Saline-alkali soil water-salt-fertilizer real-time monitoring system and method based on Internet of Things

Abstract

The invention discloses a saline-alkali soil water-salt-fertilizer real-time monitoring system and method based on the Internet of Things, and relates to the technical field of saline-alkali soil monitoring. Multi-parameter sensor nodes are arranged in a preset area to collect soil salinity, moisture, conductivity, pH value and ground temperature information; acquiring a remote sensing image, terrain elevation and land utilization type data, and performing preprocessing and spatial registration; based on the spatial position relation, the hydrological connectivity and the irrigation and drainage structure, establishing a spatial graph structure required by graph neural network input; deploying a model on the constructed graph, and predicting the future salinity trend or saline-alkali risk grade of each monitoring unit; utilizing adjacent node features and an edge weight propagation mechanism to realize prediction extrapolation of a data sparse region; based on a prediction result and an agronomic rule, generating a multi-objective optimization regulation and control scheme through an evolutionary strategy algorithm; and displaying the prediction map, the risk map and the regulation and control suggestions, and triggering early warning and control linkage when the threshold value exceeds the limit.

Description

Technical Field

[0001] This invention relates to the field of saline-alkali land monitoring technology, specifically to a real-time monitoring system and method for water-salt-fertilizer in saline-alkali land based on the Internet of Things. Background Technology

[0002] Real-time monitoring of water, salt, and fertilizer in saline-alkali land based on the Internet of Things (IoT) refers to the use of IoT technology to collect key data such as soil moisture, salinity, and nutrients (fertility) in real time by deploying sensors and communication devices in saline-alkali land. This data is then transmitted to a monitoring platform via a wireless network, enabling continuous monitoring and intelligent management of the soil's water, salt, and fertilizer status. This technology helps to precisely regulate irrigation and fertilization programs, improve soil structure in saline-alkali land, and enhance agricultural production efficiency and sustainability.

[0003] However, in the management of saline-alkali land, the poor spatial generalization ability leads to a particularly serious problem of "prediction failure in blind spots." Since the model mainly relies on data from the sensor coverage area for training, areas at the edge or far from the sensor layout (such as near villages, reclaimed areas, or reservoirs) often suffer from distorted predictions due to data sparsity. Taking a pilot area in the Northeast Plain as an example, the model performed well in the main plot, but incorrectly predicted "normal salinity" in peripheral areas, while field surveys showed severe salinization. This misjudgment can easily lead to an imbalance in the remediation zoning scheme, causing pollution sources to go unidentified, ultimately resulting in salt seeping back into core farmland and undermining the overall remediation effectiveness. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time monitoring system and method for water-salt-fertilizer in saline-alkali land based on the Internet of Things, so as to solve the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a real-time monitoring system for water-salt-fertilizer in saline-alkali land based on the Internet of Things, comprising:

[0006] The sensor acquisition module is used to deploy multiple sensor nodes in a preset area of ​​saline-alkali land to collect information on soil salinity, moisture, electrical conductivity, pH value and ground temperature in the target area.

[0007] The remote sensing and geographic information module is used to acquire remote sensing images, topographic elevation data and land use type data covering the target area, and to preprocess and spatially register the data.

[0008] The spatial graph construction module is used to establish a spatial graph structure based on the positional relationship, geographical adjacency, groundwater flow path or irrigation and drainage channel between the sensor nodes and remote sensing pixels. Each monitoring unit is represented as a node of the graph neural network, and the spatial connection relationship between nodes is represented as an edge.

[0009] The graph neural network prediction module is used to deploy a graph neural network model on the constructed spatial map. By performing graph convolution calculations on the features of sensor nodes, it predicts the soil salinity evolution trend or salinity risk level of target nodes within a preset time range in the future.

[0010] The edge plot prediction module is used to estimate the salinity trend of edge plots or spatial blind spots without sensor coverage by using information transmission between adjacent nodes and graph structure learning.

[0011] The dynamic regulation decision module is used to generate a dynamic regulation strategy based on the prediction results of the graph neural network, combined with historical management parameters and agronomic rules, and output control commands to the field control terminal.

[0012] The visualization and early warning module is used to display in real time the salinity prediction map, risk level map and control suggestions of each plot in the target area. If the predicted value exceeds the safety threshold, an early warning signal will be automatically triggered.

[0013] Preferably, the sensor acquisition module includes:

[0014] Multiple multi-parameter soil sensor nodes are deployed in representative areas within the target saline-alkali land. These areas include the main cultivated area, marginal plots, areas adjacent to drainage ditches, and areas near water storage areas. Each sensor node is vertically deployed at three depth levels: 020cm, 2050cm, and 50–100cm. Each node periodically collects data on electrical conductivity, volumetric water content, soil pH, and ground temperature at a sampling frequency of 5–15 minutes. The data is transmitted to an edge gateway or cloud platform via LoRa, NB-IoT, or 4G communication modules.

[0015] Preferably, the spatial map construction module includes:

[0016] Abstract sensor monitoring points, remote sensing pixel centers, or farmland boundary centers into nodes in a graph neural network;

[0017] The edge connections between nodes are determined based on geographical proximity, consistency of water flow paths, and similarity between the irrigation and drainage system structure and the environment.

[0018] Each edge is assigned a weight, which is calculated based on factors such as geographical distance, consistency of water flow paths, and differences in crop salt sensitivity coefficients.

[0019] Output the node feature matrix, adjacency matrix, and edge weight matrix.

[0020] Preferably, the geographical proximity is calculated by using the Euclidean distance inverse ratio or Gaussian kernel function; the water flow path matrix is ​​derived based on the DEM to determine whether the nodes belong to the same catchment unit, and if so, the weight is doubled; the differences in salt sensitivity of the crop root zone of the nodes are compared, the difference value of crop root salt sensitivity coefficient (CSSG) is calculated, and normalized to the edge weight inverse adjustment factor; finally, all edge weights are normalized to the range of 0 to 1 to form a weighted adjacency matrix.

[0021] Preferably, the method for calculating the difference in crop root salt sensitivity coefficient is as follows: For each node, read the average conductivity within the current depth range of 0–40 cm or 0–60 cm; for each plot of crop, extract the threshold conductivity (Threshold ECe), i.e., the salt starting point at which crop yield begins to decrease, from the agricultural standard database according to its crop type and growth stage; slope parameter: the percentage decrease in yield for every 1 dS / m increase in salt content; for each node, substitute into the crop response model to calculate the difference in crop root salt sensitivity coefficient (CSSG) between the two plots.

[0022] Preferably, the dynamic control decision module includes:

[0023] Receive soil salinity trend and risk level results; call historical regulation data, crop type and agronomic rule base to construct regulation objective function; introduce evolutionary strategy optimization algorithm to generate optimal control variable combination based on multi-objective function, including irrigation amount, drainage duration and application cycle; convert the optimal strategy into executable control command format and send it to the corresponding field terminal equipment.

[0024] Preferably, multiple control strategy sample combinations are initialized;

[0025] Calculate the salt control score, water resource efficiency score, crop response score, and cost penalty factor for each combination in the future forecast period;

[0026] New strategy combinations are generated by automatically adjusting the sampling covariance matrix and step size based on the fitness function.

[0027] Output the strategy with the highest score after multiple iterations;

[0028] If the constraints are not met, return to the evolutionary loop until they are met.

[0029] This invention also provides a method for real-time monitoring of water-salt-fertilizer in saline-alkali land based on the Internet of Things, comprising:

[0030] Multiple sensor nodes are deployed in a pre-defined area of ​​saline-alkali land to collect information on soil salinity, moisture, electrical conductivity, pH value and ground temperature in the target area.

[0031] Acquire remote sensing images, topographic elevation data, and land use type data covering the target area, and perform preprocessing and spatial registration on the data;

[0032] A spatial graph structure is established based on the positional relationship, geographical adjacency, groundwater flow path, or irrigation and drainage channel between the sensor nodes and remote sensing pixels. Each monitoring unit is represented as a node of the graph neural network, and the spatial connection relationship between nodes is represented as an edge.

[0033] A graph neural network model is deployed on the constructed spatial map. By performing graph convolution calculations on the features of sensor nodes, the soil salinity evolution trend or salinity risk level of the target node within a future preset time range is predicted.

[0034] For edge plots or spatial blind spots without sensor coverage, salinity trend estimation is completed through information transmission between adjacent nodes and graph structure learning;

[0035] Based on the prediction results of the graph neural network, combined with historical management parameters and agronomic rules, a dynamic regulation strategy is generated, and control commands are output to the field control terminal.

[0036] It displays real-time salinity prediction maps, risk level maps, and control recommendations for various plots within the target area. If the predicted value exceeds the safety threshold, it automatically triggers an early warning signal.

[0037] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0038] 1. This invention achieves precise perception and control of the spatial heterogeneity and dynamic evolution of saline-alkali land by introducing multi-source data fusion, graph neural network modeling, and intelligent optimization decision-making. The system combines sensor-measured data and remote sensing imagery to construct a spatial graph structure, and uses graph convolution and attention mechanisms to effectively predict edge blind areas. This solves the problems of prediction failure and frequent governance blind spots in traditional methods in data-sparse areas, and significantly improves prediction accuracy and spatial generalization ability.

[0039] 2. This invention employs a covariance matrix adaptive evolutionary strategy (CMA-ES) to automatically generate multi-objective control schemes, effectively balancing salinity control, water resource utilization, crop growth requirements, and control costs. It achieves a closed-loop control process from data perception to risk prediction, strategy generation, and execution feedback. The system features visualization and intelligent early warning capabilities, and is widely applicable to scenarios such as agricultural digital governance, smart farmland management, and ecological restoration in complex saline-alkali land environments, demonstrating significant promotional value and engineering application potential. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0041] Figure 1 This is a flowchart of the system modules of the present invention.

[0042] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Example 1, please refer to Figure 1 As shown in this embodiment, the IoT-based real-time monitoring system for water, salt, and fertilizer in saline-alkali land includes:

[0045] The sensor acquisition module is used to deploy multiple sensor nodes in a preset area of ​​saline-alkali land to collect information on soil salinity, moisture, electrical conductivity, pH value and ground temperature in the target area.

[0046] The remote sensing and geographic information module is used to acquire remote sensing images, topographic elevation data and land use type data covering the target area, and to preprocess and spatially register the data.

[0047] The spatial graph construction module is used to establish a spatial graph structure based on the positional relationship, geographical adjacency, groundwater flow path or irrigation and drainage channel between the sensor nodes and remote sensing pixels. Each monitoring unit is represented as a node of the graph neural network, and the spatial connection relationship between nodes is represented as an edge.

[0048] The graph neural network prediction module is used to deploy a graph neural network model on the constructed spatial map. By performing graph convolution calculations on the features of sensor nodes, it predicts the soil salinity evolution trend or salinity risk level of target nodes within a preset time range in the future.

[0049] The edge plot prediction module is used to estimate the salinity trend of edge plots or spatial blind spots without sensor coverage by using information transmission between adjacent nodes and graph structure learning.

[0050] The dynamic regulation decision module is used to generate a dynamic regulation strategy based on the prediction results of the graph neural network, combined with historical management parameters and agronomic rules, and output control commands to the field control terminal.

[0051] The visualization and early warning module is used to display in real time the salinity prediction map, risk level map and control suggestions of each plot in the target area. If the predicted value exceeds the safety threshold, an early warning signal will be automatically triggered.

[0052] To achieve dynamic sensing and precise control of soil conditions in saline-alkali land, the sensor acquisition module, as the front-end sensing layer of this system, is responsible for acquiring raw data on key environmental factors. This module needs to operate stably for a long time under natural conditions of high salinity, high alkalinity, and strong corrosivity to collect representative soil and environmental parameters, providing reliable support for subsequent modeling analysis and control decisions.

[0053] In this embodiment of the invention, the sensor acquisition module includes, but is not limited to, the following components:

[0054] The multi-parameter soil sensor component includes: a soil salinity sensor for monitoring soil electrical conductivity (ECe) in dS / m; a soil moisture sensor for real-time volumetric water content based on FDR or TDR technology; a soil pH sensor for monitoring pH fluctuations and assisting in assessing alkali damage risk; and a soil temperature sensor for monitoring root activity layer temperature and providing correction factors for model time-series fitting. The node controller and acquisition unit are configured with a microcontroller (such as STM32 or ESP32) integrating data acquisition, preprocessing, and transmission functions. The communication module supports wireless transmission protocols such as LoRa, NB-IoT, and 4G / 5G to send acquired data to edge computing nodes or cloud platforms. The power supply unit uses a combination of solar panels and lithium batteries to ensure continuous operation in remote, power-deprived areas. The protective housing is encapsulated in IP67-rated corrosion-resistant engineering plastic, adaptable to extreme outdoor environments such as salinity, high humidity, and high temperature.

[0055] Based on historical salinity distribution, topography, irrigation and drainage system distribution, and remote sensing image analysis, representative sampling locations were selected, including main cultivated areas, marginal plots, areas adjacent to drainage ditches, and areas near water storage areas, to ensure comprehensive data coverage. Each monitoring point employed a multi-layered soil probe layout: 0–20 cm (cultivated layer); 20–50 cm (root layer); 50–100 cm (permeable layer / latent salt layer); soil parameter sampling cycle: once every 5–15 minutes (adjustable); automatic uploading of daily average values ​​and weekly rate of change statistics to avoid redundant data transmission and power waste; multi-point calibration was performed before the equipment left the factory, and a remote calibration command interface was provided; periodic recalibration can be performed based on subsequent sample verification results.

[0056] Saline-alkali land exhibits significant spatial heterogeneity, with its salt distribution and accumulation influenced by various spatial factors such as topographic relief, land use patterns, and hydrological characteristics. Point-based sensors alone cannot fully reflect the salinity status of a plot. Therefore, the remote sensing and geographic information module proposed in this invention aims to introduce large-scale, continuous, and reproducible spatial data to compensate for blind spots in sensor deployment, providing key spatial feature support for subsequent graph neural network modeling and edge plot prediction.

[0057] The remote sensing and geographic information module includes the following sub-modules:

[0058] The remote sensing image acquisition unit acquires multi-source remote sensing images covering the target saline-alkali land, including: medium-resolution satellite data such as Sentinel-2 (10m resolution) and Landsat-8 (30m); high-resolution images (such as GF-1 and WorldView) for detailed investigation of key areas; and multi-temporal data to support time series analysis.

[0059] The remote sensing parameter extraction unit extracts characterization indices related to salinity from remote sensing images, including but not limited to: NDVI (Normalized Difference Vegetation Index): reflecting vegetation cover and indirectly inferring the impact of salt damage; NDSI (Normalized Dioxide Index): used to identify exposed saline soil; LST (Land Surface Temperature): reflecting soil evaporation intensity; and soil reflectance characteristic curves.

[0060] The terrain elevation data processing unit acquires DEMs (Digital Elevation Models) with a preferred resolution of 30m or higher. This data is used to: construct the terrain slope, aspect, and depression distribution of a site; and infer groundwater accumulation or drainage paths.

[0061] The land use type identification unit acquires or classifies land types within the target area, including cultivated land, forest land, water bodies, salt flats, residential areas, etc., and inputs them as node attributes into the GNN model.

[0062] The spatial registration and interpolation submodule performs unified coordinate transformation and spatial registration on the collected multi-source data: based on the WGS84 or CGCS2000 geographic coordinate system; it uses affine transformation, image resampling and other methods to achieve consistency of images with different resolutions; and it performs spatial interpolation (such as Kriging, IDW) between remote sensing data and sensor deployment points to improve local feature density.

[0063] Remote sensing parameters and geographic information data are mapped to each node of the graph neural network as input to the structured feature vector; for nodes without sensor coverage, remote sensing and DEM data serve as their main input sources to achieve boundary extrapolation of salinity trends; data time series stacking is supported for dynamic trend modeling and change detection.

[0064] The formation and evolution of saline-alkali land are influenced by a variety of complex spatial factors, exhibiting high spatial heterogeneity and structural correlation. In traditional agricultural information systems, plots are usually treated as independent samples without considering their interactions, which leads to a significant decrease in prediction accuracy in marginal plots or monitoring blind spots.

[0065] To overcome the above problems, this invention proposes a spatial graph construction module. By introducing the idea of ​​graph structure modeling, the "monitoring unit" in saline-alkali land is regarded as a node in a graph neural network (GNN). The edge connections between nodes are established based on multi-dimensional factors such as spatial relationships, hydrological structure or farmland management units, thereby realizing the propagation and generalization of spatial information in model training.

[0066] The spatial graph construction module specifically includes the following five sub-units, which work together to complete the construction process of the spatial graph structure:

[0067] Node Definition and Generation Unit: This unit is responsible for uniformly abstracting the original data sources (including sensor points, remote sensing pixel centers, and land parcel division results) to generate nodes in the spatial map. Each node corresponds to a physical monitoring unit or its approximate representative area and is assigned the following attributes:

[0068] Spatial location attributes: latitude and longitude coordinates, or x / y coordinates under a unified projection; monitoring data characteristics: such as electrical conductivity, moisture content, soil pH, surface temperature, etc.; spatial characteristic variables: such as NDVI value, DEM elevation, slope, aspect; plot label attributes: such as crop type, management unit number, soil type category. Node numbers are named according to rules (e.g., N001, N002...) to support structured input to the graph neural network.

[0069] Spatial Edge Connection Determination Unit: This unit is used to determine whether an edge connection exists between nodes and to define the criteria for connection. The criteria for determination are categorized as follows:

[0070] Geographic adjacency is used to connect nodes; if the Euclidean or Manhattan distance between nodes is less than a set threshold (e.g., 50 meters), they are considered to be adjacent; the K-nearest neighbor (KNN) strategy can be used to generate a fixed number of adjacency edges for each node to ensure the connectivity of the graph structure.

[0071] Topographic and hydrological connectivity is established; water flow paths are constructed based on DEM analysis, and upstream and downstream nodes are determined using the water flow direction matrix and runoff data; if nodes are located in the same hydrological runoff unit, a strong connection is established, and the edge weight is doubled.

[0072] The structure is connected by manual operations; if nodes belong to the same irrigation system, salt drainage network, or farm road connection area, they are considered agricultural structure connectivity units and edge connections are established; the corresponding farmland GIS layer is used as a reference source for matching.

[0073] Environmental similarity connections: If the similarity of variables such as node NDVI, vegetation index, surface temperature, and soil type is greater than the threshold (e.g., cosine similarity > 0.9), it is considered to have behavioral consistency and a "soft edge" connection is established.

[0074] Edge weight calculation and encoding unit; after connecting nodes, each edge is assigned a numerical weight for weighted information propagation in the graph neural network. Common calculation methods are as follows:

[0075]

[0076]

[0077] All edge weights are standardized to the range of 0 to 1, and the resulting edge weight matrix is ​​used as input to the neural network.

[0078] For areas without sensor deployment but covered by remote sensing data, this module can automatically generate virtual nodes, whose characteristics are derived from:

[0079] Remote sensing image inversion parameters (such as NDVI, NDSI, and surface temperature); topographic and land use information; spatial adjacency node aggregation characteristics. After virtual nodes are incorporated into the graph structure, they will automatically obtain contextual propagation information through graph neural networks, which can be used for trend prediction in blind areas where no measured data is available.

[0080] Finally, all node and edge sets are uniformly converted into: node feature matrix, adjacency matrix, and edge weight matrix (if a weighted graph neural network is used); the output structure serves as the input data structure for the graph neural network model. A graph database (such as Neo4j) can be selected to store the spatial graph for visualization and querying.

[0081] The graph neural network prediction module is used to deploy graph neural network models on the constructed spatial map structure of saline-alkali land. By aggregating the soil and environmental characteristics of each node and its neighboring nodes, it can predict the trend of salinity evolution or the level of salinity risk. It is particularly suitable for the prediction needs of sparse data and marginal plots.

[0082] Receive output from the "Spatial Graph Construction Module", including: Node Feature Matrix: containing soil salinity, moisture, pH value, NDVI, terrain information, etc. for each plot; Adjacency Relationship: defining whether plots are adjacent nodes to each other; Edge Weight Matrix: measuring the influence strength between adjacent nodes, such as based on geographical distance, water flow path or irrigation connectivity.

[0083] Choose a suitable graph neural network architecture, such as: Graph Convolutional Network (GCN): for basic spatial aggregation tasks; Graph Attention Network (GAT): suitable for scenarios with heterogeneous connections and requiring differentiated processing; GraphSAGE: suitable for scenarios with a large number of nodes and dynamically changing structures. Initialize network parameters, including the number of neurons in each layer, activation function type, learning rate, etc.

[0084] For example, this invention uses a Graph Attention Network (GAT) as an example. For a node i in the l-th layer of a graph neural network, the representation of its next layer is... The calculation method is as follows: This represents the feature vector of node i at layer l in the graph; Represents the set of neighboring nodes of node i; W represents the attention weight from neighbor node j to node i in layer l, which is learned through training; (l) σ is the trainable weight matrix of the l-th layer, used for linear transformation of node features; σ is a non-linear activation function, such as ReLU (Rectified Linear Unit).

[0085] For each target node, the graph neural network model is invoked to perform the following operations:

[0086] Extract features from the neighboring nodes of the current node; weight the features of the neighboring nodes using edge weights; fuse the weighted result with the features of the current node itself; generate an updated representation (embedding vector) of the node through a graph convolutional layer.

[0087] Stacking multiple layers of graph convolution operations, each layer extracts higher-order spatial structure information (e.g., neighbors of neighbors); the output of each layer updates the embedding representation of each node, enabling the model to learn deeper levels of inter-plot influence relationships.

[0088] For each target node, output one of the following prediction results:

[0089] Soil salinity trend values ​​(e.g., salinity forecasts for the next 3 or 7 days); salinity risk levels (e.g., mild, moderate, severe); output the confidence or uncertainty index of the forecast for subsequent regulation strategy decisions.

[0090] Training data includes historical time-series monitoring data and remote sensing data labels; model parameters are trained using error minimization methods (e.g., minimizing the mean square error between predicted and measured values); early stopping strategies, overfitting prevention mechanisms, and validation set evaluations are introduced to improve the model's generalization ability.

[0091] The model can be deployed on edge servers or cloud platforms; it can automatically receive the latest sensor and remote sensing data at a set frequency for prediction updates; and it can quickly adapt to new land parcels or changed areas and output prediction results.

[0092] The edge plot prediction module addresses the problem of prediction failure in saline-alkali land edge areas or spatial blind spots caused by insufficient or nonexistent sensor deployment. Based on a pre-constructed spatial map structure, this module leverages the adjacency information propagation capability of graph neural networks to estimate the salinity variation trend or salinity risk level of plots lacking actual measurement data.

[0093] The system scans the current spatial graph structure and identifies the following types of nodes:

[0094] Nodes without actual sensor data coverage; nodes with missing data due to data transmission interruption or acquisition failure; nodes located at the edge of the graph structure and with few spatially adjacent nodes are marked as "edge nodes" or "blind zone nodes".

[0095] Feature completion or initialization of edge nodes can be performed in the following ways: extracting NDVI, surface temperature, soil reflectance, etc. from remote sensing images as basic inputs; introducing static or semi-dynamic variables such as land use type, elevation, slope, and meteorological data; if the above data is still empty, the average value of adjacent nodes in the graph is used as the initial estimate (weak supervision).

[0096] In graph neural network models, the specific process of performing graph convolution or graph attention mechanisms is as follows:

[0097] Each edge node receives feature information from its neighboring nodes; the feature information is weighted according to edge weights, such as geographical proximity, consistency of water flow paths, and differences in the salinity sensitivity coefficient of the crop root activity layer; the system updates the hidden representation (embedding) of the edge nodes through graph structure learning.

[0098] Geographic proximity refers to the actual physical proximity of two nodes (i.e., two monitoring plots or spatial units) in two-dimensional geographic space, usually calculated based on their latitude and longitude coordinates or projected coordinates. This parameter measures whether two plots are close to each other on the surface, thereby inferring their environmental homogeneity in terms of climate conditions, farming management, human intervention, etc. The closer the plots are geographically, the more consistent their trends in soil salinity, moisture, temperature, etc., so they should be assigned higher edge weights when disseminating information; it can be used to capture short-distance spatial correlations in local microenvironments.

[0099] The calculation methods include: calculating the Euclidean distance (straight-line distance) between two nodes; the shorter the distance, the higher the proximity. To avoid excessive weighting of distant nodes, normalization can be performed using a Gaussian kernel function or inverse distance weighting (IDW).

[0100] The consistency of water flow paths refers to the hydrological connectivity between two plots in terms of terrain slope and surface runoff paths, and is used to determine whether they are in the same natural drainage direction or irrigation and drainage system. This parameter measures the connectivity of two nodes in the hydrodynamic system, that is, whether water is likely to flow from one plot to another; in saline-alkali land, the water flow path directly determines the direction and rate of salt migration. If two nodes are in the same catchment area or water flow path, the salt dynamics may also affect each other; it is applicable to identifying remotely influential plots that are not physically adjacent but "hydraulically related", which is a key feature that cannot be captured by geographical proximity.

[0101] Based on digital elevation model (DEM) data, construct a flow direction matrix and a catchment area matrix; determine whether two nodes are on the same water flow path or in an upstream-downstream relationship; nodes with short distances, connected flow directions, and similar catchment area values are given higher consistency scores; automatic analysis can be performed using hydrological modeling software (such as ArcGIS Hydrology tools, TauDEM).

[0102] The difference value of the salt sensitivity coefficient of crop roots refers to the difference in the response ability of crops to changes in salt concentration in the root activity layer between two plots, which is reflected as the relative difference in the degree of impact on growth, yield, or physiological activities. This parameter reflects whether the two plots have similar regulatory requirements and salt tolerance capabilities in agricultural management. The smaller the difference, the closer the two nodes should be connected for feature transfer.

[0103] For each node, read the average value of the electrical conductivity (ECe) within the current depth range of 0–40 cm or 0–60 cm; if multiple-layer sensors are installed, perform weighted averaging; denote it as the ECe value of plot i: ECei, and that of plot j as ECej.

[0104] For the crops in each plot, extract the following parameters from the agricultural standard database according to their crop types and growth stages:

[0105] Threshold electrical conductivity (Threshold ECe), which represents the starting point of salt at which crop production begins to decline;

[0106] Slope parameter (Slope), which represents the percentage of yield decline per 1 dS / m increase in salt;

[0107] Example (wheat): Threshold = 6.0 dS / m; Slope = 7.1% per dS / m;

[0108] For each node, substitute into the crop response model (such as a linear yield decline model): if ECe < Threshold ECe, then the sensitivity coefficient = 0;

[0109] If ECe ≥ Threshold, then:

[0110] Si=(ECei-Thresholdi)×Slopei;

[0111] Sj=(ECej-Thresholdj)×Slopej;

[0112] Where Si and Sj represent the current salinity sensitivity values ​​of plots i and j, respectively, in units of the percentage decrease in yield; Slopei is the slope value of plot i, and Slopej is the slope value of plot j.

[0113] The sensitivity difference value CSSG between the two plots is calculated as follows: CSSG=∣Si-Sj∣; the smaller the value, the more similar the degree of response of the crops in the two plots to the current salinity environment; the larger the value, the greater the difference in the tolerance of the crops in the two plots to salinity changes, and it is not advisable to force migration prediction.

[0114] Normalize all CSSG values ​​to between 0 and 1 for easy combination with other edge weight parameters; it can be used as a reverse factor of edge weights in the calculation of propagation weights in graph neural networks: the smaller the CSSG, the larger the edge weight; the edge weight EdgeWeightCSSG can be: EdgeWeightCSSG = 1 - CSSGnorm; CSSGnorm is the normalized CSSG value.

[0115] The final embedding vector of the edge node is input into the model output layer to predict target indicators, including: salinity trend prediction (continuous value), such as the trend of conductivity rising or falling in the next 3 days; salinity risk level prediction (categorical value), such as three categories: "mild, moderate and severe"; at the same time, confidence level can be output, which is used as the basis for judging whether the system will trigger manual review in the future.

[0116] Once sensor data is subsequently deployed or manually sampled data is uploaded from the edge plots, the system will automatically perform the following operations:

[0117] Compare the model's predicted values ​​with the measured values; if the error exceeds a set threshold, mark the node for model retraining; update the model weights with new data to gradually enhance the model's generalization ability on edge plots.

[0118] The dynamic regulation and decision-making module is used to dynamically generate specific regulation strategies such as irrigation, salt removal, and application of soil amendments based on the prediction results of graph neural networks, combined with historical management records, soil evolution trends, crop types and agronomic rules. It then sends control commands to smart terminals (such as pump station controllers, solenoid valves, and pesticide application devices) to achieve precise management of saline-alkali land.

[0119] The following multi-source information was collected as the set of input variables for regulatory decisions:

[0120] Graph neural network prediction results: Soil salinity evolution trend for each plot over the next 3–15 days; salinity risk level; prediction confidence level (used for error tolerance adjustment of control strategies). Historical management parameters: Irrigation frequency, water consumption, salt discharge frequency and method in recent years; records of soil amendment use (e.g., gypsum, organic matter); previous management effects (assessed by feedback from remote sensing / ground monitoring). Agronomic rules and constraints: Current crop species and growth stage; crop salinity tolerance threshold (ECe standard); irrigation window, operational restrictions (e.g., stopping irrigation on rainy days, prioritizing drainage, etc.).

[0121] With multi-objective optimization as the goal, a comprehensive regulatory objective function is constructed, including but not limited to: maximizing the rate of salinity reduction; minimizing irrigation and drainage water consumption; ensuring a salinity safety zone for crop growth; and controlling input costs (such as energy consumption and chemical usage).

[0122] An evolutionary strategy optimization algorithm is introduced, which is an intelligent search algorithm suitable for nonlinear, multi-objective, and complex rule spaces. It is superior to traditional genetic algorithms or grid search: it automatically adjusts control variables (such as irrigation volume, drainage start-up time, and amendment application amount); it dynamically adapts to the salinity evolution model of different plots and avoids the failure of "fixed strategy templates".

[0123] The core of the CMA-ES algorithm is the "fitness function," which determines whether the strategy combination achieves multi-objective optimality after simulation. The fitness function consists of the following weighted objectives:

[0124] To reduce soil salinity to below a safe threshold (e.g., <4 dS / m) during the future forecast period;

[0125] Minimize the amount of water consumed per unit decrease in salinity and the energy consumption for drainage;

[0126] Ensure that the root zone moisture content is moderate and ECe does not exceed the limit during the current growth stage of the crop;

[0127] The changes in control variables should not be too drastic to prevent frequent equipment start-ups and shutdowns and physiological shocks to crops.

[0128] The comprehensive function form (in words) is as follows: Fitness = Salinity control score + Water resource efficiency score + Crop safety score - Cost penalty factor.

[0129] Several "strategy combinations" are initialized randomly or based on experience, each including several control variables;

[0130] For each strategy combination, simulations predict salinity changes, crop responses, water consumption, etc. over a future period to calculate fitness scores.

[0131] Based on the fitness level, the system automatically learns the collaborative relationships between different variables and adjusts the sampling distribution (the direction of the high-fit strategy is searched first).

[0132] By sampling new combinations of variables and continuing the iterative steps, it can usually converge to the optimal solution or a near-optimal solution within 10-100 rounds.

[0133] The set of variables with the highest scores will be used as the recommended control instructions for this round of regulation.

[0134] The optimal variables will be mapped to standard control command formats (such as "irrigate 60mm every 6 days; drain for 45 minutes, apply 20kg of gypsum per round"); the commands will be sent to the corresponding field control units (solenoid valves, pumping stations, mixing devices, etc.) through the control platform; the control module will record strategy execution logs to provide data feedback for subsequent model optimization.

[0135] The optimal combination of control variables is used to generate a regulation scheme, and its rationality is verified according to the following dimensions: physiological feasibility verification: whether it is within the crop's salt tolerance range; engineering feasibility verification: whether it meets the limitations of equipment capacity, pump station flow rate, power supply capacity, etc.; site compatibility judgment: whether it conflicts with the regulation of adjacent sites (such as upstream and downstream water level logic). If it does not meet the requirements, the evolutionary strategy optimization algorithm will recalculate the strategy until a feasible solution is output.

[0136] The generated strategy is converted into an executable instruction format, such as Modbus or LoRa control instructions; the communication module sends the commands to the following field smart terminals: irrigation valve controller (automatic opening and closing, flow regulation); drainage pump station (start and stop according to time periods); amendment mixing equipment (setting the application ratio and time); the system can be set with a local buffer mechanism to ensure that the previous strategy is still executed when offline.

[0137] During the implementation of the control measures, real-time feedback data (such as the decrease in electrical conductivity and the recovery of soil moisture) is continuously acquired. If the actual effect deviates too much from the prediction (exceeding the set error threshold), then: the "strategy fine-tuning mechanism" is activated: the local control strategy is updated using a reinforcement learning fine-tuning model (such as DDPG); or a manual review reminder is triggered to improve the robustness and interpretability of the system.

[0138] The visualization and early warning module is used to summarize, analyze and display the system's prediction and control results in real time, including salinity trend maps, risk level distribution maps and dynamic control suggestions for each plot. When key indicators (such as predicted salinity values) exceed the set threshold, it automatically triggers early warning signals and links control responses.

[0139] Receive intermediate result data output from the following modules:

[0140] The graph neural network prediction module outputs: salinity prediction values ​​for each plot in the future time window (e.g., 3 days, 7 days); risk level (e.g., mild, moderate, severe); prediction confidence index (uncertainty distribution); the dynamic regulation decision module outputs: recommended regulation schemes generated for each plot (irrigation volume, salt discharge frequency, amount of soil amendment, etc.); control execution status and feedback (e.g., in progress, completed, abnormal); and real-time data from edge sensors (used to compare prediction deviations and dynamically correct early warnings).

[0141] Using GIS visualization technology and a heatmap engine, the following map layers were generated:

[0142]

[0143] It can be displayed in real time on a web platform, mobile terminal or central control screen, and supports mouse / touch interaction. Click on the plot to view the prediction curve, regulation suggestions and historical data.

[0144] The system supports setting safety thresholds manually or according to model recommendations (the default setting is based on crop salinity tolerance parameters); thresholds can be set according to the following dimensions: ECe value (e.g., >6dS / m for warning); rate of salinity increase (e.g., >1.5dS / m / 7 days); consistency of multiple prediction results (continuous increase); risk level transition (e.g., from mild to severe); users can also set personalized threshold strategies (e.g., setting stricter limits for crop seedlings).

[0145] The system scans the entire map for prediction results at a set frequency (e.g., every 6 hours); for plots exceeding the threshold, the following operations are performed:

[0146] Pop-up / Marker Layer: Displays warning plots in red or flashing; Push Notification: Sends push notifications to the administrator's App or SMS (including plot number, exceeding standard items, and control suggestions); Automatic Trigger Linkage Control: Calls the corresponding scheme in the "Control Strategy Library" and directly issues irrigation and drainage instructions; Records Alarm Logs: Includes trigger time, predicted value, response status, final effect, etc., for use in model iteration training.

[0147] Users can view the following for any time period: salinity evolution curve; risk level changes; control implementation records; support exporting charts (Excel / PDF / images) for expert diagnosis or project report generation; and use the time slider to view the "prediction vs. actual" deviation to evaluate model effectiveness.

[0148] Example 2, please refer to Figure 2 As shown in this embodiment, the real-time monitoring method for water-salt-fertilizer in saline-alkali land based on the Internet of Things includes:

[0149] Multiple sensor nodes are deployed in a pre-defined area of ​​saline-alkali land to collect information on soil salinity, moisture, electrical conductivity, pH value and ground temperature in the target area.

[0150] Acquire remote sensing images, topographic elevation data, and land use type data covering the target area, and perform preprocessing and spatial registration on the data;

[0151] A spatial graph structure is established based on the positional relationship, geographical adjacency, groundwater flow path, or irrigation and drainage channel between the sensor nodes and remote sensing pixels. Each monitoring unit is represented as a node of the graph neural network, and the spatial connection relationship between nodes is represented as an edge.

[0152] A graph neural network model is deployed on the constructed spatial map. By performing graph convolution calculations on the features of sensor nodes, the soil salinity evolution trend or salinity risk level of the target node within a future preset time range is predicted.

[0153] For edge plots or spatial blind spots without sensor coverage, salinity trend estimation is completed through information transmission between adjacent nodes and graph structure learning;

[0154] Based on the prediction results of the graph neural network, combined with historical management parameters and agronomic rules, a dynamic regulation strategy is generated, and control commands are output to the field control terminal.

[0155] It displays real-time salinity prediction maps, risk level maps, and control recommendations for various plots within the target area. If the predicted value exceeds the safety threshold, it automatically triggers an early warning signal.

[0156] The above description is merely a specific embodiment of this application, but the scope of protection of this application is limited.

[0157] The scope is not limited thereto; any person skilled in the art may apply to any application disclosed herein.

[0158] Within the scope of the technology, any changes or substitutions that can be easily conceived should be covered by the protection provisions of this application.

[0159] Within the protected area.

Claims

1. A real-time monitoring system for water-salt-fertilizer in saline-alkali land based on the Internet of Things, characterized in that: include: The sensor acquisition module is used to deploy multiple sensor nodes in a preset area of ​​saline-alkali land to collect information on soil salinity, moisture, electrical conductivity, pH value and ground temperature in the target area. The remote sensing and geographic information module is used to acquire remote sensing images, topographic elevation data and land use type data covering the target area, and to preprocess and spatially register the data. The spatial graph construction module is used to establish a spatial graph structure based on the positional relationship, geographical adjacency, groundwater flow path or irrigation and drainage channel between the sensor nodes and remote sensing pixels. Each monitoring unit is represented as a node of the graph neural network, and the spatial connection relationship between nodes is represented as an edge. The graph neural network prediction module is used to deploy a graph neural network model on the constructed spatial map. By performing graph convolution calculations on the features of sensor nodes, it predicts the soil salinity evolution trend or salinity risk level of target nodes within a preset time range in the future. The edge plot prediction module is used to estimate the salinity trend of edge plots or spatial blind spots without sensor coverage by means of information transmission between adjacent nodes and graph structure learning. The dynamic regulation decision module is used to generate a dynamic regulation strategy based on the prediction results of the graph neural network, combined with historical management parameters and agronomic rules, and output control commands to the field control terminal. The visualization and early warning module is used to display in real time the salinity prediction map, risk level map and control suggestions of each plot in the target area. If the predicted value exceeds the safety threshold, an early warning signal will be automatically triggered.

2. The IoT-based real-time monitoring system for water-salt-fertilizer in saline-alkali land according to claim 1, characterized in that: The sensor acquisition module includes: Multiple multi-parameter soil sensor nodes are deployed in representative areas within the target saline-alkali land. These areas include the main cultivated area, marginal plots, areas adjacent to drainage ditches, and areas near water storage areas. Each sensor node is vertically deployed at three depth levels: 020cm, 2050cm, and 50–100cm. Each node periodically collects data on electrical conductivity, volumetric water content, soil pH, and ground temperature at a sampling frequency of 5–15 minutes. The data is transmitted to an edge gateway or cloud platform via LoRa, NB-IoT, or 4G communication modules.

3. The IoT-based real-time monitoring system for water-salt-fertilizer in saline-alkali land according to claim 1, characterized in that: The spatial map construction module includes: Abstract sensor monitoring points, remote sensing pixel centers, or farmland boundary centers into nodes in a graph neural network; The edge connections between nodes are determined based on geographical proximity, consistency of water flow paths, and similarity between the irrigation and drainage system structure and the environment. Each edge is assigned a weight, which is calculated based on factors such as geographical distance, consistency of water flow paths, and differences in crop salt sensitivity coefficients. Output the node feature matrix, adjacency matrix, and edge weight matrix.

4. The IoT-based real-time monitoring system for water-salt-fertilizer in saline-alkali land according to claim 3, characterized in that: in, Geographic proximity can be calculated using the inverse Euclidean distance or the Gaussian kernel function. Based on the DEM, the water flow path matrix is ​​derived, and it is determined whether the nodes belong to the same catchment unit. If so, the weight is doubled. The differences in salt sensitivity of the crop root zone of the nodes are compared, and the difference value of the crop root salt sensitivity coefficient (CSSG) is calculated and normalized into the edge weight reverse adjustment factor. Finally, all edge weights are normalized to the range of 0 to 1 to form a weighted adjacency matrix.

5. The IoT-based real-time monitoring system for water-salt-fertilizer in saline-alkali land according to claim 4, characterized in that: The method for calculating the difference in crop root salt sensitivity coefficient is as follows: For each node, read the average conductivity within the current depth range of 0–40 cm or 0–60 cm; for each plot of crop, extract the threshold conductivity from the agricultural standard database according to its crop type and growth stage, which is the starting point of salt content at which crop yield begins to decrease; slope parameter: the percentage decrease in yield for every 1 dS / m increase in salt content; for each node, substitute it into the crop response model to calculate the difference in crop root salt sensitivity coefficient (CSSG) between the two plots.

6. The IoT-based real-time monitoring system for water-salt-fertilizer in saline-alkali land according to claim 1, characterized in that: The dynamic control decision-making module includes: Receive soil salinity trend and risk level results; call historical regulation data, crop type and agronomic rule base to construct regulation objective function; introduce evolutionary strategy optimization algorithm to generate optimal control variable combination based on multi-objective function, including irrigation amount, drainage duration and application cycle; convert the optimal strategy into executable control command format and send it to the corresponding field terminal equipment.

7. The IoT-based real-time monitoring system for water-salt-fertilizer in saline-alkali land according to claim 6, characterized in that: The optimization process includes: Initialize multiple control strategy sample combinations; Calculate the salt control score, water resource efficiency score, crop response score, and cost penalty factor for each combination in the future forecast period; New strategy combinations are generated by automatically adjusting the sampling covariance matrix and step size based on the fitness function. Output the strategy with the highest score after multiple iterations; If the constraints are not met, return to the evolutionary loop until they are met.

8. A method for real-time monitoring of water-salt-fertilizer in saline-alkali land based on the Internet of Things, used to implement the real-time monitoring system for water-salt-fertilizer in saline-alkali land based on the Internet of Things as described in any one of claims 1-7, characterized in that: include: Multiple sensor nodes are deployed in a pre-defined area of ​​saline-alkali land to collect information on soil salinity, moisture, electrical conductivity, pH value and ground temperature in the target area. Acquire remote sensing images, topographic elevation data, and land use type data covering the target area, and perform preprocessing and spatial registration on the data; A spatial graph structure is established based on the positional relationship, geographical adjacency, groundwater flow path, or irrigation and drainage channel between the sensor nodes and remote sensing pixels. Each monitoring unit is represented as a node of the graph neural network, and the spatial connection relationship between nodes is represented as an edge. A graph neural network model is deployed on the constructed spatial map. By performing graph convolution calculations on the features of sensor nodes, the soil salinity evolution trend or salinity risk level of the target node within a future preset time range is predicted. For edge plots or spatial blind spots without sensor coverage, salinity trend estimation is completed through information transmission between adjacent nodes and graph structure learning; Based on the prediction results of the graph neural network, combined with historical management parameters and agronomic rules, a dynamic regulation strategy is generated, and control commands are output to the field control terminal. It displays real-time salinity prediction maps, risk level maps, and control recommendations for various plots within the target area. If the predicted value exceeds the safety threshold, it automatically triggers an early warning signal.

Images