation to some extent. However, limitations remain in areas such as complex background interference, significant differences in target morphology, and cross-regional adaptability.

[0004] Although the automatic identification technology of open-pit mines based on convolutional neural networks has made some progress in recent years and can replace the manual interpretation process to a certain extent, it still exposes several key bottlenecks when facing complex mining areas and large-scale monitoring applications. These bottlenecks are mainly reflected in the following aspects: (1) Due to the limitations of the locality of convolutional operators and the way the receptive field is expanded, the model is not capable of representing the overall characteristics of cross-scale, strong structural targets such as "large-scale morphology - multi-level steps - continuous slopes" in open-pit mines. For scenarios with significant changes in the scale of the pit, clear step levels, and continuous geometric features of the slope, the network often cannot simultaneously take into account local details and global spatial dependencies, resulting in insufficient capture of cross-scale correlation features. Consequently, in large-scale remote sensing images, the identification accuracy is limited and the generalization stability is insufficient. (2) The discrimination criteria of related methods are mostly concentrated on texture, color spectrum, and local geometric clues. They are insufficient in depicting the spatial organization relationship between typical structural units of the mine, especially lacking topological constraints and hierarchical logical modeling of structural units. For example, there should be a clear hierarchical correspondence and spatial adjacency between the mining slope and each level of steps, and there should also be certain connectivity and zoning rules between different work areas; when the model does not explicitly learn the above relationships, it is easy to produce incorrect segmentation, discontinuous structure or even breakage when the background interference is strong or the structure is locally missing. (3) Under the conditions of vegetation cover, shadow occlusion, slag / soil pile interference and limited image spatial resolution, the mining boundary often presents a weak contrast, irregular and fragmented form. The existing convolutional network has insufficient response ability and noise suppression ability to weak boundaries, and the discrimination of boundary areas is not precise enough, which easily leads to missed detection and false detection, and is accompanied by problems such as boundary non-closure, local discontinuity and jaggedness, thereby affecting the geometric accuracy and stability of the recognition results, and making it difficult to meet the business needs of accurately delineating the disturbance range. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, equipment, medium, and product for target identification in open-pit mines, which can improve the geometric accuracy and stability of target identification results in open-pit mines.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for target identification in open-pit mines, including: Acquire drone remote sensing images and lidar point cloud data of the area to be identified in the open-pit mine; Digital orthophotos are obtained based on the UAV remote sensing images; slope data is obtained based on the lidar point cloud data. Acquire sample UAV remote sensing images and sample lidar point cloud data of open-pit mines; A digital orthophoto image of the sample is generated based on the sample UAV remote sensing image; the slope data of the sample is obtained based on the sample lidar point cloud data; An open-pit mine dataset is constructed based on the sample digital orthophotos and the sample slope data; An initial semantic segmentation model is constructed based on the U-Net network and a two-layer routing attention module; Using the aforementioned open-pit mine dataset, the initial semantic segmentation model is trained to obtain the open-pit mine semantic segmentation model; Using the aforementioned open-pit mine semantic segmentation model, an open-pit mine semantic segmentation mask is obtained based on the digital orthophoto and the slope data; The target recognition result of the open-pit mine is obtained based on the semantic segmentation mask of the open-pit mine.

[0007] In one embodiment, obtaining sample slope data based on the sample lidar point cloud data includes: The sample lidar point cloud data is classified and filtered to extract ground point cloud data and construct a sample digital elevation model. The sample slope data is determined based on the sample digital elevation model.

[0008] In one embodiment, an open-pit mine dataset is constructed based on the sample digital orthophotos and the sample slope data, including: Based on the sample digital elevation model, the spatial distribution of the open-pit mine area boundary and typical structural units is calibrated in the sample digital orthophoto image to obtain the annotation results; the typical structural units include the pit area, the stope slope, the mining bench and the stope bottom. Using the sample digital orthophoto and the sample slope data as input, and the annotation results corresponding to the sample digital orthophoto as output, a sample pair is constructed. The open-pit mine dataset is constructed based on the sample pairs.

[0009] In one implementation, an initial semantic segmentation model is constructed based on a U-Net network and a two-layer routing attention module, including: A dual-layer routing attention module is introduced into the bottleneck layer of the U-Net network to obtain the initial semantic segmentation model.

[0010] In one embodiment, the initial semantic segmentation model is trained using the open-pit mine dataset to obtain the open-pit mine semantic segmentation model, including: A training set and a validation set were constructed based on the open-pit mine dataset. The initial semantic segmentation model is trained based on the training set using the backpropagation algorithm to obtain the trained initial semantic segmentation model. The initial semantic segmentation model after training is validated using a loss function on the validation set until the loss function value on the validation set reaches the set requirement, thus obtaining the open-pit mine semantic segmentation model.

[0011] In one embodiment, the loss function is a cross-entropy loss function, a Dice loss function, or a combined loss function; the combined loss function includes both the cross-entropy loss function and the Dice loss function.

[0012] In one embodiment, obtaining the target recognition result of the open-pit mine based on the semantic segmentation mask of the open-pit mine includes: The semantic segmentation mask of the open-pit mine is subjected to regularization and vectorization processing to obtain the boundary vector data of the open-pit mine; Based on the geographic reference information of the UAV remote sensing image, the boundary vector data of the open-pit mine is mapped to a standard coordinate reference to determine the morphological index value corresponding to the UAV remote sensing image. The target identification result of the open-pit mine is obtained based on the boundary vector data of the open-pit mine and the morphological index value corresponding to the UAV remote sensing image.

[0013] Secondly, this application provides a target identification device for open-pit mines, comprising: The data acquisition and processing unit is used to acquire UAV remote sensing images and lidar point cloud data of the area to be identified in the open-pit mine; the data acquisition and processing unit is also used to obtain digital orthophotos based on the UAV remote sensing images; the data acquisition and processing unit is also used to obtain slope data based on the lidar point cloud data; the data acquisition and processing unit is also used to acquire sample UAV remote sensing images and sample lidar point cloud data of the open-pit mine; the data acquisition and processing unit is also used to generate sample digital orthophotos based on the sample UAV remote sensing images; the data acquisition and processing unit is also used to obtain sample slope data based on the sample lidar point cloud data. The model building unit, connected to the data acquisition and processing unit, is used to build an open-pit mine dataset based on the sample digital orthophoto and the sample slope data; the model building unit is also used to build an initial semantic segmentation model based on a U-Net network and a two-layer routing attention module; the model building unit is also used to train the initial semantic segmentation model based on the open-pit mine dataset to obtain an open-pit mine semantic segmentation model; The identification unit is connected to the data acquisition and processing unit and the model building unit respectively, and is used to obtain the open-pit mine semantic segmentation mask based on the digital orthophoto and the slope data using the open-pit mine semantic segmentation model. The output unit, connected to the recognition unit, is used to obtain the target recognition result of the open-pit mine based on the semantic segmentation mask of the open-pit mine.

[0014] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the open-pit mine target identification method described in any one of the above.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the open-pit mine target identification method described above.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, apparatus, equipment, and medium for target recognition in open-pit mines. It constructs an open-pit mine dataset based on sample digital orthophotos and sample slope data, and builds an initial semantic segmentation model based on a U-Net network and a two-layer routing attention module. The initial semantic segmentation model is further trained on the open-pit mine dataset to obtain a semantic segmentation model. This effectively alleviates the shortcomings of traditional convolutional neural networks, which are limited by local receptive fields and struggle to consider the overall spatial structure of the mining area. It also solves the problems of missed and false detections in traditional convolutional neural networks, as well as the tendency to produce incorrect segmentation, structural incoherence, or even breaks in situations with strong background interference or local structural defects. By employing a semantic segmentation model, a semantic segmentation mask for the open-pit mine is obtained based on the digital orthophotos to be recognized, thereby obtaining the target recognition result. This improves the geometric accuracy and stability of the recognition result, thus meeting the business requirement of accurately delineating the disturbance range. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a target identification method in an open-pit mine according to an embodiment of this application; Figure 2 This is a schematic diagram illustrating the process of constructing an open-pit mine dataset according to an embodiment of this application. Figure 3 This is a schematic diagram of the structure of an open-pit mine semantic segmentation model provided in an embodiment of this application; Figure 4A schematic diagram of the structure of an open-pit mine target identification device provided in another embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] In one exemplary embodiment, such as Figure 1 As shown, a target identification method for open-pit mines is provided, including: Step S1: Obtain UAV remote sensing images and lidar point cloud data of the area to be identified in the open-pit mine.

[0022] Step S2: Obtain digital orthophotos based on UAV remote sensing imagery. Obtain slope data based on lidar point cloud data.

[0023] Step S3: Using the open-pit mine semantic segmentation model, a semantic segmentation mask for the open-pit mine is obtained based on digital orthophotos and slope data.

[0024] The construction process of the semantic segmentation model for open-pit mines includes: Step S21: Obtain sample UAV remote sensing images and sample lidar point cloud data of the open-pit mine.

[0025] Step S22: Generate sample digital orthophotos based on sample UAV remote sensing images. Obtain sample slope data based on sample lidar point cloud data.

[0026] Step S23: Construct an open-pit mine dataset based on sample digital orthophotos and sample slope data.

[0027] Step S24: Construct an initial semantic segmentation model based on the U-Net network and a two-layer routing attention module.

[0028] Step S25: Using the open-pit mine dataset, train the initial semantic segmentation model to obtain the open-pit mine semantic segmentation model.

[0029] Step S4: Obtain the target recognition result of the open-pit mine based on the semantic segmentation mask of the open-pit mine.

[0030] In one embodiment, the process of obtaining sample slope data based on sample lidar point cloud data in step S22 includes: classifying and filtering the sample lidar point cloud data, extracting ground point clouds, and constructing a sample digital elevation model (DEM). The sample slope data is then determined based on the DEM. Similarly, the process of obtaining slope data based on lidar point cloud data in step S2 is the same as the process of obtaining sample slope data based on sample lidar point cloud data.

[0031] Based on this, the implementation process of step S23 includes: Based on the sample digital elevation model, the spatial distribution of the open-pit mine area boundary and typical structural units is calibrated in the sample digital orthophoto image to obtain annotation results. Typical structural units include the pit area, stope slope, mining bench, and stope bottom. Using the sample digital orthophoto image and sample slope data as input, and the annotation results corresponding to the sample digital orthophoto image, output sample pairs are constructed. An open-pit mine dataset is then constructed based on these sample pairs.

[0032] For example, to acquire high-precision, multi-source sample data, an unmanned aerial vehicle (UAV) platform equipped with a Light Detection and Ranging (LiDAR) system and a visible light camera was used to conduct aerial photogrammetry operations in the open-pit mine. This was supplemented by handheld RTK devices for high-precision positioning and control, thereby acquiring high-resolution UAV remote sensing images and LiDAR point cloud data. The aerial photogrammetry operation covered the entire open-pit mine survey area, and data acquisition could be completed through multiple flights, ultimately obtaining a complete, high-resolution, and reliable sample set (including sample UAV remote sensing images and sample LiDAR point cloud data).

[0033] Aerial triangulation and orthorectification processing (which can be performed in image processing software) are performed on sample UAV remote sensing images to generate a Digital Orthophoto Map (DOM) of the open-pit mine survey area. Subsequently, the generated DOM is cropped at the survey area boundaries and its quality is checked to form true-color image data (i.e., sample DOM) that can be used for target interpretation and model training. Simultaneously, the synchronously acquired sample LiDAR point cloud data is classified and filtered to extract ground point clouds from the sample LiDAR point cloud data, thereby constructing a high-precision sample Digital Elevation Model (DEM). Based on the sample DOM, sample slope data for the corresponding area is calculated. The sample slope data is used to characterize the intensity of surface slope changes and local terrain slope characteristics to assist in target identification in the open-pit mine.

[0034] In the data annotation phase, high-resolution sample digital orthophotos are primarily used as the foundation, combined with 3D models and sample digital elevation models for comprehensive interpretation. Interpreters manually interpret and annotate the data, calibrating the spatial distribution of open-pit mine area boundaries and typical structural units within the sample digital orthophotos. Typical structural units include the pit area, stope slopes, mining benches, and stope floor. Based on these annotation results, the sample digital orthophotos, and sample slope data, an open-pit mine dataset is constructed. Figure 2 As shown, the construction process of the open-pit mine dataset includes: first, acquiring sample digital elevation models (DEMs) and high-resolution sample digital orthophotos; generating a 3D model based on the sample DEMs, and using the 3D model as auxiliary data for manual interpretation (i.e., manual visual delineation of the open-pit mine area); then, interpreters combining the sample DEMs, the 3D model, and the sample digital orthophotos to manually visually delineate the open-pit mine area, obtaining the initial open-pit mine area boundary; expert review of the initial open-pit mine area boundary to obtain vector data of the open-pit mine area; vector-to-raster processing of the open-pit mine area vector data to generate raster label files consistent with the spatial location of the sample digital orthophotos; based on the raster label files, extracting the open-pit mine area from the sample digital orthophotos and the sample slope data determined by the sample DEMs; finally, converting the extracted results into a dataset format, cropping and organizing them into a deep learning open-pit mine dataset (i.e., the open-pit mine dataset).

[0035] In this embodiment, each sample pair includes a sample digital orthophoto with a fixed size of 512×512 pixels, sample slope data of the corresponding area, and a semantic segmentation label file (i.e., annotation result, PNG format) corresponding to the sample digital orthophoto.

[0036] In open-pit mine identification using UAV imagery, targets often possess both large-scale contours and fragmented boundaries, and are significantly affected by interference from shadows, vegetation, and spoil heaps. Traditional segmentation networks are prone to problems such as inconsistent overall structure, coarse boundaries, and missed detections due to weak boundaries. Therefore, to address these issues, step S24 involves introducing a dual-layer routing attention module into the bottleneck layer of the U-Net network to obtain an initial semantic segmentation model. By embedding the dual-layer routing attention module into the bottleneck layer of the U-Net network, its core dual-layer routing attention mechanism enables dynamic feature selection at both local and global scales: on the one hand, local attention can accurately capture texture changes and boundary gradient information within the open-pit mine, improving the model's edge localization capabilities under vegetation cover, shadow interference, and high-noise backgrounds; on the other hand, the global routing mechanism can establish long-range structural relationships between typical structural units such as pits, mining steps, and slopes over a larger spatial range, maintaining the continuous expression of the overall scope and structural morphology of the open-pit mine. This direction-sensitive and adaptive attention allocation strategy (i.e., the two-layer routing attention mechanism) effectively compensates for the shortcomings of the isotropic nature of traditional convolution, enabling the model to more accurately characterize the complex spatial structure and internal organizational form of open-pit mines.

[0037] The implementation process of step S25 includes: constructing a training set and a validation set based on the open-pit mine dataset; using the backpropagation algorithm, training an initial semantic segmentation model based on the training set to obtain the trained initial semantic segmentation model; and using a loss function to validate the trained initial semantic segmentation model based on the validation set until the loss function value on the validation set reaches the set requirement, thus obtaining the open-pit mine semantic segmentation model.

[0038] The loss function can be the cross-entropy loss function, the Dice loss function, or a combination loss function. The combination loss function includes both the cross-entropy loss function and the Dice loss function.

[0039] In this embodiment, in addition to constructing training and validation sets based on the open-pit mine dataset, a test set is also constructed based on the open-pit mine dataset.

[0040] For example, the open-pit mine dataset is divided into a training set, a validation set, and a test set according to a set ratio (8:1:1 in this embodiment). Sample digital orthophotos and sample slope data from the training set are used as input, and the annotation results corresponding to the sample digital orthophotos are used as supervision signals. The initial semantic segmentation model is trained according to a set number of training rounds, and the weight parameters in the model are optimized using the backpropagation algorithm. After each training round, the loss function value of the initial semantic segmentation model on the validation set is calculated to monitor the model's generalization performance on the validation set. A dynamic learning rate adjustment strategy controls the model's training convergence process. When the loss function value on the validation set does not decrease for multiple consecutive rounds (e.g., 4 or 5 rounds), or when the loss function value on the validation set meets the preset convergence condition (i.e., the loss function value is less than the preset convergence threshold, which can be set according to actual needs), an early stopping strategy is initiated to prevent overfitting and ensure optimal robustness, ultimately yielding the open-pit mine semantic segmentation model. The test set is used to independently test the open-pit mine semantic segmentation model after model training, to evaluate the model's generalization ability and actual recognition performance on unknown samples. Specifically, the digital orthophoto images of samples in the test set and the corresponding sample slope data are input into the open-pit mine semantic segmentation model. The model outputs the target category corresponding to each pixel in the digital orthophoto image of the samples, generating a semantic segmentation result map (i.e., a sample binary segmentation mask). The sample binary segmentation mask is compared pixel by pixel with the pre-labeled ground truth map (i.e., the labeling result) in the test set to obtain the number of correctly predicted pixels, the number of false positives, and the number of missed pixels. Based on this, evaluation indicators such as pixel accuracy, recall, F1 score, and mean intersection-over-union ratio are calculated to evaluate the recognition accuracy of the open-pit mine semantic segmentation model for the open-pit mine area and its boundaries.

[0041] In one embodiment, in step S3 above, slope data obtained from the digital elevation model (DEM) is used as a terrain auxiliary feature, rather than directly using the entire DEM as input to the open-pit mine semantic segmentation model. This is because the focus of open-pit mine identification is not on the absolute elevation value of the surface itself, but on the abrupt topographical structures such as pit boundaries, mining steps, slope steepness zones, and disturbance areas. The DEM mainly reflects the absolute elevation distribution of a region. Different mining areas often have differences in overall terrain undulation and elevation benchmarks. If the DEM is directly input into the open-pit mine semantic segmentation model, it easily introduces background elevation information that is not directly related to the open-pit mine target identification, affecting the cross-regional generalization ability of the open-pit mine semantic segmentation model. Slope data can convert elevation changes into local slope gradient features, more directly highlighting the structural characteristics of the open-pit mine, weakening the impact of overall terrain differences on the identification results, and making the open-pit mine semantic segmentation model more focused on the geometric shape and local terrain change features of the open-pit mine target. Therefore, it is more suitable as the terrain auxiliary input data in this application. Digital orthophotos are used to characterize the texture, color, boundary and surface morphology of open-pit mining areas, while slope data is used to characterize the intensity of surface slope changes, topographic slope and local undulation features. Both are used as inputs to the semantic segmentation model of open-pit mining to enhance the joint characterization ability of structures such as open-pit mine pits, mining benches and mining slopes.

[0042] Step S4 includes: performing regularization and vectorization processing on the semantic segmentation mask of the open-pit mine to obtain the open-pit mine boundary vector data. Based on the georeferenced information of the UAV remote sensing imagery, the open-pit mine boundary vector data is mapped to a standard coordinate datum to determine the morphological index values ​​corresponding to the UAV remote sensing imagery. Based on the open-pit mine boundary vector data and the morphological index values ​​corresponding to the UAV remote sensing imagery, the open-pit mine target recognition result is obtained.

[0043] For example, aerial triangulation and orthorectification are performed on UAV remote sensing imagery to obtain digital orthophotos. LiDAR point cloud data is classified and filtered to extract ground point clouds, constructing a digital elevation model (DEM), and slope data is determined based on the DEM. The digital orthophotos and slope data are then input into a semantic segmentation model for the open-pit mine.

[0044] The semantic segmentation model for open-pit mines has an overall U-shaped structure, including an encoder, a bottleneck layer, and a decoder. For example... Figure 3 As shown, Figure 3 The scene diagrams provided are for illustrative purposes only and are not intended to limit the scope of this application. Figure 3The numerical values ​​(e.g., 4, 64, 128, 256, etc.) are used as feature dimensions. The overall semantic segmentation model for open-pit mines has a U-shaped structure, including an encoder, a bottleneck layer, and a decoder. Digital orthophotos and slope data serve as multi-source inputs, and are first subjected to joint feature extraction via the encoder path. In the encoder, through successive layers of convolutional layers and downsampling operations (such as max pooling), multi-scale spatial features from shallow detail features to deep semantic features are extracted from the multi-source inputs, resulting in deep joint features. Among them, shallow features mainly characterize the texture and edge changes near the boundary of the open-pit mine, mid-level features mainly characterize the morphological features of the pit outline, step extension zone, and slope transition area, and deep features mainly characterize the overall spatial distribution pattern of the open-pit mine and its semantic differences with the surrounding background area, in order to adapt to the recognition needs of open-pit mines with irregular target boundaries, obvious structural hierarchy, strong background interference, and significant cross-regional continuity. Specifically, Figure 3 The 3×3 convolutional layer and ReLU in the model represent the convolution operation performed on the input features using a 3×3 convolutional kernel, and the introduction of non-linear expressive power using the ReLU activation function; the 2×2 max pooling layer is used to downsample the feature map to expand the receptive field and extract deeper semantic features; copying and pruning represent copying the high-resolution features of the corresponding layer of the encoder to the decoder, and then pruning them according to the feature map size before splicing and fusing them with the decoder features; the 2×2 upsampling layer is used to gradually restore the spatial resolution of the feature map; the 1×1 convolutional layer is used to map the feature channels in the output stage to obtain the pixel-level prediction results corresponding to the open-pit mine category; the dual-layer routing attention module is set at the bottleneck layer between the encoder and decoder, and is used to enhance the local window attention of the deep joint features and interact with cross-window routing information, thereby improving the model's ability to express structural features such as open-pit mine pits, steps, slopes and boundaries.

[0045] The bottleneck layer, located between the encoder and decoder, is where the deep semantic features (i.e., joint features) extracted jointly from digital orthophotos and slope data are most concentrated. Since this location has a large receptive field after multiple downsampling steps, it can more fully aggregate the joint features extracted based on the DOM and slope. Therefore, this application improves the convolutional layer at this location with a dual-layer routing attention module (including local window attention and cross-window routing information interaction) to simultaneously enhance key local details and overall spatial relationships, addressing the characteristics of open-pit mine identification such as fragmented local boundaries, distinct step levels, continuous slope distribution, and significant cross-regional structural correlations. Specifically, local window attention enhances the recognition ability of key areas such as pit boundaries, step edges, and slope transition zones, jointly represented by DOM-based texture information and slope topographic slope information; cross-window routing information interaction establishes cross-regional structural correlations and spatial constraints between pits, steps, and slopes, thereby improving the accuracy and stability of the open-pit mine semantic segmentation model in recognizing the overall morphology, irregular boundaries, and multiple consecutive mines.

[0046] The dual-layer routing attention module embedded in the bottleneck layer does not change the size of the input and output feature maps, but through the dual-layer routing attention mechanism and sparse attention computation, it performs global context modeling on the joint features extracted by the encoder, thereby enhancing the ability to capture long-distance dependencies of large-scale structures such as mine pits and steps.

[0047] The dual-layer routing attention module divides the bottleneck features (i.e., the joint features of the input) into multiple windows. Within each window, attention weights are calculated between the query vector Q, key vector K, and value vector V to obtain local enhancement features that highlight responses related to step edges, slope textures, and pit edge transition zones. During the global routing computation phase, several key tokens are selected from each window based on learnable routing scores. Cross-window sparse information interaction is then performed between these key tokens to establish long-distance dependencies between pits, steps, and slopes, suppressing the adverse effects of interference areas such as roads, bare land, and spoil heaps on the recognition results.

[0048] The specific implementation process is as follows: Input feature map The implementation process is divided into For each region, a linear projection is used to obtain the query Q, key K, and value V. The calculation formula is as follows: .

[0049] in, , and Let represent the learnable mapping matrices for the query, key, and value, respectively. This process projects the joint features of the input into the attention computation space, providing a foundation for subsequent region selection and fine-grained feature interaction. Then, the top-k most relevant regions are selected based on their relevance scores, denoted as: .

[0050] in, Indicates the relationship with the first The set of most relevant candidate regions for each query region. Indicates the total number of regions. Indicates the first The query range and the first The regional correlation score between the candidate areas of landslide collapse is a scalar value; the Top-k operation represents the regional correlation score between the candidate areas of landslide collapse and the candidate areas of landslide collapse. Select the candidate regions with the highest relevance scores Each region, therefore Includes Each region is indexed. Through this process, the open-pit mine semantic segmentation model can first filter out the context regions most relevant to the current query region at a coarse regional scale, thereby reducing the participation of irrelevant background regions in subsequent attention calculations.

[0051] After completing the region filtering, then aggregate the candidate regions. Fine-grained token-level attention computation is performed internally, and its output can be represented as: .

[0052] in, Indicates the first Bottleneck characteristics of the query region after enhancement by a two-layer routing attention module Indicates the first Token query matrix within a query range and These represent the key and value matrices obtained by concatenating the candidate region sets, respectively. Representing feature dimension, This represents the normalized exponential function, used to convert attention scores into attention weights.

[0053] At the bottleneck layer of the model, the convolutional layer is replaced by a dual-layer routing attention module. This module performs global information modeling and cross-regional feature interaction on the joint features of the multi-scale data extracted by the digital orthophoto (DOM) and slope data. The resulting enhanced bottleneck features are output, thereby effectively improving the overall perception and representation capabilities of the open-pit mine semantic segmentation model for features such as the mining pit range, mining bench strip structure, and continuous boundary of the mining slope in the open-pit mine.

[0054] The decoder fuses the enhanced features output from the bottleneck layer with the shallow high-resolution features (including edge and texture details) of the corresponding layer of the encoder through step-by-step upsampling and skip connections. It also performs step-by-step upsampling operations in conjunction with the skip connection mechanism to make up for the spatial location information lost during downsampling, restore the spatial details of the image, and finally output a multi-channel feature map.

[0055] In addition, the open-pit mine semantic segmentation model can also include an output layer, which maps the multi-channel feature map to a pixel-level probability map of the open-pit mine area through a 1×1 convolutional layer at the end of the decoder. Finally, it performs binarization processing by using a fixed threshold or an adaptive threshold, and outputs a binary segmentation mask (i.e., the open-pit mine semantic segmentation mask) with the same size as the input image. In this mask, a pixel value of 1 represents the open-pit mine area and a pixel value of 0 represents the background area.

[0056] The binary segmentation mask output from the semantic segmentation model of the open-pit mine is regularized, vectorized, matched coordinates, and its morphological parameters are statistically analyzed to transform pixel-level results into engineering outputs. Specifically, the binary segmentation mask is first regularized, including hole filling, morphological closing operations, connected component area threshold filtering, and removal of fine noise. Then, the outer contour is extracted and the boundary is smoothed to convert the raster to vector and obtain the mine area range (i.e., open-pit mine boundary vector data). Further, based on the georeferenced information of the UAV remote sensing image, the open-pit mine boundary vector data is mapped to a standard coordinate reference (selected or set according to the user's actual needs, such as the geodetic coordinate system), resulting in open-pit mine boundary vector data. For each mine patch in the open-pit mine boundary vector data with standard coordinate information, morphological indicators such as area, perimeter, and compactness / shape index are calculated. At the same time, based on the slope data, the average slope, slope distribution characteristics, or local topographic undulation information within the patches are statistically analyzed to form integrated spatial element and attribute table output data.

[0057] In addition, during the quality control process, the semantic segmentation mask of the open-pit mine can be spatially overlaid with the sample annotation or sampling area to calculate evaluation indicators such as mean intersection over union (mIoU), precision, recall, or F1 score; at the same time, low-confidence patches can be automatically removed or marked by combining confidence thresholds and area thresholds, and finally a quality control report is output.

[0058] .

[0059] .

[0060] .

[0061] .

[0062] Wherein, TP represents the number of open-pit mine pixels correctly identified as open-pit mines; TN represents the number of background pixels correctly identified as backgrounds; FP represents the number of background pixels incorrectly identified as open-pit mines; and FN represents the number of open-pit mine pixels incorrectly identified as backgrounds.

[0063] The final engineering deliverables package (i.e., open-pit mine target identification results) includes: the vector boundary of the open-pit mine area, a morphological parameter attribute table (i.e., morphological index values ​​corresponding to UAV remote sensing imagery, including area, perimeter, shape, etc.), and a quality control report. This engineering deliverables package not only characterizes the mine boundary location but can also be integrated into the regulatory business system as the basis for spatial overlay analysis with mining permit boundaries, ecological protection red lines, or other regulatory thematic layers, providing technical support for the continuous updating and dynamic monitoring of open-pit mine spatial data.

[0064] By improving the convolutional layer with a dual-layer routing attention module at the bottleneck layer of the Bi-UNet semantic segmentation network (i.e., the open-pit mine semantic segmentation model), and combining digital orthophotos with slope data derived from the digital elevation model, a complete technology chain was constructed, from the raw data collection of UAVs, intelligent identification of open-pit mines to the output of engineering application results. This improved the accuracy and boundary precision of open-pit mine identification in the context of complex mining areas.

[0065] Based on the above embodiments, this application introduces a dual-layer routing attention module at the bottleneck layer of the U-Net network to perform global information modeling and cross-regional feature interaction on the multi-scale joint features output by the encoder, and outputs enhanced bottleneck features. In the context of complex mining areas, this improves the accuracy and boundary precision of open-pit mine identification, effectively enhances the global information modeling and long-distance dependency capture and representation capabilities of complex pit-step-slope structures in open-pit mine areas, and can output vector results and parameter indicators (i.e., engineering result packages) that can be directly used for regulatory verification and statistical analysis. By simultaneously introducing local window attention and global interaction with cross-window routing information, the shortcomings of traditional convolutional neural networks, which are limited by local receptive fields and have difficulty taking into account the overall spatial structure of the mining area, are effectively alleviated. The resulting mask is more likely to maintain the overall connectivity of the mining target in a large-scale scene and is more robust to weak boundaries under shadow or vegetation cover conditions, thereby reducing geometric errors caused by segmentation and boundary drift. It performs well in terms of open-pit mine identification accuracy, misjudgment area control, and mine boundary segmentation quality, providing a reliable data foundation and technical support for open-pit mine supervision, illegal mining identification, ecological restoration assessment, and safety risk analysis.

[0066] Based on the same inventive concept, this application also provides an open-pit mine target identification device for implementing the above-mentioned open-pit mine target identification method. The solution provided by this device is similar to the implementation scheme described in the above-described method. Therefore, the specific limitations in one or more embodiments of the open-pit mine target identification device provided below can be found in the limitations of the open-pit mine target identification method described above, and will not be repeated here.

[0067] In one exemplary embodiment, such as Figure 4 As shown, an open-pit mine target identification device is provided, comprising: The data acquisition and processing unit is used to acquire UAV remote sensing imagery and LiDAR point cloud data of the area to be identified in the open-pit mine. The data acquisition and processing unit is also used to obtain digital orthophotos based on the UAV remote sensing imagery. The data acquisition and processing unit is also used to obtain slope data based on the LiDAR point cloud data. The data acquisition and processing unit is also used to acquire sample UAV remote sensing imagery and sample LiDAR point cloud data of the open-pit mine. The data acquisition and processing unit is also used to generate sample digital orthophotos based on the sample UAV remote sensing imagery. The data acquisition and processing unit is also used to obtain sample slope data based on the sample LiDAR point cloud data.

[0068] The model building unit, connected to the data acquisition and processing unit, is used to construct an open-pit mine dataset based on sample digital orthophotos and sample slope data. The model building unit is also used to construct an initial semantic segmentation model based on a U-Net network and a two-layer routing attention module. Furthermore, the model building unit is used to train the initial semantic segmentation model based on the open-pit mine dataset, resulting in an open-pit mine semantic segmentation model.

[0069] The identification unit is connected to the data acquisition and processing unit and the model building unit, respectively, and is used to obtain the open-pit mine semantic segmentation mask based on digital orthophotos and slope data using the open-pit mine semantic segmentation model.

[0070] The output unit, connected to the recognition unit, is used to obtain the target recognition result of the open-pit mine based on the semantic segmentation mask of the open-pit mine.

[0071] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data related to open-pit mine target identification methods. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an open-pit mine target identification method.

[0072] Those skilled in the art will understand that Figure 5 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0073] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0074] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0075] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0076] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0077] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for target identification in an open-pit mine, characterized in that, include: Acquire drone remote sensing images and lidar point cloud data of the area to be identified in the open-pit mine; Digital orthophotos are obtained based on the aforementioned UAV remote sensing images; Slope data is obtained based on the lidar point cloud data; Acquire sample UAV remote sensing images and sample lidar point cloud data of open-pit mines; Generate sample digital orthophotos based on the sample UAV remote sensing images; The sample slope data is obtained based on the sample lidar point cloud data. An open-pit mine dataset is constructed based on the sample digital orthophotos and the sample slope data; An initial semantic segmentation model is constructed based on the U-Net network and a two-layer routing attention module; Using the aforementioned open-pit mine dataset, the initial semantic segmentation model is trained to obtain the open-pit mine semantic segmentation model; Using the aforementioned open-pit mine semantic segmentation model, an open-pit mine semantic segmentation mask is obtained based on the digital orthophoto and the slope data; The target recognition result of the open-pit mine is obtained based on the semantic segmentation mask of the open-pit mine.

2. The open-pit mine target identification method according to claim 1, characterized in that, Based on the sample lidar point cloud data, sample slope data is obtained, including: The sample lidar point cloud data is classified and filtered to extract ground point cloud data and construct a sample digital elevation model. The sample slope data is determined based on the sample digital elevation model.

3. The open-pit mine target identification method according to claim 2, characterized in that, An open-pit mine dataset is constructed based on the sample digital orthophotos and the sample slope data, including: Based on the sample digital elevation model, the spatial distribution of the open-pit mine area boundary and typical structural units is calibrated in the sample digital orthophoto image to obtain the annotation results; the typical structural units include the pit area, the stope slope, the mining bench and the stope bottom. Using the sample digital orthophoto and the sample slope data as input, and the annotation results corresponding to the sample digital orthophoto as output, a sample pair is constructed. The open-pit mine dataset is constructed based on the sample pairs.

4. The open-pit mine target identification method according to claim 1, characterized in that, An initial semantic segmentation model is constructed based on the U-Net network and a two-layer routing attention module, including: A dual-layer routing attention module is introduced into the bottleneck layer of the U-Net network to obtain the initial semantic segmentation model.

5. The open-pit mine target identification method according to claim 1, characterized in that, Using the aforementioned open-pit mine dataset, the initial semantic segmentation model is trained to obtain the open-pit mine semantic segmentation model, which includes: A training set and a validation set were constructed based on the open-pit mine dataset. The initial semantic segmentation model is trained based on the training set using the backpropagation algorithm to obtain the trained initial semantic segmentation model. The initial semantic segmentation model after training is validated using a loss function on the validation set until the loss function value on the validation set reaches the set requirement, thus obtaining the open-pit mine semantic segmentation model.

6. The open-pit mine target identification method according to claim 5, characterized in that, The loss function is the cross-entropy loss function, the Dice loss function, or a combined loss function; the combined loss function includes the cross-entropy loss function and the Dice loss function.

7. The open-pit mine target identification method according to claim 1, characterized in that, The target recognition result of the open-pit mine is obtained based on the semantic segmentation mask of the open-pit mine, including: The semantic segmentation mask of the open-pit mine is subjected to regularization and vectorization processing to obtain the boundary vector data of the open-pit mine; Based on the geographic reference information of the UAV remote sensing image, the boundary vector data of the open-pit mine is mapped to a standard coordinate reference to determine the morphological index value corresponding to the UAV remote sensing image. The target identification result of the open-pit mine is obtained based on the boundary vector data of the open-pit mine and the morphological index value corresponding to the UAV remote sensing image.

8. A target identification device for open-pit mines, characterized in that, include: The data acquisition and processing unit is used to acquire UAV remote sensing images and lidar point cloud data of the area to be identified in the open-pit mine; the data acquisition and processing unit is also used to obtain digital orthophotos based on the UAV remote sensing images; the data acquisition and processing unit is also used to obtain slope data based on the lidar point cloud data; the data acquisition and processing unit is also used to acquire sample UAV remote sensing images and sample lidar point cloud data of the open-pit mine; the data acquisition and processing unit is also used to generate sample digital orthophotos based on the sample UAV remote sensing images; the data acquisition and processing unit is also used to obtain sample slope data based on the sample lidar point cloud data. The model building unit, connected to the data acquisition and processing unit, is used to build an open-pit mine dataset based on the sample digital orthophoto and the sample slope data; the model building unit is also used to build an initial semantic segmentation model based on a U-Net network and a two-layer routing attention module; the model building unit is also used to train the initial semantic segmentation model based on the open-pit mine dataset to obtain an open-pit mine semantic segmentation model; The identification unit is connected to the data acquisition and processing unit and the model building unit respectively, and is used to obtain the open-pit mine semantic segmentation mask based on the digital orthophoto and the slope data using the open-pit mine semantic segmentation model. The output unit, connected to the recognition unit, is used to obtain the target recognition result of the open-pit mine based on the semantic segmentation mask of the open-pit mine.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the open-pit mine target identification method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the open-pit mine target identification method according to any one of claims 1-7.

Images