Intelligent agricultural intelligent monitoring system and method based on big data

Abstract

The invention discloses an intelligent agricultural intelligent monitoring system and method based on big data, relates to the technical field of agricultural intelligent monitoring, and aims at the core defect of a terrace monitoring scene, through curvature analysis and a microclimate attenuation model, the monitoring of a physically unreachable area is converted into a computable problem of a digital twin space, and the real-time monitoring of the terrace monitoring scene is realized. The sensor deployment requirements of abrupt slopes and ridges are obviously reduced; the terrain continuity coefficient accurately depicts the integrity of the stepped structure, and the humidity gradient quantifies the environment propagation rule of the adjacent plots, so that the virtual sensor network can dynamically generate blind area soil moisture content data, and the terrain adaptability defect of an OODA closed loop is overcome; besides, physical nodes and virtual nodes are integrated by adopting a heterogeneous topological graph, effective neighborhood information is dynamically screened by a graph attention mechanism through spatial attenuation and microclimate similarity, and strong dependence of traditional federal learning on data homogenization is avoided.

Description

Technical Field

[0001] The present invention relates to the field of agricultural intelligent monitoring technology, and in particular to a smart agricultural intelligent monitoring system and method based on big data. Background Art

[0002] In recent years, smart agricultural monitoring systems have achieved remarkable results in large-scale plain farmland through the integration of the Internet of Things, drone aerial photography, and artificial intelligence technologies. Mainstream solutions rely on fixed sensor networks and multispectral imaging technology to achieve real-time monitoring of parameters such as crop growth and soil moisture. At the algorithm level, most of them are multimodal fusion frameworks based on deep learning.

[0003] However, in special terrain scenarios such as terraces and mountains, the stepped structure of terraces limits the density of sensor deployment, forming monitoring blind spots in areas such as ridges and steep slopes. Traditional drone inspections are prone to image distortion and data faults due to the undulating terrain. The latest research reconstructs terrain features through three-dimensional point cloud modeling (such as the digital twin solution proposed in "Fuxi Farm"), but the modeling process relies on high-precision lidar, and the cost and computing power thresholds hinder large-scale application. In addition, existing algorithms are mostly trained based on plain data sets, and lack adaptability to characteristics such as crop shadow occlusion and microclimate mutations in terraced scenes, resulting in a high false alarm rate for pest warnings.

[0004] To address the monitoring deficiencies caused by terrain complexity, some solutions use dynamic path planning algorithms to optimize drone flight trajectories through reinforcement learning to improve the integrity of image acquisition in steep slope areas; or use lightweight three-dimensional reconstruction technology (such as the agricultural-adapted version of the Neural Radiation Field (NeRF)) to generate terrain models based solely on visible light images, reducing hardware dependence; however, such solutions rely on drone hardware upgrades, and lightweight reconstruction models are not stable enough under complex lighting conditions. Federated learning faces the real problem of farmers' low willingness to share data. Therefore, there is an urgent need for a smart agricultural intelligent monitoring solution based on big data to solve such problems. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a big data-based intelligent monitoring system and method for smart agriculture to solve the problems that terraced field monitoring is limited by the complexity of the terrain, and existing solutions have sensor blind spots, inefficient data fusion, and insufficient model generalization.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides a smart agricultural intelligent monitoring system based on big data, which includes:

[0009] The terrain feature migration module is used to build a 3D digital twin model based on historical aerial data of the target area and extract terrain-related features;

[0010] A virtual sensor network module generates predicted environmental parameters of areas where no physical sensors are deployed based on the terrain association features and the graph attention network;

[0011] A multi-region federated learning module aggregates monitoring data from multiple adjacent terraced fields and dynamically updates the weight distribution of the terrain-related features.

[0012] Edge computing nodes are deployed on drones or fixed base stations to perform lightweight inference tasks of the virtual sensor network module.

[0013] As a preferred solution of the intelligent monitoring system for smart agriculture based on big data described in the present invention, the terrain feature migration module includes:

[0014] 3D point cloud reconstruction unit, which generates terrace terrain surface model based on visible light image sequence;

[0015] Microclimate correlation analysis unit, which explores the nonlinear mapping relationship between soil moisture, light intensity and terrain elevation;

[0016] The extraction of the terrain-related features includes:

[0017] Perform curvature analysis on the terrace steps to generate terrain continuity coefficient;

[0018] A microclimate propagation attenuation model is constructed based on the soil moisture gradient differences between adjacent steps.

[0019] As a preferred solution of the intelligent monitoring system of smart agriculture based on big data described in the present invention, in the process of extracting terrain-related features in the terrain feature migration module, the three-dimensional point cloud E i =(x i ,y i ,z i ) Establish the neighborhood covariance matrix and construct the local curvature according to the eigenvalue ratio:

[0020]

[0021] in, is the curvature of point i, are the first, second, and third eigenvalues ​​of the covariance matrix, in units of m 2 , x i ,y i ,z i Represents coordinates;

[0022] Filter a collection of step edges using a threshold:

[0023]

[0024] in, is the curvature threshold, and the mean and variance are calculated on ε:

[0025]

[0026] in, is the mean curvature, is the curvature variance, |ε| is the number of edge points;

[0027] The terrain continuity coefficient is given based on statistics:

[0028]

[0029] Among them, G C is the continuity coefficient, ε is the stability constant;

[0030] Establish a gradient for the soil moisture difference between adjacent steps j and j+1:

[0031]

[0032] in, is the humidity gradient, in %m -1 , is the step j humidity, unit is %, is the elevation difference between the two steps, in m;

[0033] Summarizing all gradients yields the microclimate attenuation index:

[0034]

[0035] Among them, M β is the decay index, J is the total number of steps;

[0036] Finally, the terrain correlation feature vector F is constructed terrain :

[0037]

[0038] As a preferred solution of the intelligent monitoring system for smart agriculture based on big data described in the present invention, the virtual sensor network module is implemented in the following manner:

[0039] Construct a heterogeneous topology graph containing known sensor nodes and virtual nodes;

[0040] Use graph attention mechanism to dynamically calculate feature propagation weights between nodes;

[0041] Output the predicted values ​​of blind spot environmental parameters and confidence assessment results.

[0042] The heterogeneous topology graph includes:

[0043] Physical sensor nodes, including soil temperature and humidity, and light intensity detection units;

[0044] Virtual sensor nodes,associate terrain elevation with historical disaster event labels;

[0045] The edge weight is calculated by combining the Euclidean distance between nodes and the microclimate similarity.

[0046] As a preferred solution of the intelligent monitoring system for smart agriculture based on big data described in the present invention, the virtual sensor network module adopts a graph attention mechanism to dynamically calculate the feature propagation weights between nodes, including:

[0047] Construct concatenated features for physical and virtual nodes and calculate content relevance. The formula is:

[0048]

[0049] in, is the content relevance score of node i→j, N W is a trainable linear mapping matrix, dimension: d×2d, is the node feature vector, dimension d, [·‖·] represents the vector concatenation operation, Leaky ReLU activation function;

[0050] Combined with spatial attenuation:

[0051]

[0052] in, is the spatial attenuation coefficient, is the Euclidean distance between nodes, unit is m, γ is the adaptive decay exponent;

[0053] Adding microclimate similarity:

[0054]

[0055] in, is the humidity similarity, is the node humidity;

[0056] The attention weight is obtained comprehensively:

[0057]

[0058] in, is the normalized weight, is the neighborhood set of node i;

[0059] Update the node representation using weights:

[0060] in, is the updated feature vector.

[0061] As a preferred solution of the intelligent monitoring system for smart agriculture based on big data described in the present invention, the multi-region federated learning module includes:

[0062] Privacy protection unit, which performs differential privacy encryption on the terraced block data participating in federated learning;

[0063] The dynamic weight allocation unit adjusts the model aggregation weight according to the data volume of each block and the disaster history records.

[0064] In a second aspect, the present invention provides a smart agricultural intelligent monitoring method based on big data, comprising:

[0065] Step S1, collecting historical aerial images and physical sensor time series data of the target terraced area;

[0066] Step S2, adapting the plain farmland pre-training model to the terraced terrain characteristics through transfer learning;

[0067] Step S3: fusing multi-block data based on the federated learning framework and updating the parameters of the virtual sensor network;

[0068] Step S4: Dynamically adjust the prediction range of the virtual sensor node according to the real-time UAV inspection path.

[0069] As a preferred solution of the intelligent monitoring method of smart agriculture based on big data described in the present invention, the transfer learning adaptation process of step S2 includes:

[0070] Extract the unique crop shadow distribution pattern and microclimate mutation characteristics of terraced fields;

[0071] Inject terrain elevation constraints into the fully connected layer of the model;

[0072] Freeze the parameters of the basic feature extraction layer and only fine-tune the high-level decision network.

[0073] As a preferred solution of the intelligent monitoring method of smart agriculture based on big data described in the present invention, in step S2, the elevation information is introduced into the multiplicative gate in the fully connected layer, and the backbone network is set to output the feature vector at the grid point p.

[0074] To normalize the elevation:

[0075]

[0076] in, is the standardized elevation, H z is the original elevation of the grid point, in m, is the mean elevation of the target block, in m. is the elevation standard deviation, in meters, and ε is a numerical stability constant, which is 10 -6 ;

[0077] Generate the elevation gating vector:

[0078]

[0079] Among them, H g is the gate vector, dimension d f , is the elevation mapping weight matrix, size d f ×1, is the gate bias vector, dimension d f ,σ is the Sigmoid activation function;

[0080] Perform characteristic modulation and adjust the formula as follows:

[0081] F elev =F in ⊙(1+H g ),

[0082] Among them, F elev is the feature after elevation modulation, with dimension d f ⊙ is element-by-element multiplication, 1 is multiplication with H g All-1 vectors of the same dimension;

[0083] Perform constrained full connection mapping, the mapping formula is:

[0084]

[0085] Among them, Y is the output vector, dimension d o , is the fully connected layer weight matrix, size d o ×d f , is the fully connected layer bias vector, dimension d o , φ is the ReLU activation function;

[0086] The elevation sparse regularization term is expressed as:

[0087] L elev =H θ ‖H g ‖1,L total =L task +Lelev ,

[0088] Among them, L elev is the elevation gated sparse loss, H θ is the regularization coefficient, ‖·‖1 is the vector 1 norm, L task is the main task loss, L total For comprehensive losses.

[0089] As a preferred solution of the intelligent monitoring method of smart agriculture based on big data described in the present invention, the step of dynamically adjusting the prediction range includes:

[0090] Analyze the real-time image data sent back by the drone and identify the boundaries of the current shooting blind area;

[0091] Calculate the coverage compensation radius of the virtual sensor network based on the flight altitude and camera viewing angle;

[0092] Activate the prediction node of the corresponding area and output the visual heat map to the monitoring terminal.

[0093] The calculation of the coverage compensation radius includes:

[0094] Determine the ground sampling distance GSD based on the UAV’s flight altitude;

[0095] The horizontal coverage overlap rate threshold is derived based on the camera field of view FOV;

[0096] Dynamically expand the compensation range based on the current image stitching integrity.

[0097] The present invention addresses the core flaws of terrace monitoring scenarios. Through curvature analysis and microclimate attenuation models, it transforms monitoring of physically inaccessible areas into a computable problem in digital twin space, significantly reducing the need for sensor deployment on steep slopes and ridges. The terrain continuity coefficient accurately characterizes the structural integrity of terraces, and the humidity gradient quantifies the environmental propagation patterns of adjacent plots, enabling the virtual sensor network to dynamically generate moisture content data in blind areas, thus resolving the terrain adaptability flaw of the OODA closed loop.

[0098] Furthermore, a heterogeneous topological graph is used to integrate physical and virtual nodes. A graph attention mechanism dynamically filters effective neighborhood information through spatial attenuation and microclimate similarity, avoiding the strong reliance of traditional federated learning on data homogeneity. A combination of differential privacy encryption and dynamic weight distribution enables the sharing of disaster patterns across multiple terraced fields while safeguarding the data sovereignty of smallholder farmers. The proposed transfer learning method introduces elevation gating modulation, embedding terrain elevation into the fully connected layer decision logic using multiplicative constraints, enabling rapid convergence of pre-trained plain models to terraced field scenarios. Dynamic calculation of the coverage compensation radius during drone inspections enables real-time linkage between flight blind spots and virtual predictions, overcoming the bottleneck of existing 3D reconstruction technology's reliance on lidar hardware. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0100] Figure 1 This is a schematic diagram of the framework of the big data-based smart agricultural intelligent monitoring system in Example 1.

[0101] Figure 2 This is a flow chart of the intelligent monitoring method for smart agriculture based on big data in Example 1. DETAILED DESCRIPTION

[0102] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0103] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0104] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0105] Example 1, with reference to Figure 1 and Figure 2 This embodiment provides a smart agricultural intelligent monitoring system based on big data, including:

[0106] The terrain feature migration module is used to build a 3D digital twin model based on historical aerial data of the target area and extract terrain-related features;

[0107] The terrain feature migration module includes:

[0108] 3D point cloud reconstruction unit, which generates terrace terrain surface model based on visible light image sequence;

[0109] Microclimate correlation analysis unit, which explores the nonlinear mapping relationship between soil moisture, light intensity and terrain elevation;

[0110] The extraction of terrain-related features includes:

[0111] Perform curvature analysis on the terrace steps to generate terrain continuity coefficient;

[0112] Based on the soil moisture gradient differences between adjacent steps, a microclimate transmission attenuation model is constructed;

[0113] In the terrain feature migration module, during the process of terrain related feature extraction, the three-dimensional point cloud E i =(x i ,y i ,z i ) Establish the neighborhood covariance matrix and construct the local curvature according to the eigenvalue ratio:

[0114]

[0115] in, is the curvature of point i, are the first, second, and third eigenvalues ​​of the covariance matrix, in units of m 2 , x i ,y i ,z i Represents coordinates;

[0116] Filter a collection of step edges using a threshold:

[0117]

[0118] in, is the curvature threshold, and the mean and variance are calculated on ε:

[0119]

[0120] in, is the mean curvature, is the curvature variance, |ε| is the number of edge points;

[0121] The terrain continuity coefficient is given based on statistics:

[0122]

[0123] Among them, G C is the continuity coefficient, ε is the stability constant, and its value is 10 -6 ;

[0124] Establish a gradient for the soil moisture difference between adjacent steps j and j+1:

[0125]

[0126] in, is the humidity gradient, in %m -1, is the step j humidity, unit is %, is the elevation difference between the two steps, in m;

[0127] Summarizing all gradients yields the microclimate attenuation index:

[0128]

[0129] Among them, M β is the decay index, J is the total number of steps;

[0130] Finally, the terrain correlation feature vector F is constructed terrain :

[0131]

[0132] Specifically, the curvature ratio depends only on the proportion of eigenvalues, avoiding disturbances caused by changes in point cloud density. The exponential decay based on variance amplifies the discreteness of cliffs or collapses, making the continuity coefficient extremely sensitive to terrain gaps. The humidity gradient is normalized by elevation to eliminate the interference of altitude differences. The microclimate attenuation index is converted into a global-scale energy decay rate, which can intuitively reflect the evapotranspiration distribution of terraces. The four-dimensional vector covers both geometric continuity and environmental gradients, with compact dimensions and sufficient information, providing stable input for cross-block transfer learning and significantly reducing the noise introduced by seasonality and sensor errors in measured data.

[0133] The virtual sensor network module generates predicted environmental parameters in areas where no physical sensors are deployed based on terrain correlation features and graph attention networks;

[0134] The virtual sensor network module is implemented in the following ways:

[0135] Construct a heterogeneous topology graph containing known sensor nodes and virtual nodes;

[0136] Use graph attention mechanism to dynamically calculate feature propagation weights between nodes;

[0137] Output the predicted values ​​of blind spot environmental parameters and confidence assessment results.

[0138] Heterogeneous topology diagrams include:

[0139] Physical sensor nodes, including soil temperature and humidity, and light intensity detection units;

[0140] Virtual sensor nodes,associate terrain elevation with historical disaster event labels;

[0141] Edge weights are calculated jointly by the Euclidean distance between nodes and microclimate similarity;

[0142] In the virtual sensor network module, a graph attention mechanism is used to dynamically calculate the feature propagation weights between nodes, including:

[0143] Construct concatenated features for physical and virtual nodes and calculate content relevance. The formula is:

[0144]

[0145] in, is the content relevance score of node i→j, N W is a trainable linear mapping matrix, dimension: d×2d, is the node feature vector, dimension d, [·‖·] represents the vector concatenation operation, Leaky ReLU activation function;

[0146] Combined with spatial attenuation:

[0147]

[0148] in, is the spatial attenuation coefficient, is the Euclidean distance between nodes, unit is m, γ is the adaptive decay exponent;

[0149] Adding microclimate similarity:

[0150]

[0151] in, is the humidity similarity, is the node humidity;

[0152] The attention weight is obtained comprehensively:

[0153]

[0154] in, is the normalized weight, is the neighborhood set of node i;

[0155] Update the node representation using weights:

[0156] in, is the updated feature vector;

[0157] Specifically, attention calculation introduces three factors: content, distance, and climate. The product structure achieves dynamic recalibration within the same channel. The distance decay exponent is adaptively inferred from the data distribution, which not only suppresses the weight dilution of sparse networks but also limits the explosion of clusters in dense areas. The humidity similarity term blocks the information spillover of different moisture segments and maintains microenvironmental differences. The normalized denominator limits the total amount of information flow per node, providing numerical stability for large-scale graphs. During online training, each factor can be updated in reverse separately. The model shows rapid convergence characteristics for seasonal changes and extreme weather. The confidence interval of blind spot prediction continues to narrow with the addition of samples, meeting the needs of real-time field monitoring.

[0158] A multi-region federated learning module aggregates monitoring data from multiple adjacent terraced fields and dynamically updates the weight distribution of terrain-related features.

[0159] The multi-region federated learning module includes:

[0160] Privacy protection unit, which performs differential privacy encryption on the terraced block data participating in federated learning;

[0161] The dynamic weight allocation unit adjusts the model aggregation weight according to the data volume of each block and the disaster history records.

[0162] Edge computing nodes, deployed on drones or fixed base stations, perform lightweight inference tasks for virtual sensor network modules.

[0163] This embodiment also provides a smart agriculture intelligent monitoring method based on big data, including:

[0164] Step S1, collecting historical aerial images and physical sensor time series data of the target terraced area;

[0165] Step S2, adapting the plain farmland pre-training model to the terraced terrain characteristics through transfer learning;

[0166] The transfer learning adaptation process in step S2 includes:

[0167] Extract the unique crop shadow distribution pattern and microclimate mutation characteristics of terraced fields;

[0168] Inject terrain elevation constraints into the fully connected layer of the model;

[0169] Freeze the parameters of the basic feature extraction layer and only fine-tune the high-level decision network;

[0170] In step S2, the elevation information is introduced into the multiplicative gate in the fully connected layer, and the backbone network is set to output the feature vector at the grid point p.

[0171] To normalize the elevation:

[0172]

[0173] in, is the standardized elevation, H z is the original elevation of the grid point, in m, is the mean elevation of the target block, in m. is the elevation standard deviation, in meters, and ε is a numerical stability constant, which is 10 -6 ;

[0174] Generate the elevation gating vector:

[0175]

[0176] Among them, H g is the gate vector, dimension d f , is the elevation mapping weight matrix, size d f ×1, is the gate bias vector, dimension d f ,σ is the Sigmoid activation function;

[0177] Perform characteristic modulation and adjust the formula as follows:

[0178] F elev =F in ⊙(1+H f ),

[0179] Among them, F elev is the feature after elevation modulation, with dimension d f ⊙ is element-by-element multiplication, 1 is multiplication with H g All-1 vectors of the same dimension;

[0180] Perform constrained full connection mapping, the mapping formula is:

[0181]

[0182] Among them, Y is the output vector, dimension d o , is the fully connected layer weight matrix, size d o ×d f , is the fully connected layer bias vector, dimension d o , φ is the ReLU activation function;

[0183] The elevation sparse regularization term is expressed as:

[0184] L elev =H θ ‖H g ‖1,L total =L task +Lelev ,

[0185] Among them, L elev is the elevation gated sparse loss, H θ is the regularization coefficient, ‖·‖1 is the vector 1 norm, L task is the main task loss, L total is comprehensive loss;

[0186] Specifically, the standardization process brings the absolute elevations of different terraced areas into a unified dimension, preventing the model from being sensitive to overall uplift or subsidence. The gating vector is mapped to a single elevation channel and then multiplied point by point with the feature. Low-lying areas are suppressed by the curve, while convex areas receive an enhanced response. The network acquires local geometric adaptability in a multiplicative manner without introducing additional positional encoding. The sparse regularization term suppresses ineffective gating, prompting the model to adjust weights only where elevation changes are significant.

[0187] Step S3: fusing multi-block data based on the federated learning framework and updating the parameters of the virtual sensor network;

[0188] Step S4: Dynamically adjust the prediction range of the virtual sensor node according to the real-time UAV inspection path.

[0189] The steps to dynamically adjust the forecast range include:

[0190] Analyze the real-time image data sent back by the drone and identify the boundaries of the current shooting blind area;

[0191] Calculate the coverage compensation radius of the virtual sensor network based on the flight altitude and camera viewing angle;

[0192] Activate the prediction node of the corresponding area and output the visual heat map to the monitoring terminal.

[0193] The calculation of the coverage compensation radius includes:

[0194] Determine the ground sampling distance GSD based on the UAV’s flight altitude;

[0195] The horizontal coverage overlap rate threshold is derived based on the camera field of view FOV;

[0196] Dynamically expand the compensation range based on the current image stitching integrity.

[0197] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A smart agricultural intelligent monitoring system based on big data, characterized by: include, The terrain feature migration module is used to build a 3D digital twin model based on historical aerial data of the target area and extract terrain-related features; A virtual sensor network module generates predicted environmental parameters of areas where no physical sensors are deployed based on the terrain association features and the graph attention network; A multi-region federated learning module aggregates monitoring data from multiple adjacent terraced fields and dynamically updates the weight distribution of the terrain-related features. Edge computing nodes are deployed on drones or fixed base stations to perform lightweight inference tasks of the virtual sensor network module.

2. The intelligent monitoring system for smart agriculture based on big data according to claim 1, characterized in that: The terrain feature migration module includes: 3D point cloud reconstruction unit, which generates terraced terrain surface model based on visible light image sequence; Microclimate correlation analysis unit, which explores the nonlinear mapping relationship between soil moisture, light intensity and terrain elevation; The extraction of the terrain-related features includes: Perform curvature analysis on the terraced edge to generate the terrain continuity coefficient; A microclimate propagation attenuation model is constructed based on the soil moisture gradient differences between adjacent steps.

3. The intelligent monitoring system for smart agriculture based on big data according to claim 2, characterized in that: In the terrain feature migration module, during the process of terrain related feature extraction, the three-dimensional point cloud E i =(x i ,y i ,z i ) Establish the neighborhood covariance matrix and construct the local curvature according to the eigenvalue ratio: in, is the curvature of point i, are the first, second, and third eigenvalues ​​of the covariance matrix, in units of m 2 , x i ,y i ,z i Represents coordinates; Filter a collection of step edges using a threshold: in, is the curvature threshold, and the mean and variance are calculated on ε: in, is the mean curvature, is the curvature variance, |ε| is the number of edge points; The terrain continuity coefficient is given based on statistics: Among them, G C is the continuity coefficient, ε is the stability constant; Establish a gradient for the soil moisture difference between adjacent steps j and j+1: in, is the humidity gradient, in %m -1 , is the step j humidity, unit is %, is the elevation difference between the two steps, in m; Summarizing all gradients yields the microclimate attenuation index: Among them, M β is the decay index, J is the total number of steps; Finally, the terrain correlation feature vector F is constructed terrain :

4. The intelligent monitoring system for smart agriculture based on big data according to claim 1, characterized in that: The virtual sensor network module is implemented in the following ways: Construct a heterogeneous topology graph containing known sensor nodes and virtual nodes; Use graph attention mechanism to dynamically calculate feature propagation weights between nodes; Output the predicted values ​​of blind area environmental parameters and confidence assessment results; The heterogeneous topology graph includes: Physical sensor nodes, including soil temperature and humidity, and light intensity detection units; Virtual sensor nodes,associate terrain elevation with historical disaster event labels; The edge weight is calculated by combining the Euclidean distance between nodes and the microclimate similarity.

5. The intelligent monitoring system for smart agriculture based on big data according to claim 4, characterized in that: In the virtual sensor network module, a graph attention mechanism is used to dynamically calculate the feature propagation weights between nodes, including: Construct concatenated features for physical and virtual nodes and calculate content relevance. The formula is: in, is the content relevance score of node i→j, N W is a trainable linear mapping matrix, dimension: d×2d, is the node feature vector, dimension d, [·‖·] represents the vector concatenation operation, Leaky ReLU activation function; Combined with spatial attenuation: in, is the spatial attenuation coefficient, is the Euclidean distance between nodes, unit is m, γ is the adaptive decay exponent; Adding microclimate similarity: in, is the humidity similarity, is the node humidity; The attention weight is obtained comprehensively: in, is the normalized weight, is the neighborhood set of node i; Update the node representation using weights: in, is the updated feature vector.

6. The intelligent monitoring system for smart agriculture based on big data according to claim 1, characterized in that: The multi-region federated learning module includes: Privacy protection unit, which performs differential privacy encryption on the terraced block data participating in federated learning; The dynamic weight allocation unit adjusts the model aggregation weight according to the data volume of each block and the disaster history records.

7. A method for intelligent monitoring of smart agriculture based on big data, based on a system for intelligent monitoring of smart agriculture based on big data according to any one of claims 1 to 6, characterized in that: include: Step S1, collecting historical aerial images and physical sensor time series data of the target terraced area; Step S2, adapting the plain farmland pre-training model to the terraced terrain characteristics through transfer learning; Step S3: fusing multi-block data based on the federated learning framework and updating the parameters of the virtual sensor network; Step S4: Dynamically adjust the prediction range of the virtual sensor node according to the real-time UAV inspection path.

8. The intelligent monitoring method for smart agriculture based on big data according to claim 7, characterized in that: The transfer learning adaptation process of step S2 includes: Extract the unique crop shadow distribution pattern and microclimate mutation characteristics of terraced fields; Inject terrain elevation constraints into the fully connected layer of the model; Freeze the parameters of the basic feature extraction layer and only fine-tune the high-level decision network.

9. The intelligent monitoring method for smart agriculture based on big data according to claim 8, characterized in that: In step S2, the elevation information is introduced into the multiplicative gate in the fully connected layer, and the backbone network is set to output the feature vector at the grid point p. To normalize the elevation: in, is the standardized elevation, H z is the original elevation of the grid point, in m, is the mean elevation of the target block, in m. is the elevation standard deviation, in meters, and ε is a numerical stability constant, which is 10 -6 ; Generate the elevation gating vector: Among them, H g is the gate vector, dimension d f , is the elevation mapping weight matrix, size d f ×1, is the gate bias vector, dimension d f ,σ is the Sigmoid activation function; Perform characteristic modulation and adjust the formula as follows: F elev =F in ⊙(1+H g ), Among them, F elev is the feature after elevation modulation, with dimension d f ⊙ is element-by-element multiplication, 1 is multiplication with H g All-1 vectors of the same dimension; Perform constrained full connection mapping, the mapping formula is: Among them, Y is the output vector, dimension d o , is the fully connected layer weight matrix, size d o ×d f , is the fully connected layer bias vector, dimension d o , φ is the ReLU activation function; The elevation sparse regularization term is expressed as: L elev =H θ ‖H g ‖1,L total =L task +L elev , Among them, L elev is the elevation gated sparse loss, H θ is the regularization coefficient, ‖·‖1 is the vector 1 norm, L task is the main task loss, L total For comprehensive losses.

10. The intelligent monitoring method for smart agriculture based on big data according to claim 9, characterized in that: The step of dynamically adjusting the prediction range includes: Analyze the real-time image data sent back by the drone to identify the boundaries of the current shooting blind area; calculate the coverage compensation radius of the virtual sensor network based on the flight altitude and camera viewing angle; Activate the prediction node of the corresponding area and output the visual heat map to the monitoring terminal; The calculation of the coverage compensation radius includes: Determine the ground sampling distance GSD based on the UAV’s flight altitude; The horizontal coverage overlap rate threshold is derived based on the camera field of view FOV; Dynamically expand the compensation range based on the current image stitching integrity.

Images