Karst key zone soil profile intelligent monitoring system and method thereof

Abstract

The invention relates to the technical field of karst soil monitoring, in particular to a karst key zone soil profile intelligent monitoring system and method. The intelligent monitoring system for the karst soil profile integrates a multi-layer sensor network and edge calculation and specially attacks the soil monitoring problem of a karst area, and the sensor network accurately collects moisture, temperature, pH value, organic matter and microorganism data; the data processing module receives and preprocesses in real time to ensure efficient transmission of data; the deep learning module constructs evolution, community and moisture models to improve the analysis precision; the decision support module generates health assessment and irrigation suggestions and assists in scientific decision making; the system overcomes the bottleneck of poor adaptability of traditional monitoring, reduces energy consumption, relieves communication burden, provides powerful technical support for research and management of karst soil, and realizes accurate, efficient and intelligent monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of karst soil monitoring, and particularly to an intelligent monitoring system and method for the soil profile in the karst critical zone. Background Art

[0002] Karst areas refer to regions mainly composed of soluble rocks (such as limestone, dolomite, etc.). Their unique geological characteristics have led to the formation of complex surface and groundwater hydrological networks. The karst critical zone is the interface connecting the atmosphere, biosphere, hydrosphere, and lithosphere, and is crucial for the stability of the ecosystem. However, the soil in karst areas is often thin, with a complex structure, vulnerable to erosion, and particularly sensitive to climate change and human activities.

[0003] Traditional soil monitoring methods often rely on manual sampling and laboratory analysis, suffering from problems such as poor timeliness, limited spatial coverage, and difficulty in achieving long-term continuous monitoring. With the development of sensor technology, the Internet of Things, and artificial intelligence, intelligent soil monitoring systems have gradually been applied to soil environmental monitoring. However, existing monitoring systems are mainly designed for plain and hilly areas and are difficult to adapt to the complex topography and soil structure characteristics of karst regions.

[0004] Therefore, there is an urgent need to develop an intelligent monitoring system for the soil profile in the karst critical zone to achieve comprehensive, real-time, intelligent monitoring and analysis of soil profile parameters in karst regions. Summary of the Invention

[0005] The purpose of the present invention is to solve the deficiencies in the prior art and provide an intelligent monitoring system and method for the soil profile in the karst critical zone, which can achieve comprehensive, real-time, intelligent monitoring and analysis of soil profile parameters according to the special topography and soil structure characteristics of karst regions.

[0006] The present invention proposes an intelligent monitoring system for the soil profile in the karst critical zone, including:

[0007] A sensor network module for collecting moisture, temperature, pH value, organic matter content, and microbial data of the karst soil profile;

[0008] A data processing module communicatively connected to the sensor network module for receiving the data collected by the sensor network module and performing edge computing preprocessing and data analysis;

[0009] A deep learning module communicatively connected to the data processing module for constructing a soil profile evolution model, a microbial community analysis model, and a water movement model;

[0010] A decision support module communicatively connected to the deep learning module for generating a soil health assessment report and irrigation strategy recommendations based on the analysis results.

[0011] Preferably, the sensor network module includes:

[0012] A multi-level sensor node unit for deploying data acquisition nodes, intermediate nodes, and edge nodes in soil layers at different depths in a karst area to form a tree-shaped topology network;

[0013] An adaptive power management unit for providing electrical energy to the multi-level sensor node unit through a solar panel, a battery management sub-module, and a voltage stabilizing and rectifying circuit;

[0014] A networking communication unit for establishing a distributed data transmission channel among the data acquisition nodes, intermediate nodes, and edge nodes.

[0015] Preferably, the data processing module includes:

[0016] A heterogeneous data spatio-temporal alignment unit for precisely aligning data from different sensors, at different time and space scales;

[0017] A data quality assessment unit for performing real-time quality assessment on data using statistical methods, and automatically identifying and processing abnormal data;

[0018] An edge computing unit for performing data preprocessing near the data source to reduce the amount of data transmission;

[0019] An adaptive sampling control unit for dynamically adjusting the sampling frequency of each sensor node based on an environmental parameter evaluation model.

[0020] Preferably, the deep learning module includes:

[0021] A recurrent neural network unit for analyzing time-series data of a karst soil profile and predicting the trend of parameter changes;

[0022] A deep autoencoder unit for establishing a soil profile evolution model, and simulating the variation laws of data in different soil layers through an autoencoder network layer and a feature extraction layer;

[0023] A convolutional neural network unit for performing image recognition on soil profile pictures and analyzing the microbial community structure;

[0024] A recurrent neural network unit for combining meteorological data to analyze the moisture migration characteristics of the soil profile, including infiltration characteristics during the rainy season and evaporation characteristics during the dry season.

[0025] Preferably, the deep autoencoder unit adopts a hierarchical nested coding structure, with each layer of autoencoder corresponding to a soil layer position, and incorporating soil physics laws as constraint conditions in the neural network training.

[0026] Preferably, the convolutional neural network unit constructs a microorganism-environment interaction network, which can analyze the microbial population quantity, the dynamic changes of functional communities, and the community succession trend.

[0027] Preferably, the decision support module includes:

[0028] A soil health assessment unit, which is used to generate a soil health status assessment report based on the analysis results of the deep learning module;

[0029] An irrigation strategy optimization unit, which is used to generate an optimal irrigation strategy through a multi-objective optimization algorithm and deep reinforcement learning technology;

[0030] A three-dimensional visualization unit, which is used to realize the interactive three-dimensional visualization display of soil profile data through WebGL technology;

[0031] A warning recommendation unit, which is used to generate hierarchical warning information and improvement measure suggestions.

[0032] Preferably, it further includes a federated learning module, which is used for:

[0033] Performing local model training on different monitoring terminal devices;

[0034] Only sharing model parameters instead of original data to protect data privacy;

[0035] Protecting individual monitoring data through differential privacy technology;

[0036] Automatically splitting the deep learning model according to the computing capabilities of different terminal devices.

[0037] Preferably, data is transmitted between the sensor network module and the data processing module through a wireless communication method, and data transmission and processing are performed between the data processing module and the deep learning module through an edge-cloud collaborative computing architecture.

[0038] The monitoring method based on the above-mentioned intelligent monitoring system for the soil profile of the karst critical zone includes the following steps:

[0039] According to the geomorphic characteristics of the karst area, a multi-level sensor network is arranged at multiple soil profile depths;

[0040] Collect the moisture, temperature, pH value, organic matter content and microbial data of the karst soil profile;

[0041] Perform spatio-temporal alignment, quality assessment and edge computing preprocessing on the collected heterogeneous data;

[0042] Analyze the temporal variation of soil parameters through a recurrent neural network and establish a temporal prediction model of soil parameters;

[0043] Establish a soil profile evolution model through a deep autoencoder to simulate the variation laws of data in different soil layers;

[0044] Use a convolutional neural network to analyze the microbial community of soil profile pictures;

[0045] Combine meteorological data and use a recurrent neural network to analyze the characteristics of soil water movement;

[0046] Generate a soil health assessment report and an optimal irrigation strategy based on the analysis results;

[0047] Dynamically adjust the sensor sampling frequency to optimize the energy consumption of the system;

[0048] Optimize the overall model performance through federated learning while protecting data privacy.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. A dedicated soil profile intelligent monitoring system is provided for the special topographical and soil structure characteristics of karst areas, solving the problem of poor adaptability of traditional monitoring systems in karst areas;

[0051] 2. By adopting a multi-level sensor network and edge computing technology, efficient data acquisition and preprocessing are realized, reducing the system energy consumption and communication burden;

[0052] 3. By integrating deep learning and soil physical models, the accuracy of soil profile evolution, microbial community analysis and water movement prediction is improved;

[0053] 4. Through adaptive sampling strategy and irrigation strategy optimization, efficient utilization of resources is realized;

[0054] 5. By introducing federated learning technology, the overall model performance is optimized while protecting data privacy.

[0055] The present invention is applicable to fields such as soil monitoring, agricultural irrigation management, and ecological environment protection in karst areas, and has important application value. Description of the Drawings

[0056] Figure 1 It is the structural block diagram of the intelligent monitoring system for the soil profile of the karst critical zone of the present invention;

[0057] Figure 2 It is the structural schematic diagram of the sensor network module of the present invention;

[0058] Figure 3 It is the working flow chart of the deep learning module of the present invention;

[0059] Figure 4 It is the flow chart of the monitoring method of the present invention. Detailed implementation manners

[0060] Please refer to the appended Figures 1-4 , to achieve the above object, the present invention provides the following technical solutions: an intelligent monitoring system for the soil profile of the karst critical zone, including a sensor network module 1, a data processing module 2, a deep learning module 3, and a decision support module 4.

[0061] The sensor network module 1 is used to collect moisture, temperature, pH value, organic matter content, and microbial data of the karst soil profile. The data processing module 2 is communicatively connected to the sensor network module 1 and is used to receive the data collected by the sensor network module 1 and perform edge computing preprocessing and data analysis. The deep learning module 3 is communicatively connected to the data processing module 2 and is used to construct a soil profile evolution model, a microbial community analysis model, and a water movement model. The decision support module 4 is communicatively connected to the deep learning module 3 and is used to generate a soil health assessment report and irrigation strategy suggestions based on the analysis results.

[0062] Preferably, the sensor network module 1 includes a multi-level sensor node unit 11, an adaptive power management unit 12, and a networking communication unit 13. The multi-level sensor node unit 11 is used to deploy data acquisition nodes, intermediate nodes, and edge nodes in different depth soil layers in the karst area to form a tree-shaped topology network. The adaptive power management unit 12 is used to provide electrical energy for the multi-level sensor node unit 11 through a solar panel, a battery management sub-module, and a voltage stabilizing and rectifying circuit. The networking communication unit 13 is used to establish a distributed data transmission channel between the data acquisition nodes, intermediate nodes, and edge nodes.

[0063] In an embodiment of the present invention, the data processing module 2 includes a heterogeneous data spatio-temporal alignment unit 21, a data quality assessment unit 22, an edge computing unit 23, and an adaptive sampling control unit 24. The heterogeneous data spatio-temporal alignment unit 21 is used to accurately align data from different sensors, different time, and space scales. The data quality assessment unit 22 is used to perform real-time quality assessment on the data using statistical methods and automatically identify and process abnormal data. The edge computing unit 23 is used to perform data preprocessing near the data source to reduce the amount of data transmission. The adaptive sampling control unit 24 is used to dynamically adjust the sampling frequency of each sensor node based on an environmental parameter evaluation model.

[0064] Preferably, the deep learning module 3 includes a recurrent neural network unit 31, a deep autoencoder unit 32, a convolutional neural network unit 33, and a recurrent neural network unit 34. The recurrent neural network unit 31 is used to analyze the time-series data of the karst soil profile and predict the trend of parameter changes. The deep autoencoder unit 32 is used to establish a soil profile evolution model and simulate the variation law of data in different soil layers through an autoencoder network layer and a feature extraction layer. The convolutional neural network unit 33 is used to perform image recognition on soil profile pictures and analyze the microbial community structure. The recurrent neural network unit 34 is used to analyze the soil profile water migration characteristics in combination with meteorological data, including infiltration characteristics during the rainy season and evaporation characteristics during the dry season.

[0065] In an embodiment of the present invention, the deep autoencoder unit 32 adopts a hierarchical nested coding structure, and each layer of the autoencoder corresponds to a soil layer position, and soil physics laws are incorporated as constraint conditions in the neural network training. The autoencoder network can be expressed as:

[0066] h = f(W·x + b),

[0067] where x represents parameters such as organic matter content, pH value, cation exchange capacity, and hydraulic properties of a sampling point in the karst soil profile, h is the output of the autoencoder network layer of the model; f represents the feature extraction layer of the model, usually an activation function such as ReLU or sigmoid; W and b are the corresponding weight and bias terms of the model.

[0068] The hierarchical nested coding structure can be implemented in the following way:

[0069] h i = f i (W i ·x i + b i ),

[0070] where i represents the soil layer position number, increasing sequentially from the surface layer to the bottom layer; x i is the parameter vector of the i-th layer of soil; h i is the output of the autoencoder of this layer; f i , W i and b i are the activation function, weight matrix, and bias vector of the autoencoder of this layer, respectively.

[0071] The loss function of the physical constraint can be expressed as:

[0072] L total = L recon + λ·L phys ,

[0073] where L recomis the reconstruction loss, which measures the accuracy of the autoencoder in reconstructing data; L phys is the physical constraint loss, which ensures that the model predictions comply with soil physics laws such as Darcy's law, Richards equation, etc.; λ is the weight coefficient that balances the two losses, usually taking values between 0.1 and 1, and the empirical value is 0.5.

[0074] Preferably, the convolutional neural network unit 33 constructs a microorganism-environment interaction network, which can analyze the microbial population quantity, the dynamic changes of functional communities, and the community succession trend. The construction of the microorganism-environment interaction network is based on the combination of a convolutional neural network and a graph neural network, and can be expressed as:

[0075] G=(V, E, A),

[0076] where G represents the constructed interaction network; V is the set of nodes, including microbial populations and environmental factors; E is the set of edges, representing the interactions between nodes; A is the adjacency matrix, describing the strength of the connection relationship between nodes.

[0077] The model for analyzing the microbial population quantity can be expressed as:

[0078] M = CNN(I),

[0079] where I is the soil profile picture, CNN is the convolutional neural network operation, and M is the identified microbial quantity matrix.

[0080] The model for predicting the community succession trend can be expressed as:

[0081] S t+1 = f(S t , E t , G),

[0082] where S t is the community structure at time t, E t is the environmental factor at time t, G is the microorganism-environment interaction network, and S t +1 is the predicted community structure at time t + 1.

[0083] In an embodiment of the present invention, the decision support module 4 includes a soil health assessment unit 41, an irrigation strategy optimization unit 42, a three-dimensional visualization unit 43, and a warning recommendation unit 44. The soil health assessment unit 41 is used to generate a soil health status assessment report based on the analysis results of the deep learning module 3. The irrigation strategy optimization unit 42 is used to generate an optimal irrigation strategy through a multi-objective optimization algorithm and deep reinforcement learning technology. The three-dimensional visualization unit 43 is used to realize the interactive three-dimensional visualization display of soil profile data through WebGL technology. The warning recommendation unit 44 is used to generate hierarchical warning information and improvement measure suggestions.

[0084] Preferably, the system of the present invention further includes a federated learning module 5 for performing local model training on different monitoring terminal devices; only sharing model parameters instead of raw data to protect data privacy; protecting individual monitoring data through differential privacy technology; and automatically splitting the deep learning model according to the computing capabilities of different terminal devices.

[0085] The core algorithm of the federated learning module 5 can be expressed as:

[0086]

[0087] where θ global represents the global model parameters, θ i represents the local model parameters of the i-th terminal, and N is the number of terminals participating in the training.

[0088] The differential privacy protection mechanism can be achieved by adding noise:

[0089]

[0090] where θ' i is the model parameters after adding noise, represents Gaussian noise with a mean of 0 and a variance of σ 2 . The noise variance σ 2 is usually determined based on the privacy budget ∈, where c is a constant, usually taken as 2; Δf is the sensitivity, representing the maximum impact of a single sample on the model; ∈ is the privacy budget, and the smaller the value, the higher the degree of privacy protection. Usually, the value is between 0.1 and 10, and the empirical value is 1.

[0091] In an embodiment of the present invention, data is transmitted between the sensor network module 1 and the data processing module 2 through a wireless communication method, and data transmission and processing are performed between the data processing module 2 and the deep learning module 3 through an edge-cloud collaborative computing architecture.

[0092] The protocol adopted for wireless communication can be ZigBee, LoRa, or NB-IoT, and a suitable communication method is selected according to the specific environmental characteristics of the karst area. Preferably, in a karst area with complex terrain and dense vegetation, the LoRa technology can be adopted, with a communication distance of up to 10 kilometers and low power consumption, suitable for a long-term deployed monitoring system.

[0093] The working mode of the edge-cloud collaborative computing architecture is: the edge node is responsible for the preliminary processing and timeliness analysis of data, and the cloud is responsible for complex model training and large-scale data mining. The data segmentation standard can be determined based on data timeliness and computational complexity. Usually, tasks with a response time requirement of less than 100 milliseconds are deployed on the edge side, and tasks with higher complexity and less sensitivity to latency are deployed on the cloud.

[0094] The present invention also provides a monitoring method based on the above intelligent monitoring system for the soil profile of the karst critical zone, comprising the following steps:

[0095] According to the geomorphic characteristics of the karst area, a multi-level sensor network is deployed at multiple soil profile depths;

[0096] Collect moisture, temperature, pH value, organic matter content and microbial data of the karst soil profile;

[0097] Perform spatio-temporal alignment, quality assessment and edge computing preprocessing on the collected heterogeneous data;

[0098] Analyze the temporal variation of soil parameters through a recurrent neural network and establish a temporal prediction model of soil parameters;

[0099] Establish a soil profile evolution model through a deep autoencoder to simulate the variation law of data in different soil layers;

[0100] Use a convolutional neural network to analyze the microbial community of soil profile pictures;

[0101] Combine meteorological data and use a recurrent neural network to analyze the characteristics of soil moisture migration;

[0102] Generate a soil health assessment report and an optimal irrigation strategy based on the analysis results;

[0103] Dynamically adjust the sensor sampling frequency to optimize the energy consumption of the system;

[0104] Optimize the overall model performance through federated learning while protecting data privacy.

[0105] In step 1, the deployment of the multi-level sensor network is based on the micro-topographic analysis and soil profile characteristics of the karst area. Preferably, the profile can be divided into the surface layer (0 - 20 cm), the subsurface layer (20 - 40 cm), the middle layer (40 - 80 cm) and the deep layer (> 80 cm), and corresponding sensor nodes are deployed in each layer. In complex terrain areas, a terrain adaptive grid division method can be used to determine the node positions.

[0106] In step 2, the sensor sampling frequency is dynamically adjusted according to the parameter change rate. Usually, the sampling frequencies of moisture and temperature parameters are relatively high (such as once every 15 - 30 minutes), while the sampling frequencies of pH value and organic matter content are relatively low (such as once every 24 hours). Microbial data is regularly obtained through an image acquisition device (such as once every 3 - 7 days).

[0107] In step 3, multi-scale time window method is adopted for spatio-temporal alignment of heterogeneous data, mapping data with different acquisition frequencies onto a unified time axis. Data quality assessment is achieved through statistical anomaly detection methods, such as setting the normal change range of parameters (for example, the soil pH value is usually between 4 and 9, and exceeding this range may indicate sensor failure or data anomaly). Edge computing preprocessing includes data filtering, compression, and feature extraction, which can usually reduce the data transmission volume by 70%-90%.

[0108] The structure of the recurrent neural network model in step 4 can be expressed as:

[0109] h t =tanh(W xh ·x t +W hh ·h t-1 +b h ),

[0110] y t =W hy ·h t +b y ,

[0111] where x t is the input parameter at time y, h t is the hidden state at time y, y t is the output prediction value at time y, W xh 、W hh 、W hy are weight matrices, b h 、b y are bias vectors, and tanh is the hyperbolic tangent activation function.

[0112] Preferably, long short-term memory network (LSTM) or gated recurrent unit (GRU) can be adopted to handle long-time series dependencies. The prediction time window is usually set to 7-30 days in the future, which is suitable for the needs of agricultural management and disaster warning.

[0113] The deep autoencoder soil profile evolution model in step 5 predicts the change trend of the overall state of the soil profile by monitoring the interaction and time evolution characteristics of different soil layer parameters. This model can identify key parameters and critical change points in the soil profile, providing a scientific basis for soil management.

[0114] In step 6, microbial community analysis processes soil profile pictures through a convolutional neural network to identify different microbial types and their distribution. Preferably, a multi-scale feature extraction architecture is adopted to capture morphological features at both the micron and millimeter levels simultaneously.

[0115] The analysis of water migration characteristics in Step 7 combines meteorological data (such as rainfall, temperature, evaporation, etc.) and soil moisture data to construct a water dynamic change model. The infiltration model can be expressed as:

[0116]

[0117] where I(t) is the infiltration rate at time t, K sat is the saturated hydraulic conductivity, θ(t) is the soil water content at time t, θ r is the residual water content, θ s is the saturated water content, is the hydraulic gradient.

[0118] The soil health assessment in Step 8 is based on a comprehensive index system, which usually includes soil fertility indicators (organic matter content, nutrient status, etc.), physical indicators (structure, aeration, etc.) and biological indicators (microbial diversity, activity, etc.). The assessment results are usually divided into four grades: excellent, good, medium and poor, and the corresponding thresholds are determined according to soil types and crop requirements.

[0119] The optimization of irrigation strategies adopts a multi-objective optimization method, taking into account water resource utilization efficiency, crop yield and economic benefits at the same time. Irrigation strategies usually include three key parameters: irrigation timing, irrigation amount and irrigation method. Preferably, the irrigation timing can be determined based on the soil water content threshold, usually when the available soil water content drops to 60%-70%; the irrigation amount is determined based on the soil water deficit and crop water demand characteristics; the irrigation method is selected according to the terrain conditions and crop characteristics.

[0120] The dynamic adjustment of the sensor sampling frequency in Step 9 is based on the parameter change rate and the uncertainty of the prediction model. When the parameter change rate is large or the model uncertainty is high, the sampling frequency is increased; otherwise, the sampling frequency is reduced to save energy. The adjustment algorithm can be expressed as:

[0121]

[0122] where f i (t) is the sampling frequency of the i-th sensor at time t, is the parameter change rate, U i is the uncertainty of the prediction model for this parameter, and α and β are weight coefficients. Usually, α ranges from 0.1 to 0.5, and β ranges from 0.2 to 0.6.

[0123] The optimization of federated learning in Step 10 improves the overall model performance by aggregating the model parameters of different monitoring points, while protecting the privacy of the original data. The optimization of communication overhead can be achieved through model compression and sparse updates, usually reducing the communication volume by 50%-80%.

[0124] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0125] Please refer to Figure 1 , the intelligent monitoring system for the karst critical zone soil profile provided by the present invention includes a sensor network module 1, a data processing module 2, a deep learning module 3, and a decision support module 4.

[0126] The sensor network module 1 is used to collect moisture, temperature, pH value, organic matter content, and microbial data of the karst soil profile. The data processing module 2 is communicatively connected to the sensor network module 1 and is used to receive the data collected by the sensor network module 1, perform edge computing preprocessing and data analysis. The deep learning module 3 is communicatively connected to the data processing module 2 and is used to construct a soil profile evolution model, a microbial community analysis model, and a moisture migration model. The decision support module 4 is communicatively connected to the deep learning module 3 and is used to generate a soil health assessment report and irrigation strategy suggestions based on the analysis results.

[0127] Please refer to Figure 2 , the sensor network module 1 includes a multi-level sensor node unit 11, an adaptive power management unit 12, and a networking communication unit 13.

[0128] The multi-level sensor node unit 11 is used to deploy data acquisition nodes, intermediate nodes, and edge nodes in different depth soil layers in the karst area to form a tree-shaped topology network. In a specific embodiment, for a typical karst area soil profile, moisture, temperature, pH value, and organic matter sensors can be deployed on the surface layer (0 - 20 cm), moisture and temperature sensors can be deployed in the middle layer (40 - 60 cm), and moisture sensors can be deployed in the deep layer (> 80 cm). The number of data acquisition nodes corresponds to the number of soil profile layers and adopts a modular design, which can be flexibly configured according to monitoring requirements.

[0129] The adaptive power management unit 12 is used to provide electrical energy for the multi-level sensor node unit 11 through a solar panel, a battery management sub-module, and a voltage stabilizing rectifier circuit. In practical applications, the power of the solar panel is usually 5 - 20 W, which is determined according to the node energy consumption requirements; the battery capacity is usually 2000 - 10000 mAh, which can support the system to work continuously for 3 - 7 days under sunlight-free conditions. The battery management sub-module realizes overcharge, over-discharge protection, and charge and discharge control to extend the battery life.

[0130] The networking communication unit 13 is used to establish a distributed data transmission channel between the data acquisition nodes, intermediate nodes, and edge nodes. Under the complex terrain conditions of the karst area, LoRa communication technology is preferably adopted, and its communication distance can reach 10 kilometers, and it has strong penetration and is suitable for complex terrain environments. The communication between nodes adopts a star-tree hybrid topology structure to improve the network reliability and coverage.

[0131] The data processing module 2 includes a heterogeneous data spatio-temporal alignment unit 21, a data quality assessment unit 22, an edge computing unit 23, and an adaptive sampling control unit 24.

[0132] The heterogeneous data spatio-temporal alignment unit 21 is used to precisely align data from different sensors, at different time and space scales. In practical applications, a multi-scale time window method can be adopted. By setting a basic time unit (such as 1 hour), data with different acquisition frequencies are mapped to a unified time axis through interpolation or aggregation. Spatial alignment can be achieved through a geographic coordinate system and depth information.

[0133] The data quality assessment unit 22 is used to perform real-time quality assessment of data using statistical methods, automatically identify and process abnormal data. The assessment methods include range check, consistency check, and time continuity check. Preferably, the normal change range of each parameter can be set (such as soil moisture 0 - 100%, pH value 4 - 9, temperature -10°C to 50°C, etc.). Data outside the range is marked as abnormal. Abnormal handling strategies include elimination, correction, or marking, which are selected according to the degree of abnormality and application requirements.

[0134] The edge computing unit 23 is used to perform data preprocessing near the data source, reducing the amount of data transmission. The preprocessing includes data filtering (removing noise), data compression (reducing storage and transmission requirements), and feature extraction (extracting key information). In practical applications, filtering can adopt moving average or Kalman filtering; data compression can adopt time series compression algorithms such as PLA (Piecewise Linear Approximation); feature extraction can calculate statistical features (mean, variance, trend, etc.).

[0135] The adaptive sampling control unit 24 is used to dynamically adjust the sampling frequency of each sensor node based on an environmental parameter evaluation model. The evaluation model can be expressed as:

[0136]

[0137] where I i is the importance index of the i-th parameter, P i is the current parameter value, P ref is the reference value (such as the parameter value under the optimal growth conditions of the crop), P max and P min are the maximum and minimum values of the parameter, dP i / dt is the parameter change rate, (dP / dt) max is the maximum change rate, U i is the uncertainty of the prediction model for this parameter, w 1 、w 2 and w 3 are the weight coefficients. Usually, w 1 = 0.3, w2 = 0.4, w 3 = 0.3。

[0138] The sampling frequency adjustment strategy is as follows:

[0139] f i = f base ·(1 + γ·I i ),

[0140] where f i is the sampling frequency of the i-th parameter, f base is the base sampling frequency (usually once per hour), γ is the adjustment coefficient, with a value range of 1 - 5, and I i is the parameter importance index.

[0141] Please refer to Figure 3 , the deep learning module 3 includes a recurrent neural network unit 31, a deep autoencoder unit 32, a convolutional neural network unit 33, and a recurrent neural network unit 34.

[0142] The recurrent neural network unit 31 is used to analyze the temporal data of the karst soil profile and predict the trend of parameter changes. In practical applications, it is preferably to use LSTM (Long Short-Term Memory Network) or GRU (Gated Recurrent Unit) to handle long temporal dependencies. The core structure of LSTM includes an input gate, a forget gate, and an output gate, which can effectively handle long-term dependency problems.

[0143] The training of the temporal prediction model uses the sliding window method. The input window size is usually 7 - 30 days, and the prediction window is 1 - 14 days. The model evaluation metrics include Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Usually, RMSE is controlled within 5% of the parameter range, and MAPE is controlled within 10%, which can meet the actual application requirements.

[0144] The deep autoencoder unit 32 is used to establish a soil profile evolution model, and simulate the variation law of data of different soil layers through the autoencoder network layer and the feature extraction layer. In an embodiment of the present invention, the deep autoencoder adopts a hierarchical nested structure, each soil layer corresponds to an autoencoder module, and information interaction is carried out between the modules through residual connections.

[0145] The specific structure of the hierarchical nested autoencoder is:

[0146] z i = f enc,i (W enc,i ·x i + b enc,i ),

[0147]

[0148] z i+1 = f enc,i+1 (W enc,i+1 · [x i+1 , z i + b enc,i+1 ),

[0149] where x i is the parameter vector of the i-th layer of soil, z i is the latent representation of the i-th layer, is the reconstruction output, f enc,i and f dec,i are the encoder and decoder activation functions respectively, W enc,i , W dec,i , b enc,i and b dec,i are the corresponding weight and bias parameters. In particular, [x i+1 , z i represents concatenating the raw data of the next layer and the latent representation of the current layer to achieve information transfer between layers.

[0150] The introduction of physical constraints is achieved by modifying the loss function:

[0151] L total = L recon + λ 1 · L phys + λ 2 · L reg ,

[0152] where L recon is the reconstruction loss, usually using the mean squared error:

[0153]

[0154] L phys is the physical constraint loss, ensuring that the model predictions conform to the soil physical laws, such as the water movement equation:

[0155]

[0156] where θ i is the soil water content, K(θ i ) is the unsaturated hydraulic conductivity function, ψ(θ i ) is the soil water potential function, and z is the soil depth.

[0157] L reg is the regularization loss, preventing the model from overfitting:

[0158]

[0159] λ 1 and λ 2is the weight coefficient, usually λ 1 takes values between 0.1 and 1, λ 2 takes values between 0.001 and 0.01.

[0160] The convolutional neural network unit 33 is used to perform image recognition on soil profile pictures and analyze the microbial community structure. The typical structure of a convolutional neural network includes multiple convolutional layers, pooling layers, and fully connected layers. In the embodiments of the present invention, a multi-scale feature extraction architecture is adopted, which includes parallel convolutional branches and uses convolutional kernels of different sizes (such as 3×3, 5×5, 7×7) to extract multi-scale features.

[0161] The operation of the convolutional layer can be expressed as:

[0162] F l =σ(W l *F l-1 +b l ),

[0163] where F l is the feature map of the l-th layer, W l is the convolutional kernel, b l is the bias, * represents the convolutional operation, and σ is the activation function, usually ReLU is adopted.

[0164] To improve the accuracy of microbial recognition, the model also incorporates an attention mechanism to focus on key regions in the image:

[0165] A = σ(F conv ),

[0166] F att = F conv ⊙A,

[0167] where F conv is the convolutional feature map, A is the attention map, is the element-wise multiplication, and F att is the feature map after adding attention.

[0168] Microbial community analysis includes quantity statistics, type recognition, and diversity assessment. Quantity statistics is achieved through image segmentation and instance counting; type recognition is achieved through a classification network, and common soil microbial types include bacteria, fungi, actinomycetes, etc.; diversity assessment is calculated through diversity indices such as the Shannon index and Simpson index.

[0169] The construction of the microbe-environment interaction network uses a graph neural network. The nodes in the network include microbial populations and environmental factors, and the edges represent their interactions. The interaction intensity is determined through correlation analysis. Usually, variable pairs with a Pearson correlation coefficient greater than 0.7 or less than -0.7 are considered to have a significant correlation.

[0170] The recurrent neural network unit 34 is used to combine meteorological data to analyze the characteristics of soil profile water migration, including infiltration characteristics during the rainy season and evaporation characteristics during the dry season. The water migration model is based on the Richards equation and combines deep learning methods, and can be expressed as:

[0171]

[0172] Among them, θ is the volumetric water content of the soil, t is the time, z is the soil layer depth (positive downward), K(θ) is the unsaturated hydraulic conductivity, ψ(θ) is the soil water potential, and S(z,t) is the root water uptake term.

[0173] The unsaturated hydraulic conductivity K(θ) and the soil water potential ψ(θ) are usually determined by empirical formulas, such as the van Genuchten-Mualem model:

[0174]

[0175] Among them, K s is the saturated hydraulic conductivity, θ s is the saturated water content, θ r is the residual water content, α, n, and m are shape parameters, usually m = 1 - 1 / n, S e is the effective saturation.

[0176] The recurrent neural network is used to predict the spatio-temporal changes of soil moisture. The model inputs include soil properties (texture, density, etc.), initial moisture conditions, meteorological data (rainfall, temperature, humidity, wind speed, etc.), and boundary conditions, and the output is the soil moisture distribution at different times and depths.

[0177] The analysis of infiltration characteristics during the rainy season mainly focuses on the rainfall infiltration process, deep percolation, and surface runoff, and the analysis of evaporation characteristics during the dry season mainly focuses on the evaporation rate, evaporation depth, and soil drying process. These characteristics are of great significance for formulating scientific irrigation strategies and drought prevention and disaster reduction measures.

[0178] The decision support module 4 includes a soil health assessment unit 41, an irrigation strategy optimization unit 42, a three-dimensional visualization unit 43, and a warning and recommendation unit 44.

[0179] The soil health assessment unit 41 is used to generate a soil health status assessment report based on the analysis results of the deep learning module 3. The soil health assessment adopts a comprehensive index system, including:

[0180] 1) Soil fertility indicators: organic matter content (usually > 2% is good), pH value (usually 6 - 7.5 is suitable), nutrient status (N, P, K, trace elements, etc.);

[0181] 2) Soil physical indicators: texture, structure, bulk density (usually < 1.4 g / cm 3 is good), air permeability (usually porosity > 40% is good);

[0182] 3) Soil biological indicators: microbial diversity (Shannon index usually > 3 is good), microbial activity (usually dehydrogenase activity > 20 μg TPF / g·24h is good).

[0183] The evaluation results are usually divided into four grades: excellent, good, medium, and poor, and the corresponding comprehensive scores are 85 - 100, 70 - 85, 55 - 70, and < 55 points respectively.

[0184] The irrigation strategy optimization unit 42 is used to generate an optimal irrigation strategy through a multi - objective optimization algorithm and deep reinforcement learning technology. The irrigation optimization objectives include: 1) minimizing the irrigation water volume; 2) maximizing the crop yield; 3) maximizing the water use efficiency; 4) minimizing the energy consumption.

[0185] The multi - objective optimization problem can be expressed as:

[0186] minF(x) = [f 1 (x), f 2 (x), …, f n (x)] T

[0187] s.t.g i (x) ≤ 0, i = 1, 2, …, m

[0188] h j (x) = 0, j = 1, 2, …, p

[0189] x L ≤ x ≤ x U

[0190] where F(x) is the objective function vector, x is the decision variable vector (such as irrigation timing, irrigation volume, etc.), g i (x) and h j (x) are the inequality and equality constraints respectively, x L and x U are the lower and upper bounds of the decision variable.

[0191] Preferably, an improved NSGA - II (Non - dominated Sorting Genetic Algorithm) or MOPSO (Multi - Objective Particle Swarm Optimization Algorithm) is used to solve the multi - objective optimization problem, obtaining a series of Pareto optimal solutions, and then the most suitable solution is selected according to the actual situation as the irrigation strategy.

[0192] Deep reinforcement learning methods can be used for real-time irrigation decision-making, where the state space includes soil moisture, weather forecasts, crop growth stages, etc., the action space is irrigation decisions (irrigate or not, irrigation amount), and the reward function combines water use efficiency, crop yield, and energy consumption. Preferably, a deep Q-network (DQN) or policy gradient method is adopted to learn the optimal irrigation strategy by interacting with the environment.

[0193] The 3D visualization unit 43 is used to achieve an interactive 3D visualization display of soil profile data through WebGL technology. The visualization content includes soil profile structure, parameter distribution of each layer, microbial distribution, and water migration process. Users can view soil information at different angles and depths through operations such as zooming, rotating, and slicing. Color coding is used to represent the magnitude of parameter values, usually using a heat map or rainbow color scheme.

[0194] The early warning recommendation unit 44 is used to generate graded early warning information and improvement measure suggestions. The early warning levels are usually divided into four levels: blue (slight), yellow (medium), orange (severe), and red (extremely severe). The early warning content includes problems such as soil drought, nutrient deficiency, acidification / alkalization, salinization, and microbial imbalance. Each level of early warning corresponds to different trigger thresholds. For example, a soil water content < 50% triggers a blue early warning, < 30% triggers a yellow early warning, < 20% triggers an orange early warning, and < 10% triggers a red early warning.

[0195] The improvement measure suggestions are generated based on the early warning type and severity, including irrigation adjustment, fertilization suggestions, soil improvement measures, etc. The suggestion content takes into account local resource conditions, economic feasibility, and environmental sustainability to ensure the practicality and operability of the suggestions.

[0196] The system of the present invention further includes a federated learning module 5 for performing local model training on different monitoring terminal devices; only sharing model parameters rather than raw data to protect data privacy; protecting individual monitoring data through differential privacy technology; and automatically splitting the deep learning model according to the computing power of different terminal devices.

[0197] The workflow of federated learning includes the following steps:

[0198] 1) The central server initializes the global model parameters θ 0 ;

[0199] 2) Distribute the model parameters to the participating terminal devices;

[0200] 3) Each terminal device uses local data to train the model and calculates the parameter updates;

[0201] 4) The central server aggregates the parameter updates of all terminal devices and updates the global model;

[0202] 5) Repeat steps 2 - 4 until the model converges.

[0203] Parameter aggregation usually adopts the weighted average method:

[0204]

[0205] where, θ t +1 is the global model parameter in the (t + 1)-th round, θ i ,t+1 is the local model parameter of the i-th terminal device in the (t + 1)-th round, n i is the data volume of the i-th terminal device, n is the total amount of all data, and N is the number of terminal devices participating in the training.

[0206] Differential privacy protection is achieved by adding noise to parameter updates, usually adopting the Laplace mechanism or the Gaussian mechanism:

[0207]

[0208] where, θ ′ i,t+1 is the parameter update after adding noise, represents Gaussian noise with a mean of 0 and a variance of σ 2 . The noise size σ is inversely proportional to the privacy budget ∈. The smaller ∈ is, the higher the privacy protection level, but the model performance may decline.

[0209] Model segmentation is determined based on the device computing power. Usually, computationally intensive tasks (such as deep network training) are assigned to devices with strong computing power, and lightweight tasks (such as shallow networks or preprocessing) are assigned to devices with weak computing power.

[0210] Data is transmitted between the sensor network module 1 and the data processing module 2 through wireless communication methods. Common wireless communication technologies include ZigBee, LoRa, NB-IoT, etc. In the karst area with complex terrain and dense vegetation, LoRa technology has obvious advantages due to its long distance (up to 10 km at most), low power consumption, and strong penetration. The communication frequency usually uses the ISM band (433 MHz, 868 MHz, or 915 MHz), and the communication rate is 0.3 - 50 kbps, which is sufficient to meet the data transmission requirements of environmental monitoring.

[0211] Data processing module 2 and deep learning module 3 perform data transmission and processing through an edge-cloud collaborative computing architecture. The edge side (such as a gateway or edge server) is responsible for preliminary data processing, time-sensitive analysis, and local storage, while the cloud side is responsible for deep learning model training, large-scale data mining, and long-term data storage. The data segmentation strategy is determined based on timeliness requirements and computational complexity: tasks with a response time requirement of less than 100 ms (such as anomaly detection) are deployed on the edge side; tasks with high computational complexity (such as model training) are deployed on the cloud side.

[0212] Please refer to Figure 4 , the present invention also provides a monitoring method based on the above intelligent monitoring system for the soil profile of the karst critical zone, including the following steps:

[0213] S1: According to the geomorphic characteristics of the karst area, a multi-level sensor network is deployed at multiple soil profile depths;

[0214] S2: Collect moisture, temperature, pH value, organic matter content, and microbial data of the karst soil profile;

[0215] S3: Perform spatio-temporal alignment, quality assessment, and edge computing preprocessing on the collected heterogeneous data;

[0216] S4: Analyze the temporal variation of soil parameters through a recurrent neural network and establish a temporal prediction model of soil parameters;

[0217] S5: Establish a soil profile evolution model through a deep autoencoder to simulate the variation law of data in different soil layers;

[0218] S6: Use a convolutional neural network to analyze the microbial community of soil profile pictures;

[0219] S7: Combine meteorological data and use a recurrent neural network to analyze the characteristics of soil water movement;

[0220] S8: Generate a soil health assessment report and an optimal irrigation strategy based on the analysis results;

[0221] S9: Dynamically adjust the sensor sampling frequency to optimize the energy consumption of the system;

[0222] S10: Optimize the overall model performance through federated learning while protecting data privacy.

[0223] In step S1, the deployment of the multi-level sensor network is based on the geomorphic features and soil profile stratification in karst areas. Common geomorphic features in karst areas include karst depressions, funnels, stone buds, and solution grooves, etc. The soil profile is usually thin and has obvious stratification. Preferably, the soil profile can be divided into the surface layer (0 - 20 cm), subsurface layer (20 - 40 cm), middle layer (40 - 80 cm), and deep layer (> 80 cm), and corresponding sensor nodes are deployed in each layer. The spatial distribution of sensor nodes adopts the stratified random sampling or systematic grid deployment method to ensure the representativeness and coverage of monitoring points.

[0224] In step S2, the acquisition frequency of sensors is dynamically adjusted according to the parameter change rate and monitoring purpose. For parameters that change rapidly (such as moisture and temperature), the initial acquisition frequency is set relatively high (such as once every 15 - 30 minutes); for parameters that change slowly (such as pH value and organic matter content), the initial acquisition frequency is set relatively low (such as once every 24 hours). Microbial data is obtained regularly through image acquisition equipment (such as once every 3 - 7 days).

[0225] In step S3, the spatio-temporal alignment of heterogeneous data adopts the multi-scale time window method and spatial interpolation technology. Time alignment is achieved by setting a basic time unit (such as 1 hour), and data with different acquisition frequencies are mapped to a unified time axis through interpolation or aggregation. Spatial alignment is achieved through the geographic coordinate system and depth information, combined with Kriging interpolation or inverse distance weighting method, to realize the unified expression of spatial heterogeneous data.

[0226] Data quality assessment is achieved through range check, consistency check, and time continuity check. Range check sets the normal change range of each parameter (such as soil moisture 0 - 100%, pH value 4 - 9, temperature - 10°C to 50°C, etc.); consistency check verifies the relationship between relevant parameters (such as temperature increase is usually accompanied by moisture decrease); time continuity check identifies mutations or jumps in the data. Abnormal data processing strategies include: 1) Slight anomalies (such as noise) are smoothed through filtering; 2) Moderate anomalies (such as short-term missing values) are filled through interpolation; 3) Severe anomalies (such as long-term missing values or hardware failures) are marked and maintenance is triggered.

[0227] Edge computing preprocessing includes data filtering, compression, and feature extraction. Filtering uses moving average, median filtering, or Kalman filtering to remove noise; data compression uses time series compression algorithms such as PLA (Piecewise Linear Approximation) or SVD (Singular Value Decomposition); feature extraction calculates statistical features (mean, variance, maximum value, minimum value, trend, etc.) and domain features (such as moisture gradient, temperature integral, etc.). After preprocessing, the data volume is usually reduced by 70% - 90%, greatly reducing the communication burden.

[0228] In step S4, the recurrent neural network model uses an LSTM or GRU structure to analyze the temporal changes of soil parameters. The model inputs include historical observation data (such as parameter values in the past 7 - 30 days) and auxiliary features (such as weather conditions, agricultural activities, etc.), and the output is the predicted value of future parameters (such as the change trend of parameters in the next 1 - 14 days). The model is trained using the sliding window method. The loss function is usually the mean square error or the mean absolute error, and the optimization algorithm uses Adam or RMSprop.

[0229] Predicting the parameter change trend is of great significance for timely detecting potential problems and formulating preventive measures. For example, a continuous downward trend in soil moisture may indicate a drought risk; a rapid change in pH value may suggest soil acidification or alkalization problems; a decrease in organic matter content may reflect soil degradation. By predicting these change trends in advance, corresponding measures (such as adjusting irrigation strategies, applying lime or organic fertilizers) can be taken to prevent the problems from worsening.

[0230] In step S5, the deep autoencoder soil profile evolution model realizes the simulation and prediction of the variation laws of data in different soil layers through a hierarchical nested structure and physical constraints. The model can identify key parameters and critical change points in the soil profile and predict future evolution trends.

[0231] The soil profile evolution model pays special attention to the interactions between layers, such as the influence of nutrient leaching in the upper layer on the lower layer, and the influence of rising water supply in the lower layer on the upper layer, etc. By capturing these interactions, the model can more accurately predict the dynamic change process of the entire profile. In particular, in karst areas, the soil profile is often thin and there are various discontinuities (such as limestone fragments, karst caves, etc.), and these special structures have important effects on the transport of water and nutrients. The model captures these effects by introducing karst characteristic factors.

[0232] Model evaluation not only focuses on the overall prediction accuracy, but also particularly focuses on the prediction accuracy of key points (such as turning points, extreme points) and the degree of satisfaction of physical constraints. Generally, it is required that the root mean square error be controlled within 5% of the parameter range, and the prediction deviation of key points be controlled within ±2 days of the time point and ±10% of the parameter value.

[0233] In step S6, the convolutional neural network performs microbial community analysis on soil profile pictures. The analysis content includes microbial quantity statistics, type identification, and diversity assessment.

[0234] Microbial quantity statistics is achieved through image segmentation and instance counting. First, instance segmentation algorithms such as Mask R - CNN are used to identify microbial individuals in the image, and then the quantities of different types of microorganisms are counted. Type identification is achieved through a classification network. Common types of soil microorganisms include bacteria (such as rhizobia, nitrogen - fixing bacteria, etc.), fungi (such as mycorrhizal fungi, saprophytic fungi, etc.), actinomycetes, protozoa, etc.

[0235] Diversity assessment is achieved by calculating various diversity indices, including:

[0236] 1) Shannon index:

[0237] where S is the number of species, and p i is the relative abundance of the i-th species;

[0238] 2) Simpson index:

[0239] 3) Pielou evenness index:

[0240] Microbial community structure analysis also includes dominant population identification, functional community analysis, and network structure analysis. Dominant populations are determined by relative abundance, and usually, species with a relative abundance > 10% are regarded as dominant species; functional communities are classified and analyzed according to the ecological functions of microorganisms (such as organic matter decomposers, nitrogen-fixing bacteria, etc.); network structure analysis identifies key species and community stability by constructing an interspecies association network.

[0241] In step S7, the recurrent neural network combines meteorological data to analyze the characteristics of soil profile water migration. The analysis content includes infiltration characteristics during the rainy season (such as infiltration rate, infiltration depth, preferential flow path, etc.) and evaporation characteristics during the dry season (such as evaporation rate, evaporation depth, soil drying process, etc.).

[0242] The water migration model is based on the Richards equation and deep learning methods, and the inputs include soil physical properties (such as texture, structure, bulk density, etc.), initial water content, meteorological conditions (such as rainfall, temperature, humidity, wind speed, etc.), and boundary conditions. The model captures the spatio-temporal dynamic characteristics of water migration through a recurrent neural network structure and predicts the changes in water content at different depths and time points.

[0243] The analysis of infiltration characteristics during the rainy season focuses on the water infiltration process under heavy rain and rainstorm conditions, including surface infiltration rate (usually between 0.5 - 20 mm / h), infiltration depth, and conditions for the occurrence of preferential flow, etc. Preferential flow is a typical phenomenon in karst areas, referring to the rapid flow of water along fractures and large pores, which usually occurs when the rainfall intensity > 10 mm / h or under soil pre-wetting conditions.

[0244] The analysis of evaporation characteristics during the dry season focuses on the water loss process under drought conditions, including evaporation rate (usually between 0.1 - 5 mm / day), evaporation depth (usually between 10 - 50 cm), and the time course of the drying process. The evaporation process is usually divided into three stages: the constant rate stage (when water is sufficient), the falling rate stage (when water begins to be limited), and the diffusion stage (when the soil surface has dried).

[0245] The analysis of water movement characteristics has important guiding significance for irrigation decision-making and water resource management. By understanding the infiltration and evaporation characteristics under different conditions, the irrigation timing and amount can be optimized, the water resource utilization efficiency can be improved, and the impact of drought can be alleviated.

[0246] In step S8, based on the foregoing analysis results, a soil health assessment report and an optimal irrigation strategy are generated. The soil health assessment report comprehensively considers the soil fertility status, physical properties, and biological activities to comprehensively evaluate the soil health status. The evaluation index system includes fertility indexes (organic matter content, pH value, nutrient status, etc.), physical indexes (texture, structure, bulk density, etc.), and biological indexes (microbial diversity, activity, etc.). The weights of each index are determined by the analytic hierarchy process. Usually, the weight of the fertility index is 0.4, the weight of the physical index is 0.3, and the weight of the biological index is 0.3.

[0247] The evaluation results are divided into four grades: excellent (85 - 100 points), good (70 - 85 points), medium (55 - 70 points), and poor (<55 points), corresponding to different soil conditions and management suggestions. The excellent grade indicates that the soil health status is extremely good, and the existing management method can be maintained; the good grade indicates that the soil health status is good, but some aspects need to be improved; the medium grade indicates that there are obvious problems in the soil and improvement measures need to be taken; the poor grade indicates that the soil degradation is serious and systematic repair is needed.

[0248] The optimal irrigation strategy is generated through multi-objective optimization and reinforcement learning methods, considering multiple objectives such as water resource utilization efficiency, crop yield, and economic benefits. The irrigation strategy includes three key parameters: irrigation timing, irrigation amount, and irrigation method.

[0249] The irrigation timing is usually determined based on the soil water content threshold. Preferably, when the available soil water content drops to 60% - 70%, irrigation starts. At this time, the plants have not yet been under water stress, and the irrigation efficiency is the highest. The irrigation amount is determined based on the soil water deficit (the difference between the saturated water content and the current water content) and the root zone depth, usually 70% - 90% of the soil water deficit in the root zone, to avoid excessive irrigation causing deep percolation and nutrient loss. The irrigation method is selected according to the terrain conditions, soil characteristics, and crop types. Micro-irrigation (drip irrigation or micro-sprinkler irrigation) is usually recommended in karst areas to improve water use efficiency.

[0250] In step S9, the sensor sampling frequency is dynamically adjusted to optimize the system energy consumption. The sampling frequency adjustment is based on the parameter change rate and the uncertainty of the prediction model. When the parameter change rate is large or the model uncertainty is high, the sampling frequency is increased; conversely, the sampling frequency is decreased to save energy.

[0251] The adjustment algorithm is expressed as:

[0252]

[0253] Among them, f i (t + 1) is the sampling frequency of the i-th sensor at time t + 1, and f b ase is the base sampling frequency (usually once per hour), and P i (t) and P i (t - 1) are the parameter values at time t and t - 1 respectively, and P max and P min are the maximum and minimum values of the parameter, and U i (t) is the uncertainty of the prediction model for this parameter (which can be measured by the width of the prediction interval), and α and β are weight coefficients. Usually, α ranges from 0.5 to 2, and β ranges from 1 to 3.

[0254] To prevent the sampling frequency from being too high or too low, upper and lower limit constraints are set:

[0255] f min ≤f i (t + 1)≤f max ,

[0256] Among them, f min is usually set to once every 6 hours, and f max is usually set to once every 5 minutes.

[0257] Dynamically adjusting the sampling frequency can significantly reduce energy consumption while maintaining data quality. Experiments show that compared with fixed-frequency sampling, this method can reduce energy consumption by 30% - 60% and extend the system operation time.

[0258] In step S10, the overall model performance is optimized through federated learning while protecting data privacy. Federated learning allows multiple monitoring points to collaboratively train a model without sharing the original data, which is particularly suitable for monitoring systems with distributed deployment.

[0259] The workflow of federated learning includes: 1) The central server initializes the global model; 2) Distribute the model to each monitoring point; 3) Each monitoring point uses local data to train the model; 4) The monitoring point sends the model update (instead of the original data) to the central server; 5) The central server aggregates all updates and updates the global model; 6) Repeat steps 2 - 5 until the model converges.

[0260] To further protect privacy, differential privacy technology is introduced in the model update to add carefully calibrated noise. The size of the noise is related to the privacy budget ∈. The smaller ∈ is, the stronger the privacy protection, but it may affect the model performance. Preferably, ∈ ranges from 1 to 5 to achieve a balance between protecting privacy and maintaining model performance.

[0261] Communication efficiency is one of the key challenges in federated learning, especially in karst areas with limited network conditions. To reduce communication overhead, techniques such as model compression (e.g., parameter quantization, sparsification) and gradient compression (e.g., Top-k sparsification) are adopted, which can usually reduce the communication volume by 50%-80%. In addition, an asynchronous update strategy is adopted, allowing monitoring points to participate in training at their own pace and avoiding the overall training stagnation caused by the delay of individual nodes.

[0262] Through federated learning, monitoring points under different regions and management conditions can collaborate to build a more robust and generalizable model while protecting their respective data privacy. Experiments show that compared with single-point training, federated learning can improve the model prediction accuracy by 10%-25%, especially in data-sparse regions.

[0263] Generally speaking, the intelligent monitoring system and method for the soil profile of the karst critical zone provided by the present invention realize the comprehensive, real-time, intelligent monitoring and analysis of the soil profile in the karst area through multi-level sensor networks, edge computing, deep learning models, and federated learning technologies. The system can adapt to the complex topographical and soil structure characteristics of the karst area, provide high-precision soil parameter monitoring, profile evolution prediction, microbial community analysis, and water migration characteristic analysis, and provide a scientific basis for soil health assessment and irrigation decision-making. The adaptive sampling strategy and energy management technology adopted by the system significantly reduce energy consumption and extend the system operation time; the federated learning technology improves the model performance while protecting data privacy. These innovations make the system have broad application prospects in the fields of soil monitoring, agricultural irrigation management, and ecological environment protection in the karst area.

[0264] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. Intelligent monitoring system for soil profile in key karst zones, characterized by: include: A sensor network module for collecting moisture, temperature, pH, organic matter content, and microbial data of karst soil profiles; A data processing module, which is in communication with the sensor network module and is used to receive data collected by the sensor network module and perform edge computing preprocessing and data analysis; A deep learning module, which is in communication with the data processing module and is used to construct a soil profile evolution model, a microbial community analysis model, and a water migration model; A decision support module is in communication with the deep learning module and is used to generate a soil health assessment report and irrigation strategy recommendations based on the analysis results.

2. The intelligent monitoring system for soil profile in karst key zone according to claim 1 is characterized in that: The sensor network module comprises: Multi-level sensor node units are used to deploy data collection nodes, intermediate nodes and edge nodes in different depths of soil layers in karst areas to form a tree-type topology network; An adaptive power management unit, used to provide power to the multi-level sensor node unit through a solar panel, a battery management submodule and a voltage stabilizing rectifier circuit; The networking communication unit is used to establish a distributed data transmission channel between the data acquisition nodes, intermediate nodes and edge nodes.

3. The intelligent monitoring system for soil profile in karst key zone according to claim 1 is characterized in that: The data processing module comprises: Heterogeneous data spatiotemporal alignment unit, used to accurately align data from different sensors and different time and space scales; Data quality assessment unit, used to conduct real-time quality assessment of data using statistical methods, and automatically identify and process abnormal data; Edge computing units, which perform data preprocessing near the data source to reduce the amount of data transmission; The adaptive sampling control unit is used to dynamically adjust the sampling frequency of each sensor node based on the environmental parameter evaluation model.

4. The intelligent monitoring system for soil profile in karst key zone according to claim 1 is characterized in that: The deep learning module includes: Recursive neural network unit, used to analyze the time series data of karst soil profile and predict parameter change trends; Deep autoencoding unit, used to establish a soil profile evolution model, simulating the change law of different soil layer data through the autoencoding network layer and feature extraction layer; Convolutional neural network unit, used for image recognition of soil profile pictures and analysis of microbial community structure; Recurrent neural network unit is used to analyze the water movement characteristics of soil profiles in combination with meteorological data, including infiltration characteristics in the rainy season and volatilization characteristics in the dry season.

5. The intelligent monitoring system for soil profile in karst key zone according to claim 4 is characterized in that: The deep autoencoder unit adopts a hierarchical nested coding structure, each layer of the autoencoder corresponds to a soil layer, and the laws of soil physics are incorporated as constraints in the neural network training.

6. The intelligent monitoring system for soil profile in karst key zone according to claim 4 is characterized in that: The convolutional neural network unit constructs a microorganism-environment interaction network, which can analyze the number of microbial populations, dynamic changes in functional communities, and community succession trends.

7. The intelligent monitoring system for soil profile in karst key zone according to claim 1 is characterized in that: The decision support module includes: Soil health assessment unit, used to generate soil health assessment report based on the analysis results of deep learning module; Irrigation strategy optimization unit, used to generate the optimal irrigation strategy through multi-objective optimization algorithm and deep reinforcement learning technology; 3D visualization unit, used to realize interactive 3D visualization display of soil profile data through WebGL technology; The warning recommendation unit is used to generate graded warning information and suggestions for improvement measures.

8. The intelligent monitoring system for soil profile in karst key zone according to claim 1 is characterized in that: Also includes federated learning modules for: Conduct local model training on different monitoring terminal devices; Only model parameters are shared instead of original data to protect data privacy; Protect individual monitoring data through differential privacy technology; Automatically segment deep learning models based on the computing power of different terminal devices.

9. The intelligent monitoring system for soil profile in karst key zone according to claim 1 is characterized in that: The sensor network module and the data processing module transmit data via wireless communication, and the data processing module and the deep learning module transmit and process data via an edge-cloud collaborative computing architecture.

10. A monitoring method based on the intelligent monitoring system for karst key zone soil profile according to any one of claims 1 to 9, characterized in that: The following steps are involved: According to the geomorphic characteristics of the karst area, a multi-level sensor network is deployed at multiple soil profile depths; Collect moisture, temperature, pH, organic matter content and microbial data of karst soil profiles; Perform spatiotemporal alignment, quality assessment, and edge computing preprocessing on the collected heterogeneous data; The time series changes of soil parameters are analyzed through recursive neural network, and a time series prediction model of soil parameters is established; A soil profile evolution model is established through a deep autoencoder to simulate the changing patterns of data in different soil layers; Use convolutional neural networks to analyze microbial communities in soil profile images; Combined with meteorological data, a recurrent neural network is used to analyze soil moisture migration characteristics; Generate soil health assessment reports and optimal irrigation strategies based on analysis results; Dynamically adjust sensor sampling frequency to optimize system energy consumption; The overall model performance is optimized while protecting data privacy through federated learning.

Images