High-resolution remote sensing image target integrated machine learning interpretation method in mine environment

Abstract

The invention belongs to the crossing field of remote sensing image processing and machine learning technologies, and particularly relates to a high-resolution remote sensing image target integration machine learning interpretation method in a mine environment. Through the customized feature extraction method for the mine environment, information reflecting mine ground feature features can be accurately recognized and extracted, the method is combined with the optimized integrated machine learning model, the method better adapts to the particularity of the mine high-resolution remote sensing image, and the recognition precision of various targets in the mine environment is improved. According to the method, automation and intelligence of remote sensing image interpretation can be realized, manual participation is greatly reduced, interpretation time is shortened, dynamic changes of a mine environment can be quickly responded, remote sensing image interpretation efficiency is improved, and timeliness of mine management is improved. Mine ground feature priori knowledge is fully utilized for extraction and model optimization, so that the model can better adapt to image data of different mine environments and different time phases, and the applicability and stability of the model are effectively improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of remote sensing image processing and machine learning technology, specifically involving a high-resolution remote sensing image target interpretation method based on machine learning, which is particularly suitable for scenarios such as dynamic monitoring of the mining environment, mineral resource exploration, geological disaster early warning and ecological restoration assessment. Background Technology

[0002] In mine environmental monitoring and management, timely and accurate acquisition of mine environmental information is fundamental to effective management and sustainable development. Current remote sensing image interpretation methods primarily rely on human-computer interactive visual interpretation and computer image recognition interpretation. Human-computer interactive visual interpretation, which depends on professionals reviewing images frame by frame, is not only extremely inefficient but also suffers from inconsistent accuracy and consistency due to differences in professional background, experience, and subjective judgment among personnel. Computer image recognition interpretation methods also have limitations when dealing with complex mine features. Mine features are diverse and influenced by various factors such as geological conditions, topography, vegetation cover, and mining activities, often exhibiting complex mixed spectral characteristics that significantly impact interpretation results.

[0003] With the rapid development of high-resolution remote sensing technology, high-resolution remote sensing images can provide rich details, such as the texture and shape of mining features. However, existing interpretation techniques struggle to fully extract and utilize this information. Although machine learning techniques are increasingly being applied in remote sensing image interpretation, there are still many shortcomings in feature extraction and model building for high-resolution remote sensing images in the specific scenario of mining environments. On the one hand, current feature extraction methods fail to fully consider the uniqueness of mining features, making it difficult to effectively distinguish different types of mining targets. On the other hand, existing machine learning models have weak generalization ability when facing complex and ever-changing mining environments, failing to adapt to the differences in image data from different mines and at different times, resulting in interpretation accuracy and reliability that cannot meet practical needs. Summary of the Invention

[0004] This invention addresses the technical problems of existing mine remote sensing image interpretation technologies, such as the difficulty in capturing key distinguishing features of mine features through feature extraction methods and poor adaptability to remote sensing image data of different areas and phases in mines. It provides a high-resolution remote sensing image target integrated machine learning interpretation method for mine environments.

[0005] This invention is achieved using the following technical solution: a machine learning-integrated interpretation method for targets in high-resolution remote sensing images in a mining environment, comprising the following steps: S1. Data Collection and Preprocessing Collect high-resolution remote sensing image data of the mining area, including multispectral and panchromatic images; perform data preprocessing, including radiometric calibration, atmospheric correction, orthorectification, and data augmentation. S2. Constructing the interpretation sample library Based on the characteristics of mine features, typical mine targets are selected from the data images obtained from S1 as interpretation objects, and an interpretation sample library containing multiple categories, different scales and forms is constructed. S21 Sample Selection Principles Ground cover: Covers all elements of the mine, including: open-pit mining face, transportation roads; ancillary facilities: tailings ponds, spoil heaps, and concentrator buildings; ecological features: vegetated areas, water bodies, and bare land; disturbance features: temporary buildings, construction areas, and waste dumps. Spatiotemporal diversity: spatially covering different mine types and different topography; temporally including image samples from different seasons; S22 Sample Library Construction Method Labeling process: Manual visual interpretation: Combining field survey data, the samples are vector-labeled using ArcGIS software; Stratified sampling: The number of samples is allocated according to the proportion of land cover area to avoid missing samples of rare land cover. Cross-validation: The annotations are independently labeled by three or more professionals, and the accuracy of the annotations is ensured through consistency checks. Sample size: Construct a dataset containing ≥3000 labeled data, where the training set: validation set: test set is divided in a 7:1:2 ratio, and the number of samples for each land cover category is ≥300; Sample augmentation: Rotating, scaling, and adding noise to samples to increase sample diversity and reduce the risk of model overfitting; S3. Feature Extraction and Selection Image processing techniques are used to capture texture and shape features in remote sensing images, and combined with prior knowledge of mine features, feature optimization and selection are performed. S31 Multidimensional Feature Extraction Spectral characteristics, including band reflectance, band ratio, and vegetation index; spectral absorption characteristics, including mineral absorption peaks extracted based on continuum removal methods; Spatial features: including texture features and shape features; S32 Feature Optimization and Selection Dimensionality reduction algorithm: Principal component analysis is used to map high-dimensional features to a low-dimensional space, retaining ≥95% of the information entropy; Recursive feature elimination: removing unimportant features through model iteration; Correlation analysis: Calculate the Pearson correlation coefficient between features, remove redundant features with a correlation greater than 0.8, and reduce the computational complexity of the model; S4. Machine Learning Model Building and Training An integrated machine learning algorithm combining random forest, LightGBM, BP neural network, gradient boosting tree, and support vector machine, along with an image recognition deep learning model, is used to construct a target interpretation model suitable for high-resolution remote sensing images of mines. S41 Multi-level Integrated Learning Framework The base layer uses RF, LightGBM, GBDT, and support vector machine to extract texture features (GLCM), terrain features (slope, aspect), and statistical features (band mean, variance) respectively. The U-Net++ network is used, which combines skip connections and multi-scale feature fusion to achieve multi-scale feature fusion and context awareness, and extracts the spatial structure information of ground features as the base layer. Intermediate layer: The outputs of all models in the base layer are used as meta-features to construct a new training set. The machine learning model outputs a class probability vector (such as the Softmax output), while the deep learning model outputs a multi-scale feature map extracted by the Feature Pyramid Network (FPN). Decision layer (hybrid ensemble optimization): Meta-model architecture, designing a dual-channel neural network as the meta-model to process the probability output of the traditional model (machine learning model) and the feature map output of the CNN (deep learning) respectively; dynamic weight adjustment, introducing a meta-learning mechanism (MAML) to automatically adjust the weights of each basic model according to the spatial complexity of the input image; S42 Domain Adaptation Optimization Transfer learning: Initialize the convolutional layer parameters using a pre-trained ResNet50 model, and fine-tune the fully connected layers for the mining scenario; Multi-task learning: Simultaneously train land cover classification and boundary detection tasks, and improve the model's generalization ability by sharing a feature extraction layer; the boundary detection task focuses on identifying the spatial boundary lines between different land cover categories or the contour edges of the target itself; S5. Model Validation and Evaluation Existing mine land use information was selected as the validation dataset to validate and evaluate the trained model. The model's performance was evaluated by calculating overall accuracy, Kappa coefficient, precision, recall, F1 score, and time stability metrics. The model was then optimized and adjusted based on the evaluation results. S51 Validation Dataset The mine field survey data, independent of the training set, includes ≥300 ground verification points, covering all land cover categories and with uniform spatial distribution. The validation metrics used are divided into six types: overall precision, Kappa coefficient, precision, recall, F1 score, and time stability. The corresponding calculation formulas are as follows: Pixels correctly classified / Total pixels; Classification consistency after adjusting for opportunity consistency; TP / (TP+FP); TP / (TP+FN); 2×Precision×Recall / (Precision+Recall); The fluctuation range of classification accuracy of images from different time phases; Precision refers to accuracy, and Recall refers to recall. TP is an abbreviation for True Positive, which means a true example: a sample that the model correctly predicts as the target class, and the true label of the sample does indeed belong to the target class; FP is an abbreviation for False Positive, which means a false positive: a sample that the model incorrectly predicts as the target class, but the true label of the sample does not belong to the target class; FN is an abbreviation for False Negative, which means a false negative example: a sample that the model incorrectly predicts as not belonging to the target class, while the true label of the sample actually belongs to the target class; S52 Verification Process Qualitative assessment: By visualizing the classification results and comparing them with real-world photos, we can check the boundary matching accuracy and the recognition effect of small features; Quantitative assessment: The above indicators were calculated using a confusion matrix, requiring an overall precision ≥ 90%, Kappa ≥ 0.85, and F1 score ≥ 85% for each category; Generalization test: The model is tested on images from different mining areas and different satellite data sources to verify its cross-scene adaptability; Iterative optimization: For misclassified categories found in the evaluation, supplement samples or adjust feature weights, and retrain the model.

[0006] This invention integrates multi-source remote sensing data with machine learning technology to achieve accurate identification and automated analysis of mine features, providing technical support for the monitoring and management of the entire life cycle of mines.

[0007] Furthermore, the specific steps of S1 are as follows: S11 Data Collection Data source: Images of the mining area were acquired using high-resolution remote sensing satellites with a spatial resolution of ≤2 meters and a spectral range covering the visible and near-infrared bands. Multiple time-series data were also collected to capture dynamic changes in the mining environment. S12 Preprocessing Flow Radiometric calibration: Converting image DN values ​​into actual surface radiance values ​​using satellite-provided calibration coefficients or radiative transfer equations to eliminate sensor response errors; Atmospheric correction: Using the FLAASH model, the effects of atmospheric scattering and absorption on the spectrum are corrected to obtain the true surface reflectance; Orthorectification: Based on DEM data, rational function models or collinear equations are used to eliminate image distortion caused by terrain undulations and generate geocoded orthophotos. Data augmentation: Image quality is improved and mine features are highlighted through contrast stretching and wavelet denoising.

[0008] Furthermore, in the spatiotemporal diversity of step S21, there are different types of mines, including metal mines, coal mines, and non-metal mines; and the topography includes mountains, plains, and hills.

[0009] Furthermore, in S31, the texture features are calculated using the gray-level co-occurrence matrix to determine contrast, entropy, and correlation, and the local binary mode describes the texture complexity; the shape features include extracting boundary perimeter, area, compactness, and Fourier descriptors to distinguish between regular buildings and irregular mining faces.

[0010] Furthermore, in the S41 multi-level ensemble learning framework, the application scenarios of each model are as follows: Random Forest is used for preliminary classification and feature selection; LightGBM is used for rapid iteration on large-scale data; BP Neural Network is used for complex terrain features, including mixed mineralized areas; Support Vector Machine is used for text classification and image recognition.

[0011] Furthermore, the CNN model (deep learning model) architecture in S41 selects the following parameters: Training optimization: Loss function: Combine cross-entropy loss with Dice coefficient to handle the imbalanced sample problem; Optimizer: AdamW is used, with an initial learning rate of 1e-4 and a cosine annealing decay strategy. Regularization: Add L2 regularization and Dropout to prevent overfitting. In regularization, λ=0.001 and Dropout rate=0.3.

[0012] Furthermore, in the S42 domain adaptation optimization, the specific tasks of boundary detection include: Category boundary detection: including the transition boundary between mining areas and surrounding vegetated areas, the boundary between tailings ponds and bare land, and the edge line between industrial sites and mining roads; Target contour detection: including the complete contour edge of tailings dams, the outer boundary of large mining equipment, and the contour of stockpiles; Anomaly boundary detection: This includes the edges of newly occupied land in the mining area, i.e., dynamic boundaries compared with historical images, and the boundaries of vegetation degradation areas affected by mining, to help capture dynamic changes in the mining environment.

[0013] The beneficial effects of the present invention are: (1) Improved interpretation accuracy: The present invention can accurately identify and extract information reflecting the features of mine landforms through a customized feature extraction method for the mining environment. Combined with the optimized integrated machine learning model, it can better adapt to the special characteristics of high-resolution remote sensing images of mines and improve the recognition accuracy of various targets in the mining environment.

[0014] (2) Improved interpretation efficiency: Compared with traditional interpretation methods, this invention can realize the automation and intelligence of remote sensing image interpretation, greatly reduce manual intervention, shorten interpretation time, and quickly respond to dynamic changes in the mining environment, thereby improving the efficiency of remote sensing image interpretation and enhancing the timeliness of mine management.

[0015] (3) Enhanced generalization ability: This invention makes full use of prior knowledge of mine features for extraction and model optimization, so that the model can better adapt to image data of different mine environments and different time phases, effectively improving the applicability and stability of the model. Detailed Implementation

[0016] The present invention solves the following technical problems: (1) Effective extraction and representation of features of high-resolution remote sensing images in mining environments: General feature extraction methods are difficult to capture the key distinguishing features of mining features. The present invention mainly targets high-resolution remote sensing images of mining environments to extract effective and representative target features. (2) Improved generalization ability of machine learning model: Existing models have poor adaptability to remote sensing image data of different areas and different phases in mines. This invention aims to improve the generalization ability of machine learning interpretation model on high-resolution remote sensing images of mine environment, so that it can adapt to image data of different scenes and different phases. (3) Reliability verification and evaluation method of interpretation results: This invention addresses the problem of how to scientifically and objectively verify and evaluate the reliability of machine learning interpretation results, and has formulated scientific evaluation indicators and established an effective verification mechanism.

[0017] 1. Data Collection and Preprocessing. High-resolution remote sensing image data of the mining area was collected, including multispectral and panchromatic images. The data underwent preprocessing, including radiometric calibration, atmospheric correction, and orthorectification, to eliminate image distortion and noise and improve image quality.

[0018] (1) Data collection Data source: High-resolution remote sensing satellites (such as GF-1, GF-2, GF-6, etc.) are used to acquire images of the mining area with a spatial resolution of ≤2 meters and a spectral range covering the visible and near-infrared bands. At the same time, multiple time-series data (quarterly / annual images) are collected to capture dynamic changes in the mining environment.

[0019] (2) Preprocessing process Radiometric calibration: Converting image DN values ​​into actual surface radiance values ​​using satellite-provided calibration coefficients or radiative transfer equations to eliminate sensor response errors.

[0020] Atmospheric correction: Using the FLAASH model, the effects of atmospheric scattering and absorption on the spectrum are corrected to obtain the true reflectance of the Earth's surface.

[0021] Orthorectification: Based on DEM data, rational function models (RFM) or collinear equations are used to eliminate image distortion caused by terrain undulations and generate geocoded orthophotos.

[0022] Data augmentation: Image quality is improved by methods such as contrast stretching and wavelet denoising, highlighting mine features (such as the bright tones of tailings ponds and the dark tones of vegetation cover).

[0023] 2. Construct an interpretation sample library. Based on the characteristics of mine features, select typical mine targets (mining faces, tailings ponds, mine buildings) as interpretation objects, and construct an interpretation sample library containing multiple categories, different scales, and forms.

[0024] (1) Sample selection principles Ground cover: Covers all elements of the mine, including: open-pit mining face, transportation roads; ancillary facilities: tailings ponds, spoil heaps, and concentrator buildings; ecological features: vegetated areas, water bodies, and bare land; disturbance features: temporary buildings, construction areas, and waste dumps.

[0025] Spatiotemporal diversity: Spatially, it covers different types of mines (metal mines, coal mines, non-metal mines) and topography (mountains, plains, hills); temporally, it includes image samples from different seasons.

[0026] (2) Sample library construction method Labeling process: Manual visual interpretation: Combining field survey data (GPS points, on-site photos), vector labeling is performed using ArcGIS software; Stratified sampling: The number of samples is allocated according to the area ratio of land features to avoid missing samples of rare land features (such as small tailings ponds); Cross-validation: The annotation is independently labeled by 3 or more professionals and the accuracy is ensured by passing the consistency test (Kappa coefficient ≥ 0.85).

[0027] Sample size: Construct a dataset containing ≥3000 labeled data sets, where the training set: validation set: test set is divided in a 7:1:2 ratio, and the number of samples for each land cover category is ≥300.

[0028] Sample augmentation: Data augmentation is performed on samples by rotating, scaling, adding noise, etc., to increase sample diversity and reduce the risk of model overfitting.

[0029] 3. Feature Extraction and Selection. Image processing techniques are used to capture texture, shape, and other features in remote sensing images. Combined with prior knowledge of mine features, feature optimization and selection are performed to improve the performance and efficiency of subsequent classifiers.

[0030] (1) Multidimensional feature extraction Spectral characteristics: band reflectance, band ratio (e.g., 4 / 5 band ratio for iron mineral identification), vegetation index (EVI, SAVI); spectral absorption characteristics: extraction of mineral absorption peaks based on continuum removal method (e.g., hydroxyl, iron ion characteristic bands).

[0031] Spatial features: Texture features: Contrast, entropy, and correlation are calculated using the Gray-Level Co-occurrence Matrix (GLCM), and texture complexity is described using Local Binary Pattern (LBP); Shape features: Boundary perimeter, area, compactness, and Fourier descriptors are extracted to distinguish between regular buildings and irregular mining faces.

[0032] (2) Feature optimization and selection dimensionality reduction algorithm: Principal component analysis (PCA): Maps high-dimensional features to low-dimensional space, retaining ≥95% of information entropy; Recursive Feature Elimination (RFE): This method eliminates unimportant features through model iteration, such as feature importance ranking in random forests.

[0033] Correlation analysis: Calculate the Pearson correlation coefficient between features, remove redundant features with a correlation greater than 0.8, and reduce the computational complexity of the model.

[0034] 4. Machine Learning Model Construction and Training. An ensemble of machine learning algorithms, including Random Forest, LightGBM, BP neural network, and Gradient Boosting Tree (GBDT), combined with a deep learning model for image recognition, was used to construct a target interpretation model suitable for high-resolution remote sensing images of mines. The model was trained using an interpretation sample library, and its classification accuracy and generalization ability were improved by adjusting model parameters and optimizing the algorithm.

[0035] (1) Integrated learning framework .

[0036] A multi-layered ensemble learning framework. The base layer uses RF, LightGBM, GBDT, and support vector machines to extract texture features (GLCM), terrain features (slope, aspect), and statistical features (band mean, variance), respectively. A U-Net++ network is employed, combining skip connections and multi-scale feature fusion to achieve multi-scale feature fusion and context awareness, extracting spatial structure information of ground features, serving as the base layer. Intermediate layer: The outputs of all models in the base layer are used as meta-features to construct a new training set. The machine learning model outputs a class probability vector (such as the Softmax output), while the deep learning model outputs a multi-scale feature map extracted by the Feature Pyramid Network (FPN). Decision layer (hybrid ensemble optimization): Meta-model architecture, designing a dual-channel neural network as the meta-model to process the probability output of the traditional model and the feature map output of the CNN respectively; dynamic weight adjustment, introducing a meta-learning mechanism (MAML) to automatically adjust the weights of each basic model according to the spatial complexity of the input image; (2) Deep learning architecture CNN model architecture: It adopts the U-Net++ network, which combines skip connections and multi-scale feature fusion to adapt to multi-scale ground features in remote sensing images (such as small buildings and large mining areas).

[0037] Training optimization: Loss function: Combine cross-entropy loss with Dice coefficient to handle the imbalanced sample problem; Optimizer: AdamW is used, with an initial learning rate of 1e-4 and a cosine annealing decay strategy. Regularization: Add L2 regularization (λ=0.001) and Dropout (rate=0.3) to prevent overfitting.

[0038] (3) Domain adaptation optimization Transfer learning: Initialize the convolutional layer parameters using a pre-trained ResNet50 model, and fine-tune the fully connected layers for the mining scenario; Multi-task learning: Simultaneously train land cover classification and boundary detection (referring to the land cover type task in the sample selection principle of "2. Constructing the interpretation sample library"), and improve the model's generalization ability by sharing the feature extraction layer.

[0039] Boundary detection tasks focus on identifying spatial boundaries between different land cover categories or the contour edges of a target itself. Specific tasks include: Category boundary detection: such as the transition boundary between mining areas and surrounding vegetated areas, the boundary between tailings ponds and bare land, and the edge line between industrial sites and mining roads; Target contour detection: such as the complete contour edge of tailings dams, the outer boundary of large mining equipment, and the contour of stockpiles; Anomaly boundary detection: such as the edge of newly occupied land in the mining area (dynamic boundary compared with historical images), the boundary of vegetation degradation area affected by mining, etc., to help capture dynamic changes in the mining environment.

[0040] 5. Model Validation and Evaluation. Existing mine land occupation information was selected as the validation dataset to validate and evaluate the trained model. The model's performance was evaluated by calculating metrics such as classification accuracy, recall, and F1 score, and the model was optimized and adjusted based on the evaluation results.

[0041] (1) Validate the dataset The data used is independent of the training set from field surveys of mines, containing ≥300 ground verification points, covering all land cover categories, and with uniform spatial distribution.

[0042] .

[0043] (2) Verification process Qualitative assessment: By visualizing the classification results and comparing them with real-world photos, we can check the boundary matching degree and the recognition effect of small features (such as pipelines and small buildings). Quantitative assessment: The above indicators were calculated using a confusion matrix, requiring OA ≥ 90%, Kappa ≥ 0.85, and F1 score ≥ 85% for each category; Generalization test: The model is tested on images from different mining areas and different satellite data sources to verify its cross-scene adaptability; Iterative optimization: For misclassifications found in the evaluation (such as confusion between tailings ponds and bare land), supplement samples or adjust feature weights, and retrain the model.

[0044] Further explanation: The above technical solution solves the problems of insufficient feature representation and weak model generalization ability in the interpretation of high-resolution remote sensing images of mines through a closed-loop design of "data-feature-model-validation".

[0045] The present invention will be further described below with reference to the embodiments.

[0046] Example 1: Application Scenario of Mine Land Feature Classification and Monitoring A mining area needs to classify and monitor different types of land features (such as ore bodies, tailings ponds, exposed surfaces, etc.) in order to assess changes in the mining environment. Implementation steps: 1. Data collection and preprocessing. (1) Collect high-resolution remote sensing images (multispectral and panchromatic images) of the mining area and perform radiometric calibration, atmospheric correction and orthorectification. (2) Enhance the images to highlight the features of the land features. 2. Feature extraction and selection. (1) Extract texture features (such as gray-level co-occurrence matrix), shape features (such as boundary regularity) and spectral features (such as NDVI index) from the images. (2) Use recursive feature elimination (RFE) to select the feature subset that contributes the most to the classification. 3. Model construction and training. (1) Construct a random forest classification model, use the extracted features as input, and classify the land feature types. (2) Use a labeled sample library to train the model and adjust parameters such as the number and depth of decision trees. 4. Result verification and evaluation. The classification results were validated using a test dataset, and overall classification accuracy (OA), Kappa coefficient, and other indicators were calculated. Application results: The overall classification accuracy (OA) reached 92.3%, an improvement of 15.7% compared to traditional manual interpretation methods; classification efficiency was significantly improved, reducing the time from several weeks for manual interpretation to several hours for machine learning interpretation; and changes in the area of ​​tailings ponds in mining areas were successfully monitored, providing a scientific basis for environmental governance.

[0047] Example 2: Application Scenarios for Mineralization Alteration Zone Identification In a certain metal mining area, high-resolution remote sensing images were used to identify mineralization alteration zones to assist in mineral resource exploration. Implementation steps: 1. Data collection and preprocessing. (1) Acquire high-resolution remote sensing images of the mining area and perform image correction in conjunction with geological survey data. (2) Enhance the spectral information of the images to highlight the characteristics of altered minerals. 2. Feature extraction and selection. (1) Extract spectral features (such as band ratios and principal component analysis results) and spatial features (such as texture and edge information). (2) Use principal component analysis (PCA) to reduce dimensionality and reduce redundant features. 3. Model construction and training. (1) Construct a support vector machine (SVM) classification model and use radial basis function (RBF) kernel function to identify mineralization alteration zones. (2) Adjust the kernel function parameters and regularization parameters to optimize model performance. 4. Result verification and evaluation. Compare the model prediction results with the field sampling data to evaluate the identification accuracy. Application results: The accuracy of mineralization alteration zone identification reached 89.5%, which is 20.1% higher than that of traditional methods; multiple potential mineralization alteration areas were discovered, providing important clues for subsequent mineral resource exploration; the model running time was reduced from several days in traditional methods to several minutes.

[0048] Example 3: Application Scenario of Mine Ecological Restoration Monitoring An ecological restoration project is underway in a mining area, requiring regular monitoring of vegetation cover changes and evaluation of the restoration effect. Implementation steps: 1. Data collection and preprocessing. (1) Collect high-resolution remote sensing images of the mining area at different time phases and perform registration and correction. (2) Calculate vegetation indices (such as NDVI, EVI) to reflect vegetation cover status. 2. Feature extraction and selection. (1) Extract vegetation index change rate, texture features, and spectral features. (2) Use correlation analysis to screen out the features most relevant to vegetation cover changes. 3. Model construction and training. (1) Construct a deep learning model (such as convolutional neural network, CNN) to classify vegetation cover changes. (2) Use transfer learning technology to accelerate model training using a pre-trained model. 4. Result verification and evaluation. Compare the model prediction results with ground measured data to evaluate the accuracy of vegetation cover change monitoring. Application results: The accuracy of vegetation cover change monitoring reached 95.6%, which is 18.2% higher than that of traditional methods; the vegetation coverage rate in the mining area increased by 12.3% within one year, indicating that the ecological restoration project has achieved significant results; the deep learning model has strong generalization ability and performs better than traditional methods in different scenarios.

[0049] Example 4: Application Scenario of Water Pollution Monitoring in Mining Areas A mining area has water pollution problems, and it is necessary to monitor the pollution level of the mining area water through remote sensing images. Implementation steps: 1. Data collection and preprocessing. (1) Collect high-resolution remote sensing images of the water area of ​​the mining area and perform atmospheric correction and water extraction. (2) Enhance the spectral information of the images to highlight the characteristics of pollutants. 2. Feature extraction and selection. (1) Extract the spectral characteristics of the water (such as band reflectance and absorption peak position) and pollution characteristics (such as suspended matter concentration). (2) Use cluster analysis to screen out features with significant differences in pollution level. 3. Model construction and training. (1) Construct a random forest (RF) model to classify the degree of water pollution (mild, moderate, severe). (2) Adjust the model parameters to optimize the classification effect. 4. Result verification and evaluation. Analyze and compare the model prediction results with the water quality sampling data to evaluate the accuracy of pollution monitoring. Application results: The accuracy of water pollution level classification reached 91.8%, which is 16.4% higher than that of traditional methods; multiple heavily polluted water areas were successfully identified, providing precise location for pollution control; the monitoring cycle was shortened from several months to several days.

[0050] As can be seen from the four specific embodiments above, this novel technical solution has broad application prospects in the interpretation of high-resolution remote sensing images of mining environments. Significant improvements have been achieved in various aspects, including land cover classification, identification of mineralization and alteration zones, ecological restoration monitoring, and water pollution monitoring. These results not only improve interpretation accuracy and efficiency but also provide important technical support for mine environmental protection and resource management.

[0051] Technical features of the present invention: (1) Innovative remote sensing image interpretation method: The present invention integrates machine learning technology into the interpretation of high-resolution remote sensing images of mines. Compared with traditional interpretation methods, it has stronger automation and intelligence features and can handle more complex mine land feature identification tasks.

[0052] (2) Customized feature extraction method: In view of the uniqueness of high-resolution remote sensing images of mining environment, this invention innovatively combines prior knowledge of mining features to develop a set of exclusive feature extraction and selection methods. This method can accurately extract features that play a key role in the interpretation of mining targets, effectively improve the model's recognition of mining targets, and is one of the key technologies for achieving high-precision interpretation.

[0053] (3) Optimized machine learning model system: This invention constructs a set of high-resolution remote sensing image target integrated machine learning interpretation methods suitable for mining environments, covering multiple models such as machine learning and deep learning, and optimizing the models according to the characteristics of mining environments. By reasonably adjusting the model parameters and improving the algorithm structure, the interpretation accuracy and generalization ability of the model are significantly improved, enabling it to better adapt to the complex and ever-changing mining environment.

[0054] (4) A comprehensive interpretation result verification and evaluation system: This invention establishes a scientific and systematic interpretation result verification and evaluation mechanism. By comprehensively utilizing indicators such as classification accuracy, recall rate, and F1 score, the interpretation results are evaluated in a comprehensive and objective manner, ensuring the accuracy and reliability of the interpretation results and providing strong support for practical applications.

Claims

1. A machine learning-integrated interpretation method for targets in high-resolution remote sensing images in a mining environment, characterized in that: Includes the following steps: S1. Data Collection and Preprocessing Collect high-resolution remote sensing image data of the mining area, including multispectral and panchromatic images; perform data preprocessing, including radiometric calibration, atmospheric correction, orthorectification, and data augmentation. S2. Constructing the interpretation sample library Based on the characteristics of mine features, typical mine targets are selected from the data images obtained from S1 as interpretation objects, and an interpretation sample library containing multiple categories, different scales and forms is constructed. S21 Sample Selection Principles Ground cover: Covers all elements of the mine, including: open-pit mining face, transportation roads; ancillary facilities: tailings ponds, spoil heaps, and concentrator buildings; ecological features: vegetated areas, water bodies, and bare land; disturbance features: temporary buildings, construction areas, and waste dumps. Spatiotemporal diversity: spatially covering different mine types and different topography; temporally including image samples from different seasons; S22 Sample Library Construction Method Labeling process: Manual visual interpretation: Combining field survey data, the samples are vector-labeled using ArcGIS software; Stratified sampling: The number of samples is allocated according to the proportion of land cover area to avoid missing samples of rare land cover. Cross-validation: The annotations are independently labeled by three or more professionals, and the accuracy of the annotations is ensured through consistency checks. Sample size: Construct a dataset containing ≥3000 labeled data, where the training set: validation set: test set is divided in a 7:1:2 ratio, and the number of samples for each land cover category is ≥300; Sample augmentation: Rotating, scaling, and adding noise to samples to increase sample diversity and reduce the risk of model overfitting; S3. Feature Extraction and Selection Image processing techniques are used to capture texture and shape features in remote sensing images, and combined with prior knowledge of mine features, feature optimization and selection are performed. S31 Multidimensional Feature Extraction Spectral characteristics, including band reflectance, band ratio, and vegetation index; spectral absorption characteristics, including mineral absorption peaks extracted based on continuum removal methods; Spatial features: including texture features and shape features; S32 Feature Optimization and Selection Dimensionality reduction algorithm: Principal component analysis is used to map high-dimensional features to a low-dimensional space, retaining ≥95% of the information entropy; Recursive feature elimination: removing unimportant features through model iteration; Correlation analysis: Calculate the Pearson correlation coefficient between features, remove redundant features with correlation > 0.8, and reduce the computational complexity of the model; S4. Machine Learning Model Building and Training An integrated machine learning algorithm, including Random Forest (RF), LightGBM, Backpropagation Neural Network (BPNN), Gradient Boosting Tree (GBDT), and Support Vector Machine (SVM), combined with an image recognition deep learning model, was used to construct a target interpretation model suitable for high-resolution remote sensing images of mines. S41 Multi-Level Integrated Learning Framework Base layer: RF, LightGBM, GBDT, and SVM are used to extract texture features, terrain features, and statistical features respectively. U-Net++ network is used, combined with skip connections and multi-scale feature fusion to achieve multi-scale feature fusion and context awareness, and extract spatial structure information of ground features as the base layer; the terrain features include slope and aspect, and the statistical features include band mean and variance. Intermediate layer: The outputs of all models in the base layer are used as meta-features to construct a new training set; the machine learning model outputs class probability vectors, and the deep learning outputs multi-scale feature maps extracted by the Feature Pyramid Network (FPN); Decision layer: Meta-model architecture, designing a dual-channel neural network as the meta-model to process the probability output of the machine learning model and the feature map output of the deep learning model respectively; dynamic weight adjustment, introducing the meta-learning mechanism MAML to automatically adjust the weights of each basic model according to the spatial complexity of the input image; S42 Domain Adaptation Optimization Transfer learning: Initialize the convolutional layer parameters using a pre-trained ResNet50 model, and fine-tune the fully connected layers for the mining scenario; Multi-task learning: Simultaneously train land cover classification and boundary detection tasks, and improve the model's generalization ability by sharing a feature extraction layer; the boundary detection task focuses on identifying the spatial boundary lines between different land cover categories or the contour edges of the target itself; S5. Model Validation and Evaluation Existing mine land use information was selected as the validation dataset to validate and evaluate the trained model. The model's performance was evaluated by calculating overall accuracy, Kappa coefficient, precision, recall, F1 score, and time stability metrics. The model was then optimized and adjusted based on the evaluation results. S51 Validation Dataset The mine field survey data, independent of the training set, includes ≥300 ground verification points, covering all land cover categories and with uniform spatial distribution. The validation metrics used are divided into six types: overall precision, Kappa coefficient, precision, recall, F1 score, and time stability. The corresponding calculation formulas are as follows: Pixels correctly classified / Total pixels; Classification consistency after adjusting for opportunity consistency; TP / (TP+FP); TP / (TP+FN); 2×Precision×Recall / (Precision+Recall); The fluctuation range of classification accuracy of images from different time phases; Precision refers to accuracy, and Recall refers to recall. TP is an abbreviation for True Positive, which means a true example: a sample that the model correctly predicts as the target class, and the true label of the sample does indeed belong to the target class; FP is an abbreviation for False Positive, which means a false positive: a sample that the model incorrectly predicts as the target class, but the true label of the sample does not belong to the target class; FN is an abbreviation for False Negative, which means a false negative example: a sample that the model incorrectly predicts as not belonging to the target class, while the true label of the sample actually belongs to the target class; S52 Verification Process Qualitative assessment: By visualizing the classification results and comparing them with real-world photos, we can check the boundary matching accuracy and the recognition effect of small features; Quantitative assessment: The above indicators were calculated using a confusion matrix, requiring an overall precision ≥ 90%, Kappa ≥ 0.85, and F1 score ≥ 85% for each category; Generalization test: The model is tested on images from different mining areas and different satellite data sources to verify its cross-scene adaptability; Iterative optimization: For misclassified categories found in the evaluation, supplement samples or adjust feature weights, and retrain the model.

2. The method for ensemble machine learning interpretation of high-resolution remote sensing images in a mining environment as described in claim 1, characterized in that, The specific steps for S1 are as follows: S11 Data Collection Data source: Images of the mining area were acquired using high-resolution remote sensing satellites with a spatial resolution of ≤2 meters and a spectral range covering the visible and near-infrared bands. Multiple time-series data were also collected to capture dynamic changes in the mining environment. S12 Preprocessing Flow Radiometric calibration: Converting image DN values ​​into actual surface radiance values ​​using satellite-provided calibration coefficients or radiative transfer equations to eliminate sensor response errors; Atmospheric correction: Using the FLAASH model, the effects of atmospheric scattering and absorption on the spectrum are corrected to obtain the true surface reflectance; Orthorectification: Based on DEM data, rational function models or collinear equations are used to eliminate image distortion caused by terrain undulations and generate geocoded orthophotos. Data augmentation: Image quality is improved and mine features are highlighted through contrast stretching and wavelet denoising.

3. The method for ensemble machine learning interpretation of high-resolution remote sensing images in a mining environment as described in claim 1, characterized in that, In the spatiotemporal diversity of step S21, there are different types of mines, including metal mines, coal mines, and non-metal mines; and the topography includes mountains, plains, and hills.

4. The method for ensemble machine learning interpretation of high-resolution remote sensing images in a mining environment as described in claim 1, characterized in that, In S31, the texture features are calculated using the gray-level co-occurrence matrix to determine contrast, entropy, and correlation, and the local binary mode describes the texture complexity; the shape features include extracting boundary perimeter, area, compactness, and Fourier descriptors to distinguish between regular buildings and irregular mining faces.

5. The method for ensemble machine learning interpretation of high-resolution remote sensing images in a mining environment as described in claim 1, characterized in that, In the S41 multi-level ensemble learning framework, the application scenarios of each model are as follows: Random Forest is used for preliminary classification and feature selection; LightGBM is used for rapid iteration on large-scale data; BP neural network is used for complex terrain features, including mixed mineralized areas; and Support Vector Machine is used for text classification and image recognition.

6. The method for ensemble machine learning interpretation of high-resolution remote sensing images in a mining environment as described in claim 1, characterized in that, The deep learning model architecture selected in S41 uses the following parameters: Training optimization: Loss function: Combine cross-entropy loss with Dice coefficient to handle the imbalanced sample problem; Optimizer: AdamW is used, with an initial learning rate of 1e-4 and a cosine annealing decay strategy. Regularization: Add L2 regularization and Dropout to prevent overfitting. In regularization, λ=0.001 and Dropout rate=0.

3.

7. The method for ensemble machine learning interpretation of high-resolution remote sensing images in a mining environment as described in claim 1, characterized in that, In S42 domain adaptation optimization, the specific tasks of boundary detection include: Category boundary detection: including the transition boundary between mining areas and surrounding vegetated areas, the boundary between tailings ponds and bare land, and the edge line between industrial sites and mining roads; Target contour detection: including the complete contour edge of tailings dams, the outer boundary of large mining equipment, and the contour of stockpiles; Anomaly boundary detection: This includes the edges of newly occupied land in the mining area, i.e., dynamic boundaries compared with historical images, and the boundaries of vegetation degradation areas affected by mining, to help capture dynamic changes in the mining environment.