Natural resource dynamic monitoring method based on deep learning

Abstract

The invention discloses a natural resource dynamic monitoring method based on deep learning, and relates to the technical field of natural resource monitoring, and the method comprises the steps: dividing a target region into a basic geographic unit, periodically updating a monitoring demand index, and grading the monitoring demand index into a core unit, an important unit and a conventional unit; differentiated scene pre-training is carried out for the monitoring priority and task characteristics of each unit, resource investment is accurately matched, and the contradiction between precision and cost and the contradiction between comprehensiveness and pertinence in traditional monitoring are effectively balanced. Meanwhile, each unit serves as a node, multi-dimensional feature vectors are extracted, edge connection is judged in combination with spatial association and attribute similarity, weights are given, deep feature fusion is achieved through an exclusive aggregation function and an association conduction mechanism, spatial linkage and time continuity of dynamic changes of natural resources are captured, one-sidedness of local analysis is avoided, and the accuracy of feature fusion is improved. And the integrity and reliability of monitoring results are improved.

Description

Technical Field

[0001] This invention relates to the field of natural resource monitoring technology, and more specifically, to a method for dynamic monitoring of natural resources based on deep learning. Background Technology

[0002] In the field of dynamic monitoring of natural resources, traditional monitoring methods often lack a scientific and refined division of monitoring areas, and mostly adopt a "uniform" monitoring model across the entire region. These methods fail to fully consider the fundamental differences in ecological value, degree of disturbance, and management needs among different regions, and fail to establish a resource input mechanism that matches monitoring needs. This leads to an imbalance in the allocation of monitoring resources—either insufficient investment in high-value, high-demand areas, making it difficult to meet the requirements of accurate monitoring; or excessive consumption of resources in low-demand areas, resulting in wasted monitoring costs and an inability to achieve an optimal balance between monitoring effectiveness and cost.

[0003] Traditional monitoring technologies lack targeted adaptation designs in the model training stage, typically employing a single model or uniform training standards to address various monitoring scenarios. Due to significant differences in the complexity, accuracy requirements, and data characteristics of monitoring tasks across different regions, a unified model architecture and training strategy struggles to simultaneously adapt to multiple needs: complex models generate redundant computations when used in routine monitoring areas, while lightweight models fail to capture fine-grained features when used in core monitoring areas. This results in monitoring results that are either insufficiently accurate or inefficient, making it difficult to simultaneously address monitoring objectives at different levels.

[0004] In terms of feature processing and monitoring analysis, traditional methods often focus on interpreting local data from individual monitoring points or isolated areas, neglecting the spatial interconnectedness and temporal continuity of dynamic changes in natural resources. These methods fail to effectively integrate inter-regional correlations and lack consistent tracking of cross-period trends, resulting in monitoring results limited to local state descriptions and unable to reflect coordinated changes across the entire region. Furthermore, due to the lack of a scientific feature fusion and transmission mechanism, the multidimensional value of monitoring data is not fully explored, easily leading to biased results and insufficient reliability, making it difficult to support refined management and dynamic protection decisions for natural resources. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a method for dynamic monitoring of natural resources based on deep learning.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for dynamic monitoring of natural resources, comprising the following steps: S1: For the natural resource target monitoring area, the natural resource target monitoring area is divided into multiple basic geographic units, and the monitoring demand index of each basic geographic unit is updated periodically. S2: Based on the monitoring demand index, the corresponding basic geographic units are divided into core monitoring units, key monitoring units and regular monitoring units. The model initialization for scenario pre-training is performed on the core monitoring units, key monitoring units and regular monitoring units. Then, differentiated scenario pre-training is performed on the core monitoring units, key monitoring units and regular monitoring units. The monitoring needs of the core monitoring units, key monitoring units and regular monitoring units are met through differentiated training. S3: Treat each core monitoring unit, key monitoring unit, and regular monitoring unit as a node, obtain the node feature vector of each node, compare all nodes pairwise, determine whether there is an edge connection between the two compared nodes, and if there is an edge connection between the two compared nodes, determine the corresponding edge weight. S4: Design aggregation functions and correlation transmission mechanisms to output the monitoring results of core monitoring units, key monitoring units, and regular monitoring units.

[0007] Furthermore, the monitoring demand index of a basic geographic unit is obtained as follows: The sensitivity score of the monitoring object for a basic geographic unit is obtained. and real-time requirement score Then, the data quality weight Qw of the basic geographic unit is obtained through the formula. The monitoring demand index of this basic geographic unit was calculated. ; , All are weighting coefficients.

[0008] Furthermore, the sensitivity score of the monitoring objects in the basic geographic unit. The method for obtaining the index is as follows: determine the ecological type of the basic geographic unit, determine the ecological type index LW based on the ecological type, and then use the formula... The sensitivity score of the monitored objects for this basic geographic unit was calculated. ; This is the threshold for the ecological type index.

[0009] Furthermore, the real-time requirement score for basic geographic units. The acquisition method is as follows: Determine the real-time demand type of the monitoring data corresponding to the basic geographic unit, determine the demand type score YT based on the real-time demand type, and then use the formula... The real-time requirement score of this basic geographic unit is calculated. ; This is the threshold for the demand type score.

[0010] Furthermore, the data quality weight Qw of the basic geographic unit: This is obtained by calculating the image effectiveness rate le and sensor normality rate sc corresponding to the basic geographic unit, using the formula... The data quality weight Qw of this basic geographic unit is calculated; y1 and y2 are both weight coefficients.

[0011] Furthermore, based on the monitoring demand index, the corresponding basic geographic units are divided into core monitoring units, key monitoring units, and routine monitoring units, as follows: when the monitoring demand index is higher than the upper threshold of monitoring demand, the corresponding basic geographic unit is classified as a core monitoring unit; when the monitoring demand index is between the upper threshold and the lower threshold of monitoring demand, the corresponding basic geographic unit is classified as a key monitoring unit; and when the monitoring demand index is lower than the lower threshold of monitoring demand, the corresponding basic geographic unit is classified as a routine monitoring unit.

[0012] Furthermore, edge weights ; Spatial association weights, based on a distance decay model: , The distance between the center points of the two units; Attribute association weight, i.e., node feature vector similarity. ; This is the balance coefficient.

[0013] Furthermore, aggregate functions ; Let i be the node feature vector of node i in the l-th layer; A weight matrix specific to the unit type; Let i be the set of neighboring nodes of node i; These are the edge weights already calculated in S2; The node type weight coefficient highlights the transmission priority of core unit characteristics; Use an activation function to avoid gradient vanishing; To perform a summation operation on all neighboring nodes j of node i; Let be the input feature vector of the neighboring node j in the l-th layer network; Definition: For all neighboring nodes k of node i Sum.

[0014] Furthermore, the correlation propagation mechanism: the triggering logic for spatial propagation: during each round of GNN forward computation, based on the edge weights... It automatically completes the feature propagation in the neighborhood, and the higher the weight of the edge, the stronger the feature propagation.

[0015] Furthermore, the time-based transmission formula incorporates the node characteristics of the previous cycle into the current cycle, ensuring the continuity of the trend: ; Let i be the initial feature vector of the GNN in the current period t; For time propagation weighting coefficients; The final aggregated feature vector of the GNN for node i in the previous period t−1; Let be the original feature vector of node i in the current period t.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The method of this invention adapts resource investment to the essential differences in monitoring needs, while differentiated scenario pre-training further enhances adaptability. For different levels of units, it completes model initialization and customized training, ensuring both monitoring depth in high-value areas and monitoring efficiency in conventional areas. This effectively solves the problem of balancing accuracy and cost, comprehensiveness and specificity in traditional monitoring. By extracting multi-dimensional node feature vectors, combining spatial correlation and attribute similarity to determine edge connections and assign reasonable edge weights, and then achieving deep feature fusion through a dedicated aggregation function and correlation transmission mechanism, this design not only allows monitoring to move beyond isolated analysis of individual units, but also captures the spatial linkage and temporal continuity of natural resource dynamics through priority transmission of neighborhood features and effective connection of cross-period time-series features. This enables monitoring results to reflect the overall coordinated change pattern, avoiding the one-sidedness caused by local analysis and significantly improving the completeness and reliability of monitoring results. The scientific division of basic geographic units and the periodic updates of requirements ensure that monitoring resources are tilted towards high-value areas, achieving a precise match between monitoring needs and technological investment. On the other hand, node-based modeling and multi-dimensional feature aggregation and transmission endow monitoring results with stronger dynamic adaptability and logical correlation. Overall, this not only improves the monitoring accuracy of core areas but also ensures the efficiency and continuity of monitoring across the entire region, enhancing the scientific nature of dynamic monitoring of natural resources. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for dynamic monitoring of natural resources based on deep learning. Figure 2 Flowchart for initializing a model pre-trained for a scene. Detailed Implementation

[0018] Reference Figures 1-2 A deep learning-based method for dynamic monitoring of natural resources is proposed, and the method steps are as follows: S1: For the natural resource target monitoring area, the natural resource target monitoring area is divided into multiple basic geographic units, and the monitoring demand index of each basic geographic unit is updated periodically (the periodic interval for updating the basic geographic unit is set based on the minimum satellite revisit period).

[0019] Based on the monitoring demand index, the corresponding basic geographic units are divided into core monitoring units, key monitoring units, and routine monitoring units, as follows: When the monitoring demand index is higher than the upper threshold of monitoring demand, the corresponding basic geographic unit is classified as a core monitoring unit; when the monitoring demand index is between the upper and lower thresholds of monitoring demand, the corresponding basic geographic unit is classified as a key monitoring unit; when the monitoring demand index is lower than the lower threshold of monitoring demand, the corresponding basic geographic unit is classified as a routine monitoring unit (the upper and lower thresholds of monitoring demand are based on historical data). The data distribution range is determined (e.g., by using the quantile method: take the 75th quantile of the historical Mdi as the upper threshold and the 25th quantile as the lower threshold).

[0020] The natural resource target monitoring area is divided into multiple basic geographic units, as follows: An electronic base map of the natural resource target monitoring area (including latitude and longitude, topography, administrative boundaries, etc.) is imported using GIS technology (such as ArcGIS and QGIS) to accurately define the geographical scope of the entire area; core auxiliary data are collected: natural geographic identifiers (mountains, rivers, lakes, etc.), management boundaries (townships, forest farms, watersheds, etc.), and existing monitoring point distribution data to avoid the units being disconnected from actual management needs; following the principles of geographic continuity, granularity adaptation, and management convenience: ① Granularity is set according to monitoring accuracy requirements (10km×10km grid for routine monitoring, 2-5km×2-5km grid for high-precision monitoring); ② Priority is given to conforming to natural / management boundaries, avoiding units crossing mountains, rivers, or administrative regions. The mainstream approach to partitioning is the grid method combined with boundary adaptation: regular square / rectangular grids are generated in GIS, overlaid with the boundary of the target area, and then cropped to form basic geographic units; special area adaptation: for areas with complex terrain (such as mountainous areas), the natural boundary method is used (split by mountains and rivers); for areas with clear management boundaries (such as forest farms and nature reserves), the management boundary is directly used as the boundary of the basic geographic unit.

[0021] The monitoring demand index of a basic geographic unit is obtained as follows: The sensitivity score of the monitoring object for a basic geographic unit is obtained. and real-time requirement score Then, the data quality weight Qw of the basic geographic unit is obtained through the formula. The monitoring demand index of this basic geographic unit was calculated. ; , All are weighting coefficients. Because the sensitivity of monitored objects reflects the ecological value, scarcity, and vulnerability of natural resources, it is the fundamental basis for determining whether monitoring is necessary. The higher the ecological value, the stronger the core necessity of monitoring, and this attribute is long-term stable and irreplaceable. The real-time requirement, however, is about ensuring the monitoring is timely enough to serve risk management or dynamic tracking, and must be based on the premise that monitoring is worthwhile. The weight should be higher than ,so The value can be 0.6. The value can be 0.4.

[0022] Sensitivity score of monitoring objects in basic geographic units The acquisition method is as follows: Determine the ecological type of the basic geographic unit, and determine the ecological type index LW based on the ecological type (according to the natural hierarchical differences in the ecological value, scarcity, and vulnerability of natural resources, ecological types are divided into three types: highly sensitive ecological types (such as habitats of rare plants and animals, and core areas of drinking water sources, which are irreplaceable and vulnerable, and difficult to restore once damaged), moderately sensitive ecological types (such as general forests and agricultural irrigation areas, which have a certain self-repair capacity and require regular dynamic monitoring), and lowly sensitive ecological types (such as deserts and unused land with single ecological functions and little impact from disturbance). The ecological type index of highly sensitive ecological types > the ecological type index of moderately sensitive ecological types > the ecological type index of lowly sensitive ecological types. In this embodiment, the ecological type index of highly sensitive ecological types is 3, the ecological type index of moderately sensitive ecological types is 2, and the ecological type index of lowly sensitive ecological types is 1. This value allocation is used to accurately match the core characteristics of the three types of ecological types with an equidistant gradient. The management priority of core protection - routine management - low intensity attention is intuitively reflected by the numerical difference (3-2=1, 2-1=1). The sensitivity score of the monitored objects for this basic geographic unit was calculated. ; This is the threshold value for the ecological type index (i.e., the theoretical maximum value of the ecological type index; in this embodiment, the value is 3).

[0023] Real-time requirement score for basic geographic units The acquisition method is as follows: Determine the real-time demand type of the monitoring data corresponding to the basic geographic unit (the real-time demand type is determined by the core management objective corresponding to the monitoring data; if the monitoring data is used for emergency response to sudden risks (such as fire, flood, pollution spread), it is determined to be an emergency response type; if the monitoring data is used for tracking dynamic trends of resources (such as vegetation growth, water level changes, land use evolution), it is determined to be a regular dynamic type; if the monitoring data is used for "compliance statistics, reserve accounting or effectiveness verification" (such as mineral reserves, wetland area, arable land area), it is determined to be a periodic verification type). Determine the demand type score YT based on the real-time demand type (because the real-time demand of the emergency response type > the real-time demand of the regular dynamic type > the real-time demand of the periodic verification type, the demand type score of the emergency response type > the regular dynamic type). The demand type score is greater than the demand type score for periodic verification. If the monitoring data of a basic geographic unit includes multiple real-time demand types, the demand type score YT is the YT value corresponding to the highest real-time demand. For example, if the monitoring data of basic geographic unit 1 includes both emergency response and routine dynamic demand types, the demand type score YT is the YT value corresponding to the emergency response type. In this embodiment, the demand type score YT for emergency response is 3, the demand type score YT for routine dynamic demand types is 2, and the demand type score YT for periodic verification demand types is 1, using equidistant gradients to accurately match the core characteristics of the three types of real-time demand. If the monitoring data of basic geographic unit 1 includes both emergency response and routine dynamic demand types, the demand type score YT for basic geographic unit 1 is 3. The real-time requirement score of this basic geographic unit is calculated. ; This is the threshold for the demand type score (i.e., the theoretical maximum value of the demand type score; in this embodiment, the value is 3).

[0024] Data quality weight Qw for basic geographic units: This involves obtaining the image effectiveness rate le and sensor normality rate sc corresponding to the basic geographic unit. The steps for obtaining image effectiveness rate le are: acquiring satellite images of the basic geographic unit taken in the previous period, simultaneously retrieving meteorological data (such as precipitation and cloud cover) and regional topographic data (to assist in judging topographic shadows) at the time of satellite image capture, and automatically identifying invalid occluded pixels in the satellite images using remote sensing interpretation systems (such as ArcGIS and ENVI) or AI algorithms (such as threshold segmentation and semantic segmentation). This mainly includes cloud and rain occluded pixels (determined based on grayscale values ​​and humidity thresholds) and topographic shadow pixels (determined by combining topographic elevation data and solar altitude angle calculations). The total number of image pixels is counted (directly extracted from image metadata), and the number of unoccluded valid pixels is calculated using the formula... The calculation results show that the sensor normality rate (sc) is obtained through the following steps: Obtain the sensor logs of each sensor in the previous period for the basic geographic unit, and mark the corresponding period duration as... Then, the normal duration of each sensor is calculated (for each sensor, the cumulative abnormal duration of all its abnormal data periods is summed to obtain the cumulative abnormal duration of that sensor). Abnormal data periods are automatically determined based on three types of anomalies in the corresponding sensor logs: ① Disconnection period (continuous time without data upload); ② Value exceeding limits period (data values ​​exceeding preset reasonable ranges, such as soil moisture > 100%); ③ Signal drift period (data fluctuation amplitude exceeding normal thresholds, such as water level fluctuation > 0.5m within 1 minute). Normal duration for a single sensor: According to the formula Calculate the normal operation rate of a single sensor. If multiple sensors are deployed in the basic geographic unit, take the normal operation rate of all sensors. The arithmetic mean of the values ​​is taken as the final sensor normality (sc) of the unit, using the formula... The data quality weight Qw for this basic geographic unit is calculated; y1 and y2 are both weight coefficients, y1+y2=1, because the core value of satellite imagery data is macroscopic coverage, which can reflect the overall and holistic state of the basic geographic unit (such as vegetation cover change, land use pattern, and large-scale disaster risks), and is the core of obtaining overall monitoring data for the unit; the core value of ground sensor data is microscopic precision—it can provide fixed-point, real-time quantitative data (such as soil moisture, water level, and pollutant concentration), and is the key to verifying imagery data and capturing precise local changes. One is macroscopic and the other microscopic, one is holistic and the other is fixed-point; their functions are complementary and cannot be substituted for each other, and their impact on data quality is equally important. Therefore, the values ​​of y1 and y2 can both be set to 0.5.

[0025] S2: Based on the monitoring demand index, the corresponding basic geographic units are divided into core monitoring units, key monitoring units, and routine monitoring units. The model is initialized for scenario pre-training for core monitoring units, key monitoring units, and routine monitoring units. Then, differentiated scenario pre-training is performed for core monitoring units, key monitoring units, and routine monitoring units to meet the monitoring needs of core monitoring units, key monitoring units, and routine monitoring units through differentiated training.

[0026] The model initialization for scenario pre-training of core monitoring units, key monitoring units, and regular monitoring units is as follows: Determine the corresponding model, training parameter set, loss function, initial weights, and training termination condition for each core monitoring unit, key monitoring unit, and regular monitoring unit. (Due to fundamental differences in the monitoring requirements (accuracy / efficiency / cost), task complexity, and data characteristics of core monitoring units, key monitoring units, and regular monitoring units, the adaptation logic for the configuration differences is as follows: Model structure: Core units use complex models (such as DeepLabV3+) for fine-grained feature extraction; regular units use lightweight models (such as ShuffleNetV2) for fast inference; key monitoring units use medium complexity.) The training parameters are designed to match different task complexity levels. The core approach uses a small learning rate, multiple iterations, and small batches to avoid training oscillations and fully fit complex features. The standard approach uses a large learning rate, fewer iterations, and large batches for faster convergence and reduced computational cost, aligning with the priority of accuracy / efficiency. The loss function is designed to address imbalanced samples (e.g., low proportion of high-value targets) using combined loss (cross-entropy + Dice), while a simple loss is used for balanced samples, adapting to different data characteristics. Weight initialization is designed to load pre-trained remote sensing weights to improve generalization ability, while random initialization reduces data dependence, matching different requirements for training effectiveness. The training termination condition is designed to set a longer "no-improvement waiting period" (15 rounds) to ensure sufficient training, while a shorter waiting period (8 rounds) is used to avoid ineffective computation, adapting to different training sufficiency requirements.

[0027] Examples are as follows: Model: Core monitoring unit (high accuracy requirements, such as rare plant and animal habitat identification, small-scale disaster hazard monitoring): Select complex deep learning models (such as DeepLabV3+, ResNet50+U-Net), retain fully connected layers + multi-scale feature fusion modules to ensure fine-grained feature extraction capabilities; Key monitoring unit (balancing accuracy and efficiency, such as general forest cover change, farmland use monitoring): Select medium-complexity models (such as MobileNetV3+U-Net lightweight version, SegNet), simplify some redundant convolutional layers, and retain key feature fusion modules; Regular monitoring unit (high efficiency requirements, such as desert, unused land status verification): Select lightweight models (such as ShuffleNetV2+ simple segmentation head, CNN basic models), remove complex feature fusion modules, and only retain the core convolution-pooling structure to reduce computational overhead.

[0028] Training parameter set: A strong data augmentation strategy is adopted (random horizontal / vertical flip, rotation ±15°, brightness / contrast ±20% adjustment, Gaussian noise addition), while pixel-level cropping is set (crop size 512×512, overlap rate 30%) to improve the model's generalization ability; Key monitoring unit: A medium data augmentation strategy is adopted (random horizontal flip, rotation ±10°, brightness / contrast ±15% adjustment), crop size 512×512, overlap rate 20%; Regular monitoring unit: A weak data augmentation strategy is adopted (random horizontal flip only), crop size 256×256, no overlap, to speed up data preprocessing.

[0029] Loss function: Core monitoring units (which may have imbalanced samples, such as a low proportion of rare species areas): a weighted combination of "cross-entropy loss + Dice loss" (weight ratio 7:3) is used to alleviate model bias caused by sample imbalance; Key monitoring units (relatively balanced sample distribution): standard cross-entropy loss is used; Regular monitoring units (simple tasks, balanced samples): simplified cross-entropy loss (removing complex regularization terms) is used to speed up computation.

[0030] Initializing weights: Core monitoring unit: Load pre-trained weights from remote sensing images (such as ResNet50 weights trained based on Sentinel-2 satellite images), fine-tune only the fully connected layer and output layer, and improve training effect by utilizing transfer learning; Key monitoring unit: Load pre-trained weights from general images (such as ImageNet pre-trained weights), and fine-tune the intermediate layer and output layer; Regular monitoring unit: Use He normal distribution to randomly initialize weights, no pre-training loading is required, reducing data dependence.

[0031] Training termination conditions: Core monitoring unit: terminates when the validation set IoU (Intersection over Union) shows no improvement for 15 consecutive rounds, or when the number of iterations reaches 200 rounds; Key monitoring unit: terminates when the validation set IoU shows no improvement for 10 consecutive rounds, or when the number of iterations reaches 150 rounds; Regular monitoring unit: terminates when the validation set accuracy shows no improvement for 8 consecutive rounds, or when the number of iterations reaches 100 rounds.

[0032] S3: Treat each core monitoring unit, key monitoring unit, and routine monitoring unit as a node, and obtain the node feature vector for each node (example of node feature vector: i = 1, 2, ..., N, where N is the total number of core monitoring units, key monitoring units, and regular monitoring units; This is the intermediate feature vector of the i-th unit (such as the last convolutional feature of DeepLabV3+, with a dimension of 256). Assign unit type identifiers (unit type identifier = 3 for core monitoring units, unit type identifier = 2 for key monitoring units, and unit type identifier = 1 for regular monitoring units, used to quantify unit priority) to perform pairwise comparisons of all nodes, determining whether there is an edge connection between each pair of nodes (there are two types of associated edges between nodes, and an edge connection is established if either condition is met: spatial proximity edge: the units corresponding to the two nodes share a physical boundary (such as adjacent grids); attribute similarity edge: the similarity of the node feature vectors of the two nodes is calculated ( If the similarity of node feature vectors is higher than the similarity threshold, it is determined that an edge connection exists. If the two compared nodes are connected by an edge, the corresponding edge weight is determined. ; Spatial association weights, based on a distance decay model: ( (distance between the center points of the two units); Attribute association weight, i.e., node feature vector similarity. ; The balance coefficient (because cross-unit changes in natural resources (such as fire spread, flood diffusion, and pollution migration) are essentially processes dominated by spatial correlation, attribute similarity (such as consistent ecological type and monitoring needs) can help screen functionally related units (such as cultivated land units in the same watershed), therefore The value can be 0.7).

[0033] S4: Design aggregation functions and correlation propagation mechanisms (aggregation functions) ; The node feature vector of node i in layer l (initially the node feature vector of S3). ); A weight matrix specific to each unit type is used (a high-dimensional matrix (256×256) is used for core monitoring units to ensure feature fitting, while a low-dimensional matrix (64×64) is used for regular monitoring units to improve efficiency). Let i be the set of neighboring nodes of node i (i.e., all nodes connected by edges). These are the edge weights already calculated in S3; The node type weight coefficients (the node type weight coefficients adopt equal gradient differences to ensure that the priority differences of the three types of units are significant (avoiding feature confusion caused by similar weights) and to prevent the features of the regular monitoring units from becoming completely ineffective due to excessive gradients. In this embodiment, the node type weight coefficient corresponding to the core monitoring unit is 1, the node type weight coefficient corresponding to the key monitoring unit is 0.7, and the node type weight coefficient corresponding to the regular monitoring unit is 0.4) highlight the transmission priority of the core unit features. Use an activation function to avoid gradient vanishing; To perform a summation operation on all neighboring nodes j of node i; Let be the input feature vector of the neighboring node j in the l-th layer network; Definition: For all neighboring nodes k of node i Summation (normalization); Correlation propagation mechanism: Triggering logic for spatial propagation: During each round of GNN forward computation, based on edge weights... The system automatically performs neighborhood feature propagation; edges with higher weights (such as adjacent core units) exhibit greater feature propagation strength. The time-based propagation formula integrates node features from the previous cycle into the current cycle, ensuring the continuity of the trend. ; The initial feature vector of node i in the current period t (as input to the 0th layer of the GNN); The time-transmission weighting coefficient is set based on experience in natural resource monitoring scenarios. The core logic is to prioritize retaining current real-time features (70%) while also considering historical trends (30%) to avoid over-reliance on history leading to monitoring delays. The value can be set to 0.3). Let L be the final aggregated feature vector of the GNN for node i in the previous period t−1 (L is the total number of layers in the GNN). Given the original feature vector of node i in the current period t (the basic features that have not been aggregated by GNN), the monitoring results of the core monitoring unit, key monitoring unit and regular monitoring unit are output.

[0034] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0035] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic monitoring of natural resources based on deep learning, characterized in that, A deep learning-based method for dynamic monitoring of natural resources, comprising the following steps: S1: For the natural resource target monitoring area, the natural resource target monitoring area is divided into multiple basic geographic units, and the monitoring demand index of each basic geographic unit is updated periodically. S2: Based on the monitoring demand index, the corresponding basic geographic units are divided into core monitoring units, key monitoring units and regular monitoring units. The model initialization for scenario pre-training is performed on the core monitoring units, key monitoring units and regular monitoring units. Then, differentiated scenario pre-training is performed on the core monitoring units, key monitoring units and regular monitoring units. The monitoring needs of the core monitoring units, key monitoring units and regular monitoring units are met through differentiated training. S3: Treat each core monitoring unit, key monitoring unit, and regular monitoring unit as a node, obtain the node feature vector of each node, compare all nodes pairwise, determine whether there is an edge connection between the two compared nodes, and if there is an edge connection between the two compared nodes, determine the corresponding edge weight. S4: Design aggregation functions and correlation transmission mechanisms to output the monitoring results of core monitoring units, key monitoring units, and regular monitoring units.

2. The method for dynamic monitoring of natural resources based on deep learning according to claim 1, characterized in that, The monitoring demand index of a basic geographic unit is obtained as follows: The sensitivity score of the monitoring object for a basic geographic unit is obtained. and real-time requirement score Then, the data quality weight Qw of the basic geographic unit is obtained through the formula. The monitoring demand index of this basic geographic unit was calculated. ; , All are weighting coefficients.

3. The method for dynamic monitoring of natural resources based on deep learning according to claim 2, characterized in that, Sensitivity score of monitoring objects in basic geographic units The method for obtaining the index is as follows: determine the ecological type of the basic geographic unit, determine the ecological type index LW based on the ecological type, and then use the formula... The sensitivity score of the monitored objects for this basic geographic unit was calculated. ; This is the threshold for the ecological type index.

4. The method for dynamic monitoring of natural resources based on deep learning according to claim 2, characterized in that, Real-time requirement score for basic geographic units The acquisition method is as follows: Determine the real-time demand type of the monitoring data corresponding to the basic geographic unit, determine the demand type score YT based on the real-time demand type, and then use the formula... The real-time requirement score of this basic geographic unit is calculated. ; This is the threshold for the demand type score.

5. The method for dynamic monitoring of natural resources based on deep learning according to claim 2, characterized in that, Data quality weight Qw for basic geographic units: This is calculated using the image effectiveness rate le and sensor normality rate sc corresponding to the basic geographic unit, obtained through the formula... The data quality weight Qw of this basic geographic unit is calculated; y1 and y2 are both weight coefficients.

6. The method for dynamic monitoring of natural resources based on deep learning according to claim 1, characterized in that, Based on the monitoring demand index, the corresponding basic geographic units are divided into core monitoring units, key monitoring units, and regular monitoring units, as follows: when the monitoring demand index is higher than the upper threshold of monitoring demand, the corresponding basic geographic unit is classified as a core monitoring unit; when the monitoring demand index is between the upper threshold and the lower threshold of monitoring demand, the corresponding basic geographic unit is classified as a key monitoring unit; and when the monitoring demand index is lower than the lower threshold of monitoring demand, the corresponding basic geographic unit is classified as a regular monitoring unit.

7. The method for dynamic monitoring of natural resources based on deep learning according to claim 1, characterized in that, edge weight ; Spatial association weights, based on a distance decay model: , The distance between the center points of the two units; Attribute association weight, i.e., node feature vector similarity. ; This is the balance coefficient.

8. The method for dynamic monitoring of natural resources based on deep learning according to claim 1, characterized in that, aggregate functions ; Let i be the node feature vector of node i in the l-th layer; A weight matrix specific to the unit type; Let i be the set of neighboring nodes of node i; These are the edge weights already calculated in S2; The node type weight coefficient highlights the transmission priority of core unit characteristics; Use an activation function to avoid gradient vanishing; To perform a summation operation on all neighboring nodes j of node i; Let be the input feature vector of the neighboring node j in the l-th layer network; Definition: For all neighboring nodes k of node i Sum.

9. The method for dynamic monitoring of natural resources based on deep learning according to claim 1, characterized in that, Association propagation mechanism: Triggering logic for spatial propagation: During each round of GNN forward computation, based on edge weights... It automatically completes the feature propagation in the neighborhood, and the higher the weight of the edge, the stronger the feature propagation.

10. A method for dynamic monitoring of natural resources based on deep learning according to claim 9, characterized in that, The time-based transmission formula incorporates the node characteristics of the previous cycle into the current cycle, ensuring the continuity of the trend: ; Let i be the initial feature vector of the GNN in the current period t; For time propagation weighting coefficients; The final aggregated feature vector of the GNN for node i in the previous period t−1; Let be the original feature vector of node i in the current period t.

Images