Surface matrix map modification data processing method

Abstract

The invention relates to the technical field of geology, in particular to a surface matrix map modification data processing method which comprises the steps of data dynamic acquisition, intelligent preprocessing, three-dimensional modeling, multi-scale nested analysis, decision deduction and dynamic visualization. Compared with the prior art that a regular manual acquisition and static database updating scheme is adopted, the method has the defects that the data updating period is long, and sudden environmental changes are difficult to capture; according to the scheme, a multi-source data real-time access and AI dynamic cleaning scheme is adopted, the method has the advantages of real-time updating capability and automatic abnormal value correction, the surface matrix model can dynamically respond to environmental changes, the timeliness of disaster early warning and emergency decision making is remarkably improved, deep learning is adopted to fuse multi-modal data and a Bayesian uncertainty quantification scheme, and the method has the advantages of being high in practicability and easy to popularize. The method has the advantages of high-resolution three-dimensional modeling and risk area visualization, a complex geological structure can be accurately described, and clear priority guidance is provided for artificial verification.

Description

Technical Field

[0001] The present invention relates to the field of geological technology, and in particular to a surface matrix map modification data processing method. Background Art

[0002] In the field of surface matrix surveys and ecological restoration, traditional technologies have long been limited by static data integration and experience-driven analysis methods, making it difficult to meet the needs of accurate decision-making in dynamic environments. Existing methods typically rely on the regular manual collection of laboratory data and satellite imagery, which has a long data update cycle. Multi-source data requires a significant amount of time for manual preprocessing due to differences in format and coordinate system, resulting in inefficient data fusion and an inability to respond in real time to dynamic changes in the surface matrix caused by emergencies such as floods and landslides. Furthermore, three-dimensional modeling technologies often use kriging interpolation or two-dimensional profile extrapolation methods, which have a smoothing effect on the depiction of complex geological structures, resulting in a significant loss of detailed information and a lack of quantification of the uncertainty of model prediction results. This makes it difficult for decision makers to assess risks, often leading to blindness in the implementation of restoration plans.

[0003] In terms of scale coordination, existing technologies often separate macro-trend predictions from micro-plot analysis. For example, the long-term weathering trends output by basin-level climate models are difficult to guide plot-level planting planning, while high-precision plot data cannot reversely optimize macro-model parameters, resulting in a disconnect between prediction and implementation. In addition, the design of ecological restoration plans relies on historical case libraries and expert experience, lacking the ability to dynamically deduce the "measure-response" relationship. The preview of the plan's effects often relies on static drawings. Decision makers cannot intuitively perceive the vegetation recovery cycle or changes in soil water holding capacity, resulting in high trial and error costs and insufficient public participation.

[0004] Despite the recent advancements in the application of satellite remote sensing, the Internet of Things, and deep learning technologies in the geological field, technological gaps remain in the real-time collaboration of multi-source data, high-precision cross-scale modeling, and intelligent decision-making. At the data level, there's a lack of a unified framework for the real-time access and standardization of satellite, sensor, and laboratory data, and outlier cleaning relies on manual rules. At the model level, a single algorithm struggles to balance 3D attribute reconstruction and uncertainty quantification. At the application level, the generation and validation of ecological restoration plans haven't yet achieved a closed "data-model-decision" loop, and visualization technology remains limited to two-dimensional drawings, hindering the efficiency of scientific decision-making. Summary of the Invention

[0005] In order to overcome the problems raised in the above background technology, the present invention proposes a surface matrix map modification data processing method.

[0006] The technical solution of the present invention is: a surface matrix map modification data processing method, comprising the following steps: S11: Dynamic data collection, real-time integration of laboratory data, satellite data, sensor data and geophysical data; S12: Intelligent preprocessing, using AI to automatically clean outliers and dynamically update the database; S13: 3D modeling, based on 3D U-Net fusion of multimodal data, generates high-precision 3D matrix attribute volumes and quantifies prediction uncertainty; S14: Multi-scale nested analysis, coupling basin-level macro-forecasting with parcel-level refined modeling; S15: Decision deduction, simulating the effects of repair plans through reinforcement learning to generate the optimal repair plan; S16: Dynamic visualization, building a WebGL interactive platform and knowledge graph to achieve 3D sectioning query and intelligent repair solution recommendation.

[0007] Preferably, when integrating laboratory data, satellite data, sensor data and geophysical data in real time, the following are specifically included: A11: Laboratory data, which experimentally determines the physical and chemical characteristics of different types of surface substrates; A12: Satellite data stream, access to Planet Labs daily 3m resolution imagery, including NDVI and land surface temperature data; A13: IoT sensors, deploying soil temperature, humidity, and groundwater level sensor networks, and using LoRaWAN to transmit real-time data. A14: Geophysical data, integrating high-density electrical method and ground penetrating radar data to obtain bedrock surface depth and fracture distribution.

[0008] Preferably, when using AI to automatically clean outliers and dynamically update the database, for laboratory data, the specific steps include: S21: Data standardization, Z-score standardization of laboratory data, and unified data format and dimension; S22: Outlier marking, using the isolation forest algorithm to automatically identify outliers, where the parameters of the isolation forest algorithm are set as: the number of trees is 100 and the contamination ratio is set to 0.05; S23: Data correction: laboratory retesting of high-risk outlier points marked by the algorithm, and interpolation of adjacent points for non-critical fields.

[0009] Preferably, when automatically cleaning outliers and dynamically updating the database through AI, for satellite data, the specific steps include: S31: UAV data acquisition, based on the preset map, automatically generates a cruise route and uses the camera to collect surface images in the target area; S32: Image preprocessing: Agisoft Metashape is used to perform brightness homogenization and eliminate shadow effects. Then, orthophotos are generated and registered to the CGCS2000 coordinate system. S33: Boundary recognition, using the YOLOv5 algorithm to analyze the registered image and identify the bounding box and category probability in the image.

[0010] Preferably, when automatically cleaning outliers through AI and dynamically updating the database, the following steps are also included: S41: Spatial alignment, registering geophysical data and remote sensing images using control points; S42: Data enhancement, based on known data, uses a generative adversarial network to synthesize virtual samples in sparsely sampled areas; S43: Attribute association, associates the laboratory data with the drone-annotated polygons through spatial join to generate an attribute table.

[0011] Preferably, when fusing multimodal data based on 3D U-Net to generate a high-precision three-dimensional matrix attribute volume and quantify prediction uncertainty, the method specifically includes: S51: Data preparation and fusion, processing the preprocessed data to generate a multi-channel three-dimensional grid stack to build the model input; S52: Model construction and training, using a 3D U-Net architecture, designing a 4-level encoder and decoder, preserving multi-scale features through skip connections, and introducing transfer learning during model initialization. Pre-trained weights from global drilling data are loaded, and some layers are frozen to accelerate fine-tuning of local data to generate high-resolution 3D attribute volumes. The loss function used in the model is: ; in, is the total loss, For Dice Loss, is the mean square loss error, and , where N is the number of samples, is the true value of the i-th sample, is the predicted value of the i-th sample; S53: Uncertainty quantification, using the MC Dropout technique in Bayesian deep learning, performs 50 random forward propagations during the test phase, calculates the mean and standard deviation of each voxel attribute, and generates a 95% confidence interval; S54: Model validation and application, using an independent test set to verify model accuracy, where evaluation metrics include F1-score and MAE; S55: Iterative optimization and deployment. Based on active learning strategies, new drilling data is added to high-uncertainty areas, the model is retrained, and the model is compressed into a lightweight version through knowledge distillation technology for deployment.

[0012] As a preference, when coupling basin-level macro-forecasting with plot-level refined modeling, the following should be considered: S61: Macro-level climate prediction and downscaling, focusing on large-scale long-term trends, integrating global climate models with basin geological data, simulating matrix weathering trends in target areas through process-driven models, and using dynamic downscaling techniques to refine climate data to 1km grids; S62: Real-time dynamic risk monitoring at the meso-level, integrating macro-level outputs with real-time sensor data to build dynamic risk models and generate early warnings in real time; S63: High-precision decision support at the micro level, based on high-precision models, integrating drone imagery and laboratory data, extracting plot-level attributes, and predicting changes in water holding capacity in combination with macroclimate trends; S64: Cross-scale feedback and iterative optimization, establish a bottom-up feedback mechanism, use plot-level measured data to reversely correct meso-level parameters and macro-models, and achieve multi-source data assimilation through ensemble Kalman filtering to improve the overall accuracy of the model.

[0013] Preferably, when simulating the effect of the repair solution through reinforcement learning and generating the optimal repair solution, the following steps are specifically included: S71: Knowledge graph construction, integrating substrate properties, historical restoration cases, and ecological response data, constructing knowledge graph nodes and defining associations to form a structured decision-making knowledge base; S72: Reinforcement learning environment modeling, defining environmental states and optional repair actions, setting reward functions, and quantifying the balance mechanism between ecological and economic goals; S73: Reinforcement learning model training, using the PPO algorithm to input historical case training models, simulate long-term responses to different measures, output the optimal strategy, and verify the reliability of the solution through Monte Carlo simulation; S74: AR virtual-reality fusion deduction, loading the 3D model of the plot through Hololens 2, superimposing the virtual restoration plan in real time, and sliding the timeline to dynamically display the restoration effect after a certain period; S75: Dynamic monitoring and feedback optimization: deploy sensors to monitor soil indicators in the restoration area, trigger model re-evaluation when data anomalies occur, and update the knowledge graph and fine-tune the model every month.

[0014] As a preferred approach, after building a WebGL interactive platform and knowledge graph to implement 3D sectioning queries and intelligent repair solution recommendations, AI full-scale detection and blockchain evidence storage are also used to ensure data quality and process traceability. The specific steps include: A21: AI quality inspection closed loop, which uses AI algorithms to test data, detect abnormal data, and generate data repair tasks; A22: Blockchain verification: key data is hashed and uploaded to the blockchain, verified through the PBFT consensus mechanism, and the entire data process is traced through timestamps and hash values.

[0015] Preferably, when using AI algorithms to detect data, detect abnormal data, and generate data repair tasks, the following steps are specifically included: S81: Automatically detect laboratory data anomalies using the isolation forest algorithm, associate the GPS coordinates of the anomaly points to generate an early warning layer, and dispatch drone retests to verify data authenticity; S82: Use GAN to generate adversarial examples to simulate model error-prone scenarios, detect areas where the confidence level of the prediction results is below a threshold, automatically mark high-uncertainty areas, and prioritize manual verification by risk level.

[0016] Beneficial effects of the present invention: 1. Compared with existing technologies that use regular manual data collection and static database update solutions, which have the disadvantages of long data update cycles and difficulty in capturing sudden environmental changes, this solution uses real-time multi-source data access and AI dynamic cleaning solutions. It has the advantages of real-time update capabilities and automatic correction of outliers, enabling the surface matrix model to dynamically respond to environmental changes, significantly improving the timeliness of disaster warning and emergency decision-making. 2. Compared to existing technologies that use traditional interpolation or 2D profile modeling, which suffer from loss of detailed information and lack of credibility assessment, this solution uses deep learning to integrate multimodal data and Bayesian uncertainty quantification. It offers the advantages of high-resolution 3D modeling and risk area visualization, accurately depicting complex geological structures and providing clear priority guidance for manual verification. 3. Compared to existing technologies that use isolated single-scale modeling schemes, which suffer from the disadvantage of disconnecting macro-trends from micro-decision-making, this scheme adopts a multi-scale nested model and data assimilation feedback scheme, which has the advantages of cross-level data integration and dynamic parameter correction. It realizes multi-level coordinated prediction from watershed to plot, significantly improving the pertinence and feasibility of ecological governance solutions. 4. Compared with the existing technology that uses manual experience planning and static scheme design, which has the disadvantages of high trial and error costs and insufficient decision support, this scheme adopts reinforcement learning optimization and augmented reality dynamic deduction scheme, which has the advantages of multi-objective intelligent optimization and virtual-reality fusion visualization, making the ecological restoration plan both scientific and intuitive, greatly improving implementation efficiency and public acceptance. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Shown is a flow chart of the surface matrix map modification data processing method of the present invention; Figure 2 What is shown is a flow chart of the intelligent preprocessing of the surface matrix map modification data processing method of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described below with reference to the accompanying drawings and examples.

[0019] See also Figure 1-Figure 2 The present invention provides an embodiment: a surface matrix map modification data processing method, comprising the following steps: S11: Dynamic data collection, real-time integration of laboratory data, satellite data, sensor data and geophysical data; S12: Intelligent preprocessing, using AI to automatically clean outliers and dynamically update the database; S13: 3D modeling, based on 3D U-Net fusion of multimodal data, generates high-precision 3D matrix attribute volumes and quantifies prediction uncertainty; S14: Multi-scale nested analysis, coupling basin-level macro-forecasting with parcel-level refined modeling; S15: Decision deduction, simulating the effects of repair plans through reinforcement learning to generate the optimal repair plan; S16: Dynamic visualization, building a WebGL interactive platform and knowledge graph to achieve 3D sectioning query and intelligent repair solution recommendation.

[0020] As described above, the present invention integrates multi-source data in real time through dynamic data acquisition to ensure data comprehensiveness and timeliness; intelligent preprocessing uses AI to automatically clean outliers and update the database, improving data quality and processing efficiency; three-dimensional modeling uses 3D U-Net to fuse multimodal data to generate high-precision three-dimensional matrix attribute volumes and quantify prediction uncertainties, providing a basis for in-depth analysis; multi-scale nested analysis couples basin-level macro-forecasting with plot-level refined modeling to meet the needs of analysis at different scales; decision-making deduction uses reinforcement learning to simulate the effects of restoration plans and generate optimal plans, enhancing the scientific nature of decision-making; dynamic visualization constructs a WebGL interactive platform and knowledge graph to realize three-dimensional section query and intelligent restoration plan recommendation, improving the intuitiveness of data display and the convenience of plan recommendation, and providing comprehensive, efficient, and scientific support for surface matrix research and restoration.

[0021] Preferably, when integrating laboratory data, satellite data, sensor data and geophysical data in real time, the following are specifically included: A11: Laboratory data, which experimentally determines the physical and chemical characteristics of different types of surface substrates; A12: Satellite data stream, access to Planet Labs daily 3m resolution imagery, including NDVI and land surface temperature data; A13: IoT sensors, deploying soil temperature, humidity, and groundwater level sensor networks, and using LoRaWAN to transmit real-time data. A14: Geophysical data, integrating high-density electrical method and ground penetrating radar data to obtain bedrock surface depth and fracture distribution.

[0022] As described above, this invention experimentally measures the physical and chemical characteristics of different types of surface matrices using laboratory data, providing a precise foundation for a deeper understanding of their essential properties. Access to satellite data streams, including daily 3m-resolution imagery from Planet Labs, including NDVI and surface temperature data, enables timely acquisition of large-scale surface dynamics, providing strong support for macro-monitoring and analysis. Deploying an IoT sensor network and utilizing LoRaWAN to transmit real-time soil temperature, moisture, and groundwater level data enables real-time, accurate monitoring of key surface parameters, helping to capture subtle changes. Furthermore, integrating high-density electrical and ground-penetrating radar data to obtain geophysical data such as bedrock depth and fracture distribution further reveals the deep structural characteristics of the surface matrix. The real-time integration of multi-source data enables comprehensive and timely information from the microscopic to the macroscopic, from the surface to the deep, providing a robust and comprehensive data foundation for surface matrix research, monitoring, and related decision-making.

[0023] Preferably, when using AI to automatically clean outliers and dynamically update the database, for laboratory data, the specific steps include: S21: Data standardization, Z-score standardization of laboratory data, and unified data format and dimension; S22: Outlier marking, using the isolation forest algorithm to automatically identify outliers, where the parameters of the isolation forest algorithm are set as: the number of trees is 100 and the contamination ratio is set to 0.05; S23: Data correction: laboratory retesting of high-risk outlier points marked by the algorithm, and interpolation of adjacent points for non-critical fields.

[0024] As described above, the present invention utilizes data standardization, employing the Z-score method to unify data formats and dimensions, eliminating interference caused by different dimensions and formats. This provides a standardized and unified foundation for subsequent data processing and analysis, ensuring data consistency and comparability. The isolation forest algorithm automatically identifies outliers, and rationally sets parameters such as the number of trees to 100 and the contamination ratio to 0.05. This algorithm accurately and efficiently detects outliers, improving the accuracy and reliability of outlier identification. Laboratory retesting of marked high-risk outliers ensures data accuracy, and interpolation of adjacent points is used for non-critical fields, ensuring data integrity while avoiding information loss caused by outlier deletion. This series of steps effectively improves the quality of laboratory data, provides a reliable data source for dynamic database updates, and provides solid data support for relevant research and decision-making.

[0025] Preferably, when automatically cleaning outliers and dynamically updating the database through AI, for satellite data, the specific steps include: S31: UAV data acquisition, based on the preset map, automatically generates a cruise route and uses the camera to collect surface images in the target area; S32: Image preprocessing: Agisoft Metashape is used to perform brightness homogenization and eliminate shadow effects. Then, orthophotos are generated and registered to the CGCS2000 coordinate system. S33: Boundary recognition, using the YOLOv5 algorithm to analyze the registered image and identify the bounding box and category probability in the image.

[0026] As described above, the present invention realizes the intelligent upgrade of surface matrix data acquisition and processing through automatic cruising of UAV to collect surface images and perform AI-driven image preprocessing and boundary recognition, which significantly improves the utilization rate and analysis accuracy of remote sensing data; the UAV automatically generates a cruising route based on a preset map and collects surface images in real time, and cooperates with Agisoft Metashape software to perform brightness uniformization and shadow elimination processing, effectively solving the recognition error problem caused by uneven illumination in traditional remote sensing images. At the same time, the orthophoto image is aligned to the CGCS2000 coordinate system to ensure the spatial consistency of multi-source data; the YOLOv5 algorithm performs automatic boundary recognition and classification probability analysis on the preprocessed image, which not only greatly improves the efficiency of boundary extraction and reduces the workload of manual interpretation, but also provides a quality assessment basis for subsequent data fusion through category probability output, which significantly enhances the usability of satellite data in three-dimensional matrix modeling and lays a solid data foundation for high-precision surface matrix property prediction.

[0027] Preferably, when automatically cleaning outliers through AI and dynamically updating the database, the following steps are also included: S41: Spatial alignment, registering geophysical data and remote sensing images using control points; S42: Data enhancement, based on known data, uses a generative adversarial network to synthesize virtual samples in sparsely sampled areas; S43: Attribute association, associates the laboratory data with the drone-annotated polygons through spatial join to generate an attribute table.

[0028] As described above, the present invention constructs a complete preprocessing system covering multi-source data fusion and feature enhancement through the collaborative processing of spatial alignment, data enhancement and attribute association; the spatial alignment link uses control point registration technology to accurately match the spatial coordinates of geophysical data and remote sensing images, eliminating the geometric deviation between different data sources and laying a precise spatial benchmark for subsequent fusion analysis; the data enhancement stage innovatively uses generative adversarial networks to simulate virtual samples of sparse sampling areas, significantly improving the spatial continuity and integrity of the data without increasing the cost of field acquisition, and effectively solving the defect that traditional interpolation methods are susceptible to noise interference; attribute association uses spatial connection technology to accurately match high-precision laboratory measurement data with polygonal areas annotated by drone images, automatically generate structured attribute tables and establish multi-dimensional feature associations, which not only realizes the efficient integration of heterogeneous data, but also enhances the model's ability to interpret complex geological conditions through attribute fusion, forming a full-process data governance closed loop from spatial correction to feature enhancement, providing a high-quality data foundation for high-precision three-dimensional matrix modeling.

[0029] Preferably, when fusing multimodal data based on 3D U-Net to generate a high-precision three-dimensional matrix attribute volume and quantify prediction uncertainty, the method specifically includes: S51: Data preparation and fusion, processing the preprocessed data to generate a multi-channel three-dimensional grid stack to build the model input; S52: Model construction and training, using a 3D U-Net architecture, designing a 4-level encoder and decoder, preserving multi-scale features through skip connections, and introducing transfer learning during model initialization. Pre-trained weights from global drilling data are loaded, and some layers are frozen to accelerate fine-tuning of local data to generate high-resolution 3D attribute volumes. The loss function used in the model is: ; in, is the total loss, For Dice Loss, is the mean square loss error, and , where N is the number of samples, is the true value of the i-th sample, is the predicted value of the i-th sample; S53: Uncertainty quantification, using the MC Dropout technique in Bayesian deep learning, performs 50 random forward propagations during the test phase, calculates the mean and standard deviation of each voxel attribute, and generates a 95% confidence interval; S54: Model validation and application, using an independent test set to verify model accuracy, where evaluation metrics include F1-score and MAE; S55: Iterative optimization and deployment. Based on active learning strategies, new drilling data is added to high-uncertainty areas, the model is retrained, and the model is compressed into a lightweight version through knowledge distillation technology for deployment.

[0030] As described above, the present invention significantly improves the accuracy and efficiency of three-dimensional matrix property modeling by fusing multimodal data through a 3D U-Net architecture and introducing a transfer learning mechanism. It innovatively adopts a composite loss function to balance classification and regression tasks, achieving millimeter-level high-resolution attribute volume generation while retaining multi-scale geological features. The MCDropout technology is used to quantify model uncertainty and generate a 95% confidence interval to provide risk warning capabilities for decision-making, solving the key defect of traditional modeling methods that lacks credibility assessment. The combination of active learning and knowledge distillation not only achieves continuous optimization and lightweight deployment of the model, but also significantly reduces field sampling costs through targeted data enhancement strategies, reducing the model's prediction error under complex geological conditions by more than 50%, while ensuring that the inference efficiency meets real-time decision-making needs, providing a complete technical solution for the transformation of geological surveys from experience-driven to data intelligence.

[0031] As a preference, when coupling basin-level macro-forecasting with plot-level refined modeling, the following should be considered: S61: Macro-level climate prediction and downscaling, focusing on large-scale long-term trends, integrating global climate models with basin geological data, simulating matrix weathering trends in target areas through process-driven models, and using dynamic downscaling techniques to refine climate data to 1km grids; S62: Real-time dynamic risk monitoring at the meso-level, integrating macro-level outputs with real-time sensor data to build dynamic risk models and generate early warnings in real time; S63: High-precision decision support at the micro level, based on high-precision models, integrating drone imagery and laboratory data, extracting plot-level attributes, and predicting changes in water holding capacity in combination with macroclimate trends; S64: Cross-scale feedback and iterative optimization, establish a bottom-up feedback mechanism, use plot-level measured data to reversely correct meso-level parameters and macro-models, and achieve multi-source data assimilation through ensemble Kalman filtering to improve the overall accuracy of the model.

[0032] As described above, the present invention achieves a technological leap from rough estimation to precise dynamic simulation in surface matrix prediction by constructing a three-level coupled model system: macro-meso-micro. At the macro level, global climate models are integrated with watershed geological data, and dynamic downscaling techniques are used to refine climate predictions down to a 1km grid. This provides reliable boundary conditions for meso-level risk monitoring and addresses the key issue of the disconnect between climate and geological elements in traditional methods. At the meso level, real-time sensor data is integrated to construct a dynamic risk model, breaking through the limitations of traditional static analysis and enabling a shift in disaster warning from "post-event response" to "pre-event prevention." At the micro level, plot-level attributes are extracted based on drone imagery and laboratory data, and changes in water holding capacity are predicted based on macroclimate trends, addressing the difficulty of traditional single-scale models in balancing regional patterns and local characteristics. A cross-scale feedback mechanism achieves multi-source data assimilation through ensemble Kalman filtering, using measured plot data to reverse-correct macro- and meso-level model parameters, forming an "observe-simulate-correct" closed-loop optimization system. This significantly improves the accuracy and reliability of cross-scale model coupling and provides scientific decision-making support for watershed ecological governance.

[0033] Preferably, when simulating the effect of the repair solution through reinforcement learning and generating the optimal repair solution, the following steps are specifically included: S71: Knowledge graph construction, integrating substrate properties, historical restoration cases and ecological response data, constructing knowledge graph nodes and defining association relationships to form a structured decision-making knowledge base; S72: Reinforcement learning environment modeling, defining environmental states and optional repair actions, setting reward functions, and quantifying the balance mechanism between ecological and economic goals; S73: Reinforcement learning model training, using the PPO algorithm to input historical case training models, simulate long-term responses to different measures, output the optimal strategy, and verify the reliability of the solution through Monte Carlo simulation; S74: AR virtual-reality fusion deduction, loading the 3D model of the plot through Hololens 2, superimposing the virtual restoration plan in real time, and sliding the timeline to dynamically display the restoration effect after a certain period; S75: Dynamic monitoring and feedback optimization: deploy sensors to monitor soil indicators in the restoration area, trigger model re-evaluation when data anomalies occur, and update the knowledge graph and fine-tune the model every month.

[0034] As described above, this invention achieves a leapfrog upgrade from experience-driven to data-driven ecological restoration solutions by constructing an intelligent decision-making system that combines knowledge graphs with reinforcement learning. The knowledge graph integrates substrate properties, historical cases, and ecological response data to form a structured decision-making knowledge base, addressing the problem of traditional restoration solutions lacking systematic knowledge support. Reinforcement learning environmental modeling achieves multi-objective balanced optimization of ecological benefits and economic costs by defining states, actions, and quantified reward functions, overcoming the limitations of traditional methods that struggle to balance multiple constraints. The PPO algorithm, combined with Monte Carlo simulation verification, not only improves model training efficiency and strategy stability, but also ensures the scientific nature and sustainability of restoration solutions through long-term response simulation. AR fusion deduction technology visualizes restoration effects, allowing users to interactively view dynamic changes at different time points, significantly improving solution communication efficiency and public participation. A dynamic monitoring and feedback mechanism uses sensors to track soil indicators in real time and triggers model iterations. Combined with continuous knowledge graph updates and model fine-tuning, this creates a complete closed loop of "perception-decision-execution-feedback," significantly enhancing the environmental adaptability and implementation accuracy of restoration solutions, providing an intelligent decision-making support platform for ecological restoration under complex geological conditions.

[0035] As a preferred approach, after building a WebGL interactive platform and knowledge graph to implement 3D sectioning queries and intelligent repair solution recommendations, AI full-scale detection and blockchain evidence storage are also used to ensure data quality and process traceability. The specific steps include: A21: AI quality inspection closed loop, which uses AI algorithms to test data, detect abnormal data, and generate data repair tasks; A22: Blockchain verification: key data is hashed and uploaded to the blockchain, verified through the PBFT consensus mechanism, and the entire data process is traced through timestamps and hash values.

[0036] As described above, the present invention has built a dual guarantee system covering data quality control and process trusted traceability through the deep integration of AI full-scale detection and blockchain evidence storage technology; the AI ​​quality inspection closed loop uses machine learning algorithms to perform real-time monitoring and anomaly identification on multi-source heterogeneous data, automatically generates data repair tasks and triggers the verification process, fundamentally solving the defects of traditional manual quality inspection such as low efficiency, narrow coverage and easy omissions, and ensuring the integrity and accuracy of data assets; the blockchain verification link uses hashing processing and PBFT consensus mechanism to store key data on the chain and establish an unalterable timestamp certificate, and cooperates with hash value traceability technology to realize trusted traceability of the entire data life cycle, which not only prevents the risk of data tampering but also meets the requirements of judicial audits. At the same time, the smart contract-driven evidence storage process greatly improves the transparency and trust of cross-departmental collaboration, provides solid technical support for the standardized implementation of surface matrix surveys and ecological restoration projects, and forms a full-link trusted closed loop from data collection to decision-making applications.

[0037] Preferably, when using AI algorithms to detect data, detect abnormal data, and generate data repair tasks, the following steps are specifically included: S81: Automatically detect laboratory data anomalies using the isolation forest algorithm, associate the GPS coordinates of the anomaly points to generate an early warning layer, and dispatch drone retests to verify data authenticity; S82: Use GAN to generate adversarial examples to simulate model error-prone scenarios, detect areas where the confidence level of the prediction results is below a threshold, automatically mark high-uncertainty areas, and prioritize manual verification by risk level.

[0038] As described above, the present invention constructs a dual protection mechanism covering data quality detection and model credibility assessment through the coordinated application of the isolation forest algorithm and the generative adversarial network; the isolation forest algorithm relies on the statistical characteristics of laboratory data to automatically identify anomalies, and associates GPS coordinates to generate a visual warning layer, and cooperates with the drone rapid retest verification mechanism to achieve accurate positioning and efficient verification of abnormal data, fundamentally solving the problems of low efficiency and high risk of missed detection in traditional manual screening; GAN adversarial sample generation technology focuses on the weak links of model prediction, accurately locates areas with low prediction confidence by simulating error-prone scenarios, and sorts manual verification priorities according to risk levels, which not only enhances the interpretability of model decisions but also reduces the risk of key decision-making errors. At the same time, the dynamically updated data quality feedback mechanism continuously optimizes the detection algorithm performance, forming a complete data governance closed loop covering "anomaly detection-verification correction-model tuning", significantly improving the reliability of surface matrix survey data and the robustness of model predictions.

[0039] Example 1: Surface matrix survey and ecological restoration of the Longwu River Valley in Tongren City, Qinghai Province Technical implementation steps: 1. Dynamic data collection: Laboratory data: 200 soil samples were collected from the river valley and measured for pH (mean 6.8), organic matter (1.2%-2.5%), and particle size (60% sand).

[0040] Satellite data: Access Planet Labs daily 3m imagery to extract NDVI (0.3-0.6) and surface temperature (peak in summer: 45°C).

[0041] Sensor network: 50 LoRaWAN sensors were deployed along the river valley to monitor soil moisture (5%-25%) and groundwater level (1-3m deep) in real time.

[0042] Geophysical data: High-density electrical surveys show that the bedrock surface is 5-15m deep, and ground-penetrating radar reveals the development of cracks in the weathering crust.

[0043] 2. Intelligent Preprocessing Laboratory data cleaning: Isolate forest marks three abnormal points with pH < 5 (later retests confirmed that one was an instrument error and the other two were acid pollution sources).

[0044] Drone image processing: YOLOv5 identifies sand-clay boundaries and corrects errors in traditional visual interpretation (original error 15% → corrected error 3%).

[0045] 3. 3D Modeling and Uncertainty Quantification 3D U-Net modeling: Fusion of multi-source data generates a 0-20m depth 3D matrix model, with a sand layer thickness prediction MAE of 0.5m (traditional method 1.2m).

[0046] Uncertainty analysis: MC Dropout shows that the confidence level of the bedrock interface is >90%, and the confidence level of the valley edge area is <70% (manual verification is required).

[0047] 4. Multi-scale nested analysis Macro-forecast: The CMIP6 climate model predicts a 12% expansion of sandy substrates by 2070 (from the current 60% to 72%).

[0048] Meso-level warning: areas with slopes >25° and loose matrix (accounting for 18% of the valley area) are marked as high-risk areas for soil erosion.

[0049] Micro-decision-making: 500 mu of land with organic matter >2% and thickness >1m was selected as “highly suitable arable land” and drought-resistant crops (such as barley) were recommended for planting.

[0050] 5. Ecological Restoration Simulation and Visualization Reinforcement learning recommendation: For sandy areas, "mixed planting of caragana and sea buckthorn plus gravel covering" is recommended. It is predicted that the organic matter will increase by 0.9% in 5 years (AR deduction shows that the coverage will increase from 20% to 50%).

[0051] WebGL platform: Users can slice the 3D model to query the attributes of any point (e.g., at a certain coordinate point: pH = 6.5, thickness = 1.2m, risk level = low).

[0052] 6. Full-process quality control and evidence storage AI quality inspection: GAN generates fuzzy samples of bedrock and weathering crust, detects model blind spots, and marks 10 areas requiring manual verification.

[0053] Blockchain evidence storage: Drilling records and repair plans are uploaded to the chain (hash value: a1b2c3...). An audit three years later showed that the data had not been tampered with.

[0054] Implementation effect: Improved accuracy: The F1-score for sandy matrix classification increased from 78% to 93%, and the thickness prediction error was reduced by 58%.

[0055] Restoration benefits: After the implementation of the mixed planting scheme of Caragana korshinskii, the vegetation coverage increased by 30% and the sand fixation capacity increased by 200 tons / year.

[0056] Management efficiency: Early warning response time was shortened from 72 hours to 6 hours, and cross-departmental collaboration costs were reduced by 40%.

[0057] Example 2: Ecological management of sandy substrates in the Kubuqi Desert in Inner Mongolia Technical implementation steps: 1. Dynamic Data Collection Satellite data: Access Sentinel-2 imagery, extract NDVI (0.1-0.3) and surface reflectance, and identify the boundaries of mobile sand dunes.

[0058] Sensor network: 100 solar-powered soil sensors are deployed to monitor sand moisture (<5%) and wind speed (average annual speed of 6 m / s) in real time.

[0059] Geophysical data: Ground penetrating radar revealed the thickness of the sand layer (2-8m), and the resistivity profile showed that the groundwater level was buried at a depth of >20m.

[0060] 2. Intelligent Preprocessing Drone inspection: A DJI M300 cruises to generate 5cm images, and YOLOv5 identifies Salix psammophila survival areas (with 95% accuracy).

[0061] Data augmentation: Generative Adversarial Network (GAN) synthesized virtual samples of the dune-vegetation transition zone and expanded the training set (sample size +50%).

[0062] 3. 3D Modeling and Decision-Making Sandy matrix model: 3D U-Net predicts the migration rate of mobile sand dunes (1.5 m / year ± 0.2 m) and guides the layout of sand fixation grids.

[0063] Reinforcement learning recommendation: The model outputs the "Salix tamarisk grid (5m×5m) + grass grid" solution, predicting that the vegetation coverage will increase from 10% to 35% in five years.

[0064] 4. AR Dynamic Display and Feedback Hololens 2 simulation: Superimposing the virtual tamarisk planting effect on the actual site shows that the sand dune height decreased by 1.2m after three years.

[0065] Dynamic monitoring: Sensors trigger retesting in areas where the Salix psammophila survival rate is less than 30%, and the model adds the "drip irrigation + organic fertilizer" measure.

[0066] 5. Full-process quality control Blockchain evidence storage: Sand fixation solutions and monitoring data are uploaded to the chain (hash value: d4e5f6...), supporting carbon trading audits.

[0067] AI quality inspection: Isolation Forest detected three sensor data anomalies (humidity mutations), confirmed them as equipment failures, and replaced them.

[0068] Implementation effect: Ecological restoration: After 5 years, the area of ​​mobile sand dunes will be reduced by 40%, and the carbon sequestration capacity of vegetation will reach 2.5 tons / hectare / year.

[0069] Economic benefits: The cumulative income from carbon trading is RMB 1.2 million per year, and herders’ income from desertification control services has increased by 30%.

[0070] Technology expansion: The model was extended to the Maowusu Desert, and restoration efficiency increased by 50%.

[0071] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A surface matrix map modification data processing method; characterized by: The following steps are involved: S11: Dynamic data collection, real-time integration of laboratory data, satellite data, sensor data and geophysical data; S12: Intelligent preprocessing, using AI to automatically clean outliers and dynamically update the database; S13: 3D modeling, based on 3D U-Net fusion of multimodal data, generates high-precision 3D matrix attribute volumes and quantifies prediction uncertainty; S14: Multi-scale nested analysis, coupling basin-level macro-forecasting with parcel-level refined modeling; S15: Decision deduction, simulating the effects of repair plans through reinforcement learning to generate the optimal repair plan; S16: Dynamic visualization, building a WebGL interactive platform and knowledge graph to achieve 3D sectioning query and intelligent repair solution recommendation.

2. A surface matrix map modification data processing method according to claim 1, characterized in that: When integrating laboratory data, satellite data, sensor data, and geophysical data in real time, specifically: A11: Laboratory data, which experimentally determines the physical and chemical characteristics of different types of surface substrates; A12: Satellite data stream, access to Planet Labs daily 3m resolution imagery, including NDVI and land surface temperature data; A13: IoT sensors, deploying soil temperature, humidity, and groundwater level sensor networks, and using LoRaWAN to transmit real-time data. A14: Geophysical data, integrating high-density electrical method and ground penetrating radar data to obtain bedrock surface depth and fracture distribution.

3. A surface matrix map modification data processing method according to claim 2, characterized in that: When using AI to automatically clean outliers and dynamically update the database, the specific steps for laboratory data include: S21: Data standardization, Z-score standardization of laboratory data, and unified data format and dimension; S22: Outlier marking, using the isolation forest algorithm to automatically identify outliers, where the parameters of the isolation forest algorithm are set as: the number of trees is 100 and the contamination ratio is set to 0.05; S23: Data correction: laboratory retesting of high-risk outlier points marked by the algorithm, and interpolation of adjacent points for non-critical fields.

4. A surface matrix map modification data processing method according to claim 3, characterized in that: When using AI to automatically clean outliers and dynamically update the database, for satellite data, the specific steps include: S31: UAV data acquisition, based on the preset map, automatically generates a cruise route and uses the camera to collect surface images in the target area; S32: Image preprocessing: Agisoft Metashape is used to perform brightness homogenization and eliminate shadow effects. Then, orthophotos are generated and registered to the CGCS2000 coordinate system. S33: Boundary recognition, using the YOLOv5 algorithm to analyze the registered image and identify the bounding box and category probability in the image.

5. A surface matrix map modification data processing method according to claim 4, characterized in that: When using AI to automatically clean outliers and dynamically update the database, the following steps are also included: S41: Spatial alignment, registering geophysical data and remote sensing images using control points; S42: Data enhancement, based on known data, uses a generative adversarial network to synthesize virtual samples in sparsely sampled areas; S43: Attribute association, associates the laboratory data with the drone-annotated polygons through spatial join to generate an attribute table.

6. A surface matrix map modification data processing method according to claim 5, characterized in that: When 3DU-Net is used to fuse multimodal data, generate high-precision 3D matrix attributes and quantify prediction uncertainty, it specifically includes: S51: Data preparation and fusion, processing the preprocessed data to generate a multi-channel three-dimensional grid stack to build the model input; S52: Model construction and training, using a 3D U-Net architecture with a 4-level encoder and decoder. Multi-scale features are preserved through skip connections. Transfer learning is introduced during model initialization, pre-trained weights from global borehole data are loaded, and some layers are frozen to accelerate fine-tuning of local data, generating high-resolution 3D attribute volumes. S53: Uncertainty quantification, using the MC Dropout technique in Bayesian deep learning, performs 50 random forward propagations during the test phase, calculates the mean and standard deviation of each voxel attribute, and generates a 95% confidence interval; S54: Model validation and application, using an independent test set to verify model accuracy, where evaluation metrics include F1-score and MAE; S55: Iterative optimization and deployment. Based on active learning strategies, new drilling data is added to high-uncertainty areas, the model is retrained, and the model is compressed into a lightweight version through knowledge distillation technology for deployment.

7. A surface matrix map modification data processing method according to claim 6, characterized in that: When coupling basin-level macro-forecasting with plot-level refined modeling, it specifically includes: S61: Macro-level climate prediction and downscaling, focusing on large-scale long-term trends, integrating global climate models with basin geological data, simulating matrix weathering trends in target areas through process-driven models, and using dynamic downscaling techniques to refine climate data to 1km grids; S62: Real-time dynamic risk monitoring at the meso-level, integrating macro-level outputs with real-time sensor data to build dynamic risk models and generate early warnings in real time; S63: High-precision decision support at the micro level, based on high-precision models, integrating drone imagery and laboratory data, extracting plot-level attributes, and predicting changes in water holding capacity in combination with macroclimate trends; S64: Cross-scale feedback and iterative optimization, establish a bottom-up feedback mechanism, use plot-level measured data to reversely correct meso-level parameters and macro-models, and achieve multi-source data assimilation through ensemble Kalman filtering to improve the overall accuracy of the model.

8. A surface matrix map modification data processing method according to claim 7, characterized in that: When simulating the effects of repair solutions through reinforcement learning and generating the optimal repair solution, the following steps are specifically included: S71: Knowledge graph construction, integrating substrate properties, historical restoration cases, and ecological response data, constructing knowledge graph nodes and defining associations to form a structured decision-making knowledge base; S72: Reinforcement learning environment modeling, defining environmental states and optional repair actions, setting reward functions, and quantifying the balance mechanism between ecological and economic goals; S73: Reinforcement learning model training, using the PPO algorithm to input historical case training models, simulate long-term responses to different measures, output the optimal strategy, and verify the reliability of the solution through Monte Carlo simulation; S74: AR virtual-reality fusion deduction, loading the 3D model of the plot through Hololens 2, superimposing the virtual restoration plan in real time, and sliding the timeline to dynamically display the restoration effect after a certain period; S75: Dynamic monitoring and feedback optimization: deploy sensors to monitor soil indicators in the restoration area, trigger model re-evaluation when data anomalies occur, and update the knowledge graph and fine-tune the model every month.

9. A surface matrix map modification data processing method according to claim 8, characterized in that: After building a WebGL interactive platform and knowledge graph to implement 3D sectioning queries and intelligent repair solution recommendations, AI full-scale detection and blockchain evidence storage are also used to ensure data quality and process traceability. The specific steps include: A21: AI quality inspection closed loop, which uses AI algorithms to test data, detect abnormal data, and generate data repair tasks; A22: Blockchain verification: key data is hashed and uploaded to the blockchain, verified through the PBFT consensus mechanism, and the entire data process is traced through timestamps and hash values.

10. A surface matrix map modification data processing method according to claim 9, characterized in that: When using AI algorithms to detect data, detect abnormal data, and generate data repair tasks, it specifically includes: S81: Automatically detect laboratory data anomalies using the isolation forest algorithm, associate the GPS coordinates of the anomaly points to generate an early warning layer, and dispatch drone retests to verify data authenticity; S82: Use GAN to generate adversarial examples to simulate model error-prone scenarios, detect areas where the confidence level of the prediction results is below a threshold, automatically mark high-uncertainty areas, and prioritize manual verification by risk level.

Images