Mining area land intelligent identification method and system based on mimicry self-learning
Through the mimic self-learning method, the multi-scale feature information and data enhancement technology are integrated, and the problem of time-consuming labels and insufficient multi-scale information extraction in land identification in mining areas is solved, and efficient and accurate land identification and resource monitoring are achieved.
Patent Information
- Application Number
- CN202411271072.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2025-01-03
AI Technical Summary
The prior art has problems such as time-consuming and labor-intensive production of label images, insufficient multi-scale information extraction, and poor model robustness in land use in mining areas, resulting in low recognition accuracy.
Adopting the intelligent identification method of land for mining areas based on mimicry self-learning, by constructing an intensive feature fusion segmentation network and an intelligent identification network model based on PscRecNet, using a small amount of tag data to integrate multi-scale feature information, and improving through data augmentation and pseudo-label generation Model learning efficiency.
It realizes intelligent and accurate identification of land in mining areas, improves the adaptability and generalization capabilities of the model, reduces manual labeling time and resource consumption, and improves recognition accuracy and robustness.
Smart Images
Figure CN119251672A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mining land utilization, and in particular to a mining land intelligent identification method and system based on mimetic self-learning. Background Art
[0002] While the development and utilization of mineral resources have driven the development of the national economy, they have also inevitably caused a series of serious ecological and environmental problems. In particular, in the process of open-pit mining, the mining area and surrounding land resources are susceptible to the direct impact of land occupation and damage caused by mining, as well as the indirect impact of land degradation caused by the deterioration of the ecological environment. Therefore, achieving efficient and accurate identification of mining land is an urgent problem to be solved, which is crucial for mastering mining activities, preventing environmental and geological risks, and achieving sustainable green development in mining areas.
[0003] With the rapid development of remote sensing technology, the spatial resolution of remote sensing images is getting higher and higher. High-resolution remote sensing images can provide clearer details and texture features of mining land, and are increasingly widely used in the extraction and identification of mining land feature information. Although the traditional visual interpretation method has a high accuracy rate for land identification, it requires a lot of manpower and material resources and has low work efficiency. In recent years, research on mining land identification methods combining deep learning and remote sensing technology has made certain progress. After pre-processing, high-resolution images are artificially produced to produce original images and label images of sample data sets, so as to train the constructed deep learning network, extract feature information, and realize the identification of mining land. However, the current methods of mining land identification based on deep learning methods still have the following problems that need to be solved:
[0004] (1) The training of deep learning networks requires labeled images, and the production of labeled images usually takes a lot of time and human resources. At the same time, due to the complex environment of mining areas, there is often a problem of difficulty in labeling.
[0005] (2) Existing methods do not take into account the multi-scale information contained in the image and are unable to extract feature information at different scales in layers. In addition, there is the problem of the diversity of features of the same type of objects and the similarity of information features between different types of objects, which leads to poor land use identification accuracy.
[0006] (3) The traditional deep learning network structure cannot effectively integrate the multi-scale feature information extracted from high-resolution images, and the model robustness is poor. Summary of the invention
[0007] In view of this, the present invention proposes a method and system for intelligent identification of mining land based on mimetic self-learning, which can extract, organize and integrate the characteristic information of mining land in layers, and use a mining land intelligent identification network constructed by training a small amount of labeled data to realize intelligent and accurate identification of mining land, which is beneficial to the monitoring and protection of land resources in mining areas.
[0008] Specifically, the purpose of the present invention is achieved through the following technical solutions:
[0009] A first aspect of the present invention provides a method for intelligently identifying mining land based on mimetic self-learning, the method comprising:
[0010] Obtain and preprocess remote sensing images of the mining area, mark the mining area land, and rasterize and crop it into image blocks as labeled samples. The labeled samples include training sets, validation sets, and test sets.
[0011] Construct a dense feature fusion segmentation network to extract features from labeled samples to obtain a mining area land feature map;
[0012] Construct an intelligent recognition network model based on mimetic self-learning, input labeled and unlabeled samples of land use into the intelligent recognition network model for model training, and obtain the trained intelligent recognition network model;
[0013] The image of the mining area to be identified is input into the trained intelligent recognition network model to obtain the recognition result of the mining area.
[0014] It should be noted that the unlabeled samples and the labeled samples are image blocks with the same pixel size obtained by cropping.
[0015] According to one embodiment of the present invention, a method for constructing a dense feature fusion segmentation network includes:
[0016] Design DenseUNet based on the U-Net architecture, integrate the dense connection characteristics of DenseNet, and build a preliminary network model;
[0017] Dense Blocks are added to the encoding and decoding stages of the preliminary network model, and multi-layer convolution and direct connection techniques are used to extract deep and surface features of labeled samples;
[0018] In the decoding stage, the deep feature map is upsampled and fused with the feature map of the corresponding encoding stage to restore the spatial information and strengthen the integration of features;
[0019] A skip connection strategy is used to directly transfer features between different layers in the network;
[0020] Use labeled samples to train the network and apply the cross entropy loss function to optimize the classification accuracy;
[0021] The model is evaluated using the validation set, the network parameters are adjusted, and the model effect is evaluated using performance indicators to obtain a dense feature fusion segmentation network.
[0022] According to one embodiment of the present invention, a method for constructing an intelligent recognition network model based on mimetic self-learning includes:
[0023] Construct the student model and teacher model framework based on the PscRecNet network model;
[0024] Perform data augmentation on unlabeled samples and generate pseudo labels;
[0025] The student model and the teacher model are subjected to multiple iterations of interactive training and model verification to obtain the final intelligent recognition network model.
[0026] According to one embodiment of the present invention, a method for performing data enhancement on unlabeled samples and generating pseudo labels includes:
[0027] Weakly enhanced samples are obtained by performing data enhancement on unlabeled samples through weak enhancement;
[0028] Generate reliable pseudo labels using weakly augmented samples through the teacher model;
[0029] By performing data enhancement on unlabeled samples, strong enhancement samples are obtained;
[0030] Generate unreliable pseudo labels using strongly augmented samples via the student model.
[0031] According to one embodiment of the present invention, a method for performing multiple iterations of interactive training on a student model and a teacher model includes:
[0032] The student model is trained using labeled samples and reliable pseudo labels;
[0033] After each training cycle, the parameters of the teacher model are updated by taking an exponential moving average of the student model parameters;
[0034] Multiple iterative updates generate new reliable pseudo-labels and unreliable pseudo-labels.
[0035] According to one embodiment of the present invention, a method for performing model verification on a student model and a teacher model includes:
[0036] Regularly use the validation set to test the intelligent recognition network model, and use performance indicators to evaluate the performance of the intelligent recognition network model;
[0037] Adjust training strategies and model architecture based on feedback results;
[0038] After training is complete, the final accuracy of the student model is calculated using an independent test set, which is calculated as:
[0039] L=L S +λL u ;
[0040] Among them, L is the overall loss function of the model, reflecting the performance of the mining land identification model; L s It is the loss function for labeled sample recognition. By comparing the difference between the model's prediction of labeled samples and the true label, it guides the model to learn the correct recognition of labeled data. u It is the loss function for unlabeled sample recognition. By comparing the difference between the model's prediction of unlabeled samples and the pseudo-label, the model is guided to learn the recognition of unlabeled data. λ is used to adjust the weight of the unsupervised segmentation loss in the overall loss.
[0041]
[0042] Where N is the number of labeled samples, N u is the number of unsupervised samples, H and W are the height and width of the segmentation mask, C is the number of categories, and m ijc is the probability value of the true segmentation label, is the corresponding model prediction value, is the predicted value of the pseudo label, β is the weight of the consistency loss, It is the model's predicted probability of the cth category in the jth row and kth column of the same sample. Consistency Loss is the consistency loss, which is used to minimize the difference between the model's predictions at different positions of the same sample.
[0043] According to one embodiment of the present invention, a method for marking mining land and rasterizing and clipping it into image blocks includes:
[0044] Identify and mark the boundaries of different land use types in remote sensing images of mining areas to obtain vector data of mining land;
[0045] Rasterizing the vector data obtained by annotation to obtain label raster data;
[0046] The original remote sensing image data and label raster data are cropped to obtain image blocks with the same pixel size and resolution.
[0047] According to one embodiment of the present invention, the land use types include open pits, spoil dumps, coal piles, mine buildings, trees, shrubs, high / medium / low coverage grasslands, mixed grass, shrubs and trees, water bodies, roads and bare land.
[0048] According to one embodiment of the present invention, the ratio of the training set, the validation set and the test set is 6:4:1.
[0049] A second aspect of the present invention provides a mining land intelligent identification system based on mimetic self-learning, comprising:
[0050] Sample acquisition module: obtain remote sensing images of mining areas and preprocess them, mark mining areas, and rasterize and crop them into image blocks as labeled samples. Labeled samples include training sets, validation sets, and test sets.
[0051] Feature extraction module: construct a dense feature fusion segmentation network to extract features from labeled samples to obtain a mining area land feature map;
[0052] Model construction and training module: construct an intelligent recognition network model based on mimetic self-learning, input labeled and unlabeled samples of land use into the intelligent recognition network model for model training, and obtain the trained intelligent recognition network model;
[0053] Mining land identification module: The mining area image to be identified is input into the trained intelligent recognition network model to obtain the identification result of the mining land.
[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0055] (1) The present invention designs a multi-level differentiated intensive feature extraction module. The mining area contains various land types, such as stopes, spoil dumps, vegetation, water bodies, etc. The features of these land types at different levels are different, and traditional methods are difficult to extract; using a multi-level differentiated intensive feature extraction module to process feature information at different levels can more accurately extract multi-level features, thereby enhancing the ability to represent complex data, helping to improve the model's understanding and expression of abstract features at different levels, making the model more adaptable and generalizable.
[0056] (2) The present invention constructs a dense U-shaped feature fusion segmentation network. Different land use types have different scales in mining areas, and the integration of global and local information is crucial for the classification of mining land. The traditional network structure cannot effectively integrate this information, while the dense U-shaped feature fusion segmentation network can simultaneously focus on features of different scales, effectively integrate global and local information, help improve the model's comprehensive understanding of input data, enhance the comprehensive grasp of information at different scales, enable the model to better adapt to various complex scenarios and changes, and improve the robustness of the model.
[0057] (3) The present invention has built an intelligent identification framework for mining land based on mimetic self-learning. Obtaining label data for mining land usually requires a lot of time and human resources, and there are also problems of difficult labeling and low learning efficiency. The intelligent identification framework for mining land guides the student network through the teacher network to make the student network converge more quickly, effectively improving the learning efficiency of the model, so that the model can adapt to new tasks more quickly; and by effectively integrating labeled and unlabeled data, the framework effectively reduces the time and resources required for manual labeling, making the model more feasible in practical applications, especially in the scenario of large-scale data sets, providing obvious advantages for improving work efficiency and reducing costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 A flow chart of a method for intelligent identification of mining land based on mimetic self-learning provided by an embodiment of the present invention;
[0059] Figure 2 A schematic diagram of a network for intelligent identification of mining land based on mimetic self-learning provided in an embodiment of the present invention.
[0060] Figure 3 A schematic diagram of a dense feature fusion segmentation network provided by an embodiment of the present invention;
[0061] Figure 4 A schematic diagram of the DenseU1 Block-DenseU6 Block structure provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0062] The present disclosure will now be discussed with reference to several exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and thus implement the present disclosure, rather than implying any limitation on the scope of the present disclosure.
[0063] Example
[0064] See also Figure 1-4 As shown, a method and system for intelligent identification of mining land based on mimetic self-learning is provided in an embodiment of the present application, wherein, Figure 1 As shown, a method for intelligent identification of mining land based on mimetic self-learning in this application includes:
[0065] Obtain and preprocess remote sensing images of the mining area, mark the mining area land, and rasterize and crop it into image blocks as labeled samples. The labeled samples include training sets, validation sets, and test sets.
[0066] Construct a dense feature fusion segmentation network to extract features from labeled samples to obtain a mining area land feature map;
[0067] Construct an intelligent recognition network model based on mimetic self-learning, input labeled and unlabeled samples of land use into the intelligent recognition network model for model training, and obtain the trained intelligent recognition network model;
[0068] The image of the mining area to be identified is input into the trained intelligent recognition network model to obtain the recognition result of the mining area.
[0069] It should be noted that the unlabeled samples and the labeled samples are image blocks with the same pixel size obtained by cropping.
[0070] Specifically, the method of the present invention is implemented in the following manner during practical application:
[0071] like Figure 2 As shown in the figure, high-resolution remote sensing images of the mining area are obtained and data preprocessed. The geographic information software ArcGIS is used to annotate the high-resolution image data of some mining areas, that is, to outline the boundaries of different land use types in the image. The annotated objects include open-pit mines, spoil dumps, coal piles, mine buildings, trees, shrubs, high / medium / low coverage grasslands, mixed grass, shrubs and trees, water bodies, roads and bare land, and the vector data of the mining area land is obtained. ArcGIS is used to rasterize the annotated vector data to obtain labeled raster data. Finally, the original remote sensing image data and the labeled raster data are cropped to obtain image blocks with the same resolution and pixel size of 512×512.
[0072] Preferably, the cropped remote sensing image blocks and their corresponding labeled raster data blocks are randomly divided into training set, validation set and test set in a ratio of 6:4:1. The training set and validation set are used to train the model, and the test set is used to test the accuracy and generalization ability of the final model. Most of the unlabeled high-resolution image data of the mining area are also cropped to obtain image blocks with the same resolution and pixel size of 512×512, which are used for self-learning of the model.
[0073] Construct a dense feature fusion segmentation network. In this embodiment, a dense U-type feature fusion segmentation network DenseUNet is used, wherein DenseUNet is composed of 5 differentiated DenseU Blocks. The feature maps generated after the input image passes through each DenseU Block are upsampled and fused with the feature maps generated by the previous layer of DenseU Block, and then the fused feature maps of each level are obtained. The 5 layers of feature maps are all upsampled to the size of the first layer feature map and channel stacking is performed. Finally, a 3×3 convolution operation is performed to obtain a mining area land recognition feature map consistent with the input data size dimension.
[0074] In some embodiments, Figure 3 , Figure 4As shown in the figure, the DenseU1 Block receives the input data (512×512×3) and obtains a 512×512×8 feature map C0 through a conv(3x3) convolution operation. The feature map is subjected to five consecutive conv(3x3) convolutions and downsampling operations in the encoding stage to obtain feature maps of different depths C1(256×256×16), C2(128×128×32), C3(64×64×64), C4(32×32×128), C5(16×16×256), C6(8×8×512), C7(8 ×8×512) is convolved with conv(3x3) to obtain L0, L0 is convolved with conv(3x3) and fused with C6 to obtain P6(8×8×512), and then five consecutive upsampling and feature fusion operations are performed in the decoding stage to obtain P5(16×16×256), P4(32×32×128), P3(64×64×64), P2(128×128×32), and P1(256×256×16). P1 is upsampled and fused with C0 features (512×512×3) and input into the subsequent feature extraction operation. It should be noted that the structures of DenseU2 Block, DenseU3 Block, DenseU4 Block, DenseU5 Block, and DenseU6 Block are similar to those of DenseU1 Block, but DenseU6 Block and DenseU1 Block replace the downsampling operation with a 3×3 hole convolution with a dilation of 2, and replace the upsampling operation with a connection operation.
[0075] An intelligent recognition network model PscRecNet based on mimetic self-learning is constructed. PscRecNet inputs labeled mining land data into the student model for supervised training, and then performs exponential moving average (EMA) on the student network model after preliminary training to obtain the teacher model. The structures of the teacher model and the student model are exactly the same.
[0076] In some embodiments, unlabeled mining land data (512×512×3) are weakly enhanced WeakAug and strongly enhanced Strong Aug respectively; specifically, weak enhancement includes horizontal flipping, random resizing, etc., and strong enhancement includes occlusion, cutting, rotation, brightness jitter, etc.; the weakly enhanced data is input into the teacher model to obtain a reliable pseudo label (ReliablePseudoLabel), and the strongly enhanced data is input into the student model to obtain an unreliable pseudo label (UncertainPseudoLabel). The reliable pseudo labels are used to guide the training of the student model and the unlabeled data, so that the unlabeled data can be more fully utilized in the training of the student model, thereby improving the utilization efficiency of the data and reducing the workload of sample labeling.
[0077] S4. Input the labeled training set and unlabeled data set of land use into the PscRecNet network model for model training. In this embodiment, the Pingzhuang West Open-pit Mine is taken as an example. The experimental parameter settings of the mining area land intelligent identification network PscRecNet are shown in Table 1, and the experimental server performance is shown in Table 2.
[0078] Table 1 Network parameter settings
[0079]
[0080] Table 2 Server configuration
[0081]
[0082] In some embodiments, the PscRecNet network model uses the validation set to verify the model accuracy of the student model after training, uses the test set to identify the mining land, and calculates the final accuracy, and the calculation method is:
[0083] L=L S +λL u ;
[0084] Among them, L is the overall loss function of the model, reflecting the performance of the mining land identification model; L s It is the loss function for labeled sample recognition. By comparing the difference between the model's prediction of labeled samples and the true label, it guides the model to learn the correct recognition of labeled data. u It is the loss function for unlabeled sample recognition. By comparing the difference between the model's prediction of unlabeled samples and the pseudo-label, the model is guided to learn the recognition of unlabeled data. λ is used to adjust the weight of the unsupervised segmentation loss in the overall loss.
[0085]
[0086] Where N is the number of labeled samples, N u is the number of unsupervised samples, H and W are the height and width of the segmentation mask, C is the number of categories, and m ijc is the probability value of the true segmentation label, is the corresponding model prediction value, is the predicted value of the pseudo label, β is the weight of the consistency loss, It is the predicted probability of the cth category of the model at different positions (jth row and kth column) of the same sample. Consistency Loss is the consistency loss, which is used to minimize the difference between the predictions of the model at different positions of the same sample.
[0087] The images of the training set and the validation set are used to train the PscRecNet network model. The present invention inputs the test set of the mining land image sample set into the PscRecNet network model to obtain the recognition result of the mining land; the accuracy of the recognition result is evaluated, and the specific evaluation indicators include: confusion matrix; precision P (Precision): the ratio of the number of a certain land type that can be correctly identified to the actual number predicted to be the land type; recall rate R (Recall): the ratio of the number of a certain land type that can be correctly identified to the actual number of the land type; F1 score (F1-score): the harmonic mean of precision and recall, which is an indicator for evaluating the performance of the mining land intelligent recognition model.
[0088] The present invention also provides a mining land intelligent identification system based on mimetic self-learning, comprising:
[0089] Sample acquisition module: obtain remote sensing images of mining areas and preprocess them, mark mining areas, and rasterize and crop them into image blocks as labeled samples. Labeled samples include training sets, validation sets, and test sets.
[0090] Feature extraction module: construct a dense feature fusion segmentation network to extract features from labeled samples to obtain a mining area land feature map;
[0091] Model construction and training module: construct an intelligent recognition network model based on mimetic self-learning, input labeled and unlabeled samples of land use into the intelligent recognition network model for model training, and obtain the trained intelligent recognition network model;
[0092] Mining land identification module: The mining area image to be identified is input into the trained intelligent recognition network model to obtain the identification result of the mining land.
[0093] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0094] (1) By constructing a dense feature fusion segmentation network (DenseUNet), the scheme can effectively extract deep and surface features from labeled samples, ensuring high-precision identification of mining land types. Dense fusion of features helps restore and enhance the spatial information of images, which is particularly important for remote sensing images of complex terrain and variable landforms.
[0095] (2) Using a pseudo-self-learning framework (PscRecNet), we increase the diversity and quantity of training data by using unlabeled samples through data augmentation and pseudo-label generation, which not only expands the training set but also improves the generalization ability of the model in practical applications. In this way, the model can better adapt to new or unlabeled data and reduce the risk of overfitting.
[0096] (3) Through interactive training of the student model and the teacher model, the system can dynamically learn from new data and update model parameters in real time. The exponential moving average (EMA) strategy of the teacher model ensures the stability and reliability of model updates, which helps to continuously learn and adapt to environmental changes.
[0097] (4) By performing the same processing and cropping on labeled and unlabeled samples, the format of input data is unified, the data preprocessing process is simplified, and it helps to optimize the use of computing resources and improve data processing speed.
[0098] (5) The structured management of training, validation, and test sets (in a ratio of 6:4:1) allows for comprehensive training and rigorous evaluation of the model. This approach ensures that each model update can be verified by performance indicators, allowing for real-time adjustments to training strategies and model parameters.
[0099] Those of ordinary skill in the art will appreciate that the modules and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0100] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and equipment can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0101] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.
[0102] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.
[0103] In addition, each functional module in the embodiment of the present invention may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.
[0104] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the energy-saving signal sending / receiving method of each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.
[0105] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the present application.
[0106] It should be understood that the size of the sequence number of each step in the content of the invention and the embodiments of the present invention does not absolutely mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention. For the purpose of example and description, the foregoing description of the implementation of the present disclosure has been given. The foregoing description is not exhaustive and is not intended to limit the present disclosure to the exact form disclosed. Various deformations and modifications may exist according to the above teachings, or various deformations and modifications may be obtained from the practice of the present disclosure. These embodiments are selected and described to illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can use the present disclosure in various embodiments and various modifications suitable for the specific purpose conceived.
Claims
1. A method for intelligent identification of mining land based on mimetic self-learning, characterized in that: The methods include: Obtain and preprocess remote sensing images of the mining area, mark the mining area land, and rasterize and crop it into image blocks as labeled samples. The labeled samples include training sets, validation sets, and test sets. Construct a dense feature fusion segmentation network to extract features from labeled samples to obtain a mining area land feature map; Construct an intelligent recognition network model based on mimetic self-learning, input labeled and unlabeled samples of land use into the intelligent recognition network model for model training, and obtain the trained intelligent recognition network model; The image of the mining area to be identified is input into the trained intelligent recognition network model to obtain the recognition result of the mining area.
2. The method for intelligent identification of mining land based on mimetic self-learning according to claim 1, characterized in that: Methods for building dense feature fusion segmentation networks include: Design DenseUNet based on the U-Net architecture, integrate the dense connection characteristics of DenseNet, and build a preliminary network model; Dense Blocks are added to the encoding and decoding stages of the preliminary network model, and multi-layer convolution and direct connection techniques are used to extract deep and surface features of labeled samples; In the decoding stage, the deep feature map is upsampled and fused with the feature map of the corresponding encoding stage to restore the spatial information and strengthen the integration of features; A skip connection strategy is used to directly transfer features between different layers in the network; Use labeled samples to train the network and apply the cross entropy loss function to optimize the classification accuracy; The model is evaluated using the validation set, the network parameters are adjusted, and the model effect is evaluated using performance indicators to obtain a dense feature fusion segmentation network.
3. The method for intelligent identification of mining land based on mimetic self-learning according to claim 1, characterized in that: The method of constructing an intelligent recognition network model based on mimetic self-learning includes: Construct the student model and teacher model framework based on the PscRecNet network model; Perform data augmentation on unlabeled samples and generate pseudo labels; The student model and the teacher model are subjected to multiple iterations of interactive training and model verification to obtain the final intelligent recognition network model.
4. The method for intelligent identification of mining land based on mimetic self-learning according to claim 3 is characterized in that: Methods for data augmentation and pseudo-label generation for unlabeled samples include: Weakly enhanced samples are obtained by performing data enhancement on unlabeled samples through weak enhancement; Generate reliable pseudo labels using weakly augmented samples through the teacher model; By performing data enhancement on unlabeled samples, strong enhancement samples are obtained; Generate unreliable pseudo labels using strongly augmented samples via the student model.
5. The method for intelligent identification of mining land based on mimetic self-learning according to claim 4, characterized in that: Methods for interactively training the student model and the teacher model for multiple iterations include: The student model is trained using labeled samples and reliable pseudo labels; After each training cycle, the parameters of the teacher model are updated by taking an exponential moving average of the student model parameters; Multiple iterative updates generate new reliable pseudo-labels and unreliable pseudo-labels.
6. The method for intelligent identification of mining land based on mimetic self-learning according to claim 3, characterized in that: Methods for model validation of student models and teacher models include: Regularly use the validation set to test the intelligent recognition network model, and use performance indicators to evaluate the performance of the intelligent recognition network model; Adjust training strategies and model architecture based on feedback results; After training is complete, the final accuracy of the student model is calculated using an independent test set, which is calculated as: L=L S +λL u ; Among them, L is the overall loss function of the model, reflecting the performance of the mining land identification model; L s It is the loss function for labeled sample recognition. By comparing the difference between the model's prediction of labeled samples and the true label, it guides the model to learn the correct recognition of labeled data. u It is the loss function for unlabeled sample recognition. By comparing the difference between the model's prediction of unlabeled samples and the pseudo-label, the model is guided to learn the recognition of unlabeled data. λ is used to adjust the weight of the unsupervised segmentation loss in the overall loss. Where N is the number of labeled samples, N u is the number of unsupervised samples, H and W are the height and width of the segmentation mask, C is the number of categories, and m ijc is the probability value of the true segmentation label, is the corresponding model prediction value, is the predicted value of the pseudo label, β is the weight of the consistency loss, It is the model's predicted probability of the cth category in the jth row and kth column of the same sample. Consistency Loss is the consistency loss, which is used to minimize the difference between the model's predictions at different positions of the same sample.
7. The method for intelligent identification of mining land based on mimetic self-learning according to claim 1, characterized in that: Methods for marking mining areas and rasterizing and clipping them into image blocks include: Identify and mark the boundaries of different land use types in remote sensing images of mining areas to obtain vector data of mining land; Rasterizing the vector data obtained by annotation to obtain label raster data; The original remote sensing image data and label raster data are cropped to obtain image blocks with the same pixel size and resolution.
8. The method for intelligent identification of mining land based on mimetic self-learning according to claim 7, characterized in that: Land use types include open-pit mines, spoil dumps, coal piles, mine buildings, trees, shrubs, high / medium / low coverage grasslands, mixed grass, shrubs and trees, water bodies, roads and bare land.
9. The method for intelligent identification of mining land based on mimetic self-learning according to claim 1, characterized in that: The ratio of training set, validation set and test set is 6:4:
1.
10. An intelligent identification system for mining land based on mimetic self-learning, characterized in that: include: Sample acquisition module: obtain remote sensing images of mining areas and preprocess them, mark mining areas, and rasterize and crop them into image blocks as labeled samples. Labeled samples include training sets, validation sets, and test sets. Feature extraction module: construct a dense feature fusion segmentation network to extract features from labeled samples to obtain a mining area land feature map; Model construction and training module: construct an intelligent recognition network model based on mimetic self-learning, input labeled and unlabeled samples of land use into the intelligent recognition network model for model training, and obtain the trained intelligent recognition network model; Mining land identification module: The mining area image to be identified is input into the trained intelligent recognition network model to obtain the identification result of the mining land.