Deep learning mining network construction method based on MINE-Net architecture

Abstract

The invention discloses a deep learning mining network construction method based on a MINE-Net architecture, and the method comprises the steps: obtaining SAR amplitude diagrams and coherence diagrams at different time points, carrying out the preprocessing and enhancement of data, and calculating the change characteristics of the junction of a high-coherence region and a low-coherence region; combining the amplitude and coherence characteristics by using an MINE-Net model, and performing mine network road identification; and a training set is expanded through pseudo tag generation and a user-defined data enhancement technology, so that the generalization ability and the recognition accuracy of the model are improved. The method can effectively reduce the false alarm rate of the low-coherence area, rapidly and accurately recognize the open-pit mine area road, and improve the mine area transportation and planning efficiency.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and remote sensing technology, and in particular to a deep learning mining network construction method based on the MINE-Net architecture. Background Art

[0002] With the continuous expansion and development of open-pit mines, traditional mine management and road identification methods are difficult to meet the needs of efficient and accurate management. The construction and maintenance of mine roads are key links in mining operations, which not only affect the efficiency of resource transportation, but also have an important impact on the ecological environment and operating costs. In traditional mine management, manual observation and experience judgment are mostly relied on, which is not only time-consuming and labor-intensive, but also difficult to accurately capture the dynamic changes of mine roads.

[0003] On the other hand, with the rapid development of remote sensing technology and data science, high-resolution SAR (synthetic aperture radar) images have become an effective tool for monitoring changes in mining roads. SAR images have the advantages of all-weather and cloud penetration, and can provide information on the physical characteristics of the terrain. However, due to the complex terrain of the mining area and the presence of a large number of low-coherence areas (such as vegetation, wetlands, etc.), traditional methods based on SAR amplitude characteristics are prone to false alarms in these areas, making it difficult to accurately identify mining roads.

[0004] Existing technologies have low recognition accuracy in low-coherence areas, and the model's generalization ability between different data sets is also weak. Due to the lack of labeled data and the limited size of training sets, traditional methods perform poorly when dealing with highly heterogeneous data, often resulting in the accuracy and robustness of model predictions failing to meet the actual needs of mining areas. Summary of the invention

[0005] In view of this, the present invention provides a deep learning mining network construction method based on the MINE-Net architecture to solve the problems of low recognition accuracy and insufficient model generalization ability of traditional mining area management methods in low coherence areas.

[0006] The present invention provides a method for constructing a deep learning mining network based on the MINE-Net architecture, the method comprising:

[0007] Acquire high-resolution SAR image data of open-pit mine areas;

[0008] Preprocessing of SAR images, including calculation of coherence features;

[0009] The deep learning model is used to identify the boundary areas between high coherence and low coherence, and the mine network road annotations are generated based on the identification results.

[0010] The present invention also introduces pseudo-label technology, which adds pseudo-label data to the training set by making high-confidence predictions on unlabeled data to expand the data volume and improve the generalization ability of the model. While expanding the scale of the training set, the robustness and generalization ability of the model are improved, so that the dynamic changes of mining area roads can be captured more accurately, providing technical support for mining area management and transportation optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is the MINE-Net mining network road identification flow chart;

[0012] Figure 2 This is a schematic diagram of the training network structure of the mining network road recognition model in the present invention.

[0013] Figure 3 This is the validation set result of the model of the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] This paper proposes a deep learning mining network construction method based on the MINE-Net architecture, which is mainly used for monitoring and optimizing the road network in open-pit mines. By combining synthetic aperture radar (SAR) images and deep learning methods, it is possible to accurately identify changes in mining roads and provide efficient road planning and management solutions. Unlike traditional optical sensors, SAR images have the ability to penetrate clouds and can obtain more accurate terrain information in complex environments.

[0016] The core technical solution of the present invention is based on the MINE-Net architecture, which processes SAR images through a deep learning model to identify changes in roads within the mining area.

[0017] The combination of amplitude and coherence features can effectively distinguish actual road changes from pseudo-changes (such as equipment movement, humidity fluctuations, etc.), significantly improving the accuracy and robustness of mine road identification.

[0018] This paper introduces data enhancement and pseudo-label generation technology:

[0019] The present invention enhances the adaptability of the model under different landforms and terrain types by randomly scaling the low coherence area;

[0020] At the same time, the present invention further improves the recognition ability and accuracy of the model by expanding the training set through the use of unlabeled data through pseudo-labeling technology.

[0021] In order to make the technical solution and operation process of the present invention clearer, the training and application of the MINE-Net model are described in detail below with reference to a specific flowchart.

[0022] Figure 1 The overall workflow of the deep learning mining network construction method based on the MINE-Net architecture is demonstrated.

[0023] Starting from the left, the input part includes two SAR amplitude images x at different locations. position1 and x position2 These are the basic data for model training and recognition:

[0024] In view of the limitations of optical sensors, the present invention chooses to use synthetic aperture radar (SAR) data to enhance the model training, verification and testing of mine network roads.

[0025] SAR's high-resolution imagery plays a fundamental role in accurately detecting and monitoring changes in mining roads. It is particularly useful in mining scenarios because it can discern details and minor changes.

[0026] The present invention uses high-resolution SAR images acquired by the X-band natural resources satellite remote sensing platform. The indicators are as follows:

[0027] The present invention selects an antenna length of 4.8 meters, an antenna width of 0.7 meters, and a bandwidth range of 150MHz / 300MHz;

[0028] The present invention selects the SAR imaging mode as StripMap;

[0029] The present invention selects a resolution of 3 meters and a scene size of 30 kilometers × 50 kilometers.

[0030] The natural resources satellite remote sensing platform is particularly suitable for areas with steep and dynamic terrain due to its narrower orbit tube, which can alleviate the geometric decorrelation effect.

[0031] To create the dataset, the present invention considers two different open pit mines, namely Site A and Site B. Site A uses three sets of acquisition pairs, and Site B uses two sets.

[0032] Each group represents two SAR acquisitions at the same location at different times. The images were registered and the flat and terrain phase components were compensated. All image pairs were selected and divided into 256×256 tiles.

[0033] Data collection at location A was conducted during the ascending trajectory from January 4 to May 5, 2022.

[0034] Data collection at Site B takes place during the descent trajectory from August 10 to September 1, 2023.

[0035] Data annotation and use: In mine network road identification, it is necessary to clearly define the types of changes of interest and ensure consistency in annotation.

[0036] The present invention focuses on marking changes related to road status and layout, such as:

[0037] The opening of new roads, the widening of existing roads or the closure of abandoned roads.

[0038] The present invention will not mark pseudo-change areas (such as short-term parking of equipment, humidity fluctuations, etc.) as road changes.

[0039] During the annotation process, SAR amplitude maps and coherence measurements are used by experts to identify road changes.

[0040] The present invention also reveals the change pattern of the mining area roads by comparing the radar images of different regional points and the coherence measurement, wherein the intersection of high coherence value and low coherence value change often represents the existence of roads.

[0041] The input data first enters a data preprocessing module, which is responsible for converting the raw SAR image into a format suitable for processing by the deep learning model.

[0042] Furthermore, necessary adjustments are made in combination with physical parameters to optimize the subsequent training and recognition processes.

[0043] Definition Road Recognition: In this paper, road recognition is defined as a binary segmentation task that aims to detect road boundaries through changes in coherence in SAR images.

[0044] The binary map indicates which locations on the ground are roads (value 1) and which are not (value 0). The key idea is to use the coherence information to identify changes at the junction of high and low coherence values, which usually indicate the presence of road structures.

[0045] The concept of coherence calculation: Coherence is an indicator that measures the similarity of radar signals at two different times. and They represent two SAR complex images at different positions and times, and their amplitudes are x position1 and x position2 , the coherence ρ is defined as the complex correlation coefficient between two images.

[0046] Areas of high coherence reflect stability of terrain or structural features, while low coherence values ​​often indicate the presence of dynamic changes, such as vegetation movement or equipment movement.

[0047] Operation of coherence calculation: In the first step of road recognition, the coherence calculation algorithm is applied to process the SAR data, which includes preprocessing the image and then calculating the coherence between the two images.

[0048] Furthermore, the obtained coherence image is used to cross-validate with the existing mining network coherence annotation results to train and verify the road recognition model.

[0049] Target area for road identification: In order to accurately identify roads, special attention is paid to areas where the coherence value changes significantly, especially the junction of high and low coherence values. Roads usually form connected paths or patterns in these areas.

[0050] After data preprocessing is completed, the data is sent to the trained deep learning model. This part extracts and learns the data features through the MINE-Net architecture, outputs the results of model training, and provides core computing power for the identification and analysis of mine network roads.

[0051] The amplitude and coherence information of the SAR image is taken as input. The input is defined as a three-channel tensor:

[0052] The model output is a binary probability map γ φ，θ (x)∈[0,1] H×W , the bivariate probability map is calculated by the given formula.

[0053] Figure 2 The network structure used to train the mine network road recognition model is shown:

[0054] The model input is three-channel image data, which are SAR amplitude map and coherence map at two different locations.

[0055] The encoder part of the network is responsible for extracting useful features from the input image and compressing it. The encoder consists of four convolutional blocks.

[0056] Each convolutional block contains a convolutional layer, a batch normalization layer, and an activation function to gradually extract multi-level features.

[0057] The decoder part gradually maps the potential representation back to the spatial resolution of the original image through four convolutional blocks symmetrical to the encoder. The decoder also combines the high-resolution features in the encoder and fuses the output of each layer in the encoder with the corresponding layer of the decoder through skip connections, thereby enhancing the detection ability of road boundaries.

[0058] Finally, the decoder outputs a binary probability map to represent the road detection results. The higher the probability value, the more likely the pixel belongs to the road area.

[0059] The edge coordinates of the middle color blocks in the binary probability map with a probability greater than 50% are selected, and all the edge coordinates that meet the conditions in the binary probability map are connected with line segments to obtain the mine network road recognition image.

[0060] Training stability and overfitting issues: Due to the small dataset, problems related to training stability and overfitting were encountered during model training.

[0061] Introducing data augmentation techniques: To address these issues, we introduced a custom data augmentation method — randomly scaling low-coherence regions:

[0062] Custom data augmentation method: Randomly scale the coherence value where the coherence value is below a certain threshold and does not belong to the road area:

[0063] By randomly scaling the local coherence values ​​that do not belong to the road area, the coherence values ​​there can be significantly homogenized to prevent model recognition errors.

[0064] Introduction of pseudo-labeling technology: Due to limited data annotation, we use pseudo-labeling technology to expand the training data and further improve model performance.

[0065] Pseudo-label generation method: The pseudo-label method uses the output of the network to generate additional training labels for unlabeled data, which can be divided into online and offline methods.

[0066] Offline pseudo-label generation: We chose an offline method, which generates pseudo labels once before training and uses them fixedly, which helps to optimize the model more stably in each iteration.

[0067] Selection of high-confidence pseudo-labels: We select those patches from the unlabeled data where at least 90% of the pixels belong to the same category as high-confidence pseudo-labels.

[0068] Expanded training set: Pseudo labels with high confidence are added to the expanded training set, but their number does not exceed the size of the original training set to maintain data balance, such as Figure 1 shown.

[0069] Training set definition and use: Through this method, the expanded training set enables the mine network road recognition model to utilize more unlabeled data in training, thereby improving the model's recognition accuracy and generalization ability for mine area roads.

[0070] Model retraining: After some pseudo-label tiles are selected, they are merged with the annotated dataset, and then the model is retrained using the merged dataset.

[0071] Verification data set detection: By using data enhancement and pseudo-label generation technology, the accuracy of mine network road recognition is significantly improved.

[0072] Figure 3 The recognition effect of the final model of the present invention is demonstrated, and the mine network roads can be clearly identified and marked.

[0073] In order to evaluate the effect of mine network road detection, the present invention uses the mine network recognition performance index, the core purpose of which is to comprehensively evaluate the accuracy, completeness and boundary consistency of the detection model in the mine network road recognition task.

[0074] The mine network identification performance index is defined as follows:

[0075] MNRPI=α·F1+β·IoU+γ·BCE

[0076] F1: F1 score, which is used to evaluate the balance between precision and recall of road detection.

[0077]

[0078] BCE (Boundary Consistency Error): Boundary consistency error measures the consistency difference between the road boundary predicted by the model and the actual road boundary.

[0079]

[0080] IoU (Intersection over Union): Intersection over Union ratio, used to measure the degree of overlap between the predicted road area and the actual road area.

[0081]

[0082] BCE (Boundary Consistency Error): Boundary consistency error measures the consistency difference between the road boundary predicted by the model and the actual road boundary.

[0083]

[0084] Among them, Boundary perd and Boundary true Represent the predicted and real road boundary pixel sets respectively. Parameters α, β, and γ are weight factors, satisfying α+β+γ=1. These weights can be adjusted according to actual needs.

[0085] Evaluation score comparison

[0086] The present invention uses the above indicators to evaluate the deep learning mining network construction method of MINE-Net architecture, the traditional optical solution and the traditional SAR data processing method. We obtained the following results:

[0087]

[0088] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a deep learning mining network based on the MINE-Net architecture, characterized in that: The method comprises: Acquire high-resolution SAR image data of the open-pit mine area and preprocess the SAR images, including calculating coherence features. Use a deep learning model to identify the boundary area between high coherence and low coherence, generate mine network road annotations based on the identification results, introduce pseudo-labeling technology to make high-confidence predictions on unlabeled data, and add pseudo-labeled data to the training set to expand the data volume and improve the model generalization ability.

2. The method according to claim 1, characterized in that The SAR image preprocessing includes calculating the coherence characteristics of the image to eliminate the influence of terrain changes on the road recognition result. Coherence boundary features are introduced, and the changes at the junction of high coherence and low coherence are used for mine network road detection. This method can accurately capture the boundary features of roads and avoid misidentifying low coherence areas as roads. The SAR amplitude map and coherence map at different locations are used as the input of the model. The coherence and amplitude information are combined through the MINE-Net model. The model's sensitivity to changes at the junction of high coherence and low coherence is used to improve the accuracy of road recognition.

3. The method according to claim 1, characterized in that The pseudo-labeling technology generates high-confidence pseudo-labels offline and selects patches whose pixels at least 90% belong to the same category from unlabeled data as high-confidence pseudo-labels. The training set is expanded using unlabeled data, and the high-coherence region enhancement strategy is combined to improve the generalization ability of the model. After the initial model training is completed, the unlabeled data is predicted, and high-confidence road prediction tiles are selected as pseudo-labels to expand the training data. Custom data enhancement: randomly scale the coherence values ​​of low-coherence regions to enhance the model's adaptability to different terrain types.

4. The method according to claim 1, characterized in that: The MINE-Net architecture includes: The encoder part is used to extract features from the input SAR image, and the decoder part maps the features back to the original image resolution through a structure symmetrical to the encoder and combines jump connections to enhance the features.

5. The method according to claim 1, characterized in that The model outputs a binary probability map to represent the road detection results. The higher the probability value, the more likely the pixel belongs to the road area.

6. The method according to claim 1, characterized in that The encoder part and the decoder part each include four convolution blocks, and each convolution block includes a convolution layer, a batch normalization layer and an activation function.

7. The method according to any one of claims 1 to 2, characterized in that The mine network road recognition model performs data enhancement by randomly scaling the coherence value in the low coherence area to enhance the robustness of the model.

8. The method according to claim 1, characterized in that: The mine network road recognition image generated by the method is obtained by performing connectivity analysis on the areas with a probability greater than 50% in the binary probability map and generating a connectivity path.

9. The method according to claim 5, characterized in that The deep learning module outputs a binary probability map of the mine network roads and generates a final road recognition result based on the probability map.

10. The method according to claim 3, characterized in that The MINE-Net architecture mining network recognition model is significantly better than traditional solutions in terms of accuracy and recall rate, especially in detecting high-confidence areas. The model of the present invention also has obvious advantages in the degree of overlap between the predicted area and the real area, and the boundary consistency error is significantly lower than other methods, indicating that it can capture the boundaries of the road more accurately.

Images